Skip to content

AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist agents (FastAPI + Next.js).

fix(prompts): scope the Google manifest to the services and tier workspace-mcp loaded (#425)

maarten-nembe279b53077d
diff --git a/.env.example b/.env.example
index a1badd86..b4d3abdb 100644
--- a/.env.example
+++ b/.env.example
@@ -296,6 +296,29 @@ GOOGLE_CHAT_SERVICE_ACCOUNT_EMAIL=
GOOGLE_OAUTH_CLIENT_ID=
GOOGLE_OAUTH_CLIENT_SECRET=
+# Which Google services the co-located workspace-mcp child loads, and how many
+# tools per service. Defaults load everything; narrow them only if you want a
+# smaller tool surface.
+#
+# WORKSPACE_MCP_TOOLS "all" (default), or comma-separated service names:
+# gmail, calendar, drive, docs, sheets (also chat,
+# forms, slides, tasks, contacts, search, appscript).
+# A service left out is never imported, so none of
+# its tools exist — the Connected Systems prompt
+# block narrows to match, and the Executive is not
+# told about tools it cannot call.
+# WORKSPACE_MCP_TOOL_TIER "complete" (default), "extended", or "core".
+# Tiers are cumulative; "core" is the smallest set
+# and leaves out ten of the pinned tools (Gmail
+# drafts/threads/attachments, Drive item listing,
+# freebusy, and the Docs/Sheets listing tools).
+#
+# Narrowing either one is a prompt-visible change: fewer services are named in
+# block 0 and fewer tool names are pinned. Leave both unset for every service
+# at the complete tier.
+WORKSPACE_MCP_TOOLS=all
+WORKSPACE_MCP_TOOL_TIER=complete
+
# Microsoft 365 MCP (optional — enables Outlook mail + calendar tools)
# Runs @softeria/ms-365-mcp-server co-located in the API image, the Microsoft
# analogue of the Google Workspace server above; both can be configured at
diff --git a/docker/workspace-mcp-launch.sh b/docker/workspace-mcp-launch.sh
index d5206550..8fb6654f 100644
--- a/docker/workspace-mcp-launch.sh
+++ b/docker/workspace-mcp-launch.sh
@@ -34,6 +34,14 @@ case "$TOOL_TIER" in '$'*) TOOL_TIER=complete ;; esac
case "$SERVICES" in '$'*) SERVICES=all ;; esac
SERVICES=$(printf '%s' "$SERVICES" | tr ',' ' ')
case "$SERVICES" in *[![:space:]]*) ;; *) SERVICES=all ;; esac
+# Re-join on whitespace so the `!= all` test below sees tokens, not raw text.
+# `--tools` is declared `nargs="*" choices=VALID_SERVICES`, and "all" is not a
+# service, so argparse exits 2 on `--tools all` and the child never starts.
+# Without this, "all," or " all " survived the two cases above as the literal
+# string "all " — which is not equal to "all", so the branch below passed
+# `--tools all` and killed the server on a value the operator clearly meant as
+# the default.
+SERVICES=$(printf '%s' "$SERVICES" | tr -s '[:space:]' ' ' | sed 's/^ //; s/ $//')
# The choice reaches workspace-mcp only as --tools below. Left in the
# environment it is read again as a fallback, and "all" is not a service name.
unset WORKSPACE_MCP_TOOLS
diff --git a/docs/deployment.md b/docs/deployment.md
index 0395905b..a2d0c16f 100644
--- a/docs/deployment.md
+++ b/docs/deployment.md
@@ -237,6 +237,52 @@ share with People on the roster.
Every Google Workspace call acts as the Executive's own account: the gateway
refuses a call whose `user_google_email` names any other address.
+#### Narrowing the Google tool surface (optional)
+
+Two variables decide how much of Google Workspace the co-located child loads.
+Both default to everything; leave them unset unless you want less.
+
+| Variable | Default | Values |
+|---|---|---|
+| `WORKSPACE_MCP_TOOLS` | `all` | `all`, or comma-separated service names: `gmail`, `calendar`, `drive`, `docs`, `sheets` (also `chat`, `forms`, `slides`, `tasks`, `contacts`, `search`, `appscript`) |
+| `WORKSPACE_MCP_TOOL_TIER` | `complete` | `core`, `extended`, `complete` |
+
+`WORKSPACE_MCP_TOOLS` reaches workspace-mcp as `--tools`, which decides which
+service **modules** it imports. A service you leave out is never imported, so
+none of its tools exist — not in the index, and not callable. The tier decides
+how many tools per service are registered; tiers are cumulative, and `core` is
+the smallest.
+
+Both are prompt-visible. The Connected Systems block in the cached system
+prompt names only the services actually loaded and pins only the tool names
+that exist at the active tier, and startup priming probes the same set — so the
+Executive is never told a tool is "already available" when it is not. If you
+narrow these, expect the block to name fewer services; that is the intended
+effect, not a fault.
+
+Two things worth knowing before narrowing:
+
+- `core` leaves out ten of the pinned tools — Gmail drafts, threads and
+ attachments, Drive item listing, calendar freebusy, and the Docs/Sheets
+ listing and info tools. `extended` and `complete` both cover all of them.
+- `all` means every service only as the **whole** value (surrounding commas or
+ spaces are fine). `gmail,all` is passed through as the two names `gmail` and
+ `all`, and since `all` is not a service name, workspace-mcp's argument parser
+ rejects it and **the child does not start at all** — you lose every Google
+ tool, not just the misspelled one. Write `all` on its own, or list only real
+ service names.
+- Both variables reach the child through the `env` block of the
+ `google_workspace` entry in `mcp_servers.json`, which must forward them
+ (`"WORKSPACE_MCP_TOOLS": "$WORKSPACE_MCP_TOOLS"` and the same for
+ `WORKSPACE_MCP_TOOL_TIER`, as
+ [mcp_servers.json.example](../packages/core/mcp_servers.json.example) does).
+ A config that forwards only one of them leaves the API filtering the prompt
+ on a value the child never saw, which hides tools that are in fact callable.
+
+A pinned tool that priming cannot find is logged once at startup, naming both
+possible causes — the service list or tier excluding it, or workspace-mcp having
+renamed it since the pin in `docker/Dockerfile`.
+
### Your own Gmail or Outlook (Act as me, optional)
Act as me lets the Executive draft replies **as the owner**, in the owner's
diff --git a/packages/core/openexecutive/architecture/architecture-facts.yaml b/packages/core/openexecutive/architecture/architecture-facts.yaml
index 911e8fbe..650a7321 100644
--- a/packages/core/openexecutive/architecture/architecture-facts.yaml
+++ b/packages/core/openexecutive/architecture/architecture-facts.yaml
@@ -7681,7 +7681,15 @@ integrations:
DiscoveredToolsFilter runs only tools one of the session's searches
returned and keeps them for the session's life, so the lifespan runs
MCPGateway.prime_pinned_tools() once in the background when
- google_workspace is configured (exact-name searches; misses logged as
+ google_workspace is configured. It probes the SAME narrowed set the
+ prompt advertises (connected_systems.pinned_google_tools, filtered by
+ WORKSPACE_MCP_TOOLS and WORKSPACE_MCP_TOOL_TIER), so a service that was
+ never imported is never probed and never warned about. A name neither
+ can account for is logged once naming both causes - excluded by those
+ two variables, or renamed by workspace-mcp since the pin - because the
+ old text named only the rename and sent operators to the manifest for
+ what was usually a service-list setting (exact-name searches; misses
+ logged as
drift), and call_tool retries a PINNED name once after discovering it when
the reply is extensible-mcp's "has not been discovered via search_tools"
refusal — after every gate has passed the same arguments. Unpinned tools
diff --git a/packages/core/openexecutive/architecture/prebuilt/agents.json b/packages/core/openexecutive/architecture/prebuilt/agents.json
index 87c8814e..928076ac 100644
--- a/packages/core/openexecutive/architecture/prebuilt/agents.json
+++ b/packages/core/openexecutive/architecture/prebuilt/agents.json
@@ -1,7 +1,7 @@
{
"section_id": "agents",
"title": "Agent Council",
- "markdown": "The Executive (`openexecutive.orchestrator.executive.Executive`) is a single coherent persona that fans out to specialist sub-agents via Anthropic tool use. Each specialist subclasses `openexecutive.agents.base.BaseAgent`, carrying a `name`, `domain`, and `model`, and exposes `analyze(...)`, which wraps the question in user-turn context and calls Claude through the per-call provider. That context is per call and never cached: `<company_stage>` (the stage from the company profile — specialists never see the profile itself), in solo mode `<principal_role>` (what kind of principal the advice is for, in plain words, with their title and remit — `router.principal_role_context`), `<department_memory>`, `<past_decisions>`, `<failure_cases>`, `<relevant_knowledge>` and `<conversation_context>`, outermost first. The specialist's own system prompt is a static domain constant under one `cache_control` block, so it stays byte-identical across companies and profile edits. A chat turn takes the stage from its session's profile (the one the Executive reasons over, and the one an eval scenario injects); workflow steps and the MCP server's `consult_specialist` read it fresh from the profile on disk. The role tag follows the turn's mode: a chat turn pins the role on its session with the mode (`pin_turn_principal_role`: the session's `principal_role` override, else the workspace settings) and sends it in solo and nothing in team, so the org block, every specialist in the turn and any workflow the turn starts see the same role; callers with no chat turn read it fresh (`load_principal_role`: the current session's mode and role, else the workspace's), and a failed read sends no tag rather than failing the consult. The CFO, CSO and sales prompts key their advice on it: for a bootstrapped, self-funded or solo company they drop venture fundraising benchmarks and work from cash, and raise only if the founder wants to.\n\nRouting uses one `consult_specialist` tool. Beside it the Executive carries two skill tools, `search_skills` and `load_skill` (skills are shown to users as **playbooks**, and every playbook belongs to a workflow: one that follows it, or its own Quick workflow). Chat finds and reads playbooks but never creates or changes one: work worth repeating is offered as a workflow through `draft_workflow`, with the method written into its steps, because workflows read playbooks at run time (approved scheduled ones included) and this loop also runs on inbound email and chat, so only a person can change what a workflow follows. A hidden built-in is invisible to search and load; each `search_skills` hit lists the `workflows` that follow it, Quick ones included, and the persona routes \"the finished deliverable\" to that workflow and quick or partial asks to the playbook inline; Tool calls made while proactive work is tagged (scheduler dispatch, reflection, research, alert review) feed the Attunement outcome ledger. Goals have three tools: `list_department_goals`, `update_department_goal` (progress the user reports) and `create_goal` (a new goal the principal states, filed under an area by slug or title and creating the area when none matches — see Org). `create_goal` runs only for the principal on a verified surface, and no unattended run is offered it: the tools in `schedule_tools.UNATTENDED_WITHHELD_TOOLS` are dropped from reflection's and research's toolkits by `unattended_toolkit`, and from the chat loop when its Session is marked `unattended` — the scheduler's PROACTIVE TRIGGER run, whose prompt quotes stored intent text; a call the model emits anyway gets an error tool_result. `record_decision_outcome` records how a past decision turned out — the weekly review lists decisions older than 30 days with no outcome as `[decision N]` — writing `decisions.outcome` (`episodic.update_decision`) and auditing the rationale; it has `create_goal`'s gate (the principal on a verified surface only; everyone else gets an audited refusal) and is in `UNATTENDED_WITHHELD_TOOLS` too. A missing or unknown id returns the decisions still waiting on an outcome, so the model can retry with the right one. Its schema is static; the solo persona has one line on it, and team mode learns it from the tool description, so the team persona is byte-identical. Attunement adds `list_open_loops` (what a person owes — everyone's for the principal only in a private conversation, otherwise the caller's own) and `close_open_loop`, which only closes a loop when the turn's resolved speaker is the principal or the loop's owner — an unrostered sender, a signed-in web user not on the roster, and unattended sessions are refused. `assign_open_loop` opens one on request (\"ask Ben to send me the Q3 numbers by Friday\") for the principal or a rostered teammate on a verified surface — never an email, a turn private to the principal or an unattended run, which are also withheld the tool — and only when `task` appears verbatim in what the speaker typed this turn, at most 5 per turn; that turn's extraction pass then opens no second loop for the same owner. The roster write tools (`upsert_person`, `archive_person`, `set_department_head`) are the principal's alone, and only on a surface that verified the speaker — the web chat, the principal's own Slack or Discord, or a private Telegram chat when `TELEGRAM_WEBHOOK_SECRET` is set to a value Telegram can send (Telegram identifies the chat rather than the sender, and without that secret its webhook accepts anyone's update). A teammate, an email (even from the principal's address), Google Chat, the CLI, the MCP server and unattended runs get `status: refused` (audited, and no action chip), because a team member's People row is also the web sign-in and outbound-email allow-list. `upsert_person` takes `kind` (`team` / `contact`; a new person defaults to `contact` in solo mode, an update keeps the current kind) and `email_aliases` (other addresses the person writes from — the full list, replacing the current one; they match mail and may be emailed but never sign in, and one already on someone else is refused), and after every write any pending roster request whose sender now matches someone is closed and its held messages replayed. `resolve_roster_request` answers a **roster request** — someone not on the People list who wrote in and is waiting (see Org): `approve` (a new person; `kind` defaults to team for a company-domain sender, else contact, when the principal did not say), `link` (someone already listed, by `person_id`; `replace_channel_id` only when the principal said so) or `decline`. It has the roster tools' gate (the principal on a verified surface; audited refusal otherwise), is in `UNATTENDED_WITHHELD_TOOLS`, and answers only an id in `Session.trusted_roster_request_ids` — the ids the server printed that turn in a `<roster_requests>` block (`briefing.context.format_roster_requests_for_prompt`, up to 10 pending requests), which `render_and_trust` appends to the `<briefing>` digest on the principal's own verified turn only and records on the session in the same step, so an id the model invents or reads out of a message adds nobody. Each line is server-derived — id, channel, address or account id, the sender's name sanitised and marked unverified, a company-domain hint, the message count — and never what they wrote. The block rides in the **user turn**, like the digest, so the cached system blocks are unchanged; the tool definition is static, and the new tool only extends the sorted tool list. The web chat route resets the trusted set every turn, and a resolved id is removed from it. and `list_people` returns both kinds with `kind` — contacts only on the principal's own verified turn (they are private to the principal; anyone else gets the team as if no contact existed) — while `lookup_person` searches the team only. `ask_about_person` reads a person's peer memory fully only on their own verified turn; anyone else's answers only to a verified principal or teammate, and only when that person turned on \"Share my work style with the team\", framed to how they work and never their view of a colleague, and otherwise returns the same \"hasn't chosen to share\" result whether or not anything is known. The same verified-principal rule, as `people_tools.contacts_reachable_now()`, decides whether a turn may reach a **contact** at all: `message_person`, the raw DM tools, `create_calendar_event` / `create_instant_meeting` attendees and the MCP gateway's Gmail / Calendar / Drive / OneDrive gates treat a contact exactly like an unknown person on every other turn, so no refusal reveals that one exists. **The untrusted-content policy** (`orchestrator.content_trust`) is one rule for text the principal did not write — inbound email, chat, attached files, watched pages — answering three questions for every surface. *Who is speaking*: the principal only when the surface proved it (`principal_speaking`: the web chat as the principal — never a signed-in email on nobody's People entry —, the CLI, their own Slack/Discord or a secret-verified Telegram chat, or their own address with Gmail's `dmarc=pass`); never a teammate, a stranger, Google Chat, the MCP server or an unattended run. *What may become memory*: episodic extraction runs only then, and only on the words outside `<untrusted_content>` blocks; peer memory and open loops are recorded under the speaker's own name, never as the principal's. *What may trigger tools*: every write on the principal's authority already refuses others in its handler; the tools with no such check that change the install for every later turn — `content_trust.PRINCIPAL_ONLY_TOOLS`: `load_mcp_server`, `list_saved_tools`, `run_python_job` and `run_script` — are not offered unless the principal is speaking on an interactive surface (not email, even authenticated), and a call the model emits anyway gets an error tool_result and a `refused: not_principal` audit row. So an inbound email, a teammate's turn, Google Chat and the scheduler's proactive trigger each get a stable tool list of their own without it. *How it is labelled*: outside text goes into the user turn inside `<untrusted_content source=… author=… author_verified=…>` with a fixed notice that it carries no authority; compatibility forms are folded (NFKC), control characters go, and the tag's name is renamed wherever the text carries it — any case, spacing or entity-escaped form — so the text cannot end its block and speak outside it.\n\nA turn about the principal's private mail (`Session.private_to_principal`) reaches the principal only. It is neither offered nor allowed to run `schedule_tools.PRIVATE_TURN_WITHHELD_TOOLS` — everything that reaches someone else, publishes where the team reads, or starts work outside the turn: the broadcasts, meeting bookings and cancellations, `run_workflow` and research, follow-ups and workflow nudges, watches, artifacts, goal updates, saved workflows, skill drafts, and `load_mcp_server` (it connects to any URL, which can carry the turn's content). Through `search_tools` and `call_tool` it reaches only `schedule_tools.PRIVATE_TURN_MCP_TOOLS`: Google Workspace reads (mail, calendar, and Drive, Docs and Sheets search and file reads, with their Microsoft 365 twins; OneDrive's `download-bytes` only for a file's content) and the Gmail send and draft the gateway narrows to the principal. A search shows it nothing else, and a call to anything else (Google Chat, a Docs or Drive write, `manage_event`, another server) is refused (`schedule_tools.private_turn_withholds`). A call the model emits anyway, MCP tools included, gets an error tool_result and writes an audit row, and every audit row the turn writes is the principal's alone. On every turn the Executive knows what is connected without searching: block 0's *Connected Systems* section names the connected Google services, other MCP servers, the messaging channels, Notion sync, Confluence sync (when on) and Act as me, and pins the Google tool names it calls directly with `call_tool`. `search_tools` is for anything not listed. When work repeats over many items, the Executive can also call `run_script` (`workflows/step_script.py`, offered when the gateway runs, unless `CHAT_SCRIPTS=false`, and only on the principal's own turns — `PRINCIPAL_ONLY_TOOLS`, withheld on an unattended pass too, because it spends the turn's whole call budget on code the model just wrote): one short Python script, run in a sandboxed Monty worker with no files, network or environment, whose only reach is the gateway tools `call_tool` could reach (those `search_tools` has returned; the gateway keeps that set for its own process-wide session, not per conversation). Each call it makes is a `call_tool` in every way that matters — the gateway's discovery, deny-list and recipient gates, Take the lead's gate on an unattended run, and the same action chip and `tool_invocation` row (marked `via: run_script`, `ok: false` when it raised), as each call happens; the script itself gets a `kind: script` row with its source and result. Its wall clock is half the turn's deadline, so a long script ends with the list of calls it made rather than being cut off with the turn, and a stopped turn closes it and its worker. A turn private to the principal is not offered it, and the Act as me lockdown withholds it whole. Besides gateway tools, a chat script may call a fixed list of the Executive's own tools (`step_script.CHAT_OWN_TOOLS`: reads, alerts, open loops, goals, the watch list, follow-ups, `message_person`, `upsert_person` — nothing that broadcasts, books calendars, edits memory or skills, or starts workflows or research), each only when the turn is offered it and through the turn's own handler (Take the lead's gate included), with the same chip and a `kind: skill`, `via: run_script` audit row. A direct call needs a `tool_use` each, so a turn's scripts share a budget instead: `CHAT_SCRIPT_MAX_CALLS` (default 200) calls in all, and per-turn caps on own tools that could otherwise act in bulk (`CHAT_OWN_TOOL_CAPS`: `message_person` 10 — messaging the whole roster would be a broadcast — `upsert_person` 25, `create_alert` 50); a refused call tells the model to say what is left. A call stopped mid-flight by the script's clock or the turn still gets a `cancelled` audit row, since it may have run. Because every round re-reads the conversation, the loop also nudges at the moment of choice: after a round whose results include a list of five or more items (`step_script.lists_many`), with `run_script` on offer and not used that round, it adds the constant `FANOUT_HINT` as a text block in that round's user message (never a cached block), once a turn, pointing the per-item work at one built tool. For library work on files (PDFs, spreadsheets, charts, images) the chat also has `run_python_job` (see Workflows → Python jobs): Python with a fixed library set in a Deno-isolated WebAssembly sandbox, fed Knowledge files by name, the message's own attachments as sent, or text it already holds, and able to keep a job that worked as a Custom tool, with result files linked for the principal to download; it is offered only on the principal's own turns and only where the sandbox is installed (not offered otherwise, so the tool prefix stays fixed per process). Agent Activity shows each run as one \"Built tool\" card (debug event `script_run`: ok, the calls it made, how many, how long, and the kept or reused tool's name). The persona's tool-building section is present only when chat scripts can actually run (the setting and Monty installed). `TOOL_BUILDING_ADDENDUM` in the cached persona tells the Executive when to build a tool — a job that looks something up and then checks or acts on each result, paged results, more than about ten items, or an answer that needs three or more reads (find, open, read; the tool returns only what answers the question, not whole documents, and only what is about to be edited or quoted at length is read in full), while a handful of items already named go as direct calls in one turn — and never to say it has none, and to talk about it as building a tool or doing it in one go. On the principal's own verified, interactive turns, a script that works can be saved as a named tool (`save_as`; elsewhere the script runs but is not saved) and run again with `run_script(tool=…, inputs=…)`; `list_saved_tools` (a constant definition, in `PRINCIPAL_ONLY_TOOLS` with `load_mcp_server`, so offered only while the principal is speaking on a verified, interactive surface; `run_script(tool=…)` is refused on any other turn) lists them, and each call a saved tool makes is a gateway call like any other (see Workflows → Saved tools). The tool's definition is a constant, so the cached tool prefix does not move. Multiple parallel `tool_use` blocks in a single assistant turn fan out concurrently through `router.route_parallel`, each running its own domain-filtered ChromaDB query (3 built-in and 2 company chunks by default, see the RAG section) before the specialist sees the question. The fan-out is failure-isolated per call: a specialist that raises or returns no text becomes an `UNAVAILABLE` tool_result (`router.specialist_unavailable_result`) instead of failing the whole turn, and its area — \"finance\", \"legal\", never the specialist's name — is reported to the user under the web reply, or by the Executive in its reply on other channels, unless another call to it answered that turn. A failed specialist does not count as consulted, so it takes no committee seat. Cancellation (Stop) still propagates. The Executive then synthesizes the returned analyses into one reply.\n\n`remember_fact`, `forget_fact` and `update_company_profile` (`orchestrator/fact_tools.py`) make a correction stick: the first two keep or retire a standing fact in `memory.facts`, and the third changes one field of the company profile. They share `record_decision_outcome`'s gate (the principal on a verified surface, an audited refusal for everyone else) and are in `UNATTENDED_WITHHELD_TOOLS`. `remember_fact` alone also admits a rostered teammate on a verified surface (web signed in, their own Slack or Discord): the fact is stored attributed (`recorded_by_role=\"teammate\"`, rendered `(per <name>)`), and is held as a `proposed` row (`awaiting_approval`) until the principal approves it, unless the principal marked that teammate trusted (\"needs my approval\" off; on by default) — and even then when it would replace the principal's fact. All three are also in `PRIVATE_TURN_WITHHELD_TOOLS`, since what they write or retire is read on everyone's turns. On the principal's own authenticated email (primary address, with Gmail's own `Authentication-Results` reporting `dmarc=pass`, not forwarded) the three tools check the request as usual, then hold it as `awaiting_confirmation`. It applies only when the principal replies to a one-time token emailed to their own address (`integrations.fact_confirmation`); the split into a check step and an `apply_*` step lets chat and the confirmed reply share one apply path. Each write also needs a `source_quote` found in the speaker's own words this turn. The active facts reach every specialist as a `<standing_facts>` user-turn tag: `route_to_specialist` reads them (None → the store, \"\" → none) for chat consults and workflow steps alike, and `BaseAgent.analyze` places the tag next to `<past_decisions>`, never in the cached specialist system prompt.\n\nThe ten entries in `SPECIALIST_REGISTRY` (`openexecutive.orchestrator.router`):\n\n| Key | Agent | Domain |\n| --- | --- | --- |\n| `cso` | `StrategyAgent` | strategy |\n| `cfo` | `FinanceAgent` | finance |\n| `chro` | `HRAgent` | hr |\n| `gc` | `LegalAgent` | legal |\n| `coo` | `OperationsAgent` | operations |\n| `cmo` | `MarketingAgent` | marketing |\n| `cpo` | `ProductAgent` | product |\n| `sales` | `SalesAgent` | sales |\n| `board_comms` | `BoardCommsAgent` | board |\n| `triage` | `TriageAgent` | triage |\n\nModels: specialists default to `claude-sonnet-5-5`; the deep-reasoning tier (CSO, CFO, GC, Board) runs on `DEEP_REASONING_MODEL`, `claude-sonnet-5-5` by default, with adaptive thinking; routing/fast classification uses `claude-haiku-5-5`. Requests are fitted to the model's generation on the way out (`feature_gate.fit_claude_generation`): 5-generation Claude models never receive `temperature` / `top_p` / `top_k`, and Haiku 5.5, which thinks unless told otherwise, gets thinking turned off when the caller asked for none. The Council's **Deep reasoning** toggle is honoured on Claude Opus/Sonnet (direct or via OpenRouter) and on any OpenRouter model whose live catalog entry supports `reasoning` (today all OpenAI, Gemini, DeepSeek and Grok entries; not Llama) — the request translator turns Anthropic adaptive thinking + effort into OpenRouter's `reasoning` parameter. It is a silent no-op on models that can't reason, and disabled for Haiku, which rejects adaptive thinking. A PDF `document` block is translated too: it becomes an OpenAI `file` part, and OpenRouter requests carrying one add the `file-parser` plugin (`native` for models whose catalog entry lists `file` input, else `PDF_OPENROUTER_ENGINE`, default `mistral-ocr`). `FeatureSpec.supports_pdf_input` gates it per model — off for a local server unless `LOCAL_PDF_INPUT` — and a gated-off PDF becomes a text note rather than vanishing. The Council picks a model in two steps: a **Provider** dropdown (Anthropic, OpenAI, Google, Meta, DeepSeek, xAI, Local, …), then a **Model** dropdown limited to that provider, with readable names (\"Claude Opus 5.5\") and the raw id and serving route shown underneath. The Anthropic group combines the direct Claude models (`claude-fable-5-1`, `claude-opus-5-5`, `claude-opus-5`, `claude-sonnet-5-5`, `claude-sonnet-5`, `claude-haiku-5-5`, `claude-haiku-4-5`) with any Claude generation that exists only on OpenRouter, and never lists a model twice. With `OPENROUTER_ENABLED`, the non-Anthropic entries come from OpenRouter's live `/models` catalog (fetched at startup, refreshed in the background, filtered to tool-capable paid text models from an allowlisted vendor set), falling back to a built-in snapshot if the fetch fails. Every Council agent also takes **additional instructions** (`agent_overrides.instructions`): text appended after its prompt by `agents.overrides.append_instructions` inside an `<additional_instructions>` block. `BaseAgent.base_system_prompt()` is the built-in prompt or its full replacement (what the Council's editor shows); `effective_system_prompt()` is that plus the instructions (what every call sends), and the Executive gets the same text through `build_system_blocks(persona_instructions=…)`. Unlike a replaced prompt, instructions leave the agent on its built-in prompt, so prompt improvements keep reaching it. A **Quality** choice above the agents (`agents.presets`) sets every agent's model and deep reasoning at once: **Fast** puts every agent on a small model (`claude-haiku-5-5`, else `claude-haiku-4-5`) with deep reasoning off, **Balanced** returns each agent to its built-in defaults, and **Thorough** puts the Executive and the domain specialists on the largest allowed model (`claude-opus-5-5`, else `claude-opus-5`) with deep reasoning on where the model supports it, leaving triage and the helpers below on their defaults. Presets change quality only; they never cap spend. Each preset resolves against the install's model allowlist at request time and is offered only when one of its models is allowed. Applying one writes ordinary per-agent overrides (model and deep reasoning only), so history, rollback and later per-agent edits keep working; an agent changed on its own afterwards shows as **Custom**. Each Council agent has a **visibility**: `core` for the Executive and the domain specialists (`BaseAgent.visibility = \"core\"`, which a new specialist must declare) and `internal` (the default) for triage and the helper agents below. Thorough moves only core agents. The Council opens in a **simple view**: the Quality choice, the voice picker, and only the core agents, each with its additional instructions and a model badge. **Show all agents** lists the internal ones too, and **Advanced** opens the full page (the browser remembers the choice). Eight Council-only agents are model/prompt-overridable in the UI but NOT callable via `consult_specialist`, because they are deliberately kept out of `SPECIALIST_REGISTRY`: `quality_judge` (committee review), `utility_fast` (gates, titles, decision parsing), `research` (the `executive_research` fan-out), `fixture_generator` (invents demo companies), `engagement_intake` (models a real client from intake material), `onboarding_interviewer` (interviews a new user about their own company and drafts its profile, people and departments), `workflow_designer` (runs the conversational New workflow wizard on `/jobs/new`: searches the available tools, asks clarifying questions, then drafts a validated dynamic-workflow definition for the user to review), and `workflow_actor` (runs a workflow's `action` steps: a bounded tool-use loop limited to the tools the user approved for that step). `fixture_generator`, `engagement_intake` and `onboarding_interviewer` share one shape — a forced tool call producing a whole company, validated for referential integrity, with a single repair retry — but opposite epistemics: the fixture author invents, the other two extract and leave unstated facts empty rather than guessing. When the workspace is in solo mode, `onboarding_interviewer`'s first user turn is prefixed with a short hint (`interview.solo_hint(role_kind)`, one static constant per role kind from the onboarding role step): the draft names only the principal, as principal, and no departments; an owner is asked about the offer, customers, pricing, cash and runway, an in-house executive about their organisation (drafted as the company — their employer), function, remit, manager, team size as context and what they are measured on, an independent about their practice, clients and pricing, and with no role given the role-neutral `SOLO_HINT` asks which they are. The hint rides in the user turn and the tool array is unchanged, so the cached system prompt and tool prefix are the same in both modes.\n\n**Voice.** The Executive answers in one voice persona, substituted into block 0 at `{VOICE_PERSONA}` (`personas/loader.get_active_body`) from the Executive override's `voice_persona_slug`. Three general voices ship in `personas/builtin/`: **Direct** (`default`, the slug used when none is set: answer first, brief and decisive), **Supportive** (a warm coach who explains the why and checks how the plan sits) and **Analytical** (numbers first, options and trade-offs, then one recommendation). They are offered by a shared picker in setup (`/onboard`, after the workspace choice) and in Settings → Executive, which saves through `PATCH /agents/executive`; a built-in voice's frontmatter carries the picker's one-line `description` and `sample` reply. The named-executive voices are marked `legacy: true`: the picker hides them, but they still load, so an install that picked one keeps it, and the Agent Council lists them under \"Legacy voices\". Voice files hold disposition only; the response-length ladder lives in `prompts/executive_persona.py`, and a test keeps numeric length rules out of every voice file. An eval scenario may set `voice_persona` to answer in a voice; it rides on `Session.voice_persona_slug`, which wins over the override for that session only.\n\nOne more structured-output utility agent sits beside `triage`: `openexecutive.agents.alert_review.AlertReviewAgent` (routing model, one forced `emit_alert_reviews` tool call, never raises). The alert review job hands it a batch of open alerts with evidence and gets back, per alert, a verdict (relevant / changed / resolved / stale), a confidence, the evidence it relied on, a one-line note, why-now, an optional deadline, and a recommended move (none / route / nudge / escalate / draft / suggest_workflow / merge / close) with its parameters. Like `triage`, it never mutates state — `alerts.review.apply_verdict` executes the verdicts deterministically within authority, so the model's job is judgment, not action.\n\n**Solo mode.** When the workspace (or the turn's session) is solo — one person using Open Executive for themselves — the Executive works for that one person, the principal, whatever their role: they may run their own business, lead a function inside a larger organisation, or work independently, and the persona reads which from the context instead of assuming it — the role lines under \"Your Principal\" in the org block, when the principal gave their role (workspace settings: kind, title, reports to, remit, measured on). Its persona swaps the team sections (\"You Are a Member of This Executive Team\", \"Choosing Who to Tell\", \"Departments & Org Coordination\") for \"You Work for One Person\", \"The People in Their World\", \"Who Hears From You\" and \"Goals and Areas\". Solo means only the principal uses Open Executive — not that they have no people: their manager, direct reports, peers, clients, investors and board are contacts. The Executive helps the principal work with them (a 1:1 or review prep, an update or budget case for their manager or board, a negotiation) but does not coordinate them, contacts them only when the principal asks and only if they are a contact, and never offers to loop the principal in. Every specialist stays relevant — HR for the principal's own reports, board and investor communications for reporting upward. Goals are grouped by *area* (the tools' `department_slug` is the area's slug). The chat loop withholds `send_department_message`, `send_company_broadcast` and `set_department_head` from its tool list, and a call to one of them anyway gets a JSON error tool_result instead of running. The mode is resolved once per turn and pinned on the session, so the persona, the org block, the tools and every tool handler agree — a switch made mid-turn applies from the next turn. Team mode is unchanged.\n\nEvery specialist *advises* — each returns domain analysis the Executive synthesizes into one voice. None of them act on the company's behalf; the side-effecting tools in the main loop (schedule, people, departments, broadcast, watchlist, workflows) belong to the Executive, not to a specialist.\n\n`recall_history` (Always in the loop, `orchestrator.history_tools`) reads the speaker's own notes, on the same terms as `ghostwrite_email` below: its own registry, offered only on a verified turn in a conversation private to them with their switch on; its `due` flag lists only dated notes, soonest first (see Memory). The ghostwriter below also gets the writer's notes from mail every recipient was on, as a consistency check (`<writer_noted>`).\n\nThe morning reflection, which has no question to look up and no specialist to consult, gets one more read-only tool of its own: `search_knowledge` (`orchestrator.knowledge_tools`), the same knowledge lookup a chat turn makes, for what the company's documents say about a customer, contract, price or policy before it acts.\n\nA fifth, `remind_me` (`orchestrator.reminder_tools`), sets a plain-text reminder sent to the speaker alone when it is due; it opens no mailbox, so it never locks a turn. A sixth, `propose_actions` (`orchestrator.action_card_tools`), does nothing itself: it leaves the speaker a card of up to five actions (a meeting, a message, a new contact) that happen only when they tap Approve, so it stays on after a turn read mail; it is where what an email itself asks for goes, since that is never done on the email's say-so. Four read-only tools beside it read the speaker's own mailbox on the same terms — `search_my_email`, `read_my_email`, `read_my_email_attachment`, `my_email_awaiting_reply` (`orchestrator.mail_read_tools`) — capped per turn, other people's messages and attached files returned as untrusted blocks (see Act as me). One tool writes under someone else's name, and it is fenced in code rather than the prompt. `ghostwrite_email` (**Act as me**, `orchestrator.delegation_tools`) writes an email as the person the Executive is speaking with and saves it as a draft in their own mailbox, Gmail or Outlook (`delegation.gmail.gmail_for` picks the client by the person's credential); chat never sends it (the person sends it themselves, or taps Send on its card). It lives in its own registry, never `_ALL_SKILL_TOOLS`, so reflection, research and workflow toolkits can never carry it. The chat loop adds it only on a turn `delegation.settings.pin_turn_delegation` offered it to — the principal, or a team member once the owner lets them, with it on, verified, in a conversation private to them (the web chat, a Slack or Discord DM, a verified private Telegram chat), on an attended turn — and builds that turn's handler map from both registries, so on any other turn a call to it is an unknown tool. From the round that calls it until the turn ends, the dispatch guard refuses only what opens an outside address, runs a script or workflow, or posts to everyone (`delegation.lockdown`), without changing the offered list; messages and invites to the roster, follow-ups, facts in the person's own words and a contact they name still run, each fenced in its own path. A later turn in the conversation stays private to its owner but holds nothing: the person asking again, having read the reply, is their own request. A contact add the lockdown refuses becomes an approval card in the chat instead (`propose_actions`). Recipients are chosen in code, never by the model (a draft's new email may go to any address, named back to check when it is neither on the roster nor typed), and the text comes from the **ghostwriter** (`delegation.ghostwriter`), a tool-less call with a constant prompt plus the person's \"How I write\" profile. The ghostwriter is deliberately not a Council agent: the Council's edit routes have no principal check. See **Act as Me (Delegation)**. A second tool joins the same way: `recall_history` (`orchestrator.history_tools`, Always in the loop) reads the speaker's own notes of what they said, by email or in chat. It has its own registry and per-turn handler map too, and the loop adds it only when `history_tools.recall_person` finds the speaker verified, in a conversation private to them, on an interactive turn that isn't private to the principal, with their notes switch on. It doesn't need Act as me. After every chat turn the Executive also hands the speaker's own words to `history_chat.schedule_chat_notes`, which takes notes in the background when their switch is on and the surface verified them (see **Memory**).\n\nThe Executive can also launch *any* workflow conversationally: `orchestrator.workflow_run_tools` adds `list_workflows` (the launchable catalog with each workflow's inputs) and `run_workflow` (runs a built-in or custom workflow to completion and returns its artifact, checkpointing approval gates as `awaiting_human`). `run_workflow` runs the weekly review, and in solo mode the morning brief, only for the principal on a verified surface (`Workflow.principal_only_modes`, `create_goal`'s rule); anyone else gets an audited refusal that does not name the rule. A run whose `result` event reports `private_to_principal` (the morning brief on the principal's own verified turn, when it read their private rows) still returns its artifact into that turn, but the stored run keeps `persistence.PRIVATE_RUN_ARTIFACT` instead of the text, since run history is readable by everyone signed in. This makes the `/jobs` catalog reachable from chat — see the Workflows section.\n\nDeliverables are first-class too. `orchestrator.artifact_tools` gives the Executive three tools: `draft_artifact` publishes a memo, web page (`html`), Word document (`docx`), spreadsheet (`xlsx`, from structured `sheets`) or a `link` to something it created in one of the principal's connected apps through the generic MCP gateway (`search_tools` → `call_tool`, any server — no vendor is named), and returns the artifact's id and link; `list_artifacts` and `get_artifact` let it find and reread its own past work (drafts and workflow outputs alike, through the same `artifact_records` read model the API uses). Every artifact belongs to someone: a draft to the person whose turn published it (`artifact_records.current_viewer`; with nobody speaking, the scheduler or the CLI, the principal), and a run someone started by hand to them, while scheduled runs are the team's. The tools, the recall, `message_person`'s link and mail attachments only ever reach the speaker's own (a link is refused when the recipient couldn't open it), not even the principal sees a teammate's, and a speaker who isn't on the People list publishes nothing. A workflow's own session inside someone's turn (the research synthesis) is pinned to that turn's viewer (`Session.documents_viewer`). Revising is `draft_artifact(supersedes=<id>)`, which archives the earlier version. When the principal asks for a deliverable the tool is always used; the \"quiet is the right default\" bar applies only to unprompted drafts. The format enum is a hard-coded sorted tuple so the cached tool schema stays byte-stable. Delivery: `message_person(person_id, text, artifact_id?)` appends the artifact's title and a `UI_BASE_URL/artifacts/<id>` link (through the same outbound guard as any DM; the link needs Open Executive access), and a Gmail send/draft can attach the file itself by `artifact_id` — see Integrations.\n\n**Extensions** (`orchestrator.extensions`). An install can add chat tools and kinds of document without changing this code. `OPENEXECUTIVE_EXTENSIONS` names Python modules. The first turn, or the first Documents read, imports each one and calls its `register()`. A module that fails is logged and skipped. `register_tool` takes an `ExtraTool`: a definition, an async handler, an optional `tool_groups` group (new or existing) and an optional chip. It refuses a name the loop already knows (`executive.builtin_tool_names`). An extension tool is never in `_ALL_SKILL_TOOLS`, so reflection, research and the synthesis never see it. The chat loop adds it per turn, like the Act as me tools, but never on an unattended run, a turn private to the principal or to a speaker who isn't on the People list (mail from a stranger), and only where its own `offered(session)` says. That predicate depends only on the kind of turn, so each kind of turn keeps one cached tool list. Act as me's lockdown refuses it once the turn has read the owner's mail, because it is unclassified and the lockdown fails closed. Its chip goes through `action_chips.summarize_action`, which applies the same no-chip rules for an error or refusal and keeps a link to a path inside the app. `register_collection` adds a kind of document: a drafted artifact tagged `collection:<name>` carries that collection on its record. The Documents page gives each collection its own filter. The collection's `on_change` hears about archive, restore and delete before they happen; if it raises, the change does not happen. `draft_artifact(supersedes=…)` refuses a collection's document. The alert pipeline strips `collection:` tags, so only the extension's own code files a document in its collection. A module that fails partway through `register()` leaves nothing registered. A tool registered with `workflow=True` is also offered to workflow action steps as `oe__<name>` (see Workflows); `read_only=True` marks one that only reads. A step runs with nobody in the turn, so `offered` isn't asked: `extensions.in_workflow_step(name)` is True for that tool, during its call only, and the handler decides what an unattended call may do.\n\nDocuments come in the other way through `read_document` (`orchestrator.document_tools`): it reads a PDF, Word, Excel, CSV, Markdown or text file by `path` — one another tool saved, such as a Gmail attachment (`get_gmail_attachment_content`) or a Drive file (`get_drive_file_download_url`) — or a company document by `filename`, with an optional PDF `pages` range. A scanned PDF is converted (`knowledge.pdf_reader`: local OCR by default; with `PDF_PROVIDER_READING=true`, the deployment's own model through its provider's PDF support, falling back to local OCR) rather than returned empty. Reads are confined to the tool download folders — the same resolve-then-contain check as `oe__read_file` — plus the company docs folder, and it is withheld on turns private to the principal (`PRIVATE_TURN_WITHHELD_TOOLS`): those come from a contact's email, whose own attachments the poller has already read into the turn, and must not steer the Executive into company documents.\n\nThe Executive also backs the **Ask OE** panel's form-fill tier via `orchestrator.form_tools.propose_form_values`: when a chat turn arrives with a `<page_context>` block describing a form on the user's screen, the Executive can propose values for it. The proposal is delivered to the UI as a `form_patch` SSE event and placed into the form as highlighted suggestions — the tool never saves or submits anything; the user reviews and saves manually. The tool is **always** in the tool list (tool definitions sit at the top of the cached prompt prefix, so a conditionally-present tool would invalidate the prompt cache whenever panel and non-panel turns alternate); its description restricts usage to turns that carry a form descriptor.\n\n**Take the lead** (`orchestrator.take_the_lead`) is how the Executive acts on its own, behind one gate. It has two switches, both off until someone turns them on. **As the Executive** is the principal's alone (`PUT /take-the-lead`). With it on, the unattended passes (reflection, research, the scheduler's proactive trigger) keep the acting tools solo mode otherwise withholds from them: booking, starting workflows, and messaging teammates rather than only the principal. `schedule_tools.unattended_toolkit` and the chat loop's unattended path then wrap every acting tool (`ACTING_TOOLS`, plus any MCP tool whose name sends, creates, updates, deletes or shares) in `gated_handlers` / `gated_call_tool`. The principal-only tools (`UNATTENDED_WITHHELD_TOOLS`) stay withheld, and every guard inside a handler still applies, including department approval levels (a booking in a department on Proposes is still a proposal) and the contact-egress rule. New mail or chat (any turn with an `origin_channel`) wakes the existing reflection early (`take_the_lead.wake`): one run waiting at most, five minutes after the first message and no more than every 30 minutes, as a one-off `take_the_lead_wake` scheduled action. Pause holds it like any scheduled work. **As you** is the Handle it for me dial's top step (delegation section).\n\n**The gate** is plain code (`take_the_lead.check`). Six kinds of action wait for a yes first, each with an Ask first switch, all on to start. The switch is the decision ledger's own mode for the class `take_the_lead_action:<kind>` (`propose` asks first, the default; `auto_execute` doesn't). The kinds are money, contracts and legal, hiring and people decisions, deleting/cancelling/sharing, someone new (an address on neither the roster nor the contacts) and big sends (a broadcast or department message, or more than five people). On top of these, the company (set by the principal) and each person (for their own As you) add structured rules: a person or address, a domain, words, or an amount at or above a figure. Added rules always hold, and an error reading them holds the action. The one thing the gate lets straight through is a direct message (`message_person` or a Slack, Telegram or Discord DM) whose only recipient is the principal (`_only_to_principal`): they are who it would wait for, so a card would only ask them to approve a message to themselves. Anything that reaches or acts on anyone else is checked, and a roster that can't be read is not the principal. A held action becomes a `take_the_lead_action` decision instance with a companion `DECISION_ALERT_SOURCE` card on Today, routed to the first teammate holding one of its People approval ranges (`take_the_lead.approval_ranges`, plain code, closest fit first: money to `spend_lt_2k`, `spend_lt_10k` or `spend_gt_10k` by its largest amount, with a higher range able to approve a lower one and no amount read as large, or to `customer_credit` for credit and payment terms; vendor contracts to `vendor_onboarding` then `legal_sign`, other contracts to `legal_sign`; hiring to `hiring_signoff`; anything else about the board or investors to `board_comms`; in Personal as in team mode), or else the principal through the authority gate's fallback (`departments.authority._route_proposal`). Deletes and shares always go to the principal. Only the principal and that approver see or resolve it (`DecisionClassSpec.principal_and_approver`). Approving it claims the row and carries out the exact call once (`take_the_lead.carry_out`: the unwrapped handler, or the gateway's `call_tool`, on an unattended session, with contact egress only when the principal approves). Declining drops it. Everything done or held is a row in `take_the_lead_log`. What it did and what was approved show in Recent activity (`today._build_activity`, kind `took_the_lead`), named but never quoted, since that feed is the team's. The held card's own `decision_resolved` row is left out because its payload quotes the message. `GET /today/activity` also adds the caller's own Handle it sends (kind `sent_as_you`, actor `you`), which are never in the shared feed.\n\n**In training.** The As the Executive switch has three positions: Off, **In training** and On (`take_the_lead.training`). In training, an action the gate would let through still waits, as a `training` hold for the principal alone, unless the principal has allowed that action with that person or thing (`take_the_lead.allowance`: \"Message Priya Nair\" covers a message to her on any channel; \"Book meetings with Priya Nair and Sam Lee\" meetings with exactly them; a workflow by id; a connected tool by name and the file it names, never one that reaches people such as a mail send or a share). The card carries Approve, **Approve + allow**, **Edit** (the message, a meeting's title, times or description, a task and its due date: `take_the_lead.apply_edits`) and Decline. Allowing is the principal's alone and needs a request the API can tie to them, as turning it on does; the six kinds and the added rules hold whatever is allowed. An edit approved with \"Do it like this next time\" is kept as an example (only the edited text, the principal's own words), and `take_the_lead.learned_note` hands the examples and the fact it's in training to the reflection and research passes in their user turn. After `SUGGEST_AFTER` (3) unchanged approvals of the same action in 30 days, Settings suggests allowing it. What it's learned is one list keyed by feature (`take_the_lead_allowed.feature`), so other features that learn from approvals can join it; each keeps its own switch. Act as me is the second (`act_as_me`, per person, `delegation.training`), with a training switch per setting (Replies, Follow-ups, Suggested actions, Drafts): its rows carry the person's id and the setting in their key, are listed and removed by that person alone, count toward their own cap, and never reach `learned_note`.",
+ "markdown": "The Executive (`openexecutive.orchestrator.executive.Executive`) is a single coherent persona that fans out to specialist sub-agents via Anthropic tool use. Each specialist subclasses `openexecutive.agents.base.BaseAgent`, carrying a `name`, `domain`, and `model`, and exposes `analyze(...)`, which wraps the question in user-turn context and calls Claude through the per-call provider. That context is per call and never cached: `<company_stage>` (the stage from the company profile — specialists never see the profile itself), in solo mode `<principal_role>` (what kind of principal the advice is for, in plain words, with their title and remit — `router.principal_role_context`), `<department_memory>`, `<past_decisions>`, `<failure_cases>`, `<relevant_knowledge>` and `<conversation_context>`, outermost first. The specialist's own system prompt is a static domain constant under one `cache_control` block, so it stays byte-identical across companies and profile edits. A chat turn takes the stage from its session's profile (the one the Executive reasons over, and the one an eval scenario injects); workflow steps and the MCP server's `consult_specialist` read it fresh from the profile on disk. The role tag follows the turn's mode: a chat turn pins the role on its session with the mode (`pin_turn_principal_role`: the session's `principal_role` override, else the workspace settings) and sends it in solo and nothing in team, so the org block, every specialist in the turn and any workflow the turn starts see the same role; callers with no chat turn read it fresh (`load_principal_role`: the current session's mode and role, else the workspace's), and a failed read sends no tag rather than failing the consult. The CFO, CSO and sales prompts key their advice on it: for a bootstrapped, self-funded or solo company they drop venture fundraising benchmarks and work from cash, and raise only if the founder wants to.\n\nRouting uses one `consult_specialist` tool. Beside it the Executive carries two skill tools, `search_skills` and `load_skill` (skills are shown to users as **playbooks**, and every playbook belongs to a workflow: one that follows it, or its own Quick workflow). Chat finds and reads playbooks but never creates or changes one: work worth repeating is offered as a workflow through `draft_workflow`, with the method written into its steps, because workflows read playbooks at run time (approved scheduled ones included) and this loop also runs on inbound email and chat, so only a person can change what a workflow follows. A hidden built-in is invisible to search and load; each `search_skills` hit lists the `workflows` that follow it, Quick ones included, and the persona routes \"the finished deliverable\" to that workflow and quick or partial asks to the playbook inline; Tool calls made while proactive work is tagged (scheduler dispatch, reflection, research, alert review) feed the Attunement outcome ledger. Goals have three tools: `list_department_goals`, `update_department_goal` (progress the user reports) and `create_goal` (a new goal the principal states, filed under an area by slug or title and creating the area when none matches — see Org). `create_goal` runs only for the principal on a verified surface, and no unattended run is offered it: the tools in `schedule_tools.UNATTENDED_WITHHELD_TOOLS` are dropped from reflection's and research's toolkits by `unattended_toolkit`, and from the chat loop when its Session is marked `unattended` — the scheduler's PROACTIVE TRIGGER run, whose prompt quotes stored intent text; a call the model emits anyway gets an error tool_result. `record_decision_outcome` records how a past decision turned out — the weekly review lists decisions older than 30 days with no outcome as `[decision N]` — writing `decisions.outcome` (`episodic.update_decision`) and auditing the rationale; it has `create_goal`'s gate (the principal on a verified surface only; everyone else gets an audited refusal) and is in `UNATTENDED_WITHHELD_TOOLS` too. A missing or unknown id returns the decisions still waiting on an outcome, so the model can retry with the right one. Its schema is static; the solo persona has one line on it, and team mode learns it from the tool description, so the team persona is byte-identical. Attunement adds `list_open_loops` (what a person owes — everyone's for the principal only in a private conversation, otherwise the caller's own) and `close_open_loop`, which only closes a loop when the turn's resolved speaker is the principal or the loop's owner — an unrostered sender, a signed-in web user not on the roster, and unattended sessions are refused. `assign_open_loop` opens one on request (\"ask Ben to send me the Q3 numbers by Friday\") for the principal or a rostered teammate on a verified surface — never an email, a turn private to the principal or an unattended run, which are also withheld the tool — and only when `task` appears verbatim in what the speaker typed this turn, at most 5 per turn; that turn's extraction pass then opens no second loop for the same owner. The roster write tools (`upsert_person`, `archive_person`, `set_department_head`) are the principal's alone, and only on a surface that verified the speaker — the web chat, the principal's own Slack or Discord, or a private Telegram chat when `TELEGRAM_WEBHOOK_SECRET` is set to a value Telegram can send (Telegram identifies the chat rather than the sender, and without that secret its webhook accepts anyone's update). A teammate, an email (even from the principal's address), Google Chat, the CLI, the MCP server and unattended runs get `status: refused` (audited, and no action chip), because a team member's People row is also the web sign-in and outbound-email allow-list. `upsert_person` takes `kind` (`team` / `contact`; a new person defaults to `contact` in solo mode, an update keeps the current kind) and `email_aliases` (other addresses the person writes from — the full list, replacing the current one; they match mail and may be emailed but never sign in, and one already on someone else is refused), and after every write any pending roster request whose sender now matches someone is closed and its held messages replayed. `resolve_roster_request` answers a **roster request** — someone not on the People list who wrote in and is waiting (see Org): `approve` (a new person; `kind` defaults to team for a company-domain sender, else contact, when the principal did not say), `link` (someone already listed, by `person_id`; `replace_channel_id` only when the principal said so) or `decline`. It has the roster tools' gate (the principal on a verified surface; audited refusal otherwise), is in `UNATTENDED_WITHHELD_TOOLS`, and answers only an id in `Session.trusted_roster_request_ids` — the ids the server printed that turn in a `<roster_requests>` block (`briefing.context.format_roster_requests_for_prompt`, up to 10 pending requests), which `render_and_trust` appends to the `<briefing>` digest on the principal's own verified turn only and records on the session in the same step, so an id the model invents or reads out of a message adds nobody. Each line is server-derived — id, channel, address or account id, the sender's name sanitised and marked unverified, a company-domain hint, the message count — and never what they wrote. The block rides in the **user turn**, like the digest, so the cached system blocks are unchanged; the tool definition is static, and the new tool only extends the sorted tool list. The web chat route resets the trusted set every turn, and a resolved id is removed from it. and `list_people` returns both kinds with `kind` — contacts only on the principal's own verified turn (they are private to the principal; anyone else gets the team as if no contact existed) — while `lookup_person` searches the team only. `ask_about_person` reads a person's peer memory fully only on their own verified turn; anyone else's answers only to a verified principal or teammate, and only when that person turned on \"Share my work style with the team\", framed to how they work and never their view of a colleague, and otherwise returns the same \"hasn't chosen to share\" result whether or not anything is known. The same verified-principal rule, as `people_tools.contacts_reachable_now()`, decides whether a turn may reach a **contact** at all: `message_person`, the raw DM tools, `create_calendar_event` / `create_instant_meeting` attendees and the MCP gateway's Gmail / Calendar / Drive / OneDrive gates treat a contact exactly like an unknown person on every other turn, so no refusal reveals that one exists. **The untrusted-content policy** (`orchestrator.content_trust`) is one rule for text the principal did not write — inbound email, chat, attached files, watched pages — answering three questions for every surface. *Who is speaking*: the principal only when the surface proved it (`principal_speaking`: the web chat as the principal — never a signed-in email on nobody's People entry —, the CLI, their own Slack/Discord or a secret-verified Telegram chat, or their own address with Gmail's `dmarc=pass`); never a teammate, a stranger, Google Chat, the MCP server or an unattended run. *What may become memory*: episodic extraction runs only then, and only on the words outside `<untrusted_content>` blocks; peer memory and open loops are recorded under the speaker's own name, never as the principal's. *What may trigger tools*: every write on the principal's authority already refuses others in its handler; the tools with no such check that change the install for every later turn — `content_trust.PRINCIPAL_ONLY_TOOLS`: `load_mcp_server`, `list_saved_tools`, `run_python_job` and `run_script` — are not offered unless the principal is speaking on an interactive surface (not email, even authenticated), and a call the model emits anyway gets an error tool_result and a `refused: not_principal` audit row. So an inbound email, a teammate's turn, Google Chat and the scheduler's proactive trigger each get a stable tool list of their own without it. *How it is labelled*: outside text goes into the user turn inside `<untrusted_content source=… author=… author_verified=…>` with a fixed notice that it carries no authority; compatibility forms are folded (NFKC), control characters go, and the tag's name is renamed wherever the text carries it — any case, spacing or entity-escaped form — so the text cannot end its block and speak outside it.\n\nA turn about the principal's private mail (`Session.private_to_principal`) reaches the principal only. It is neither offered nor allowed to run `schedule_tools.PRIVATE_TURN_WITHHELD_TOOLS` — everything that reaches someone else, publishes where the team reads, or starts work outside the turn: the broadcasts, meeting bookings and cancellations, `run_workflow` and research, follow-ups and workflow nudges, watches, artifacts, goal updates, saved workflows, skill drafts, and `load_mcp_server` (it connects to any URL, which can carry the turn's content). Through `search_tools` and `call_tool` it reaches only `schedule_tools.PRIVATE_TURN_MCP_TOOLS`: Google Workspace reads (mail, calendar, and Drive, Docs and Sheets search and file reads, with their Microsoft 365 twins; OneDrive's `download-bytes` only for a file's content) and the Gmail send and draft the gateway narrows to the principal. A search shows it nothing else, and a call to anything else (Google Chat, a Docs or Drive write, `manage_event`, another server) is refused (`schedule_tools.private_turn_withholds`). A call the model emits anyway, MCP tools included, gets an error tool_result and writes an audit row, and every audit row the turn writes is the principal's alone. On every turn the Executive knows what is connected without searching: block 0's *Connected Systems* section names the connected Google services, other MCP servers, the messaging channels, Notion sync, Confluence sync (when on) and Act as me, and pins the Google tool names it calls directly with `call_tool`. Both the service list and the pinned names are narrowed to what workspace-mcp actually loaded (`WORKSPACE_MCP_TOOLS`, `WORKSPACE_MCP_TOOL_TIER`), so a service left out is not named and its tools are not pinned — being told a tool is already available when it is not in the index costs a failed call the model cannot explain to itself. `search_tools` is for anything not listed. When work repeats over many items, the Executive can also call `run_script` (`workflows/step_script.py`, offered when the gateway runs, unless `CHAT_SCRIPTS=false`, and only on the principal's own turns — `PRINCIPAL_ONLY_TOOLS`, withheld on an unattended pass too, because it spends the turn's whole call budget on code the model just wrote): one short Python script, run in a sandboxed Monty worker with no files, network or environment, whose only reach is the gateway tools `call_tool` could reach (those `search_tools` has returned; the gateway keeps that set for its own process-wide session, not per conversation). Each call it makes is a `call_tool` in every way that matters — the gateway's discovery, deny-list and recipient gates, Take the lead's gate on an unattended run, and the same action chip and `tool_invocation` row (marked `via: run_script`, `ok: false` when it raised), as each call happens; the script itself gets a `kind: script` row with its source and result. Its wall clock is half the turn's deadline, so a long script ends with the list of calls it made rather than being cut off with the turn, and a stopped turn closes it and its worker. A turn private to the principal is not offered it, and the Act as me lockdown withholds it whole. Besides gateway tools, a chat script may call a fixed list of the Executive's own tools (`step_script.CHAT_OWN_TOOLS`: reads, alerts, open loops, goals, the watch list, follow-ups, `message_person`, `upsert_person` — nothing that broadcasts, books calendars, edits memory or skills, or starts workflows or research), each only when the turn is offered it and through the turn's own handler (Take the lead's gate included), with the same chip and a `kind: skill`, `via: run_script` audit row. A direct call needs a `tool_use` each, so a turn's scripts share a budget instead: `CHAT_SCRIPT_MAX_CALLS` (default 200) calls in all, and per-turn caps on own tools that could otherwise act in bulk (`CHAT_OWN_TOOL_CAPS`: `message_person` 10 — messaging the whole roster would be a broadcast — `upsert_person` 25, `create_alert` 50); a refused call tells the model to say what is left. A call stopped mid-flight by the script's clock or the turn still gets a `cancelled` audit row, since it may have run. Because every round re-reads the conversation, the loop also nudges at the moment of choice: after a round whose results include a list of five or more items (`step_script.lists_many`), with `run_script` on offer and not used that round, it adds the constant `FANOUT_HINT` as a text block in that round's user message (never a cached block), once a turn, pointing the per-item work at one built tool. For library work on files (PDFs, spreadsheets, charts, images) the chat also has `run_python_job` (see Workflows → Python jobs): Python with a fixed library set in a Deno-isolated WebAssembly sandbox, fed Knowledge files by name, the message's own attachments as sent, or text it already holds, and able to keep a job that worked as a Custom tool, with result files linked for the principal to download; it is offered only on the principal's own turns and only where the sandbox is installed (not offered otherwise, so the tool prefix stays fixed per process). Agent Activity shows each run as one \"Built tool\" card (debug event `script_run`: ok, the calls it made, how many, how long, and the kept or reused tool's name). The persona's tool-building section is present only when chat scripts can actually run (the setting and Monty installed). `TOOL_BUILDING_ADDENDUM` in the cached persona tells the Executive when to build a tool — a job that looks something up and then checks or acts on each result, paged results, more than about ten items, or an answer that needs three or more reads (find, open, read; the tool returns only what answers the question, not whole documents, and only what is about to be edited or quoted at length is read in full), while a handful of items already named go as direct calls in one turn — and never to say it has none, and to talk about it as building a tool or doing it in one go. On the principal's own verified, interactive turns, a script that works can be saved as a named tool (`save_as`; elsewhere the script runs but is not saved) and run again with `run_script(tool=…, inputs=…)`; `list_saved_tools` (a constant definition, in `PRINCIPAL_ONLY_TOOLS` with `load_mcp_server`, so offered only while the principal is speaking on a verified, interactive surface; `run_script(tool=…)` is refused on any other turn) lists them, and each call a saved tool makes is a gateway call like any other (see Workflows → Saved tools). The tool's definition is a constant, so the cached tool prefix does not move. Multiple parallel `tool_use` blocks in a single assistant turn fan out concurrently through `router.route_parallel`, each running its own domain-filtered ChromaDB query (3 built-in and 2 company chunks by default, see the RAG section) before the specialist sees the question. The fan-out is failure-isolated per call: a specialist that raises or returns no text becomes an `UNAVAILABLE` tool_result (`router.specialist_unavailable_result`) instead of failing the whole turn, and its area — \"finance\", \"legal\", never the specialist's name — is reported to the user under the web reply, or by the Executive in its reply on other channels, unless another call to it answered that turn. A failed specialist does not count as consulted, so it takes no committee seat. Cancellation (Stop) still propagates. The Executive then synthesizes the returned analyses into one reply.\n\n`remember_fact`, `forget_fact` and `update_company_profile` (`orchestrator/fact_tools.py`) make a correction stick: the first two keep or retire a standing fact in `memory.facts`, and the third changes one field of the company profile. They share `record_decision_outcome`'s gate (the principal on a verified surface, an audited refusal for everyone else) and are in `UNATTENDED_WITHHELD_TOOLS`. `remember_fact` alone also admits a rostered teammate on a verified surface (web signed in, their own Slack or Discord): the fact is stored attributed (`recorded_by_role=\"teammate\"`, rendered `(per <name>)`), and is held as a `proposed` row (`awaiting_approval`) until the principal approves it, unless the principal marked that teammate trusted (\"needs my approval\" off; on by default) — and even then when it would replace the principal's fact. All three are also in `PRIVATE_TURN_WITHHELD_TOOLS`, since what they write or retire is read on everyone's turns. On the principal's own authenticated email (primary address, with Gmail's own `Authentication-Results` reporting `dmarc=pass`, not forwarded) the three tools check the request as usual, then hold it as `awaiting_confirmation`. It applies only when the principal replies to a one-time token emailed to their own address (`integrations.fact_confirmation`); the split into a check step and an `apply_*` step lets chat and the confirmed reply share one apply path. Each write also needs a `source_quote` found in the speaker's own words this turn. The active facts reach every specialist as a `<standing_facts>` user-turn tag: `route_to_specialist` reads them (None → the store, \"\" → none) for chat consults and workflow steps alike, and `BaseAgent.analyze` places the tag next to `<past_decisions>`, never in the cached specialist system prompt.\n\nThe ten entries in `SPECIALIST_REGISTRY` (`openexecutive.orchestrator.router`):\n\n| Key | Agent | Domain |\n| --- | --- | --- |\n| `cso` | `StrategyAgent` | strategy |\n| `cfo` | `FinanceAgent` | finance |\n| `chro` | `HRAgent` | hr |\n| `gc` | `LegalAgent` | legal |\n| `coo` | `OperationsAgent` | operations |\n| `cmo` | `MarketingAgent` | marketing |\n| `cpo` | `ProductAgent` | product |\n| `sales` | `SalesAgent` | sales |\n| `board_comms` | `BoardCommsAgent` | board |\n| `triage` | `TriageAgent` | triage |\n\nModels: specialists default to `claude-sonnet-5-5`; the deep-reasoning tier (CSO, CFO, GC, Board) runs on `DEEP_REASONING_MODEL`, `claude-sonnet-5-5` by default, with adaptive thinking; routing/fast classification uses `claude-haiku-5-5`. Requests are fitted to the model's generation on the way out (`feature_gate.fit_claude_generation`): 5-generation Claude models never receive `temperature` / `top_p` / `top_k`, and Haiku 5.5, which thinks unless told otherwise, gets thinking turned off when the caller asked for none. The Council's **Deep reasoning** toggle is honoured on Claude Opus/Sonnet (direct or via OpenRouter) and on any OpenRouter model whose live catalog entry supports `reasoning` (today all OpenAI, Gemini, DeepSeek and Grok entries; not Llama) — the request translator turns Anthropic adaptive thinking + effort into OpenRouter's `reasoning` parameter. It is a silent no-op on models that can't reason, and disabled for Haiku, which rejects adaptive thinking. A PDF `document` block is translated too: it becomes an OpenAI `file` part, and OpenRouter requests carrying one add the `file-parser` plugin (`native` for models whose catalog entry lists `file` input, else `PDF_OPENROUTER_ENGINE`, default `mistral-ocr`). `FeatureSpec.supports_pdf_input` gates it per model — off for a local server unless `LOCAL_PDF_INPUT` — and a gated-off PDF becomes a text note rather than vanishing. The Council picks a model in two steps: a **Provider** dropdown (Anthropic, OpenAI, Google, Meta, DeepSeek, xAI, Local, …), then a **Model** dropdown limited to that provider, with readable names (\"Claude Opus 5.5\") and the raw id and serving route shown underneath. The Anthropic group combines the direct Claude models (`claude-fable-5-1`, `claude-opus-5-5`, `claude-opus-5`, `claude-sonnet-5-5`, `claude-sonnet-5`, `claude-haiku-5-5`, `claude-haiku-4-5`) with any Claude generation that exists only on OpenRouter, and never lists a model twice. With `OPENROUTER_ENABLED`, the non-Anthropic entries come from OpenRouter's live `/models` catalog (fetched at startup, refreshed in the background, filtered to tool-capable paid text models from an allowlisted vendor set), falling back to a built-in snapshot if the fetch fails. Every Council agent also takes **additional instructions** (`agent_overrides.instructions`): text appended after its prompt by `agents.overrides.append_instructions` inside an `<additional_instructions>` block. `BaseAgent.base_system_prompt()` is the built-in prompt or its full replacement (what the Council's editor shows); `effective_system_prompt()` is that plus the instructions (what every call sends), and the Executive gets the same text through `build_system_blocks(persona_instructions=…)`. Unlike a replaced prompt, instructions leave the agent on its built-in prompt, so prompt improvements keep reaching it. A **Quality** choice above the agents (`agents.presets`) sets every agent's model and deep reasoning at once: **Fast** puts every agent on a small model (`claude-haiku-5-5`, else `claude-haiku-4-5`) with deep reasoning off, **Balanced** returns each agent to its built-in defaults, and **Thorough** puts the Executive and the domain specialists on the largest allowed model (`claude-opus-5-5`, else `claude-opus-5`) with deep reasoning on where the model supports it, leaving triage and the helpers below on their defaults. Presets change quality only; they never cap spend. Each preset resolves against the install's model allowlist at request time and is offered only when one of its models is allowed. Applying one writes ordinary per-agent overrides (model and deep reasoning only), so history, rollback and later per-agent edits keep working; an agent changed on its own afterwards shows as **Custom**. Each Council agent has a **visibility**: `core` for the Executive and the domain specialists (`BaseAgent.visibility = \"core\"`, which a new specialist must declare) and `internal` (the default) for triage and the helper agents below. Thorough moves only core agents. The Council opens in a **simple view**: the Quality choice, the voice picker, and only the core agents, each with its additional instructions and a model badge. **Show all agents** lists the internal ones too, and **Advanced** opens the full page (the browser remembers the choice). Eight Council-only agents are model/prompt-overridable in the UI but NOT callable via `consult_specialist`, because they are deliberately kept out of `SPECIALIST_REGISTRY`: `quality_judge` (committee review), `utility_fast` (gates, titles, decision parsing), `research` (the `executive_research` fan-out), `fixture_generator` (invents demo companies), `engagement_intake` (models a real client from intake material), `onboarding_interviewer` (interviews a new user about their own company and drafts its profile, people and departments), `workflow_designer` (runs the conversational New workflow wizard on `/jobs/new`: searches the available tools, asks clarifying questions, then drafts a validated dynamic-workflow definition for the user to review), and `workflow_actor` (runs a workflow's `action` steps: a bounded tool-use loop limited to the tools the user approved for that step). `fixture_generator`, `engagement_intake` and `onboarding_interviewer` share one shape — a forced tool call producing a whole company, validated for referential integrity, with a single repair retry — but opposite epistemics: the fixture author invents, the other two extract and leave unstated facts empty rather than guessing. When the workspace is in solo mode, `onboarding_interviewer`'s first user turn is prefixed with a short hint (`interview.solo_hint(role_kind)`, one static constant per role kind from the onboarding role step): the draft names only the principal, as principal, and no departments; an owner is asked about the offer, customers, pricing, cash and runway, an in-house executive about their organisation (drafted as the company — their employer), function, remit, manager, team size as context and what they are measured on, an independent about their practice, clients and pricing, and with no role given the role-neutral `SOLO_HINT` asks which they are. The hint rides in the user turn and the tool array is unchanged, so the cached system prompt and tool prefix are the same in both modes.\n\n**Voice.** The Executive answers in one voice persona, substituted into block 0 at `{VOICE_PERSONA}` (`personas/loader.get_active_body`) from the Executive override's `voice_persona_slug`. Three general voices ship in `personas/builtin/`: **Direct** (`default`, the slug used when none is set: answer first, brief and decisive), **Supportive** (a warm coach who explains the why and checks how the plan sits) and **Analytical** (numbers first, options and trade-offs, then one recommendation). They are offered by a shared picker in setup (`/onboard`, after the workspace choice) and in Settings → Executive, which saves through `PATCH /agents/executive`; a built-in voice's frontmatter carries the picker's one-line `description` and `sample` reply. The named-executive voices are marked `legacy: true`: the picker hides them, but they still load, so an install that picked one keeps it, and the Agent Council lists them under \"Legacy voices\". Voice files hold disposition only; the response-length ladder lives in `prompts/executive_persona.py`, and a test keeps numeric length rules out of every voice file. An eval scenario may set `voice_persona` to answer in a voice; it rides on `Session.voice_persona_slug`, which wins over the override for that session only.\n\nOne more structured-output utility agent sits beside `triage`: `openexecutive.agents.alert_review.AlertReviewAgent` (routing model, one forced `emit_alert_reviews` tool call, never raises). The alert review job hands it a batch of open alerts with evidence and gets back, per alert, a verdict (relevant / changed / resolved / stale), a confidence, the evidence it relied on, a one-line note, why-now, an optional deadline, and a recommended move (none / route / nudge / escalate / draft / suggest_workflow / merge / close) with its parameters. Like `triage`, it never mutates state — `alerts.review.apply_verdict` executes the verdicts deterministically within authority, so the model's job is judgment, not action.\n\n**Solo mode.** When the workspace (or the turn's session) is solo — one person using Open Executive for themselves — the Executive works for that one person, the principal, whatever their role: they may run their own business, lead a function inside a larger organisation, or work independently, and the persona reads which from the context instead of assuming it — the role lines under \"Your Principal\" in the org block, when the principal gave their role (workspace settings: kind, title, reports to, remit, measured on). Its persona swaps the team sections (\"You Are a Member of This Executive Team\", \"Choosing Who to Tell\", \"Departments & Org Coordination\") for \"You Work for One Person\", \"The People in Their World\", \"Who Hears From You\" and \"Goals and Areas\". Solo means only the principal uses Open Executive — not that they have no people: their manager, direct reports, peers, clients, investors and board are contacts. The Executive helps the principal work with them (a 1:1 or review prep, an update or budget case for their manager or board, a negotiation) but does not coordinate them, contacts them only when the principal asks and only if they are a contact, and never offers to loop the principal in. Every specialist stays relevant — HR for the principal's own reports, board and investor communications for reporting upward. Goals are grouped by *area* (the tools' `department_slug` is the area's slug). The chat loop withholds `send_department_message`, `send_company_broadcast` and `set_department_head` from its tool list, and a call to one of them anyway gets a JSON error tool_result instead of running. The mode is resolved once per turn and pinned on the session, so the persona, the org block, the tools and every tool handler agree — a switch made mid-turn applies from the next turn. Team mode is unchanged.\n\nEvery specialist *advises* — each returns domain analysis the Executive synthesizes into one voice. None of them act on the company's behalf; the side-effecting tools in the main loop (schedule, people, departments, broadcast, watchlist, workflows) belong to the Executive, not to a specialist.\n\n`recall_history` (Always in the loop, `orchestrator.history_tools`) reads the speaker's own notes, on the same terms as `ghostwrite_email` below: its own registry, offered only on a verified turn in a conversation private to them with their switch on; its `due` flag lists only dated notes, soonest first (see Memory). The ghostwriter below also gets the writer's notes from mail every recipient was on, as a consistency check (`<writer_noted>`).\n\nThe morning reflection, which has no question to look up and no specialist to consult, gets one more read-only tool of its own: `search_knowledge` (`orchestrator.knowledge_tools`), the same knowledge lookup a chat turn makes, for what the company's documents say about a customer, contract, price or policy before it acts.\n\nA fifth, `remind_me` (`orchestrator.reminder_tools`), sets a plain-text reminder sent to the speaker alone when it is due; it opens no mailbox, so it never locks a turn. A sixth, `propose_actions` (`orchestrator.action_card_tools`), does nothing itself: it leaves the speaker a card of up to five actions (a meeting, a message, a new contact) that happen only when they tap Approve, so it stays on after a turn read mail; it is where what an email itself asks for goes, since that is never done on the email's say-so. Four read-only tools beside it read the speaker's own mailbox on the same terms — `search_my_email`, `read_my_email`, `read_my_email_attachment`, `my_email_awaiting_reply` (`orchestrator.mail_read_tools`) — capped per turn, other people's messages and attached files returned as untrusted blocks (see Act as me). One tool writes under someone else's name, and it is fenced in code rather than the prompt. `ghostwrite_email` (**Act as me**, `orchestrator.delegation_tools`) writes an email as the person the Executive is speaking with and saves it as a draft in their own mailbox, Gmail or Outlook (`delegation.gmail.gmail_for` picks the client by the person's credential); chat never sends it (the person sends it themselves, or taps Send on its card). It lives in its own registry, never `_ALL_SKILL_TOOLS`, so reflection, research and workflow toolkits can never carry it. The chat loop adds it only on a turn `delegation.settings.pin_turn_delegation` offered it to — the principal, or a team member once the owner lets them, with it on, verified, in a conversation private to them (the web chat, a Slack or Discord DM, a verified private Telegram chat), on an attended turn — and builds that turn's handler map from both registries, so on any other turn a call to it is an unknown tool. From the round that calls it until the turn ends, the dispatch guard refuses only what opens an outside address, runs a script or workflow, or posts to everyone (`delegation.lockdown`), without changing the offered list; messages and invites to the roster, follow-ups, facts in the person's own words and a contact they name still run, each fenced in its own path. A later turn in the conversation stays private to its owner but holds nothing: the person asking again, having read the reply, is their own request. A contact add the lockdown refuses becomes an approval card in the chat instead (`propose_actions`). Recipients are chosen in code, never by the model (a draft's new email may go to any address, named back to check when it is neither on the roster nor typed), and the text comes from the **ghostwriter** (`delegation.ghostwriter`), a tool-less call with a constant prompt plus the person's \"How I write\" profile. The ghostwriter is deliberately not a Council agent: the Council's edit routes have no principal check. See **Act as Me (Delegation)**. A second tool joins the same way: `recall_history` (`orchestrator.history_tools`, Always in the loop) reads the speaker's own notes of what they said, by email or in chat. It has its own registry and per-turn handler map too, and the loop adds it only when `history_tools.recall_person` finds the speaker verified, in a conversation private to them, on an interactive turn that isn't private to the principal, with their notes switch on. It doesn't need Act as me. After every chat turn the Executive also hands the speaker's own words to `history_chat.schedule_chat_notes`, which takes notes in the background when their switch is on and the surface verified them (see **Memory**).\n\nThe Executive can also launch *any* workflow conversationally: `orchestrator.workflow_run_tools` adds `list_workflows` (the launchable catalog with each workflow's inputs) and `run_workflow` (runs a built-in or custom workflow to completion and returns its artifact, checkpointing approval gates as `awaiting_human`). `run_workflow` runs the weekly review, and in solo mode the morning brief, only for the principal on a verified surface (`Workflow.principal_only_modes`, `create_goal`'s rule); anyone else gets an audited refusal that does not name the rule. A run whose `result` event reports `private_to_principal` (the morning brief on the principal's own verified turn, when it read their private rows) still returns its artifact into that turn, but the stored run keeps `persistence.PRIVATE_RUN_ARTIFACT` instead of the text, since run history is readable by everyone signed in. This makes the `/jobs` catalog reachable from chat — see the Workflows section.\n\nDeliverables are first-class too. `orchestrator.artifact_tools` gives the Executive three tools: `draft_artifact` publishes a memo, web page (`html`), Word document (`docx`), spreadsheet (`xlsx`, from structured `sheets`) or a `link` to something it created in one of the principal's connected apps through the generic MCP gateway (`search_tools` → `call_tool`, any server — no vendor is named), and returns the artifact's id and link; `list_artifacts` and `get_artifact` let it find and reread its own past work (drafts and workflow outputs alike, through the same `artifact_records` read model the API uses). Every artifact belongs to someone: a draft to the person whose turn published it (`artifact_records.current_viewer`; with nobody speaking, the scheduler or the CLI, the principal), and a run someone started by hand to them, while scheduled runs are the team's. The tools, the recall, `message_person`'s link and mail attachments only ever reach the speaker's own (a link is refused when the recipient couldn't open it), not even the principal sees a teammate's, and a speaker who isn't on the People list publishes nothing. A workflow's own session inside someone's turn (the research synthesis) is pinned to that turn's viewer (`Session.documents_viewer`). Revising is `draft_artifact(supersedes=<id>)`, which archives the earlier version. When the principal asks for a deliverable the tool is always used; the \"quiet is the right default\" bar applies only to unprompted drafts. The format enum is a hard-coded sorted tuple so the cached tool schema stays byte-stable. Delivery: `message_person(person_id, text, artifact_id?)` appends the artifact's title and a `UI_BASE_URL/artifacts/<id>` link (through the same outbound guard as any DM; the link needs Open Executive access), and a Gmail send/draft can attach the file itself by `artifact_id` — see Integrations.\n\n**Extensions** (`orchestrator.extensions`). An install can add chat tools and kinds of document without changing this code. `OPENEXECUTIVE_EXTENSIONS` names Python modules. The first turn, or the first Documents read, imports each one and calls its `register()`. A module that fails is logged and skipped. `register_tool` takes an `ExtraTool`: a definition, an async handler, an optional `tool_groups` group (new or existing) and an optional chip. It refuses a name the loop already knows (`executive.builtin_tool_names`). An extension tool is never in `_ALL_SKILL_TOOLS`, so reflection, research and the synthesis never see it. The chat loop adds it per turn, like the Act as me tools, but never on an unattended run, a turn private to the principal or to a speaker who isn't on the People list (mail from a stranger), and only where its own `offered(session)` says. That predicate depends only on the kind of turn, so each kind of turn keeps one cached tool list. Act as me's lockdown refuses it once the turn has read the owner's mail, because it is unclassified and the lockdown fails closed. Its chip goes through `action_chips.summarize_action`, which applies the same no-chip rules for an error or refusal and keeps a link to a path inside the app. `register_collection` adds a kind of document: a drafted artifact tagged `collection:<name>` carries that collection on its record. The Documents page gives each collection its own filter. The collection's `on_change` hears about archive, restore and delete before they happen; if it raises, the change does not happen. `draft_artifact(supersedes=…)` refuses a collection's document. The alert pipeline strips `collection:` tags, so only the extension's own code files a document in its collection. A module that fails partway through `register()` leaves nothing registered. A tool registered with `workflow=True` is also offered to workflow action steps as `oe__<name>` (see Workflows); `read_only=True` marks one that only reads. A step runs with nobody in the turn, so `offered` isn't asked: `extensions.in_workflow_step(name)` is True for that tool, during its call only, and the handler decides what an unattended call may do.\n\nDocuments come in the other way through `read_document` (`orchestrator.document_tools`): it reads a PDF, Word, Excel, CSV, Markdown or text file by `path` — one another tool saved, such as a Gmail attachment (`get_gmail_attachment_content`) or a Drive file (`get_drive_file_download_url`) — or a company document by `filename`, with an optional PDF `pages` range. A scanned PDF is converted (`knowledge.pdf_reader`: local OCR by default; with `PDF_PROVIDER_READING=true`, the deployment's own model through its provider's PDF support, falling back to local OCR) rather than returned empty. Reads are confined to the tool download folders — the same resolve-then-contain check as `oe__read_file` — plus the company docs folder, and it is withheld on turns private to the principal (`PRIVATE_TURN_WITHHELD_TOOLS`): those come from a contact's email, whose own attachments the poller has already read into the turn, and must not steer the Executive into company documents.\n\nThe Executive also backs the **Ask OE** panel's form-fill tier via `orchestrator.form_tools.propose_form_values`: when a chat turn arrives with a `<page_context>` block describing a form on the user's screen, the Executive can propose values for it. The proposal is delivered to the UI as a `form_patch` SSE event and placed into the form as highlighted suggestions — the tool never saves or submits anything; the user reviews and saves manually. The tool is **always** in the tool list (tool definitions sit at the top of the cached prompt prefix, so a conditionally-present tool would invalidate the prompt cache whenever panel and non-panel turns alternate); its description restricts usage to turns that carry a form descriptor.\n\n**Take the lead** (`orchestrator.take_the_lead`) is how the Executive acts on its own, behind one gate. It has two switches, both off until someone turns them on. **As the Executive** is the principal's alone (`PUT /take-the-lead`). With it on, the unattended passes (reflection, research, the scheduler's proactive trigger) keep the acting tools solo mode otherwise withholds from them: booking, starting workflows, and messaging teammates rather than only the principal. `schedule_tools.unattended_toolkit` and the chat loop's unattended path then wrap every acting tool (`ACTING_TOOLS`, plus any MCP tool whose name sends, creates, updates, deletes or shares) in `gated_handlers` / `gated_call_tool`. The principal-only tools (`UNATTENDED_WITHHELD_TOOLS`) stay withheld, and every guard inside a handler still applies, including department approval levels (a booking in a department on Proposes is still a proposal) and the contact-egress rule. New mail or chat (any turn with an `origin_channel`) wakes the existing reflection early (`take_the_lead.wake`): one run waiting at most, five minutes after the first message and no more than every 30 minutes, as a one-off `take_the_lead_wake` scheduled action. Pause holds it like any scheduled work. **As you** is the Handle it for me dial's top step (delegation section).\n\n**The gate** is plain code (`take_the_lead.check`). Six kinds of action wait for a yes first, each with an Ask first switch, all on to start. The switch is the decision ledger's own mode for the class `take_the_lead_action:<kind>` (`propose` asks first, the default; `auto_execute` doesn't). The kinds are money, contracts and legal, hiring and people decisions, deleting/cancelling/sharing, someone new (an address on neither the roster nor the contacts) and big sends (a broadcast or department message, or more than five people). On top of these, the company (set by the principal) and each person (for their own As you) add structured rules: a person or address, a domain, words, or an amount at or above a figure. Added rules always hold, and an error reading them holds the action. The one thing the gate lets straight through is a direct message (`message_person` or a Slack, Telegram or Discord DM) whose only recipient is the principal (`_only_to_principal`): they are who it would wait for, so a card would only ask them to approve a message to themselves. Anything that reaches or acts on anyone else is checked, and a roster that can't be read is not the principal. A held action becomes a `take_the_lead_action` decision instance with a companion `DECISION_ALERT_SOURCE` card on Today, routed to the first teammate holding one of its People approval ranges (`take_the_lead.approval_ranges`, plain code, closest fit first: money to `spend_lt_2k`, `spend_lt_10k` or `spend_gt_10k` by its largest amount, with a higher range able to approve a lower one and no amount read as large, or to `customer_credit` for credit and payment terms; vendor contracts to `vendor_onboarding` then `legal_sign`, other contracts to `legal_sign`; hiring to `hiring_signoff`; anything else about the board or investors to `board_comms`; in Personal as in team mode), or else the principal through the authority gate's fallback (`departments.authority._route_proposal`). Deletes and shares always go to the principal. Only the principal and that approver see or resolve it (`DecisionClassSpec.principal_and_approver`). Approving it claims the row and carries out the exact call once (`take_the_lead.carry_out`: the unwrapped handler, or the gateway's `call_tool`, on an unattended session, with contact egress only when the principal approves). Declining drops it. Everything done or held is a row in `take_the_lead_log`. What it did and what was approved show in Recent activity (`today._build_activity`, kind `took_the_lead`), named but never quoted, since that feed is the team's. The held card's own `decision_resolved` row is left out because its payload quotes the message. `GET /today/activity` also adds the caller's own Handle it sends (kind `sent_as_you`, actor `you`), which are never in the shared feed.\n\n**In training.** The As the Executive switch has three positions: Off, **In training** and On (`take_the_lead.training`). In training, an action the gate would let through still waits, as a `training` hold for the principal alone, unless the principal has allowed that action with that person or thing (`take_the_lead.allowance`: \"Message Priya Nair\" covers a message to her on any channel; \"Book meetings with Priya Nair and Sam Lee\" meetings with exactly them; a workflow by id; a connected tool by name and the file it names, never one that reaches people such as a mail send or a share). The card carries Approve, **Approve + allow**, **Edit** (the message, a meeting's title, times or description, a task and its due date: `take_the_lead.apply_edits`) and Decline. Allowing is the principal's alone and needs a request the API can tie to them, as turning it on does; the six kinds and the added rules hold whatever is allowed. An edit approved with \"Do it like this next time\" is kept as an example (only the edited text, the principal's own words), and `take_the_lead.learned_note` hands the examples and the fact it's in training to the reflection and research passes in their user turn. After `SUGGEST_AFTER` (3) unchanged approvals of the same action in 30 days, Settings suggests allowing it. What it's learned is one list keyed by feature (`take_the_lead_allowed.feature`), so other features that learn from approvals can join it; each keeps its own switch. Act as me is the second (`act_as_me`, per person, `delegation.training`), with a training switch per setting (Replies, Follow-ups, Suggested actions, Drafts): its rows carry the person's id and the setting in their key, are listed and removed by that person alone, count toward their own cap, and never reach `learned_note`.",
"mermaid": "flowchart TD\n User([User message]) --> Exec[\"Executive<br/>orchestrator\"]\n Exec -->|consult_specialist| Cso[\"cso: StrategyAgent\"]\n Exec -->|consult_specialist| Cfo[\"cfo: FinanceAgent\"]\n Exec -->|consult_specialist| Gc[\"gc: LegalAgent\"]\n Exec -->|consult_specialist| More[\"chro / coo / cmo / cpo / sales<br/>board_comms / triage\"]\n Mode{{\"Team or solo<br/>(persona + tools per turn)\"}} --> Exec\n Exec -->|\"ghostwrite_email (Act as me)\"| Gw[\"ghostwriter<br/>(tool-less, not a Council agent)\"]\n Cso --> Claude[\"Claude (Anthropic)\"]\n Cfo --> Claude\n Gc --> Claude\n More --> Claude\n Claude --> Exec\n class User entry\n class Exec compute\n class Mode compute\n class Cso compute\n class Cfo compute\n class Gc compute\n class More compute\n class Claude external\n class Gw compute",
"generated_at": "2026-10-09T23:01:59Z"
}
diff --git a/packages/core/openexecutive/architecture/prebuilt/caching.json b/packages/core/openexecutive/architecture/prebuilt/caching.json
index fe8907cb..e998b527 100644
--- a/packages/core/openexecutive/architecture/prebuilt/caching.json
+++ b/packages/core/openexecutive/architecture/prebuilt/caching.json
@@ -1,7 +1,7 @@
{
"section_id": "caching",
"title": "Prompt Caching",
- "markdown": "The Executive system prompt is partitioned for Anthropic prompt caching by `openexecutive.prompts.cache_manager.build_system_blocks`. It emits at most two `cache_control` blocks, consolidated down from four to stay within Anthropic's 4-block API limit alongside the tool definitions and the agent loop's intra-turn marker.\n\n- **Block 0 — persona + knowledge (1h TTL, always present):** the executive persona constant — `EXECUTIVE_PERSONA_PROMPT` for a team, `EXECUTIVE_PERSONA_SOLO_PROMPT` in solo mode — or an admin override, which wins as-is in either mode, with the `{VOICE_PERSONA}` substitution, then the admin's Council **additional instructions** for the Executive (`persona_instructions`, wrapped in `<additional_instructions>`; blank leaves the block byte-identical, and like the override it changes only on save), conditional `WEB_SEARCH_ADDENDUM` / `MCP_ADDENDUM`, the constant `TOOL_BUILDING_ADDENDUM` (when to build a tool with `run_script`, check `list_saved_tools` first, keep what will be needed again, and talk about it without saying \"script\"; present with a gateway unless `CHAT_SCRIPTS=false`, a setting, so constant per deployment), the identity addendum (exec name/email, impersonation guard), the *Connected Systems* section, the constant `DELEGATION_ADDENDUM` while anyone on the install has Act as me on, the timezone addendum, and `KNOWLEDGE_INDEX_SUMMARY`. The timezone addendum names the user's zone, read fresh on every build from the workspace settings (`memory.workspace_settings.get_user_timezone`: the workspace's zone, else `USER_TIMEZONE`, else UTC). Every component is a constant or changes only when an admin or the user changes a setting, so the TTL is long. Setting a new time zone changes the addendum, so the 1h block misses once on that change and is warm again from the next turn — accepted, since a stale zone would mis-schedule follow-ups. Act as me works the same way: its addendum is keyed on the install-level setting (`delegation.settings.block0_delegation_on`), never on the turn or the speaker, so switching it on or off misses the 1h block once; off, block 0 is byte-identical to an install without the feature.\n The *Connected Systems* section (`prompts.connected_systems.render_connected_systems`) tells the model what is connected so it never has to discover it: Google Workspace on or off (on only when `google_workspace` is among the servers the running gateway started with, `MCPGateway.server_names`, so a gateway that failed to start never reads as connected), and when on, how to act from its own account plus the pinned Google tool names by service (`GOOGLE_TOOL_MANIFEST`: Gmail, Calendar, Drive, Docs, Sheets, as workspace-mcp 1.29.0 names them) to call directly; any other configured MCP servers by name (`_comment` keys skipped); Slack / Discord / Telegram / Google Chat on or off from their settings (names only, never a token); Notion knowledge sync; a Confluence knowledge sync line only when that sync is on (a setting, so constant per deployment); and one constant Act as me line that points at `DELEGATION_ADDENDUM`, so switching Act as me still changes block 0 by that addendum alone. The \"never ask which address to send from\" guidance lives here now, shown only while Google is connected; the identity addendum keeps the name, address and impersonation guard. Every input is fixed per process and the output is sorted, so block 0 stays warm; it is appended even under a persona override. It replaces what the model used to learn from `search_tools`, an uncached proxy whose results were gone by the next turn.\n- **Block 1 — company + org (5m TTL, only if non-empty):** `company_profile.to_prompt_block()` plus `render_org_block(mode=…)` — for a team: departments, heads, principals, charters and the People roster; in solo: the principal's own line, their role lines when set (a static lead-in marking the role private to the principal, then kind in plain words, title, reports to, remit, measured on — from the role pinned for the turn: the workspace settings, or a session's `principal_role` override), and their goals grouped by area. The role is set once per install and changes only when the principal edits it, so it is as stable as the rest of the block. Shorter TTL so org edits propagate quickly. The principal's contacts are private to them, so `build_system_blocks(include_contacts=...)` adds the `## Contacts` section, after the team or solo block, only on the principal's own verified turn (`executive._contacts_in_prompt`). That gives block 1 exactly two stable variants per mode — with and without contacts — and never anything per-request; every other turn's block is byte-identical to an install with no contacts.\n\n## Solo mode: one prefix per mode\n\nSolo mode — one person using Open Executive for themselves — changes three cached parts, all resolved **once per turn** by `pin_turn_workspace_mode(session)` (the session's override, else the workspace setting) so they always agree: the block-0 persona, the block-1 org block, and the tool list, which drops `send_company_broadcast`, `send_department_message` and `set_department_head` *before* the sort, so it stays sorted and keeps its 1h marker on the last tool. The resolved mode is pinned on the session for the rest of the turn, so the tool handlers use it too and a switch made mid-turn applies from the next turn. Every part is a constant or stable per install, so each mode keeps its own warm cache and switching modes misses once.\n\nThe two personas share most of their text. `prompts/executive_persona.py` holds plain string literals — a shared head, body and tail plus team or solo sections — concatenated at import and never formatted, so both prompts are constants. The team prompt is byte-for-byte what it was before the split; a test pins its sha256.\n\nTool schemas are static: nothing in a definition depends on company data — `create_goal` takes its area as free text rather than an enum of the current departments, which would change the cached tool prefix every time an area was added, and `record_decision_outcome` takes a plain decision id. The solo persona is still one constant: its goal and decision-outcome lines (`_SOLO_CREATE_GOAL`, `_SOLO_DECISION_OUTCOME`) are literals concatenated at import, and the team persona is byte-identical (a test pins its sha256). The scheduler's proactive-trigger run (a Session marked `unattended`) is sent the list without `schedule_tools.UNATTENDED_WITHHELD_TOOLS`, so it has a stable tool prefix of its own, as each mode does. So does a turn private to the principal (`Session.private_to_principal`: mail from one of their contacts, or mail they forwarded): it is sent the list without `schedule_tools.PRIVATE_TURN_WITHHELD_TOOLS`, dropped before the sort, so private turns in a mode share one sorted prefix with its marker on the last tool. Act as me adds tools rather than dropping any: on a turn `delegation.settings.pin_turn_delegation` offered it to (the principal's own verified, attended turn in a private conversation, with it on), `ghostwrite_email`, the four mailbox reads (`search_my_email`, `read_my_email`, `read_my_email_attachment`, `my_email_awaiting_reply`), `remind_me` and `propose_actions` join the pool together before the sort, so those turns share one more sorted prefix and every other turn's list is byte-identical to before. The offer never depends on the Gmail connection, so a lapsed token never flips the prefix. `recall_history` (Always in the loop) joins the same way, before the sort, on a turn whose speaker may read their own notes; its offer reads only their switch and the surface, never the notes themselves. The chat note-taker runs after the turn, outside the cached prompt. Tool definitions are sorted by name so their serialized order is stable across requests; the `web_search` server tool is appended after the sorted list so it never receives a `cache_control` marker. The last client tool carries the third marker (1h TTL).\n\n## Tools opened on demand\n\nThe less common tools are not in the list at all. `orchestrator.tool_groups` puts them in named groups — calendar, people, goals, watch_and_alerts, workflows, documents, messaging, skills — and `tool_groups.split` takes them out of the turn's offered tools after every filter above (mode, unattended, private turn) and before the sort, adding two small tools in their place: `open_tools(group)`, whose description names each offered group and its tools, and `use_tool(name, input)`. When the model opens a group, the loop answers with that group's tools (name, description, input schema) as a **tool result**. It never adds them to the list: the list is the start of the cached prefix, so a list that grew mid-turn would miss the cache for the tools, both system blocks and the whole history. The list is the same on every call of every turn, and a group's descriptions are paid for only on a turn that opens it. That took the direct list from about 19,500 tokens to about 6,500, re-sent on every model call.\n\n`open_tools`'s description and its results name only the tools the turn offers (`tool_groups.open_tools_tool`), so a private turn never sees a withheld tool's name. The description depends only on that set, which is fixed for each kind of turn, so each still has one stable list. A `use_tool` call is rewritten into the tool's own call, same id, before dispatch (`tool_groups.unwrap`); the model's assistant block stays as it was sent, so the result pairs with the id it used. The rewritten call meets the same guards, handler, progress events and audit rows as a direct call. `use_tool` takes only names in `tool_groups.DEFERRED`; anything else, or an input that is not an object, gets an error tool result. The everyday tools (`consult_specialist`, `propose_form_values`, people and memory lookups, follow-ups, the MCP meta-tools) stay direct.\n\n## The intra-turn marker\n\nThe fourth and final breakpoint belongs to the tool-use loop. After each iteration appends its `tool_result` blocks, `executive._apply_loop_cache_marker` moves a single ephemeral marker onto the **newest non-empty tool result**. Without it the loop caches only the static tools+system prefix, so every iteration re-sends the whole accumulated transcript at full input price — the runtime signature is `cache_read_input_tokens` pinned flat while the prompt grows iteration over iteration.\n\nThe marker **moves rather than accumulates**: each pass first sweeps every `tool_result` clean, then re-marks the newest one. A marker left in place per iteration would reach `max_iterations` breakpoints and blow the 4-block limit. Sweeping first also makes the helper idempotent — there is no bookkeeping index to drift.\n\n## The history marker\n\nOnce a conversation has history, the agent loop also marks the **newest non-empty history turn** (normally the previous reply) with a **1h** breakpoint (`executive._apply_history_cache_marker`), so the next message reads the whole conversation back from cache and writes only what is new: the last exchange, this turn's context and its tool calls. History turns are rebuilt from plain text, so the previous turn's loop entries, which ended inside that turn's retrieved context and tool results, never match; without this marker every message re-wrote the whole conversation into the cache. 1h because people reply in minutes, not seconds.\n\nThe history marker takes the company block's slot: block 1 stays in the prefix, covered by the history marker, and dropping its 5m marker keeps the TTLs in the order the API requires, longer before shorter. On a first message (no history) block 1 keeps its 5m marker as before. `_build_messages` itself never marks history, so the committee revision call, which reuses those messages, keeps the plain layout. `Session.get_recent_history` moves its window's start in steps of 10 turns (keeping 20 to 30), so a long conversation's first history message, and the cached prefix with it, stays put between steps instead of shifting every turn. On the OpenRouter path the translator keeps assistant turns as typed text parts whenever a request carries the history marker (`translator._request_caches_assistant_text`), so the marker reaches Anthropic and every reply has the same shape whether or not it is the one marked. A rolling marker on every history turn used to claim a latent fifth breakpoint, which Anthropic rejects outright.\n\n**The breakpoint budget, in full:** last client tool (1h) + persona+knowledge (1h) + company+org (5m, first message only) or history (1h) + intra-turn tool result (5m) = at most 4, longer TTLs first.\n\nRAG context is **never** placed in a cached system block — retrieved hits vary per query and would poison the cache. It is injected into the **user turn** instead.\n\nSpecialist calls follow the same rule. Each specialist caches only its static domain prompt (`BaseAgent.analyze`, one ephemeral block); everything per call — retrieved knowledge, failure cases, past decisions, department memory, the company's stage (`<company_stage>`) and, in solo mode, the principal's role (`<principal_role>`) — goes in the specialist's user turn, so the cached domain prompt is byte-identical across companies and profile edits.\n\nThe ghostwriter's call (Act as me) carries no cache markers: its system prompt is the constant `GHOSTWRITER_PROMPT` plus that person's voice block — short, and sent only when a draft is written — and the thread and the intent go in its user turn.\n\n## Tool-result bound\n\n`TOOL_RESULT_MAX_CHARS` (default 20,000 characters, about 5,000 tokens) caps a single tool result before it enters the prompt, applied at the one point every result reaches the model rather than at each producer — so it also covers specialist output, tool errors and the unknown-tool fallback, while non-model consumers (the `propose_form_values` JSON parse, the audit trail) still read the full text. A turn's results also share one reading budget, `TOOL_RESULTS_TURN_MAX_CHARS` (default 40,000 characters; 0 turns it off): once the turn's results reach it, each further result shows its first 3,000 characters and its marker says, in place of the advice to ask again more narrowly (that re-read would be cut too), that the question has read as much as it can, so the model answers from what it has and says what it could not read, and it never edits or rewrites anything from that cut result, so a partial read never becomes a lossy write. The read it needs to edit may be the one that was cut, so each question gets one full re-read (`executive._TurnReadingBudget`): repeating a cut call with exactly the same input is answered with the result the turn already has, under the per-result cap, and the marker says so while that re-read is left. The tool never runs again, because the budget cuts any tool's result and a repeated write would happen twice. After it is used, the marker says to offer the edit as its own request, which starts with a fresh budget. A turn therefore reads at most the budget plus one more result (`executive._turn_result_limit`). A specialist's answer is the analysis itself, not a read, so it neither counts nor is cut. A built tool's (`run_script`) result counts but is never cut below the per-result cap: it lists the writes that already ran, and a short cut would hide some and get them repeated. Tool discovery (`open_tools`, `search_tools`) is treated the same way, since its result is the schema the next call needs. The per-result cap is a circuit breaker, not a routine clipper: ordinary tool output (a status, an email thread, a specialist's analysis) stays under it. A long document read does not, and the marker asks for a narrower request, such as a page range or a section, because a result stays in the prompt for every remaining call of the turn. A cut result carries an explicit marker naming the tool and stating that the result is incomplete — truncating silently is worse than not truncating, because the model reads the cut text as the whole answer.\n\n**The invariant:** never put dynamic content in a `cache_control` block. A single mutating value breaks the cache and drives roughly a 10x cost increase.",
+ "markdown": "The Executive system prompt is partitioned for Anthropic prompt caching by `openexecutive.prompts.cache_manager.build_system_blocks`. It emits at most two `cache_control` blocks, consolidated down from four to stay within Anthropic's 4-block API limit alongside the tool definitions and the agent loop's intra-turn marker.\n\n- **Block 0 — persona + knowledge (1h TTL, always present):** the executive persona constant — `EXECUTIVE_PERSONA_PROMPT` for a team, `EXECUTIVE_PERSONA_SOLO_PROMPT` in solo mode — or an admin override, which wins as-is in either mode, with the `{VOICE_PERSONA}` substitution, then the admin's Council **additional instructions** for the Executive (`persona_instructions`, wrapped in `<additional_instructions>`; blank leaves the block byte-identical, and like the override it changes only on save), conditional `WEB_SEARCH_ADDENDUM` / `MCP_ADDENDUM`, the constant `TOOL_BUILDING_ADDENDUM` (when to build a tool with `run_script`, check `list_saved_tools` first, keep what will be needed again, and talk about it without saying \"script\"; present with a gateway unless `CHAT_SCRIPTS=false`, a setting, so constant per deployment), the identity addendum (exec name/email, impersonation guard), the *Connected Systems* section (whose Google part is derived from `WORKSPACE_MCP_TOOLS` and `WORKSPACE_MCP_TOOL_TIER` — process-constant settings, so it is as cacheable as the rest), the constant `DELEGATION_ADDENDUM` while anyone on the install has Act as me on, the timezone addendum, and `KNOWLEDGE_INDEX_SUMMARY`. The timezone addendum names the user's zone, read fresh on every build from the workspace settings (`memory.workspace_settings.get_user_timezone`: the workspace's zone, else `USER_TIMEZONE`, else UTC). Every component is a constant or changes only when an admin or the user changes a setting, so the TTL is long. Setting a new time zone changes the addendum, so the 1h block misses once on that change and is warm again from the next turn — accepted, since a stale zone would mis-schedule follow-ups. Act as me works the same way: its addendum is keyed on the install-level setting (`delegation.settings.block0_delegation_on`), never on the turn or the speaker, so switching it on or off misses the 1h block once; off, block 0 is byte-identical to an install without the feature.\n The *Connected Systems* section (`prompts.connected_systems.render_connected_systems`) tells the model what is connected so it never has to discover it: Google Workspace on or off (on only when `google_workspace` is among the servers the running gateway started with, `MCPGateway.server_names`, so a gateway that failed to start never reads as connected), and when on, how to act from its own account plus the pinned Google tool names by service (`GOOGLE_TOOL_MANIFEST`: Gmail, Calendar, Drive, Docs, Sheets, as workspace-mcp 1.29.0 names them) to call directly; any other configured MCP servers by name (`_comment` keys skipped); Slack / Discord / Telegram / Google Chat on or off from their settings (names only, never a token); Notion knowledge sync; a Confluence knowledge sync line only when that sync is on (a setting, so constant per deployment); and one constant Act as me line that points at `DELEGATION_ADDENDUM`, so switching Act as me still changes block 0 by that addendum alone. The \"never ask which address to send from\" guidance lives here now, shown only while Google is connected; the identity addendum keeps the name, address and impersonation guard. Every input is fixed per process and the output is sorted, so block 0 stays warm; it is appended even under a persona override. It replaces what the model used to learn from `search_tools`, an uncached proxy whose results were gone by the next turn.\n- **Block 1 — company + org (5m TTL, only if non-empty):** `company_profile.to_prompt_block()` plus `render_org_block(mode=…)` — for a team: departments, heads, principals, charters and the People roster; in solo: the principal's own line, their role lines when set (a static lead-in marking the role private to the principal, then kind in plain words, title, reports to, remit, measured on — from the role pinned for the turn: the workspace settings, or a session's `principal_role` override), and their goals grouped by area. The role is set once per install and changes only when the principal edits it, so it is as stable as the rest of the block. Shorter TTL so org edits propagate quickly. The principal's contacts are private to them, so `build_system_blocks(include_contacts=...)` adds the `## Contacts` section, after the team or solo block, only on the principal's own verified turn (`executive._contacts_in_prompt`). That gives block 1 exactly two stable variants per mode — with and without contacts — and never anything per-request; every other turn's block is byte-identical to an install with no contacts.\n\n## Solo mode: one prefix per mode\n\nSolo mode — one person using Open Executive for themselves — changes three cached parts, all resolved **once per turn** by `pin_turn_workspace_mode(session)` (the session's override, else the workspace setting) so they always agree: the block-0 persona, the block-1 org block, and the tool list, which drops `send_company_broadcast`, `send_department_message` and `set_department_head` *before* the sort, so it stays sorted and keeps its 1h marker on the last tool. The resolved mode is pinned on the session for the rest of the turn, so the tool handlers use it too and a switch made mid-turn applies from the next turn. Every part is a constant or stable per install, so each mode keeps its own warm cache and switching modes misses once.\n\nThe two personas share most of their text. `prompts/executive_persona.py` holds plain string literals — a shared head, body and tail plus team or solo sections — concatenated at import and never formatted, so both prompts are constants. The team prompt is byte-for-byte what it was before the split; a test pins its sha256.\n\nTool schemas are static: nothing in a definition depends on company data — `create_goal` takes its area as free text rather than an enum of the current departments, which would change the cached tool prefix every time an area was added, and `record_decision_outcome` takes a plain decision id. The solo persona is still one constant: its goal and decision-outcome lines (`_SOLO_CREATE_GOAL`, `_SOLO_DECISION_OUTCOME`) are literals concatenated at import, and the team persona is byte-identical (a test pins its sha256). The scheduler's proactive-trigger run (a Session marked `unattended`) is sent the list without `schedule_tools.UNATTENDED_WITHHELD_TOOLS`, so it has a stable tool prefix of its own, as each mode does. So does a turn private to the principal (`Session.private_to_principal`: mail from one of their contacts, or mail they forwarded): it is sent the list without `schedule_tools.PRIVATE_TURN_WITHHELD_TOOLS`, dropped before the sort, so private turns in a mode share one sorted prefix with its marker on the last tool. Act as me adds tools rather than dropping any: on a turn `delegation.settings.pin_turn_delegation` offered it to (the principal's own verified, attended turn in a private conversation, with it on), `ghostwrite_email`, the four mailbox reads (`search_my_email`, `read_my_email`, `read_my_email_attachment`, `my_email_awaiting_reply`), `remind_me` and `propose_actions` join the pool together before the sort, so those turns share one more sorted prefix and every other turn's list is byte-identical to before. The offer never depends on the Gmail connection, so a lapsed token never flips the prefix. `recall_history` (Always in the loop) joins the same way, before the sort, on a turn whose speaker may read their own notes; its offer reads only their switch and the surface, never the notes themselves. The chat note-taker runs after the turn, outside the cached prompt. Tool definitions are sorted by name so their serialized order is stable across requests; the `web_search` server tool is appended after the sorted list so it never receives a `cache_control` marker. The last client tool carries the third marker (1h TTL).\n\n## Tools opened on demand\n\nThe less common tools are not in the list at all. `orchestrator.tool_groups` puts them in named groups — calendar, people, goals, watch_and_alerts, workflows, documents, messaging, skills — and `tool_groups.split` takes them out of the turn's offered tools after every filter above (mode, unattended, private turn) and before the sort, adding two small tools in their place: `open_tools(group)`, whose description names each offered group and its tools, and `use_tool(name, input)`. When the model opens a group, the loop answers with that group's tools (name, description, input schema) as a **tool result**. It never adds them to the list: the list is the start of the cached prefix, so a list that grew mid-turn would miss the cache for the tools, both system blocks and the whole history. The list is the same on every call of every turn, and a group's descriptions are paid for only on a turn that opens it. That took the direct list from about 19,500 tokens to about 6,500, re-sent on every model call.\n\n`open_tools`'s description and its results name only the tools the turn offers (`tool_groups.open_tools_tool`), so a private turn never sees a withheld tool's name. The description depends only on that set, which is fixed for each kind of turn, so each still has one stable list. A `use_tool` call is rewritten into the tool's own call, same id, before dispatch (`tool_groups.unwrap`); the model's assistant block stays as it was sent, so the result pairs with the id it used. The rewritten call meets the same guards, handler, progress events and audit rows as a direct call. `use_tool` takes only names in `tool_groups.DEFERRED`; anything else, or an input that is not an object, gets an error tool result. The everyday tools (`consult_specialist`, `propose_form_values`, people and memory lookups, follow-ups, the MCP meta-tools) stay direct.\n\n## The intra-turn marker\n\nThe fourth and final breakpoint belongs to the tool-use loop. After each iteration appends its `tool_result` blocks, `executive._apply_loop_cache_marker` moves a single ephemeral marker onto the **newest non-empty tool result**. Without it the loop caches only the static tools+system prefix, so every iteration re-sends the whole accumulated transcript at full input price — the runtime signature is `cache_read_input_tokens` pinned flat while the prompt grows iteration over iteration.\n\nThe marker **moves rather than accumulates**: each pass first sweeps every `tool_result` clean, then re-marks the newest one. A marker left in place per iteration would reach `max_iterations` breakpoints and blow the 4-block limit. Sweeping first also makes the helper idempotent — there is no bookkeeping index to drift.\n\n## The history marker\n\nOnce a conversation has history, the agent loop also marks the **newest non-empty history turn** (normally the previous reply) with a **1h** breakpoint (`executive._apply_history_cache_marker`), so the next message reads the whole conversation back from cache and writes only what is new: the last exchange, this turn's context and its tool calls. History turns are rebuilt from plain text, so the previous turn's loop entries, which ended inside that turn's retrieved context and tool results, never match; without this marker every message re-wrote the whole conversation into the cache. 1h because people reply in minutes, not seconds.\n\nThe history marker takes the company block's slot: block 1 stays in the prefix, covered by the history marker, and dropping its 5m marker keeps the TTLs in the order the API requires, longer before shorter. On a first message (no history) block 1 keeps its 5m marker as before. `_build_messages` itself never marks history, so the committee revision call, which reuses those messages, keeps the plain layout. `Session.get_recent_history` moves its window's start in steps of 10 turns (keeping 20 to 30), so a long conversation's first history message, and the cached prefix with it, stays put between steps instead of shifting every turn. On the OpenRouter path the translator keeps assistant turns as typed text parts whenever a request carries the history marker (`translator._request_caches_assistant_text`), so the marker reaches Anthropic and every reply has the same shape whether or not it is the one marked. A rolling marker on every history turn used to claim a latent fifth breakpoint, which Anthropic rejects outright.\n\n**The breakpoint budget, in full:** last client tool (1h) + persona+knowledge (1h) + company+org (5m, first message only) or history (1h) + intra-turn tool result (5m) = at most 4, longer TTLs first.\n\nRAG context is **never** placed in a cached system block — retrieved hits vary per query and would poison the cache. It is injected into the **user turn** instead.\n\nSpecialist calls follow the same rule. Each specialist caches only its static domain prompt (`BaseAgent.analyze`, one ephemeral block); everything per call — retrieved knowledge, failure cases, past decisions, department memory, the company's stage (`<company_stage>`) and, in solo mode, the principal's role (`<principal_role>`) — goes in the specialist's user turn, so the cached domain prompt is byte-identical across companies and profile edits.\n\nThe ghostwriter's call (Act as me) carries no cache markers: its system prompt is the constant `GHOSTWRITER_PROMPT` plus that person's voice block — short, and sent only when a draft is written — and the thread and the intent go in its user turn.\n\n## Tool-result bound\n\n`TOOL_RESULT_MAX_CHARS` (default 20,000 characters, about 5,000 tokens) caps a single tool result before it enters the prompt, applied at the one point every result reaches the model rather than at each producer — so it also covers specialist output, tool errors and the unknown-tool fallback, while non-model consumers (the `propose_form_values` JSON parse, the audit trail) still read the full text. A turn's results also share one reading budget, `TOOL_RESULTS_TURN_MAX_CHARS` (default 40,000 characters; 0 turns it off): once the turn's results reach it, each further result shows its first 3,000 characters and its marker says, in place of the advice to ask again more narrowly (that re-read would be cut too), that the question has read as much as it can, so the model answers from what it has and says what it could not read, and it never edits or rewrites anything from that cut result, so a partial read never becomes a lossy write. The read it needs to edit may be the one that was cut, so each question gets one full re-read (`executive._TurnReadingBudget`): repeating a cut call with exactly the same input is answered with the result the turn already has, under the per-result cap, and the marker says so while that re-read is left. The tool never runs again, because the budget cuts any tool's result and a repeated write would happen twice. After it is used, the marker says to offer the edit as its own request, which starts with a fresh budget. A turn therefore reads at most the budget plus one more result (`executive._turn_result_limit`). A specialist's answer is the analysis itself, not a read, so it neither counts nor is cut. A built tool's (`run_script`) result counts but is never cut below the per-result cap: it lists the writes that already ran, and a short cut would hide some and get them repeated. Tool discovery (`open_tools`, `search_tools`) is treated the same way, since its result is the schema the next call needs. The per-result cap is a circuit breaker, not a routine clipper: ordinary tool output (a status, an email thread, a specialist's analysis) stays under it. A long document read does not, and the marker asks for a narrower request, such as a page range or a section, because a result stays in the prompt for every remaining call of the turn. A cut result carries an explicit marker naming the tool and stating that the result is incomplete — truncating silently is worse than not truncating, because the model reads the cut text as the whole answer.\n\n**The invariant:** never put dynamic content in a `cache_control` block. A single mutating value breaks the cache and drives roughly a 10x cost increase.",
"mermaid": "flowchart TD\n Tools[\"Tool defs (sorted by name)<br/>less common ones behind open_tools<br/>last one: cache_control 1h\"] --> B0\n Mode{{\"effective_workspace_mode<br/>(team or solo, once per turn)\"}} --> Tools\n Mode --> B0\n Mode --> B1\n Deleg{{\"Act as me on?<br/>(install-level, not per turn)\"}} --> B0\n Offer{{\"ghostwrite_email offered?<br/>(pinned once per turn)\"}} --> Tools\n B0[\"Block 0: persona (team or solo) + knowledge<br/>cache_control 1h\"] --> B1\n B1[\"Block 1: company + org (team or solo render)<br/>cache_control 5m on a first message\"] --> Limit{{\"≤ 4 cache_control blocks\"}}\n Loop[\"Newest tool_result<br/>cache_control 5m (moves each iteration)\"] --> Limit\n Hist[\"Newest history turn<br/>cache_control 1h (replaces block 1 marker)\"] --> Limit\n Limit --> Turn[\"User turn\"]\n RAG[\"RAG hits<br/>(vary per query)\"] --> Turn\n Turn --> API[\"Anthropic API\"]\n class B0 cache\n class B1 cache\n class RAG hot\n class Turn hot\n class Tools compute\n class Mode compute\n class Loop cache\n class API external\n class Deleg compute\n class Offer compute",
"generated_at": "2026-10-09T06:10:00Z"
}
diff --git a/packages/core/openexecutive/architecture/prebuilt/integrations.json b/packages/core/openexecutive/architecture/prebuilt/integrations.json
index 2a2f63f3..7116c376 100644
--- a/packages/core/openexecutive/architecture/prebuilt/integrations.json
+++ b/packages/core/openexecutive/architecture/prebuilt/integrations.json
@@ -1,7 +1,7 @@
{
"section_id": "integrations",
"title": "Integrations",
- "markdown": "External channels connect callers to the Executive. Each inbound adapter resolves the sender to a Person — the chat adapters via `find_person_by_{slack,discord,telegram}_id`, the email poller via `people.identity.resolve_email_sender` (a primary address or an alias, then the company-domain rule; see Org) — and threads `person_id` into `Executive.chat()`. A sender who resolves to nobody is no longer simply dropped: see **New senders (roster requests)** below.\n\n### Inbound channels\n\n- **Slack** (`integrations.slack_bot`) — async Bolt socket-mode (`AsyncApp` + `AsyncSocketModeHandler`) embedded in the lifespan; responds to `@mention`s, DMs, and thread continuations, with per-conversation session memory (see **Conversation state** below). Text only: there is no interactivity endpoint and no Block Kit, so nothing Slack sends carries a button.\n- **Discord** (`integrations.discord_bot`) — discord.py gateway embedded in the lifespan; per-thread auto-titling, session-scoped memory, author attribution, configurable `@mention` thread promotion.\n- **Telegram** — webhook gating + per-chat history.\n- **Google Chat** — webhook + service-account auth. It resolves no approval gates and has no roster gate (the webhook authenticates Google, not the sender), and it never populates the session's trusted alert ids, so `ack_alert`'s trust check — which now runs on every session rather than only those declaring an `origin_channel` — refuses there by default, and its `<channel>` block says plainly that a sign-off cannot be answered from it.\n- **Email** (`integrations.email_poller`) — MCP poll loop over the Executive's own mailbox — Gmail via `workspace-mcp` or Outlook via `ms-365-mcp-server`, chosen by `EMAIL_PROVIDER` (see *Workspace providers* below); it starts only once the MCP gateway is up, never surfaces Reply-To redirects, filters auto-replies. Its sessions carry `origin_channel=\"email\"`: the untrusted-content policy (`orchestrator.content_trust`) reads that tag to tell mail from the principal's own surfaces — left empty, as it once was, every stranger's body reached the decision extractor as the principal's words. The mail goes into the turn inside an `<untrusted_content source=\"email\" author=\"<sender>\" author_verified=\"no\">` block that says it carries none of the principal's authority, with the framing line and any `[POLICY]` notice outside it; mail from the principal's primary address that Gmail authenticated (`dmarc=pass`) is their own words — but only their new text goes in unwrapped: when the mail quotes or forwards another message (`_new_text_lines` cut something), the message as received follows in an `<untrusted_content source=\"email_quoted\">` block, since a stranger's text they replied to or passed on is not theirs. While the Executive is paused (`scheduler.pause`) it skips each poll and leaves the inbox unread, so held mail is processed on the first poll after resume; the chat adapters above are not paused. The Executive sees the whole message; peer memory records only `_email_memory_text` — the subject, the sender's new text (cut at the first quoted or forwarded message) and the attachment filenames — parsed from the rendered message with any trailing `--- REPLY ---` block dropped first, without backtracking-prone regexes, since the text is sender-controlled, and only for a rostered sender, the only kind with a peer to record into. The poller's `integration_inbound` row keeps the mail's text for the audit log, the way the reply's `chat_turn` row keeps the Executive's answer: `full.message` is the sender's new text (what web chat keeps for a turn), `full.attachments` the attachment filenames, and `details.preview` the first 120 characters of the new text, so /audit shows what was asked next to what was answered (the summary still reads `Inbound email from <sender>: <subject>`). The whole body with its quoted chain (`full.body`) is kept only on a row private to the principal: the chain is other people's mail — a contact's, on the principal's reply to one — and every signed-in user reads a public row. Each field is cut by its escaped JSON length (20,000 characters of text, 8,000 of attachment names), the measure the logger's 64 KB cap uses, so a long mail in a non-Latin script loses its tail rather than the row's structure. Outbound, the principal briefs and the client-rotation digest are emailed through the same gateway with the configured backend's send tool (`google_workspace__send_gmail_message` or `microsoft_365__send-mail`, via `workspace.registry.send_from_executive`, from the Executive's own mailbox to the principal's address, subject `<brief title> — <date>`, the date in the user's zone) as formatted HTML — rendered from the brief's Markdown with raw HTML and images off, so text quoted from outside shows as text and opening the email loads nothing remote. Email is first when the principal prefers it, and otherwise the backup after every chat channel, tried only when Slack, Discord and Telegram are missing or fail; an in-band error from the tool counts as a failed send and the next channel is tried. An emailed brief is not a DM, so the per-Person outbound guard below does not apply — the gateway's roster egress gate does.\n\n Two senders get their own framing. Mail from one of the principal's **contacts** (a `kind=\"contact\"` People row — see Org) replaces the non-roster `[POLICY]` notice with a contact one: it names them (name and role/company), says not to reply unless the principal asks and not to tell anyone else, and asks for a summary for the principal instead. Both this turn and the forwarded-by-principal turn below are marked `Session.private_to_principal`: they can email, DM or invite only the principal, an alert they raise is private to the principal, and they are not offered the tools that reach anyone else or start work outside the turn (`schedule_tools.PRIVATE_TURN_WITHHELD_TOOLS`: broadcasts, meeting bookings, workflows, follow-ups, watches, artifacts, goal updates, skill drafts, connecting a new MCP server), and of the MCP tools only Google Workspace reads and the gated Gmail send (see MCP below). Every audit row about such mail is private to the principal (see Audit Log), the poller's two `integration_inbound` rows included (`_private_to_principal_mail`), so the audit log's other readers see none of it. The poller holds `audit.private_rows` for the whole handling of such a mail, because some rows are written before the turn binds its private session: the knowledge retrieval quotes the mail's From line, subject and opening. The rows themselves read exactly like an outside sender's: the `integration_inbound` row is the non-roster one (`accepted_non_roster`), and the notice opens with the non-roster notice's exact sentence. The poller handles each message with no session bound, so one message's private turn never marks the next message's rows. The turn stays unattributed (`person_id` None — no peer memory, no co-presence), and the gateway refuses mail to a contact on a poller turn anyway. Mail **forwarded by the principal** — the sender resolves to the principal and `_new_text_lines` reports a forwarded message — is wrapped in a `<forwarded_by_principal>` block: summarise it, draft a reply the principal could send (as text in the reply, never sent or saved as a Gmail draft), note any commitment or date, and offer to add the original sender as a contact (confirmed from the web app or the principal's Slack / Discord, since an ordinary email turn cannot change the People list — the one email that can is the principal's tokened reply to a roster-request email, which is parsed before any model turn; see below). The reply still goes to the principal only. The From header is unauthenticated, so this framing grants nothing: a spoofed \"forward\" just gets the real principal a summary.\n\nSlack, Discord, and Telegram are roster-gated on inbound: an unrostered sender's message never reaches the Executive (`details.outcome=\"rejected_unknown_sender\"`). The gates look up **team members only** — a contact's Slack, Discord or Telegram id is treated as unknown, so a contact can never talk to the bot (nor count as a co-present peer). What happens next is the roster request below — except for a contact, a thread continuation, a Telegram group or an unverified Telegram webhook, which stay silent. Email is not inbound-gated — non-roster mail flows through with a `[POLICY]` notice so triage can classify it. Slack/Discord thread continuations pass the shared `integrations.response_gate` (utility-fast model, biased toward replying). Slack only continues a thread it has replied in: it reads the thread with `conversations_replies` (5 s bound) and looks for its own user id, which `auth_test` resolves at startup. If that lookup fails, mentions and DMs still work and thread continuations stop until a retry succeeds (`slack_bot._resolve_bot_user_id`, at most once a minute, from the channel `message` listener). When Slack doesn't return the thread in time, the thread's stored history (`slack:thread:{channel}:{thread_ts}`, written on every turn the bot answers there) stands in: with history, a rostered sender gets an ephemeral \"I couldn't read this thread just now… @-mention me\" and the drop is audited (`outcome=dropped_thread_unreadable`). A thread the bot never answered in stays silent. A message the gate skips is audited (`outcome=skipped_gate`, with the gate's reason) and gets an :eyes: reaction (`reactions_add`, which needs the bot's `reactions:write` scope; without it the reaction is skipped), so a sender the gate misjudged can see the message was read and @-mention the bot.\n\n**No link previews on replies.** The Executive's replies go out with previews off — Telegram `sendMessage` carries `link_preview_options.is_disabled`, Slack's reply `say` passes `unfurl_links=False` / `unfurl_media=False`, and every Discord reply send passes `suppress_embeds=True`. The direct messages the Executive starts are the same: `send_slack_dm` posts with `unfurl_links=False` / `unfurl_media=False`, and Discord's `send_dm` sets the `SUPPRESS_EMBEDS` flag on each chunk, so a reminder or a brief sent by DM carries no preview either. Otherwise the chat service fetches a link in the text on its own, so a reply that quotes someone's mail (Act as me reads it) could carry data to a sender's server with no click. The web chat does the same for images: `Message.tsx` loads only an image the app itself serves (`lib/markdownImages.loadsInline`) and shows any other as a link.\n\n### New senders (roster requests)\n\nSomeone who writes in and is not on the People list — and is not one of the principal's contacts — is **held, acknowledged and asked about** instead of dropped (`integrations.roster_intake.intake` over the `people.roster_requests` ledger; lifecycle in Org). Each adapter calls it after its gate:\n\n- **Slack** — a DM or an `@mention` (not a thread continuation: the bot was not addressed there). The name and profile email come from `users_info` (5 s timeout). The acknowledgement goes in the DM, or as an **ephemeral** message for a mention, so nobody else in the channel learns who is and is not on the roster. The held event keeps only a fixed set of keys (text, user, channel, ts, thread, files).\n- **Discord** — a DM or a mention (not a thread the bot merely follows). The acknowledgement goes in the DM, or by DM for a channel mention; `/ask` answers the interaction with an ephemeral acknowledgement first (Discord allows 3 s), then holds the prompt. Its old \"ask the admin to add you\" reply is gone.\n- **Telegram** — only a **private** chat, and only when `TELEGRAM_WEBHOOK_SECRET` is set and valid: Telegram identifies the chat rather than the sender, and without the secret anyone could forge an update. A group, or an unverified webhook, stays silent as before. Attachment file ids are held and downloaded only on replay.\n- **Email** — non-roster mail that is not private to the principal (so never a contact's) and does not look machine-sent (`looks_automated`: a no-reply / notifications / billing-style local part, a `List-Id` / `List-Unsubscribe` / `Auto-Submitted` / bulk `Precedence` header, or a DMARC failure). The mail is **still triaged** by the Executive on this pass; its `[POLICY]` notice now says the sender was already told their message is waiting and the principal was asked who they are, so the model raises no second alert just to add them, and that the email will come back once they are added.\n\nEvery path skips a contact first (`find_person_by_*_id(..., include_contacts=True)`, or the email poller's private-mail rule): a contact has no chat access and gets no request and no acknowledgement, and the team-visible `rejected_unknown_sender` / `accepted_non_roster` rows read the same for a contact as for a stranger.\n\n**The acknowledgement** is one fixed text (`roster_intake.ACK_TEXT` — \"your message arrived … waiting for confirmation\"), never the model and never an echo of what they wrote, so a forged sender gains nothing by triggering it. It is sent at most once per sender per `ROSTER_ACK_WINDOW_DAYS` (7) and at most `ROSTER_ACK_DAILY_CAP` (20) a day across all senders (`claim_ack`, one transaction; a failed send releases its claim). An email acknowledgement is the one send the MCP gateway lets reach an address off the roster: `send_email_ack` sets `mcp_gateway.roster_ack_grant`, a one-shot ContextVar grant that admits a single `send_gmail_message` with exactly the granted recipient (a plain string, no list), subject and body and no other key, then is spent (audited `integration_outbound`, `kind: roster_ack`, private). An email sender is acknowledged only when Gmail authenticated them: once `claim_ack` succeeds, the poller's `send_ack` re-reads the mail as raw MIME and needs Gmail's own topmost `Authentication-Results` reporting `dmarc=pass` for the From domain (`fact_confirmation.sender_authenticated`, fail-closed). The printed headers the poller reads never carry that header, so without the second read a forged From would draw the acknowledgement to whoever it names — backscatter from the Executive's mailbox. An unauthenticated sender makes `send_ack` raise `roster_intake.AckWithheld`: the claim is released, and they are held silently. `authenticated_by_gmail` strips quoted strings and comments before reading the verdict — Gmail copies the envelope sender and DKIM tags into its own stamp, and a quoted local part could otherwise carry a fake `dmarc=pass` — and needs exactly one `dmarc=` verdict reading exactly `dmarc=pass header.from=<domain>` (Gmail may follow it with its own `dara=` resinfo, so its position is not checked). Residual: when Gmail writes no verdict at all (a From domain without DMARC) and echoes an unsanitised `;dmarc=pass header.from=…` elsewhere in its stamp, that echo would read as the verdict; bounded by the acknowledgement caps. A withheld acknowledgement is audited `roster_ack_withheld`. `intake` acknowledges before surfacing the card and telling the principal, and both follow the request's `ack_sent_at`: a withheld, failed or capped acknowledgement reads \"haven't been told anything yet\" on the Today card, in the principal's prompt and in the email `[POLICY]` notice.\n\n**Telling the principal.** A new request puts a private card on the principal's `/today` (see Today) and pushes a prompt on their delivery plan — preferred chat DM, then the others, then email — for a chat sender or an email sender on the company's own domain; an outside email sender waits on the card, as does everything past `ROSTER_REQUEST_DAILY_CAP` (30) pushes a day (the same cap bounds how many requests open a day; a declined sender is not raised again for 30 days). A chat prompt can be answered in place (\"that's Annamarie, add her to the team\"). An email prompt carries a one-time token (`RR-` + 20 base32 characters, stored hashed); the poller checks every inbound mail for one **before** triage (`try_email_roster_answer`): the reply must come from the principal's primary address exactly (not an alias, not the company-domain rule), carry the token of a pending request, not be an automatic reply, and not fail DMARC where the header is visible; and is parsed by `parse_decision`'s `roster_identity` shape — anything else falls through to the ordinary path. The token is hidden (`RR-[hidden]`) in every other Google Workspace result the gateway returns — the confirmation sits in the Executive's Sent folder, and any turn could otherwise read it — except the poller's own fetch of an inbound message (`mcp_gateway.reveal_roster_tokens`); a token-bearing mail that is not an answer is scrubbed before triage.\n\n**Replay.** Once the principal answers (web card, chat tool, email reply) — or a People-page / `upsert_person` write makes the sender match someone (`after_roster_write`) — the held messages are replayed through the adapter that received them (`register_replayer`: Slack re-runs the held event, Discord and Telegram their held text, email re-fetches the message from Gmail and handles it as mail from a known sender), oldest first, now attributed to the new Person. A teammate gets every held message; a new contact only held email (it takes the private contact path; held chat is dropped); a decline drops everything. Replays run in a **fresh `contextvars.Context`**: the answer may have come from the principal's own verified turn, and a stranger's message must never be read with the principal's session — and their contacts — still bound.\n\n### Standing facts by email (confirm by reply)\n\nThe principal can ask for a standing fact, a retirement or a company-profile edit by email, and the fact tools check it exactly as in chat (the own-words gates run on `_email_memory_text`, the sender's new text, not the quoted chain or attachments). An email's From line proves nothing, though, and a standing fact is read by every later prompt as the principal's own account. So an emailed change is **held, never applied**, until the principal's own mailbox confirms it (`integrations.fact_confirmation`, `memory.facts.fact_confirmations`):\n\n- **Who may ask.** The poller stamps every email turn with `Session.email_from` (the bare lowercased From address) and `email_authenticated`. The text the poller reads never carries `Authentication-Results` (workspace-mcp prints a fixed set of headers), so for mail claiming the principal's primary address it reads the message again as raw MIME (`fact_confirmation.sender_authenticated`) and requires Gmail's own stamp: the **topmost** `Authentication-Results` header (Gmail adds it on receipt, above anything the sender wrote), authserv-id `mx.google.com`, reporting `dmarc=pass` for the From address's domain, with the raw From matching (`authenticated_by_gmail`). It fails closed: no header, `dmarc=none`, a temporary error or a failed read all leave the turn unauthenticated. Mail from anyone else costs no second read and is never authenticated. `fact_tools._gate` admits an email turn only from the principal's **primary** address, exactly (no alias, no company-domain rule), authenticated, and not private to the principal (mail they forwarded stays refused, as do contacts, teammates and unknown senders).\n- **The hold.** The tool returns `awaiting_confirmation` (no action chip, and the model is told to say nothing has changed). `request_confirmation` stores the change with the SHA-256 of a fresh one-time token (`FC-` + 20 base32 characters, 100 bits; 7-day expiry; at most 10 waiting at once) and emails the token to the principal's roster address, outside any turn's outbound context. A profile edit is dry-run first, so one that cannot apply is reported in the turn, before any email goes out. If the email fails, nothing is held.\n- **Keeping the token from any model.** The tool result never carries it, and `mcp_gateway.hide_roster_tokens` now scrubs `FC-` tokens as well as `RR-` ones (`FC-[hidden]`) from every Google Workspace result except the poller's own fetch of an inbound message. Act as me's reads of the principal's own Gmail, where the confirmation email lands, hide them too (`delegation.gmail.parse_message` and the thread search), so the drafting model never sees a live token either.\n- **The reply.** `_handle_one_email` hands every message to `try_email_fact_confirmation` right after the roster-answer check, before any model turn. It acts only on a reply from the principal's primary address carrying the token of a **pending** confirmation. A made-up, spent, cancelled or expired reference matches nothing and falls through to the ordinary path with the token hidden, so no old or forged token can make the Executive email the principal. A reply with a live token is read again as raw MIME (`read_raw`) and must pass the same Gmail DMARC check as the request; one that doesn't (or whose raw read fails) is refused, the change keeps waiting, and the principal is told. Then an **automatic reply** is consumed quietly — no answer, no model turn, the change keeps waiting: the printed headers catch a Gmail vacation reply (`Precedence: bulk`), and the raw headers catch what they don't show (`automatic_reply`: `Auto-Submitted` other than `no`, `X-Auto-Response-Suppress` and similar, `Precedence: auto_reply`, an out-of-office subject), since an Exchange out-of-office echoes the token from the principal's own address and may well say \"I'll confirm on my return\". The raw message starts at the tool's exact separator line (`\\n\\n--- RAW MIME ---\\n`), so a sender-written Subject imitating it cannot supply the headers. CONFIRM (or yes / approve) applies the change through `fact_tools.apply_confirmed`, the same apply step chat uses. CANCEL (or no) drops it, and an unclear reply leaves it waiting. The principal gets a one-line result either way. A compare-and-set on the row means two replies can apply it only once.\n- **Audit.** Private `fact_confirmation` rows record held / applied / cancelled / refused. The applied change also writes the fact tools' usual private rows, with `source_channel=\"email\"` as its provenance.\n\n### Checking a channel is set up\n\nSettings → Setup status (`GET /setup/status`, `api/setup_checks.py`) gives each channel a light and a plain-language fix. It tests the credentials with one free read-only call — Slack `auth.test`; Telegram `getMe`, then `getWebhookInfo` (is a webhook registered, is it this app's `/webhook/telegram`, did a delivery fail in the past day); Discord `users/@me`, only while the bot is still connecting — and reads what the lifespan left on `app.state`: the Slack socket's `is_connected()`, the Discord bot's `is_ready()` together with its gateway socket being open (discord.py stays ready while it reconnects after a drop), or the exception its task stopped with (`LoginFailure`, `PrivilegedIntentsRequired`), and the MCP gateway. A listening Slack is amber while the listener hasn't resolved its own user id (`slack_bot.bot_user_id()`), since thread continuations are off until it does. Because inbound is roster-gated, a connected Slack, Discord or Telegram is amber while nobody on the roster has its id (`slack_user_id`, `discord_user_id`, `telegram_chat_id`), or when its latest `integration_inbound` row is a `Rejected: …` from the past day whose sender still isn't on the roster — the page shows the sender's id, never message text. Local login with no `TELEGRAM_WEBHOOK_SECRET` is red for Telegram, since the local-login gate turns its webhook away. Google Chat and Gmail are checked from configuration: a numeric project number and a readable service-account key; the OAuth client (or service-account key) the workspace-mcp child signs in with.\n\nA channel setting written `KEY= # note` in a `.env` file arrives as the note's text, because python-dotenv strips only a comment that follows a value. `Settings` reads such a value as unset for the Slack, Telegram, Discord and Google Chat credentials, so the line no longer switches a channel on with its own comment as the token. `TELEGRAM_WEBHOOK_SECRET` is the exception, because read as unset a junk secret would turn the webhook's check off. It is kept instead, and since Telegram can only ever send 1–256 of `A-Z a-z 0-9 _ -` (what `setWebhook` accepts), a secret outside that set — a leftover comment, a pasted `<value from Step 2>` — could only be matched by someone who guessed it: while it is set the webhook refuses every update, neither the local-login gate nor the roster tools count it as proof (`Settings.telegram_webhook_secret_valid`), and Setup status shows it red (`.env.example` shipped four such lines; the notes now sit on their own lines).\n\n### Conversation state\n\nEvery adapter keys a deterministic `session_id` off the conversation surface and replays that session's stored turns into `Session.conversation_history`, so a follow-up (\"yes, go ahead\") lands with the question still in context. Slack's ids are `slack:dm:{user}`, `slack:thread:{channel}:{thread_ts}`, and `slack:channel:{channel}:{user}` for a mention that starts a thread; because the bot's reply roots a Slack thread at that message's `ts`, the turn is double-written to that thread's session too, so the first continuation inside it is not cold. A reply inside a thread THAT PERSON ROOTED also offers the parent `slack:channel:{channel}:{user}` id when matching an approval gate, since a gate raised on the mention recorded that session while the answer arrives under the thread's. The root-author check is load-bearing: widening on \"threaded and not a DM\" alone offered one person's channel session inside any thread in that channel, so a remark they made to a colleague could resolve their open gate and post the run's title into that colleague's thread. Discord has the same shape — its auto-thread router promotes a mention reply into a bot-owned thread — and the same alias. An @-mention inside a DM is handled once, by the DM path: Slack fires both events, and taking both forked the conversation across two session ids. The same id is the audit grouping key and the scope passed to `format_for_prompt`. A per-session `asyncio.Lock` serializes turns within one conversation — the inflight semaphore caps fan-out across the workspace, which does nothing to stop two messages in the same thread from both reading history before either writes it. Persistence happens after the reply is sent and its failure is logged, never surfaced: the user already has their answer.\n\n**Conversations within a chat app.** A Telegram chat, a Slack or Discord DM and a person's @mentions in one channel (`slack:channel:…`, `discord:channel:…`) are each one endless stream, so they are cut into conversations (`memory.conversations`, shared helpers in `integrations/conversations.py`). Each adapter computes the stream's base id as before and asks `conversation_for` which conversation a message belongs to: the newest, unless the stream has been quiet for two hours (`CONVERSATION_GAP`), when the next one starts. The first conversation keeps the bare base id; later ones add `~2`, `~3`, … (`memory.conversation_ids`). A whole message of `/new`, \"new chat\" or \"new conversation\" starts a fresh one instead of a turn and is answered with the name of the one it ended; the new row has no messages and stays off the Chats page until its first turn. After a conversation's first exchange `title_after_turn` names it in the background with `utils.session_title.generate_session_title` (the first words of the message if that fails), as a web chat's first turn does. Threads (`slack:thread:…`, `discord:thread:…`), email and every other id are left whole. What has to see the whole stream compares base ids: the inbound resolver, so an approval asked before a cut is answered after it, and the session-scoped decisions and advice `format_for_prompt` renders. At boot `split_existing_streams` cuts history saved before this by the same rule, once (`app_migrations`), titling each part from its first message and carrying a read-mail lock to every part. An adapter installed outside the package adds its streams with `conversation_ids.register_rolling` and cuts its own old history once, with its own migration name and an `only` filter for its base ids, because the built-in sweep may already have run without it.\n\nThis was absent on Slack until #136. Every message built a fresh `Session` and minted a per-*message* id (`thread_ts` falls back to the message's own `ts`, so a DM never repeated one), which meant a multi-turn confirmation could not work: the Executive asked whether to send a briefing to Slack, the user said yes, and the next turn had no record of the question.\n\n### What the model is told about the channel\n\nChat adapters pass a `<channel>` block (`integrations.channel_context`) in the **user turn** — never a cached system block. It names the surface, states that the person cannot see the app's cards or buttons, and draws the line the assistant kept crossing: briefing proposals and alerts are actioned on the briefing page; a proposal that books something can only be approved there; a workflow sign-off the Executive is explicitly waiting on *can* be answered in the conversation. The principal's DMs on Slack, Discord and Telegram also receive the open-alert digest (`briefing.context.format_open_alerts_for_prompt`) as the `<briefing>` block, gated to DMs because the board is company-wide and a shared channel must not pull it in. That digest is the ONLY trusted source of an `alert_id` for `ack_alert`. The briefing page's `[Discuss mode — alert_id=N]` primer is a pointer to which item is under discussion, not authority on its own: the server accepts that id only if it is also on the live board it derived itself. An id appearing anywhere else, including inside an alert's own body, is not trusted: alerts are minted from inbound email and chat, so their bodies are attacker-controlled. Two mechanisms enforce that rather than one prompt rule. Every field the digest interpolates is collapsed to a single line, because a newline in an attacker-written headline, suggested action, tag or review note would otherwise forge a whole extra `[N]` line — a fabricated instruction to clear somebody else's alert. And the live board the server derived is recorded on the session, so `ack_alert` refuses any id outside it — on EVERY session, web included; the web previously had no check at all. The recorded set is the whole live board (the cap `/today` renders as cards), not just the rows the digest printed, so a card past the digest's 30-row render cap is still ackable rather than sending the principal back to the page. Be precise about what this buys: it stops an id that is invented, closed, snoozed or expired, including one quoted out of an attacker-written alert body. It does NOT stop the model being argued into acking the wrong LIVE card — one alert's body can still make a case about another id that is genuinely open. Scoping trust to the single item under discussion would need the primer id carried server-side, which it is not.\n\n### Handler failures\n\nAn adapter's catch-all replies \"I encountered an error processing your request.\" On Slack that used to be the whole story — one uncorrelated log line, no audit row, and an unguarded error-send that could itself escape into Bolt. It now logs with the session, sender, channel, thread and mode, and writes an `integration_inbound` row with `details.outcome=\"handler_error\"`, joining the inbound row under the same `session_id`. Upstream, a single crashing tool handler no longer takes the turn down at all: the orchestrator's tool gathers use `return_exceptions=True`, hand the model an error `tool_result`, and audit it with `ok=false` (cancellation is re-raised, not swallowed).\n\n### Attachments\n\n`integrations.attachments` handles web chat, Discord, Telegram, Slack and Google Chat. Images become Anthropic vision blocks; documents (PDF/DOCX/TXT/MD/CSV) are inlined into the turn that carried them — inside an `<untrusted_content source=\"attachment\" author=\"<name>\">` block (`content_trust.wrap_untrusted`, whose closing tag the text cannot forge), under an `[Attached: <name>]` label, which the open-loop pass reads as \"not the speaker's words\" and the decision extractor strips before it reads the turn — and indexed into the ChromaDB `inbound_attachments` collection. 20 MB/file cap; web chat uses `POST /chat/upload`.\n\n**Scanned PDFs** are read, not reported empty. A PDF with no text layer goes through `knowledge.pdf_reader`: with `PDF_PROVIDER_READING=true` (opt-in, off by default) the deployment's own model reads it through its provider's PDF support (Anthropic natively, OpenRouter's `file-parser`, a local server only with `LOCAL_PDF_INPUT`); otherwise, and as the fallback, local OCR (pypdfium2 + RapidOCR, offline, no key) — so every deployment converts them. The label then reads `(converted from scanned pages)`, and an unreadable one says why. The index takes the text already extracted for the turn, so a scan is converted once. A file a channel delivered (everything here except `POST /chat/upload`, where the signed-in user sends it) converts at most `PDF_INBOUND_MAX_PAGES` (30) pages and draws on a rolling `PDF_INBOUND_PAGES_PER_HOUR` (300) budget shared by all senders, so sending scans cannot run up unbounded model spend or CPU.\n\nPer channel:\n- **Slack** — a `file_share` message (previously dropped outright for its `subtype`) is handled like any message; files download only after the roster gate and the response gate, with the bot token sent only to `https://files.slack.com/` (needs the `files:read` scope). History and peer memory record the words plus `(Attached files: …)`, never a document's text.\n- **Google Chat** — uploaded attachments download through `media.download` with the reply credentials (`chat.bot`), checked against the 20 MB cap per chunk, 5 per message; an attachment-only message is no longer dropped. The webhook authenticates Google, not the sender, so files are read only when the event's `user.email` is on the People roster (a team member); anyone else's files are named to the Executive, never downloaded. A Drive file shared into a message is left to the Drive tools, with its id.\n- **Email** — the poller (`integrations.email_attachments`) downloads up to 5 readable attachments of the message itself via `get_gmail_attachment_content` — workspace-mcp in stdio mode saves the file and answers `Saved to: <path>`, which must start its line, be the last such line and end in the answer's `Saved filename:` (the answer echoes the sender-chosen attachment name first, and a name carrying `Saved to: /…` must not redirect the read), then resolve inside the tool download folders — and appends their text to the turn (15k chars each). Only for a sender the principal knows (team, contact, principal); a stranger's files stay a list. It runs outside the model loop, so it works on private turns without adding the attachment tool to `PRIVATE_TURN_MCP_TOOLS`.\n- Anything else (a Drive PDF, a skipped attachment) — the Executive's `read_document` tool.\n\nIndexed chunks are keyed on `attachment:<bare filename>` and tagged `type=attachment`. An attachment's name is chosen by whoever sent the message and the name is the chunk-id namespace, so without the prefix an inbound `strategy-2026.md` would upsert over the curated company document of that name. The prefix separates inbound content from curated uploads and keeps provenance visible; it does not separate senders from one another.\n\nIsolation is the **collection**: `retriever.retrieve` never queries `inbound_attachments` at all, so an attachment cannot resurface as company knowledge in a later, unrelated turn. Attachment processing is roster-gated on Discord too — it previously ran before the gate, so an unrostered sender's file was downloaded and indexed before their message was dropped. These rows still have no operator surface: they are never written to `company/docs/`, so `GET /documents` does not list them.\n\nThis replaced a domain-based scheme that did not work. The chunks lived in `company_docs` under a domain outside the specialist set, on the premise that an unrecognised domain matched no filter. It matched every one — `ChromaDBStore.query` attaches a `where` clause only when a domain filter is present, and an Executive-level retrieval passes none — so attachments came back under “From your company documents”, top-ranked and indistinguishable from a curated upload. Legacy rows are migrated across on boot.\n\n### MCP gateway & outbound\n\nMCP runs in **both directions**. As a *client*, `orchestrator.mcp_gateway.MCPGateway` proxies MCP servers (Gmail, Calendar, Drive, Outlook mail and calendar — whatever `mcp_servers.json` configures) as Executive tools, with egress gating on Gmail recipients, Calendar attendees, Drive share grantees, and Microsoft 365 recipients/attendees against the roster (and a refusal of public / whole-domain Drive sharing). The Gmail gate also accepts only a fixed set of argument keys, so workspace-mcp's `reply_all` (recipients taken from the thread, never checked) and `forward_message_id` / `include_forwarded_attachments` (forwarding any inbox message) are refused; a reply is `thread_id` plus an explicit, roster-checked `to`. Three more refusals close paths around those gates: every Apps Script tool (code running as the Executive's account, outside the gates, and since workspace-mcp 1.29.0 schedulable to keep running); any argument that makes the server fetch a URL (`fileUrl` / `file_url` on the Drive and import tools, a Gmail attachment given as a `url`), since the fetch can carry data out in its query string before any recipient gate matters; and text Google would mail to someone the roster never checked (`rsvp_comment` to an event's organizer, an out-of-office `decline_message` to whoever invites the Executive). Every gate reads its arguments by normalized spelling: workspace-mcp 1.29.0 renames a camelCase key such as `Attendees`, `userGoogleEmail` or `rsvpComment` to the real parameter, so a gate matching the exact key would never see it. Arguments that are not a JSON object are refused for the same reason. Google-side fetches are refused too: any key ending in `url` / `uri` at any depth whose value has a URL scheme (a Slides image or page background, a Docs or Forms image; a document hyperlink and a meeting link are stored, not fetched, and pass, as does a Drive file id), (a value is a URL after the characters a URL parser strips are removed, and `//host` counts), and on every tool that can write cells — the Sheets tools and the Drive create / update / import tools, which turn CSV into a Sheet — any string anywhere holding a formula call that fetches (`IMPORTDATA`, `IMPORTXML`, `IMPORTHTML`, `IMPORTFEED`, `IMAGE`), JSON-encoded or mid-CSV included: a string argument that encodes JSON is read as the server decodes it (one too large to read is refused), text is also scanned with character references decoded and comments / CDATA removed, and an upload that would become a native Sheet may only be text `content` — a `base64_content` or server-side `file_path` upload there is refused, since an XLSX hides its formulas behind zip members and XML encodings; a PNG to Drive or a DOCX to Docs is untouched. Not covered, unverified: whether Drive's HTML / Markdown import fetches an `<img src>` when it converts `content` into a Doc. A `manage_event` call that names no attendees is sent with `send_updates=\"none\"`, so the event's existing guests are not emailed text the gate never checked. The allow-list (`_egress_sets`, built by `_roster_allow_set` as a `people.identity.RosterAllow`) is every team member's email and email aliases plus the Executive's own address, always, and — for a teammate — any address on one of the company's own domains whose local part (`+tag` dropped) matches theirs, the same rule inbound mail is matched by; a **contact's** email joins it only when `people_tools.contacts_reachable_now()` is true — a turn the principal started on a verified surface (web chat, their own Slack or Discord, a verified private Telegram chat), or inside `grant_contact_egress()` (the decisions approve route, when the approver is the principal). On any other turn — an inbound email, a teammate, the scheduler, a workflow, the alert review — a contact's address is refused exactly like a stranger's, so the refusal cannot reveal that it belongs to a contact (contacts are private to the principal). The single exception to the roster rule is the fixed roster-request acknowledgement above, under its one-shot `roster_ack_grant`. A turn about the principal's private mail (`Session.private_to_principal`) narrows the list to the principal's own address and the Executive's. Such a turn may call only the MCP tools in `schedule_tools.PRIVATE_TURN_MCP_TOOLS`: Google Workspace reads (`search_gmail_messages`, `get_gmail_message_content`, `get_events`, `list_calendars`, `query_freebusy`, and the Drive, Docs and Sheets reads such as `search_drive_files`, `get_drive_file_content`, `get_doc_content` and `read_sheet_values`) and the two Gmail tools this gate narrows to the principal (`send_gmail_message`, `draft_gmail_message`). It is an exact allow-list, not a server: workspace-mcp runs at the `complete` tier, where a Google Chat message to a space, a Docs, Sheets or Drive write into a file already shared with others, `manage_event` on a shared calendar (its gate checks attendees and pins notifications, nothing else) reaches people with no recipient check; and every other server has none at all (a Slack or Notion server, the default `fetch` server, whose URL can carry the turn's content, a server loaded at runtime). Its `search_tools` results show only the allowed tools (`workflows.tool_catalog.filter_search_results`; output it cannot parse reads as no match), and a `call_tool` naming anything else is refused with an error tool_result and an audit row private to the principal (`schedule_tools.private_turn_withholds`), a missing name included. `search_tools` and `call_tool` stay offered, so the tool list and its cached prefix do not change. The typed calendar tool (`_resolve_attendees`) and the chat DM tools (`message_person`, `send_telegram_message`, `send_discord_dm` and their person-id recovery) apply the same rule — a contact's id reads as an unknown one off the principal's turn — and a private turn may DM or invite only the principal (`PRIVATE_TURN_REFUSAL` otherwise); `message_person` never falls back to a \"Needs you\" alert for a contact, who cannot sign in to see it. The Google Workspace server (`workspace-mcp`) behind those Gmail/Calendar/Drive tools runs **co-located inside the API** as a stdio child of the gateway — fitting the single-tenant model (one install = one company = one exec Google account), so there's no separate service to operate. It's baked into the API image and launched by `docker/workspace-mcp-launch.sh`, which selects single-user OAuth (token persisted at `/data/google_credentials`) or service-account domain-wide delegation from `GWORKSPACE_AUTH_MODE`. The optional `WORKSPACE_MCP_TOOLS` (comma-separated service names, default `all`) limits which Google services it loads. The gateway forwards the Google secrets to the child via `_FORWARDED_ENV_VARS`. Every `google_workspace__*` call acts as the Executive's own account. workspace-mcp takes the acting account per call (`user_google_email`) and nothing used to check it, so under domain-wide delegation a model-chosen address could open anyone's mailbox in the domain; `mcp_gateway._check_acting_account` now refuses a value naming any other address — before every other gate, with an `integration_outbound_blocked` audit row. An absent value stays allowed (the server's own account).\n\n**What the model is told is connected.** `MCPGateway.start()` records the configured server names (`server_names`), and the Executive renders them with the channel settings into block 0's *Connected Systems* section (`prompts.connected_systems`; see the caching section). While Google is connected that section pins the Google tool names the Executive uses constantly (`GOOGLE_TOOL_MANIFEST`) and tells it to call them with `call_tool` directly, with no `search_tools` round trip. extensible-mcp only runs a tool one of the session's searches has returned, and remembers it for the session's life, so the gateway does that discovery itself: the lifespan starts `MCPGateway.prime_pinned_tools()` in the background once the gateway is up with `google_workspace` configured (an exact-name search per pinned tool, the way `workflows.tool_catalog.resolve` does it; a name the search no longer finds is logged as workspace-mcp drift). A pinned call that still comes back with extensible-mcp's \"has not been discovered via search_tools\" refusal, for example before priming finished, is discovered and retried once. That happens after every gate has passed the same arguments, and only for pinned names: any other tool still needs the model's own search, and nothing here widens what a call may do, including the private-turn allow-list.\n\n**Microsoft 365** is the Outlook twin: `@softeria/ms-365-mcp-server` (a Node package, pinned) is installed in a throwaway Node build stage and only the `node` binary plus the package are copied into the API image, launched by `docker/ms365-mcp-launch.sh` as the `microsoft_365` entry with an **explicit tool allow-list** (`--enabled-tools`, an anchored regex) rather than the upstream `mail,calendar` preset — that preset also registers inbox forwarding rules, mailbox settings (external auto-reply), calendar sharing and event forwarding, which can address people the gateway never sees. Exactly the Outlook mail + calendar tools and the OneDrive tools are exposed (`microsoft_365__send-mail`, `microsoft_365__create-calendar-event`, `microsoft_365__search-onedrive-files`, …; hyphenated Graph endpoint aliases, proxied verbatim), and every admitted tool is either roster-gated or carries no recipient; the gate also covers the excluded writers in case an operator widens the list. Both servers can be configured at once. Auth is the MSAL device-code flow against the operator's own Entra app registration (`MS365_MCP_CLIENT_ID` is required; the server's built-in public client is refused); the refresh token persists at `/data/ms365_credentials` and is seeded once with `ms365-mcp-launch.sh --login` from inside the container — the server has a native device-code CLI, so there is no mint script and nothing to copy. The launcher scrubs the literal `$VAR` placeholders extensible-mcp leaves for unset variables (otherwise MSAL would be handed `$MS365_MCP_TENANT_ID` as an authority), and `MS365_MCP_EXPECTED_USERNAME` pins the bound mailbox. The egress gate for this family walks Graph's nested shapes (`body.Message.toRecipients[].emailAddress.address`, `body.attendees[]…`) — every email-shaped string outside the free-text fields must be on the roster, control characters and non-ASCII addresses are refused, unforeseen keys are still scanned (fail-closed) — and covers the whole write surface: send, reply, reply-all, forward, every draft tool, `update-mail-message` (it can PATCH recipients onto a draft) and create-/update-calendar-event; reads and `delete-calendar-event` (the only cancel path) pass through. Graph's reply, reply-all and send-draft actions carry no recipient in their arguments (the recipient is whoever the referenced `messageId` names), so for those the gate first reads the referenced message through the server and roster-checks the recipients Graph will derive (`replyTo` else `from` for a reply; `replyTo` and `from` together, plus the original to/cc, for reply-all; the draft's own to/cc/bcc for send-draft), refusing on any lookup failure; RSVP and cancel actions get the same treatment through an event lookup, because their comment is emailed to the organizer or every attendee; `download-bytes` is pinned to mail attachment bytes and OneDrive file content (`/drives/{id}/items/{id}/content`, optionally Graph's `?format=pdf` rendering), and a OneDrive result comes back as the file's extracted text rather than base64 — the attachment pin is what makes the poller's promise that an unrostered sender cannot be auto-replied to hold on Outlook as it does on Gmail; and the rendered inbound text quotes any forged `--- REPLY ---` line inside a body so the persona only ever reads the real block. Every mail write's `from`, `sender` and `replyTo` may name only the Executive's own address (the Outlook twin of the Gmail acting-account pin), so the model cannot send as a rostered colleague's mailbox or steer replies elsewhere; and `move-mail-message` into Deleted Items, Junk Email or Recoverable Items is refused, so a move cannot trash mail ungated. The server registers its account tools whatever its tool allow-list says, so the gateway itself refuses `login`, `logout`, `select-account` and `remove-account` (audited like any other refusal) and drops them from `search_tools` results — a chat turn cannot sign the mailbox out, switch accounts or start a sign-in; setup is only the launcher's `--login`, and the read-only `verify-login` / `list-accounts` stay callable. The example config's hyphenated patterns still deny `delete-mail-*`, `logout`, `remove-account` and `download-bytes-to-file` as defense in depth. A successful `send-mail` records the same outbound-context linkage Gmail does.\n\n**OneDrive in chat** (opt-in with `MS365_MCP_ONEDRIVE=true` and a fresh `--login`, since the server asks for its tools' permissions on every token refresh and a sign-in without file consent would otherwise fail mail and calendar too) matches what the Executive can do on Google Drive: search, list a folder, read a file, upload and edit, create folders, move, rename, copy and share; there is no delete tool. The launcher adds `Files.Read.All` at sign-in (`--extra-scopes`), so folders shared *with* the Executive can be read too; writes (upload, new folder, move, copy) land only in the Executive's own drives, checked against `list-drives` at call time, so a planted instruction cannot save company files into a stranger's shared folder. Sharing is gated like Drive sharing: `share-drive-item` invitees must be People on the roster and named by email (an `objectId` or `alias` grantee is refused, since nothing can check it), no `owner` role, and `create-drive-item-share-link` must be scoped to specific people (`scope: \"users\"`): an anonymous or whole-organization link, or one with no scope (the tenant default), is refused. An invite must require sign-in and grant only view or edit, a link must be a view or edit link, and `search-onedrive-files` takes plain words only, since its query goes into the request path unencoded.\n\nAs a *server*, `openexecutive.mcp_server` exposes OE's own company context and specialist council to external MCP clients at `/mcp` — the inverse direction; see the dedicated **MCP Server** section. The gateway launches extensible-mcp as a `uvx` subprocess whose fastembed tool-search index uses an embedding model baked into the API image and loaded offline, so it no longer re-downloads that model from the Hugging Face Hub on every cold start. extensible-mcp is pinned twice over: to a commit, which fixes its own code, and with `--exclude-newer`, which fixes the resolution of its dependencies — without the cutoff uvx re-resolves them against PyPI on every networked start, so a new release of any of them would reach running deployments on restart. Both live in `_EXTENSIBLE_MCP_LAUNCH_ARGS` in `orchestrator/mcp_gateway.py`, the Dockerfile's pre-warm uses the same arguments (a unit test enforces the match), so an image whose pre-warm succeeded starts the gateway with no network. The pre-warm is best-effort (`|| true`): an image built while GitHub was unreachable fetches the pinned commit on its first start instead. Updating the gateway is a deliberate bump of the commit and cutoff; `uv` itself is not pinned. Because the MCP stdio client does not inherit the parent environment, `MCPGateway.start()` explicitly forwards the embedding-cache/offline env vars (`FASTEMBED_CACHE_PATH`, `HF_HUB_OFFLINE`, …) into the child.\n\n**Emailing an artifact.** A Gmail send or draft may carry `attachments=[{\"artifact_id\": \"alert:12\", \"as\"?: \"docx\"}]`. `call_tool` runs the recipient gate first; only if it passes does `_expand_artifact_attachments` replace each such entry with the rendered file (`content` base64 + `filename` + `mime_type`, workspace-mcp's own attachment shape), produced by `artifact_records.render_artifact_file` — the same renderer as `/artifacts/{id}/download`. Only artifact rows resolve (never an arbitrary alert), links have no file, and an entry mixing `artifact_id` with other keys is refused. A stringified or single-object `attachments` is normalised first, so an artifact entry can never reach workspace-mcp unexpanded (it would silently drop it). At most 5 distinct artifacts per email (repeats are dropped) and 15 MB rendered in total; rendering runs off the event loop via `asyncio.to_thread`. A refusal writes `artifact_attachment_refused`; a successful send writes `artifact_attached` — both carry the recipients. Other attachment entries pass through unchanged. The same expansion serves the Microsoft 365 mail writes: the entry becomes a Graph `#microsoft.graph.fileAttachment` (`name`, `contentType`, `contentBytes`) placed where Graph expects it for that tool (`body.Message.attachments` for `send-mail`, reply, forward and the createReply drafts; `body.attachments` for `create-draft-email`), whether the model put it at the top level or inside the message; an entry anywhere else, one on `update-mail-message` or `send-draft-message` (their requests cannot carry attachments), or one behind a body that is not an object is refused rather than forwarded unexpanded.\n\n**Enablement, and why it degrades instead of dying.** `MCP_ENABLED` left unset is inferred from the presence of `mcp_servers_config_path` — dropping an `mcp_servers.json` next to `profile.yaml` turns MCP on with no further configuration. An **explicit** setting always wins in both directions: `MCP_ENABLED=false` keeps MCP off with the file still in place. It did not: `_resolve_mcp` tested `not self.mcp_enabled`, which cannot tell \"never set\" from \"set to false\", so the file's presence silently overrode the operator. The file is probed through `config.mcp_config_file_present`, never `Path.exists` directly — that re-raises an EACCES on a parent directory, and from inside a validator that comes out of `Settings()` itself, which `api/main.py` calls at module level. The check now reads `model_fields_set` (the fields the env actually supplied) and stores that on a private attr behind the read-only `mcp_auto_enabled` property, because the validator's own assignment joins that set and would otherwise make every auto-enable look explicit downstream.\n\nStartup is **best-effort**, in `main._start_mcp_gateway`. Two checks come before the subprocess: MCP off is a no-op, and a config defining no servers under `mcpServers` is logged and skipped — extensible-mcp's own loader raises on that and exits, and because the child is already gone the only thing that surfaces is anyio's \"Attempted to exit cancel scope in a different task\" from `stdio_client` unwinding, a traceback naming nothing about configuration. The start is also bounded (120s, generous because that child may still clone extensible-mcp on a cold start if the image's pre-warm failed): `ClientSession` carries no `read_timeout_seconds`, so `initialize()` waits forever on a child that is *alive but silent*, and an unbounded wait is a lifespan that never yields — the same healthcheck-never-passes symptom, which no `except` can see. A gateway that fails anyway is logged at ERROR with the config path, the servers it named, and whether the config file or an explicit `MCP_ENABLED` turned MCP on, then torn down under its own timeout; the API boots **without MCP tools and without the email poller** rather than refusing to boot. The guard catches `BaseException`, not `Exception`, re-raising a bare `CancelledError` first: the original failure came from anyio, which wraps a task group's failures in `BaseExceptionGroup` as soon as one is a `BaseException` — and that group is not an `Exception`, so `except Exception` would miss the very shape this contains. The teardown is contained too, since it runs from inside that `except` and would otherwise replace the actionable log with the crash it prevents. `configured_server_names` never raises either, and deliberately catches broadly: it reads an operator-owned path during startup, where an exception is the crash loop rather than a bug report — `Path.is_file` re-raises an EACCES on a parent directory, and `json.loads` answers deeply-nested input with `RecursionError`, which is not a `ValueError`. It tests `is_file()` so a directory or a symlink to a FIFO is never opened, and caps the read at 1 MiB. That exception used to propagate out of the lifespan and kill the container — under `restart: unless-stopped`, a crash loop whose only outward symptom was a healthcheck that never passed. Outbound tools: per-Person DMs (`send_slack_dm`, `send_discord_dm`, `send_telegram_message`) plus room-targeted `send_department_message` and `send_company_broadcast`. All three per-Person DMs refuse a recipient id that matches no non-archived Person row (`schedule_tools._dm_recipient_on_roster`), which is what stops untrusted content in a turn — inbound mail, a chat message, a knowledge document — from directing a DM at an arbitrary workspace member. `send_slack_dm` previously consulted the roster only when the turn was already private to the principal, so on an ordinary turn it applied no roster check at all while its Discord and Telegram siblings did; the room-targeted tools are deliberately ungated, since those channels are team-visible. The per-Person DMs are additionally rate-limited and de-duplicated by the outbound anti-spam guard (below).\n\n### Outbound anti-spam guard\n\nEvery per-Person DM passes a deterministic guard (`orchestrator.outbound_guard.check_outbound_allowed`) after the roster gate and recipient-id recovery but before the network send — the single chokepoint **all** proactive DM paths reach (ad-hoc chat sends, the scheduler's followup/nudge dispatch, principal-brief delivery, the reflection turn, and research-finding routing). On a violation it returns a human-readable reason and the handler **suppresses** the send (`{\"status\":\"suppressed\",\"reason\":…}`), writes a `tool_invocation` audit row with `suppressed=true`, and writes **no** `done` activity row — so a blocked attempt never counts toward its own rate cap, and the Executive can read the reason and hold or `schedule_followup` for the recipient's next window. Three controls: **content dedup** (a near-identical message — whitespace/case-normalized, first 160 chars — already sent to the same recipient within `OUTBOUND_DEDUP_WINDOW_MINUTES`, default 360); a **per-recipient rate cap** (`OUTBOUND_MAX_PER_RECIPIENT_PER_WINDOW`, default 5, over `OUTBOUND_RATE_WINDOW_MINUTES`, default 60); and **quiet hours / availability** (`OUTBOUND_RESPECT_QUIET_HOURS`, default on — refuse while the recipient is on leave or outside their `people.channel` availability windows, surfacing the next open window). Delivery history is read from the existing `status='done'` `scheduled_actions` rows (`episodic.recent_sends_for_channel_ref`), so there is no new table. The guard is **fail-open**: any internal lookup error returns \"allow\", so it can never turn a working send into a crash or a silently-dropped message. Principal briefs and the internal heartbeat kinds use the `__internal__` channel and never reach these handlers, so they are unaffected.\n\n### Calendar & video links\n\nThe Executive books meetings through two typed tools over the configured calendar backend (`CALENDAR_PROVIDER`: Google Calendar via `manage_event`, or Outlook via `create-calendar-event` — see *Workspace providers*). `create_calendar_event` schedules a **future** meeting (propose→approve by default, or auto-execute once that decision class is promoted) and may be used proactively when a stalled goal warrants a sync. `create_instant_meeting` spins up an **impromptu call starting now**, auto-executing without the approval step and skipping the business-hours/horizon caps, so the Executive can hand back a join link immediately. Both attach a video-meeting link by default (`add_video_link`; opt out for in-person) — a **Google Meet** link parsed from the created event's `conferenceData`, or a **Microsoft Teams** link from `onlineMeeting.joinUrl` — surfaced to the caller plus persisted on the decision ledger as `meet_link`. The tool descriptions are static, backend-neutral text, so the cached tool block never depends on settings; the pre-rename `add_google_meet` key is still honoured for old ledger rows. Either way the caps that matter still hold — roster membership, principal-protection, max attendees, and a daily booking ceiling.\n\n**Around-the-meeting participation.** The Executive doesn't join the call as a live voice; instead, on every successful booking it schedules a best-effort post-meeting follow-up (~5 min after the end) that DMs a human attendee asking for a quick recap — key decisions and action items. The reply flows back through the same cross-session DM context linkage described below, so the Executive can act on it (summarize, follow up) in a later turn. No meeting-transcript or Drive integration is required.\n\n### Workspace providers\n\nThe Executive can *discover* any mail or calendar tool the gateway exposes, but three code paths call a fixed backend with no model in the loop and used to spell Gmail tool names inline: the inbound mailbox poller (list unread → fetch → mark read), alert email dispatch plus the scheduler's email send-tool hint, and the typed booking tools plus the approval-time conflict check. `integrations.workspace` puts those behind two Protocols (`MailProvider`, `CalendarProvider`) with one module per backend (`google.py`, `microsoft.py`) and a registry keyed on `EMAIL_PROVIDER` / `CALENDAR_PROVIDER` — separate switches (Outlook mail with a Google calendar is legitimate), both defaulting to `google`, so an existing install changes nothing until an operator opts in. The Google backend is the old inline code moved verbatim; tests pin its argument dicts byte-for-byte. The Microsoft backend polls only the Inbox folder (`list-mail-folder-messages` with `isRead eq false`, never the all-folders listing that includes Junk), converts Graph's HTML bodies to text with a real parser, lists a message's attachments so the names appear in the same `N. name (mime, size KB)` shape Gmail's text uses (and reach peer memory), marks read with `update-mail-message`, sends through `send-mail`'s nested shape, books with `isOnlineMeeting` + `onlineMeetingProvider=teamsForBusiness`, cancels with `delete-calendar-event`, and answers the advisory conflict check from the Executive's **own** calendar view (delegated permissions cannot read attendees' free/busy — narrower than Google's `query_freebusy`).\n\nEvery backend renders an inbound message into the same `Subject:` / `From:` / `To:` / `Cc:` + `--- BODY ---` text the persona already expects, and the poller adds a `--- REPLY ---` block naming the exact reply tool and threading ids for *that* message (Gmail: `send_gmail_message` + `thread_id`; Outlook: `reply-mail-message` + `messageId`) after the mail's `<untrusted_content>` block, never inside it. The block rides in the user turn — which is what lets the cached persona say \"use the reply tool in the REPLY block\" instead of naming a vendor tool — and the scheduler's proactive-email trigger names the backend's send tool the same way. Providers never raise on a malformed backend response (they return empty / `None` / `{\"error\": …}`), so one bad cycle never kills a loop, and a chosen backend whose MCP server is missing from `mcp_servers.json` is logged at startup and skipped by the poller: fail-soft, never a boot failure. The poller's newer checks run on that rendered text for both backends (contacts and private turns, roster requests and their emailed answers, the audit row), with three Outlook differences that fail closed: no sender is ever authenticated (the `dmarc=pass` read is Gmail's), so there is no acknowledgement to an unknown sender, no emailed fact confirmation and no principal-authenticated mail; attachments stay a list (the download is Gmail's); and the brief's calendar block is skipped (its listing is Google Calendar's). The code's own emails to the principal (roster prompts, fact confirmations, scheduled deliveries) go through `workspace.registry.send_from_executive`, and `runner.email_ready` follows `EMAIL_PROVIDER`. A private turn's allow-list (`schedule_tools.PRIVATE_TURN_MCP_TOOLS`) and Act as me's mail-touched reads include the Outlook twins of the Gmail and Calendar tools, and roster answer tokens are hidden from Microsoft 365 results as from Google's.\n\n### Cross-session DM context\n\nWhen the principal says \"DM Alex about X\" in one session, OE sends the DM and — if a live conversation is active and the recipient is not one of the principal's contacts (a contact cannot DM the bot, and quoting the principal's conversation into their email reply's unattended turn would put it in the audit log) — records an `outbound_context` linkage row (recipient, full outbound text, originating session id). Email is covered too but recorded one layer deeper — there is no `send_email` tool, so the linkage is written in `orchestrator.mcp_gateway` after a successful `google_workspace__send_gmail_message` or `microsoft_365__send-mail`, one row per `to`/`cc` recipient keyed by the bare lowercased address (OE's own address skipped). When Alex later replies on the 1:1 DM path — or by email — the shared helper `integrations.inbound_hydration.hydrate_user_message` (called by the Discord, Telegram, and Slack handlers, and by the email poller keyed on the sender address) looks up the most-recent `open` linkage for that sender within 72h and prepends an `<outbound_reply_context>` block — the outbound text plus a backstory excerpt from the originating session — to the message handed to `Executive.chat()`. So OE knows what the reply is about instead of starting from an empty per-person session. The block is context for the LLM turn only: peer memory strips it before recording what the person said (`strip_outbound_reply_context`, applied in `sync_turn`, `sync_department_turn` and to the prefetch query), so the Executive's own DM is never attributed to the person who replied to it. The block lands in the user turn (caching rule) and is injected into a copy only — the pristine reply is persisted to history. The linkage is one-shot (consumed on first matched reply, race-safe) so an unrelated later DM isn't re-hydrated. This is deterministic and independent of the per-person `peer_memory` (Honcho) layer.\n\nEvery adapter that persists a turn (Slack, Discord, Telegram, and the web route) records the message's **resolved sender** on `chat_messages.sender_person_id` — the rostered Person the adapter matched, or NULL for an unrostered sender — independently of the session owner, which falls back to the principal on Discord and Telegram so those threads stay visible. Attunement's open loops and feedback rely on this so an outsider's words are never read as someone on the roster's.\n\n### OneDrive folder sync\n\nThe Microsoft twin of the Drive folder sync below (`knowledge.onedrive_sync`, `ONEDRIVE_SYNC_ENABLED`; setup in `docs/onedrive_sync_setup.md`). It copies the files in `ONEDRIVE_SYNC_FOLDERS` (`<drive id>/<item id>` pairs; `openexecutive onedrive-folder <link>` resolves a sharing link) into the isolated `onedrive_docs` collection on its own heartbeat, and shows them on the knowledge page as \"OneDrive\" rows with a **Sync now** button. Microsoft has no service account that sees only what is shared with it (app-only Graph access covers the whole tenant), so it reads as the Executive's own Microsoft 365 sign-in, and the folder list is the scope. It never stores a credential: each tick asks the launcher for a token (`ms365-mcp-launch.sh --access-token`, which runs `docker/ms365-access-token.mjs` through the server's own token cache and will only mint file-read tokens), keeps it in memory, and sends it only to Graph. `knowledge.onedrive_client` issues GETs only, follows a paging link only back to Graph, and fetches a file's pre-authenticated download link without the token; errors never carry that link. PowerPoint and older Office formats are read through Graph's PDF rendering.\n\n### Google Drive folder sync\n\nThe synced files show on the knowledge page's Company documents list as \"Google Drive\" rows (sync time, last edit in Drive, an Open in Drive link and a text view), and the page's source card shows the last run, any sign-in error or count of files that failed to sync, and a **Sync now** button that runs one tick on demand.\n\n\nSeparate from the Drive tools the gateway proxies. The optional folder sync (`knowledge.drive_sync`, `DRIVE_SYNC_ENABLED`; setup in `docs/drive_sync_setup.md`) copies the files in `DRIVE_SYNC_FOLDER_IDS` into the isolated `drive_docs` knowledge collection on a scheduler heartbeat, so the Executive knows those folders without searching Drive live (see RAG). It calls the Drive v3 API directly (`knowledge.drive_client`, `httpx`) as a **service account** with the `drive.readonly` scope (`DRIVE_SYNC_SERVICE_ACCOUNT_FILE`), not as the Executive's own account through workspace-mcp. Google enforces the folder scope, because the account sees only what is shared with it, and the sync can change nothing. Structured JSON means no display text to parse, and every file or folder id is checked against `[A-Za-z0-9_-]{1,128}` before it goes into a URL or a query. The two paths meet in two places. A synced chunk carries its Drive file id and sync time, so the Executive can open the live file with its own `get_drive_file_content`. The setup doc says to share each synced folder with the Executive's account too, so that live open works.\n\n### Confluence space sync\n\nThe optional Confluence sync (`knowledge.confluence_sync`, `CONFLUENCE_SYNC_ENABLED`; setup in `docs/confluence_sync_setup.md`) copies the pages of the spaces in `CONFLUENCE_SYNC_SPACE_KEYS` into the isolated `confluence_wiki` knowledge collection on a scheduler heartbeat, so the Executive knows the company wiki without a tool call (see RAG). It works with Confluence Cloud and Server/Data Center: `knowledge.confluence_client` calls the v1 REST API both serve, directly with `httpx`, signing in with a personal access token (Server/DC) or a username and API token (Cloud). It reads only, never follows a redirect (so the token stays on the configured site), streams responses with a 20 MB cap, and checks page ids and space keys before they reach a URL or a CQL query. The URL must be `https://` unless `CONFLUENCE_SYNC_ALLOW_HTTP` is set, and `CONFLUENCE_SSL_VERIFY` takes `true`, `false` or a CA bundle path for a private certificate authority. With `CONFLUENCE_SYNC_PUBLIC_HOSTS_ONLY` (off by default, for an install where whoever sets the URL must not reach the backend's own network) every request resolves the site's host, refuses it when any address is not public (loopback, private, link-local, shared), and connects to the checked address with the site's name kept for the Host header and the TLS certificate check, directly rather than through a proxy. `confluence_client.config_problem` is the same settings check the app runs at start, for a caller that writes these settings on someone's behalf.\n\nIt is separate from a Confluence MCP server such as mcp-atlassian, which a deployment can add to `mcp_servers.json` for live reads and edits in chat. The two share the env names `CONFLUENCE_URL`, `CONFLUENCE_PERSONAL_TOKEN`, `CONFLUENCE_USERNAME`, `CONFLUENCE_API_TOKEN` and `CONFLUENCE_SSL_VERIFY`, so one set of values serves both, and a synced chunk carries its page id so the Executive can open the live page with such a tool. A Confluence token is a user's, and what the sync reads becomes retrievable by everyone who talks to the Executive, so the space list is required and pages with a read restriction (their own or a parent's) are skipped unless `CONFLUENCE_SYNC_SKIP_RESTRICTED=false`. That keeps any existing member's token safe to use, without a dedicated licence. Manual runs: `openexecutive sync-confluence` / `purge-confluence`.\n\n### Your own Gmail or Outlook (Act as me)\n\nA second, separate mail connection, for **Act as me** (see **Act as Me (Delegation)**): a person's *own* mailbox, which `delegation.gmail.DelegateGmail` reads and drafts into over the Gmail REST API. It is never registered with the MCP gateway, so the model can reach it only through the typed `ghostwrite_email` handler — never by tool name through `call_tool`. The credential is one file per person in `DELEGATION_GOOGLE_CREDENTIALS_DIR` (default `company/delegation_google/`, `/data/delegation_google` in Docker), minted on a computer with a browser by `scripts/connect-own-gmail.py --email <you>` with the Executive's own OAuth client and the scopes `gmail.readonly` and `gmail.compose` — the same seed-it-yourself flow as the Executive's token, since the API serves no OAuth callback. It never lives in `WORKSPACE_MCP_CREDENTIALS_DIR`, where workspace-mcp picks a credential by address.\n\nEvery use re-checks with Google that the credential opens the mailbox on the person's People entry, and refuses outright when that address is the Executive's own: a shared mailbox would make everything the Executive writes look like theirs. Drafts are threaded (`In-Reply-To`, `References` of whole ids ending with the parent, `threadId`) and stamped `X-OE-Ghostwritten`; the only send is `drafts.send` of an existing draft by id, when the person taps Send on its reply card (`delegation.reply_send`). A Gmail rate limit, even as a 403, is an error to retry rather than a reconnect. The voice learner and the thread shown to the ghostwriter keep only each sender's own words through `email_poller.sender_new_text`, the poller's quoted-reply stripper made public. Setup status shows it as **Your own mailbox (Act as me)**.\n\nThe same file can instead hold an Outlook sign-in (`\"provider\": \"microsoft\"`), minted by `scripts/connect-own-outlook.py --email <you>` with a device-code sign-in on the Executive's own Entra app (`MS365_MCP_CLIENT_ID`, `MS365_MCP_TENANT_ID`, default `common`) and the delegated scopes `User.Read`, `Mail.ReadWrite`, `Mail.Send` and `offline_access`. `delegation.gmail.gmail_for` then hands back `delegation.outlook.DelegateOutlook`, which talks to Microsoft Graph directly — never the Microsoft 365 MCP server the Executive's own Outlook uses — with the same methods and the same one send: Graph's `send` on an existing draft, with no body. Microsoft rotates the refresh token on use and the new one is written back into the file. A sender counts as verified only from the topmost `Authentication-Results`, the one Exchange wrote, and only mail in their Sent Items counts as theirs.",
+ "markdown": "External channels connect callers to the Executive. Each inbound adapter resolves the sender to a Person — the chat adapters via `find_person_by_{slack,discord,telegram}_id`, the email poller via `people.identity.resolve_email_sender` (a primary address or an alias, then the company-domain rule; see Org) — and threads `person_id` into `Executive.chat()`. A sender who resolves to nobody is no longer simply dropped: see **New senders (roster requests)** below.\n\n### Inbound channels\n\n- **Slack** (`integrations.slack_bot`) — async Bolt socket-mode (`AsyncApp` + `AsyncSocketModeHandler`) embedded in the lifespan; responds to `@mention`s, DMs, and thread continuations, with per-conversation session memory (see **Conversation state** below). Text only: there is no interactivity endpoint and no Block Kit, so nothing Slack sends carries a button.\n- **Discord** (`integrations.discord_bot`) — discord.py gateway embedded in the lifespan; per-thread auto-titling, session-scoped memory, author attribution, configurable `@mention` thread promotion.\n- **Telegram** — webhook gating + per-chat history.\n- **Google Chat** — webhook + service-account auth. It resolves no approval gates and has no roster gate (the webhook authenticates Google, not the sender), and it never populates the session's trusted alert ids, so `ack_alert`'s trust check — which now runs on every session rather than only those declaring an `origin_channel` — refuses there by default, and its `<channel>` block says plainly that a sign-off cannot be answered from it.\n- **Email** (`integrations.email_poller`) — MCP poll loop over the Executive's own mailbox — Gmail via `workspace-mcp` or Outlook via `ms-365-mcp-server`, chosen by `EMAIL_PROVIDER` (see *Workspace providers* below); it starts only once the MCP gateway is up, never surfaces Reply-To redirects, filters auto-replies. Its sessions carry `origin_channel=\"email\"`: the untrusted-content policy (`orchestrator.content_trust`) reads that tag to tell mail from the principal's own surfaces — left empty, as it once was, every stranger's body reached the decision extractor as the principal's words. The mail goes into the turn inside an `<untrusted_content source=\"email\" author=\"<sender>\" author_verified=\"no\">` block that says it carries none of the principal's authority, with the framing line and any `[POLICY]` notice outside it; mail from the principal's primary address that Gmail authenticated (`dmarc=pass`) is their own words — but only their new text goes in unwrapped: when the mail quotes or forwards another message (`_new_text_lines` cut something), the message as received follows in an `<untrusted_content source=\"email_quoted\">` block, since a stranger's text they replied to or passed on is not theirs. While the Executive is paused (`scheduler.pause`) it skips each poll and leaves the inbox unread, so held mail is processed on the first poll after resume; the chat adapters above are not paused. The Executive sees the whole message; peer memory records only `_email_memory_text` — the subject, the sender's new text (cut at the first quoted or forwarded message) and the attachment filenames — parsed from the rendered message with any trailing `--- REPLY ---` block dropped first, without backtracking-prone regexes, since the text is sender-controlled, and only for a rostered sender, the only kind with a peer to record into. The poller's `integration_inbound` row keeps the mail's text for the audit log, the way the reply's `chat_turn` row keeps the Executive's answer: `full.message` is the sender's new text (what web chat keeps for a turn), `full.attachments` the attachment filenames, and `details.preview` the first 120 characters of the new text, so /audit shows what was asked next to what was answered (the summary still reads `Inbound email from <sender>: <subject>`). The whole body with its quoted chain (`full.body`) is kept only on a row private to the principal: the chain is other people's mail — a contact's, on the principal's reply to one — and every signed-in user reads a public row. Each field is cut by its escaped JSON length (20,000 characters of text, 8,000 of attachment names), the measure the logger's 64 KB cap uses, so a long mail in a non-Latin script loses its tail rather than the row's structure. Outbound, the principal briefs and the client-rotation digest are emailed through the same gateway with the configured backend's send tool (`google_workspace__send_gmail_message` or `microsoft_365__send-mail`, via `workspace.registry.send_from_executive`, from the Executive's own mailbox to the principal's address, subject `<brief title> — <date>`, the date in the user's zone) as formatted HTML — rendered from the brief's Markdown with raw HTML and images off, so text quoted from outside shows as text and opening the email loads nothing remote. Email is first when the principal prefers it, and otherwise the backup after every chat channel, tried only when Slack, Discord and Telegram are missing or fail; an in-band error from the tool counts as a failed send and the next channel is tried. An emailed brief is not a DM, so the per-Person outbound guard below does not apply — the gateway's roster egress gate does.\n\n Two senders get their own framing. Mail from one of the principal's **contacts** (a `kind=\"contact\"` People row — see Org) replaces the non-roster `[POLICY]` notice with a contact one: it names them (name and role/company), says not to reply unless the principal asks and not to tell anyone else, and asks for a summary for the principal instead. Both this turn and the forwarded-by-principal turn below are marked `Session.private_to_principal`: they can email, DM or invite only the principal, an alert they raise is private to the principal, and they are not offered the tools that reach anyone else or start work outside the turn (`schedule_tools.PRIVATE_TURN_WITHHELD_TOOLS`: broadcasts, meeting bookings, workflows, follow-ups, watches, artifacts, goal updates, skill drafts, connecting a new MCP server), and of the MCP tools only Google Workspace reads and the gated Gmail send (see MCP below). Every audit row about such mail is private to the principal (see Audit Log), the poller's two `integration_inbound` rows included (`_private_to_principal_mail`), so the audit log's other readers see none of it. The poller holds `audit.private_rows` for the whole handling of such a mail, because some rows are written before the turn binds its private session: the knowledge retrieval quotes the mail's From line, subject and opening. The rows themselves read exactly like an outside sender's: the `integration_inbound` row is the non-roster one (`accepted_non_roster`), and the notice opens with the non-roster notice's exact sentence. The poller handles each message with no session bound, so one message's private turn never marks the next message's rows. The turn stays unattributed (`person_id` None — no peer memory, no co-presence), and the gateway refuses mail to a contact on a poller turn anyway. Mail **forwarded by the principal** — the sender resolves to the principal and `_new_text_lines` reports a forwarded message — is wrapped in a `<forwarded_by_principal>` block: summarise it, draft a reply the principal could send (as text in the reply, never sent or saved as a Gmail draft), note any commitment or date, and offer to add the original sender as a contact (confirmed from the web app or the principal's Slack / Discord, since an ordinary email turn cannot change the People list — the one email that can is the principal's tokened reply to a roster-request email, which is parsed before any model turn; see below). The reply still goes to the principal only. The From header is unauthenticated, so this framing grants nothing: a spoofed \"forward\" just gets the real principal a summary.\n\nSlack, Discord, and Telegram are roster-gated on inbound: an unrostered sender's message never reaches the Executive (`details.outcome=\"rejected_unknown_sender\"`). The gates look up **team members only** — a contact's Slack, Discord or Telegram id is treated as unknown, so a contact can never talk to the bot (nor count as a co-present peer). What happens next is the roster request below — except for a contact, a thread continuation, a Telegram group or an unverified Telegram webhook, which stay silent. Email is not inbound-gated — non-roster mail flows through with a `[POLICY]` notice so triage can classify it. Slack/Discord thread continuations pass the shared `integrations.response_gate` (utility-fast model, biased toward replying). Slack only continues a thread it has replied in: it reads the thread with `conversations_replies` (5 s bound) and looks for its own user id, which `auth_test` resolves at startup. If that lookup fails, mentions and DMs still work and thread continuations stop until a retry succeeds (`slack_bot._resolve_bot_user_id`, at most once a minute, from the channel `message` listener). When Slack doesn't return the thread in time, the thread's stored history (`slack:thread:{channel}:{thread_ts}`, written on every turn the bot answers there) stands in: with history, a rostered sender gets an ephemeral \"I couldn't read this thread just now… @-mention me\" and the drop is audited (`outcome=dropped_thread_unreadable`). A thread the bot never answered in stays silent. A message the gate skips is audited (`outcome=skipped_gate`, with the gate's reason) and gets an :eyes: reaction (`reactions_add`, which needs the bot's `reactions:write` scope; without it the reaction is skipped), so a sender the gate misjudged can see the message was read and @-mention the bot.\n\n**No link previews on replies.** The Executive's replies go out with previews off — Telegram `sendMessage` carries `link_preview_options.is_disabled`, Slack's reply `say` passes `unfurl_links=False` / `unfurl_media=False`, and every Discord reply send passes `suppress_embeds=True`. The direct messages the Executive starts are the same: `send_slack_dm` posts with `unfurl_links=False` / `unfurl_media=False`, and Discord's `send_dm` sets the `SUPPRESS_EMBEDS` flag on each chunk, so a reminder or a brief sent by DM carries no preview either. Otherwise the chat service fetches a link in the text on its own, so a reply that quotes someone's mail (Act as me reads it) could carry data to a sender's server with no click. The web chat does the same for images: `Message.tsx` loads only an image the app itself serves (`lib/markdownImages.loadsInline`) and shows any other as a link.\n\n### New senders (roster requests)\n\nSomeone who writes in and is not on the People list — and is not one of the principal's contacts — is **held, acknowledged and asked about** instead of dropped (`integrations.roster_intake.intake` over the `people.roster_requests` ledger; lifecycle in Org). Each adapter calls it after its gate:\n\n- **Slack** — a DM or an `@mention` (not a thread continuation: the bot was not addressed there). The name and profile email come from `users_info` (5 s timeout). The acknowledgement goes in the DM, or as an **ephemeral** message for a mention, so nobody else in the channel learns who is and is not on the roster. The held event keeps only a fixed set of keys (text, user, channel, ts, thread, files).\n- **Discord** — a DM or a mention (not a thread the bot merely follows). The acknowledgement goes in the DM, or by DM for a channel mention; `/ask` answers the interaction with an ephemeral acknowledgement first (Discord allows 3 s), then holds the prompt. Its old \"ask the admin to add you\" reply is gone.\n- **Telegram** — only a **private** chat, and only when `TELEGRAM_WEBHOOK_SECRET` is set and valid: Telegram identifies the chat rather than the sender, and without the secret anyone could forge an update. A group, or an unverified webhook, stays silent as before. Attachment file ids are held and downloaded only on replay.\n- **Email** — non-roster mail that is not private to the principal (so never a contact's) and does not look machine-sent (`looks_automated`: a no-reply / notifications / billing-style local part, a `List-Id` / `List-Unsubscribe` / `Auto-Submitted` / bulk `Precedence` header, or a DMARC failure). The mail is **still triaged** by the Executive on this pass; its `[POLICY]` notice now says the sender was already told their message is waiting and the principal was asked who they are, so the model raises no second alert just to add them, and that the email will come back once they are added.\n\nEvery path skips a contact first (`find_person_by_*_id(..., include_contacts=True)`, or the email poller's private-mail rule): a contact has no chat access and gets no request and no acknowledgement, and the team-visible `rejected_unknown_sender` / `accepted_non_roster` rows read the same for a contact as for a stranger.\n\n**The acknowledgement** is one fixed text (`roster_intake.ACK_TEXT` — \"your message arrived … waiting for confirmation\"), never the model and never an echo of what they wrote, so a forged sender gains nothing by triggering it. It is sent at most once per sender per `ROSTER_ACK_WINDOW_DAYS` (7) and at most `ROSTER_ACK_DAILY_CAP` (20) a day across all senders (`claim_ack`, one transaction; a failed send releases its claim). An email acknowledgement is the one send the MCP gateway lets reach an address off the roster: `send_email_ack` sets `mcp_gateway.roster_ack_grant`, a one-shot ContextVar grant that admits a single `send_gmail_message` with exactly the granted recipient (a plain string, no list), subject and body and no other key, then is spent (audited `integration_outbound`, `kind: roster_ack`, private). An email sender is acknowledged only when Gmail authenticated them: once `claim_ack` succeeds, the poller's `send_ack` re-reads the mail as raw MIME and needs Gmail's own topmost `Authentication-Results` reporting `dmarc=pass` for the From domain (`fact_confirmation.sender_authenticated`, fail-closed). The printed headers the poller reads never carry that header, so without the second read a forged From would draw the acknowledgement to whoever it names — backscatter from the Executive's mailbox. An unauthenticated sender makes `send_ack` raise `roster_intake.AckWithheld`: the claim is released, and they are held silently. `authenticated_by_gmail` strips quoted strings and comments before reading the verdict — Gmail copies the envelope sender and DKIM tags into its own stamp, and a quoted local part could otherwise carry a fake `dmarc=pass` — and needs exactly one `dmarc=` verdict reading exactly `dmarc=pass header.from=<domain>` (Gmail may follow it with its own `dara=` resinfo, so its position is not checked). Residual: when Gmail writes no verdict at all (a From domain without DMARC) and echoes an unsanitised `;dmarc=pass header.from=…` elsewhere in its stamp, that echo would read as the verdict; bounded by the acknowledgement caps. A withheld acknowledgement is audited `roster_ack_withheld`. `intake` acknowledges before surfacing the card and telling the principal, and both follow the request's `ack_sent_at`: a withheld, failed or capped acknowledgement reads \"haven't been told anything yet\" on the Today card, in the principal's prompt and in the email `[POLICY]` notice.\n\n**Telling the principal.** A new request puts a private card on the principal's `/today` (see Today) and pushes a prompt on their delivery plan — preferred chat DM, then the others, then email — for a chat sender or an email sender on the company's own domain; an outside email sender waits on the card, as does everything past `ROSTER_REQUEST_DAILY_CAP` (30) pushes a day (the same cap bounds how many requests open a day; a declined sender is not raised again for 30 days). A chat prompt can be answered in place (\"that's Annamarie, add her to the team\"). An email prompt carries a one-time token (`RR-` + 20 base32 characters, stored hashed); the poller checks every inbound mail for one **before** triage (`try_email_roster_answer`): the reply must come from the principal's primary address exactly (not an alias, not the company-domain rule), carry the token of a pending request, not be an automatic reply, and not fail DMARC where the header is visible; and is parsed by `parse_decision`'s `roster_identity` shape — anything else falls through to the ordinary path. The token is hidden (`RR-[hidden]`) in every other Google Workspace result the gateway returns — the confirmation sits in the Executive's Sent folder, and any turn could otherwise read it — except the poller's own fetch of an inbound message (`mcp_gateway.reveal_roster_tokens`); a token-bearing mail that is not an answer is scrubbed before triage.\n\n**Replay.** Once the principal answers (web card, chat tool, email reply) — or a People-page / `upsert_person` write makes the sender match someone (`after_roster_write`) — the held messages are replayed through the adapter that received them (`register_replayer`: Slack re-runs the held event, Discord and Telegram their held text, email re-fetches the message from Gmail and handles it as mail from a known sender), oldest first, now attributed to the new Person. A teammate gets every held message; a new contact only held email (it takes the private contact path; held chat is dropped); a decline drops everything. Replays run in a **fresh `contextvars.Context`**: the answer may have come from the principal's own verified turn, and a stranger's message must never be read with the principal's session — and their contacts — still bound.\n\n### Standing facts by email (confirm by reply)\n\nThe principal can ask for a standing fact, a retirement or a company-profile edit by email, and the fact tools check it exactly as in chat (the own-words gates run on `_email_memory_text`, the sender's new text, not the quoted chain or attachments). An email's From line proves nothing, though, and a standing fact is read by every later prompt as the principal's own account. So an emailed change is **held, never applied**, until the principal's own mailbox confirms it (`integrations.fact_confirmation`, `memory.facts.fact_confirmations`):\n\n- **Who may ask.** The poller stamps every email turn with `Session.email_from` (the bare lowercased From address) and `email_authenticated`. The text the poller reads never carries `Authentication-Results` (workspace-mcp prints a fixed set of headers), so for mail claiming the principal's primary address it reads the message again as raw MIME (`fact_confirmation.sender_authenticated`) and requires Gmail's own stamp: the **topmost** `Authentication-Results` header (Gmail adds it on receipt, above anything the sender wrote), authserv-id `mx.google.com`, reporting `dmarc=pass` for the From address's domain, with the raw From matching (`authenticated_by_gmail`). It fails closed: no header, `dmarc=none`, a temporary error or a failed read all leave the turn unauthenticated. Mail from anyone else costs no second read and is never authenticated. `fact_tools._gate` admits an email turn only from the principal's **primary** address, exactly (no alias, no company-domain rule), authenticated, and not private to the principal (mail they forwarded stays refused, as do contacts, teammates and unknown senders).\n- **The hold.** The tool returns `awaiting_confirmation` (no action chip, and the model is told to say nothing has changed). `request_confirmation` stores the change with the SHA-256 of a fresh one-time token (`FC-` + 20 base32 characters, 100 bits; 7-day expiry; at most 10 waiting at once) and emails the token to the principal's roster address, outside any turn's outbound context. A profile edit is dry-run first, so one that cannot apply is reported in the turn, before any email goes out. If the email fails, nothing is held.\n- **Keeping the token from any model.** The tool result never carries it, and `mcp_gateway.hide_roster_tokens` now scrubs `FC-` tokens as well as `RR-` ones (`FC-[hidden]`) from every Google Workspace result except the poller's own fetch of an inbound message. Act as me's reads of the principal's own Gmail, where the confirmation email lands, hide them too (`delegation.gmail.parse_message` and the thread search), so the drafting model never sees a live token either.\n- **The reply.** `_handle_one_email` hands every message to `try_email_fact_confirmation` right after the roster-answer check, before any model turn. It acts only on a reply from the principal's primary address carrying the token of a **pending** confirmation. A made-up, spent, cancelled or expired reference matches nothing and falls through to the ordinary path with the token hidden, so no old or forged token can make the Executive email the principal. A reply with a live token is read again as raw MIME (`read_raw`) and must pass the same Gmail DMARC check as the request; one that doesn't (or whose raw read fails) is refused, the change keeps waiting, and the principal is told. Then an **automatic reply** is consumed quietly — no answer, no model turn, the change keeps waiting: the printed headers catch a Gmail vacation reply (`Precedence: bulk`), and the raw headers catch what they don't show (`automatic_reply`: `Auto-Submitted` other than `no`, `X-Auto-Response-Suppress` and similar, `Precedence: auto_reply`, an out-of-office subject), since an Exchange out-of-office echoes the token from the principal's own address and may well say \"I'll confirm on my return\". The raw message starts at the tool's exact separator line (`\\n\\n--- RAW MIME ---\\n`), so a sender-written Subject imitating it cannot supply the headers. CONFIRM (or yes / approve) applies the change through `fact_tools.apply_confirmed`, the same apply step chat uses. CANCEL (or no) drops it, and an unclear reply leaves it waiting. The principal gets a one-line result either way. A compare-and-set on the row means two replies can apply it only once.\n- **Audit.** Private `fact_confirmation` rows record held / applied / cancelled / refused. The applied change also writes the fact tools' usual private rows, with `source_channel=\"email\"` as its provenance.\n\n### Checking a channel is set up\n\nSettings → Setup status (`GET /setup/status`, `api/setup_checks.py`) gives each channel a light and a plain-language fix. It tests the credentials with one free read-only call — Slack `auth.test`; Telegram `getMe`, then `getWebhookInfo` (is a webhook registered, is it this app's `/webhook/telegram`, did a delivery fail in the past day); Discord `users/@me`, only while the bot is still connecting — and reads what the lifespan left on `app.state`: the Slack socket's `is_connected()`, the Discord bot's `is_ready()` together with its gateway socket being open (discord.py stays ready while it reconnects after a drop), or the exception its task stopped with (`LoginFailure`, `PrivilegedIntentsRequired`), and the MCP gateway. A listening Slack is amber while the listener hasn't resolved its own user id (`slack_bot.bot_user_id()`), since thread continuations are off until it does. Because inbound is roster-gated, a connected Slack, Discord or Telegram is amber while nobody on the roster has its id (`slack_user_id`, `discord_user_id`, `telegram_chat_id`), or when its latest `integration_inbound` row is a `Rejected: …` from the past day whose sender still isn't on the roster — the page shows the sender's id, never message text. Local login with no `TELEGRAM_WEBHOOK_SECRET` is red for Telegram, since the local-login gate turns its webhook away. Google Chat and Gmail are checked from configuration: a numeric project number and a readable service-account key; the OAuth client (or service-account key) the workspace-mcp child signs in with.\n\nA channel setting written `KEY= # note` in a `.env` file arrives as the note's text, because python-dotenv strips only a comment that follows a value. `Settings` reads such a value as unset for the Slack, Telegram, Discord and Google Chat credentials, so the line no longer switches a channel on with its own comment as the token. `TELEGRAM_WEBHOOK_SECRET` is the exception, because read as unset a junk secret would turn the webhook's check off. It is kept instead, and since Telegram can only ever send 1–256 of `A-Z a-z 0-9 _ -` (what `setWebhook` accepts), a secret outside that set — a leftover comment, a pasted `<value from Step 2>` — could only be matched by someone who guessed it: while it is set the webhook refuses every update, neither the local-login gate nor the roster tools count it as proof (`Settings.telegram_webhook_secret_valid`), and Setup status shows it red (`.env.example` shipped four such lines; the notes now sit on their own lines).\n\n### Conversation state\n\nEvery adapter keys a deterministic `session_id` off the conversation surface and replays that session's stored turns into `Session.conversation_history`, so a follow-up (\"yes, go ahead\") lands with the question still in context. Slack's ids are `slack:dm:{user}`, `slack:thread:{channel}:{thread_ts}`, and `slack:channel:{channel}:{user}` for a mention that starts a thread; because the bot's reply roots a Slack thread at that message's `ts`, the turn is double-written to that thread's session too, so the first continuation inside it is not cold. A reply inside a thread THAT PERSON ROOTED also offers the parent `slack:channel:{channel}:{user}` id when matching an approval gate, since a gate raised on the mention recorded that session while the answer arrives under the thread's. The root-author check is load-bearing: widening on \"threaded and not a DM\" alone offered one person's channel session inside any thread in that channel, so a remark they made to a colleague could resolve their open gate and post the run's title into that colleague's thread. Discord has the same shape — its auto-thread router promotes a mention reply into a bot-owned thread — and the same alias. An @-mention inside a DM is handled once, by the DM path: Slack fires both events, and taking both forked the conversation across two session ids. The same id is the audit grouping key and the scope passed to `format_for_prompt`. A per-session `asyncio.Lock` serializes turns within one conversation — the inflight semaphore caps fan-out across the workspace, which does nothing to stop two messages in the same thread from both reading history before either writes it. Persistence happens after the reply is sent and its failure is logged, never surfaced: the user already has their answer.\n\n**Conversations within a chat app.** A Telegram chat, a Slack or Discord DM and a person's @mentions in one channel (`slack:channel:…`, `discord:channel:…`) are each one endless stream, so they are cut into conversations (`memory.conversations`, shared helpers in `integrations/conversations.py`). Each adapter computes the stream's base id as before and asks `conversation_for` which conversation a message belongs to: the newest, unless the stream has been quiet for two hours (`CONVERSATION_GAP`), when the next one starts. The first conversation keeps the bare base id; later ones add `~2`, `~3`, … (`memory.conversation_ids`). A whole message of `/new`, \"new chat\" or \"new conversation\" starts a fresh one instead of a turn and is answered with the name of the one it ended; the new row has no messages and stays off the Chats page until its first turn. After a conversation's first exchange `title_after_turn` names it in the background with `utils.session_title.generate_session_title` (the first words of the message if that fails), as a web chat's first turn does. Threads (`slack:thread:…`, `discord:thread:…`), email and every other id are left whole. What has to see the whole stream compares base ids: the inbound resolver, so an approval asked before a cut is answered after it, and the session-scoped decisions and advice `format_for_prompt` renders. At boot `split_existing_streams` cuts history saved before this by the same rule, once (`app_migrations`), titling each part from its first message and carrying a read-mail lock to every part. An adapter installed outside the package adds its streams with `conversation_ids.register_rolling` and cuts its own old history once, with its own migration name and an `only` filter for its base ids, because the built-in sweep may already have run without it.\n\nThis was absent on Slack until #136. Every message built a fresh `Session` and minted a per-*message* id (`thread_ts` falls back to the message's own `ts`, so a DM never repeated one), which meant a multi-turn confirmation could not work: the Executive asked whether to send a briefing to Slack, the user said yes, and the next turn had no record of the question.\n\n### What the model is told about the channel\n\nChat adapters pass a `<channel>` block (`integrations.channel_context`) in the **user turn** — never a cached system block. It names the surface, states that the person cannot see the app's cards or buttons, and draws the line the assistant kept crossing: briefing proposals and alerts are actioned on the briefing page; a proposal that books something can only be approved there; a workflow sign-off the Executive is explicitly waiting on *can* be answered in the conversation. The principal's DMs on Slack, Discord and Telegram also receive the open-alert digest (`briefing.context.format_open_alerts_for_prompt`) as the `<briefing>` block, gated to DMs because the board is company-wide and a shared channel must not pull it in. That digest is the ONLY trusted source of an `alert_id` for `ack_alert`. The briefing page's `[Discuss mode — alert_id=N]` primer is a pointer to which item is under discussion, not authority on its own: the server accepts that id only if it is also on the live board it derived itself. An id appearing anywhere else, including inside an alert's own body, is not trusted: alerts are minted from inbound email and chat, so their bodies are attacker-controlled. Two mechanisms enforce that rather than one prompt rule. Every field the digest interpolates is collapsed to a single line, because a newline in an attacker-written headline, suggested action, tag or review note would otherwise forge a whole extra `[N]` line — a fabricated instruction to clear somebody else's alert. And the live board the server derived is recorded on the session, so `ack_alert` refuses any id outside it — on EVERY session, web included; the web previously had no check at all. The recorded set is the whole live board (the cap `/today` renders as cards), not just the rows the digest printed, so a card past the digest's 30-row render cap is still ackable rather than sending the principal back to the page. Be precise about what this buys: it stops an id that is invented, closed, snoozed or expired, including one quoted out of an attacker-written alert body. It does NOT stop the model being argued into acking the wrong LIVE card — one alert's body can still make a case about another id that is genuinely open. Scoping trust to the single item under discussion would need the primer id carried server-side, which it is not.\n\n### Handler failures\n\nAn adapter's catch-all replies \"I encountered an error processing your request.\" On Slack that used to be the whole story — one uncorrelated log line, no audit row, and an unguarded error-send that could itself escape into Bolt. It now logs with the session, sender, channel, thread and mode, and writes an `integration_inbound` row with `details.outcome=\"handler_error\"`, joining the inbound row under the same `session_id`. Upstream, a single crashing tool handler no longer takes the turn down at all: the orchestrator's tool gathers use `return_exceptions=True`, hand the model an error `tool_result`, and audit it with `ok=false` (cancellation is re-raised, not swallowed).\n\n### Attachments\n\n`integrations.attachments` handles web chat, Discord, Telegram, Slack and Google Chat. Images become Anthropic vision blocks; documents (PDF/DOCX/TXT/MD/CSV) are inlined into the turn that carried them — inside an `<untrusted_content source=\"attachment\" author=\"<name>\">` block (`content_trust.wrap_untrusted`, whose closing tag the text cannot forge), under an `[Attached: <name>]` label, which the open-loop pass reads as \"not the speaker's words\" and the decision extractor strips before it reads the turn — and indexed into the ChromaDB `inbound_attachments` collection. 20 MB/file cap; web chat uses `POST /chat/upload`.\n\n**Scanned PDFs** are read, not reported empty. A PDF with no text layer goes through `knowledge.pdf_reader`: with `PDF_PROVIDER_READING=true` (opt-in, off by default) the deployment's own model reads it through its provider's PDF support (Anthropic natively, OpenRouter's `file-parser`, a local server only with `LOCAL_PDF_INPUT`); otherwise, and as the fallback, local OCR (pypdfium2 + RapidOCR, offline, no key) — so every deployment converts them. The label then reads `(converted from scanned pages)`, and an unreadable one says why. The index takes the text already extracted for the turn, so a scan is converted once. A file a channel delivered (everything here except `POST /chat/upload`, where the signed-in user sends it) converts at most `PDF_INBOUND_MAX_PAGES` (30) pages and draws on a rolling `PDF_INBOUND_PAGES_PER_HOUR` (300) budget shared by all senders, so sending scans cannot run up unbounded model spend or CPU.\n\nPer channel:\n- **Slack** — a `file_share` message (previously dropped outright for its `subtype`) is handled like any message; files download only after the roster gate and the response gate, with the bot token sent only to `https://files.slack.com/` (needs the `files:read` scope). History and peer memory record the words plus `(Attached files: …)`, never a document's text.\n- **Google Chat** — uploaded attachments download through `media.download` with the reply credentials (`chat.bot`), checked against the 20 MB cap per chunk, 5 per message; an attachment-only message is no longer dropped. The webhook authenticates Google, not the sender, so files are read only when the event's `user.email` is on the People roster (a team member); anyone else's files are named to the Executive, never downloaded. A Drive file shared into a message is left to the Drive tools, with its id.\n- **Email** — the poller (`integrations.email_attachments`) downloads up to 5 readable attachments of the message itself via `get_gmail_attachment_content` — workspace-mcp in stdio mode saves the file and answers `Saved to: <path>`, which must start its line, be the last such line and end in the answer's `Saved filename:` (the answer echoes the sender-chosen attachment name first, and a name carrying `Saved to: /…` must not redirect the read), then resolve inside the tool download folders — and appends their text to the turn (15k chars each). Only for a sender the principal knows (team, contact, principal); a stranger's files stay a list. It runs outside the model loop, so it works on private turns without adding the attachment tool to `PRIVATE_TURN_MCP_TOOLS`.\n- Anything else (a Drive PDF, a skipped attachment) — the Executive's `read_document` tool.\n\nIndexed chunks are keyed on `attachment:<bare filename>` and tagged `type=attachment`. An attachment's name is chosen by whoever sent the message and the name is the chunk-id namespace, so without the prefix an inbound `strategy-2026.md` would upsert over the curated company document of that name. The prefix separates inbound content from curated uploads and keeps provenance visible; it does not separate senders from one another.\n\nIsolation is the **collection**: `retriever.retrieve` never queries `inbound_attachments` at all, so an attachment cannot resurface as company knowledge in a later, unrelated turn. Attachment processing is roster-gated on Discord too — it previously ran before the gate, so an unrostered sender's file was downloaded and indexed before their message was dropped. These rows still have no operator surface: they are never written to `company/docs/`, so `GET /documents` does not list them.\n\nThis replaced a domain-based scheme that did not work. The chunks lived in `company_docs` under a domain outside the specialist set, on the premise that an unrecognised domain matched no filter. It matched every one — `ChromaDBStore.query` attaches a `where` clause only when a domain filter is present, and an Executive-level retrieval passes none — so attachments came back under “From your company documents”, top-ranked and indistinguishable from a curated upload. Legacy rows are migrated across on boot.\n\n### MCP gateway & outbound\n\nMCP runs in **both directions**. As a *client*, `orchestrator.mcp_gateway.MCPGateway` proxies MCP servers (Gmail, Calendar, Drive, Outlook mail and calendar — whatever `mcp_servers.json` configures) as Executive tools, with egress gating on Gmail recipients, Calendar attendees, Drive share grantees, and Microsoft 365 recipients/attendees against the roster (and a refusal of public / whole-domain Drive sharing). The Gmail gate also accepts only a fixed set of argument keys, so workspace-mcp's `reply_all` (recipients taken from the thread, never checked) and `forward_message_id` / `include_forwarded_attachments` (forwarding any inbox message) are refused; a reply is `thread_id` plus an explicit, roster-checked `to`. Three more refusals close paths around those gates: every Apps Script tool (code running as the Executive's account, outside the gates, and since workspace-mcp 1.29.0 schedulable to keep running); any argument that makes the server fetch a URL (`fileUrl` / `file_url` on the Drive and import tools, a Gmail attachment given as a `url`), since the fetch can carry data out in its query string before any recipient gate matters; and text Google would mail to someone the roster never checked (`rsvp_comment` to an event's organizer, an out-of-office `decline_message` to whoever invites the Executive). Every gate reads its arguments by normalized spelling: workspace-mcp 1.29.0 renames a camelCase key such as `Attendees`, `userGoogleEmail` or `rsvpComment` to the real parameter, so a gate matching the exact key would never see it. Arguments that are not a JSON object are refused for the same reason. Google-side fetches are refused too: any key ending in `url` / `uri` at any depth whose value has a URL scheme (a Slides image or page background, a Docs or Forms image; a document hyperlink and a meeting link are stored, not fetched, and pass, as does a Drive file id), (a value is a URL after the characters a URL parser strips are removed, and `//host` counts), and on every tool that can write cells — the Sheets tools and the Drive create / update / import tools, which turn CSV into a Sheet — any string anywhere holding a formula call that fetches (`IMPORTDATA`, `IMPORTXML`, `IMPORTHTML`, `IMPORTFEED`, `IMAGE`), JSON-encoded or mid-CSV included: a string argument that encodes JSON is read as the server decodes it (one too large to read is refused), text is also scanned with character references decoded and comments / CDATA removed, and an upload that would become a native Sheet may only be text `content` — a `base64_content` or server-side `file_path` upload there is refused, since an XLSX hides its formulas behind zip members and XML encodings; a PNG to Drive or a DOCX to Docs is untouched. Not covered, unverified: whether Drive's HTML / Markdown import fetches an `<img src>` when it converts `content` into a Doc. A `manage_event` call that names no attendees is sent with `send_updates=\"none\"`, so the event's existing guests are not emailed text the gate never checked. The allow-list (`_egress_sets`, built by `_roster_allow_set` as a `people.identity.RosterAllow`) is every team member's email and email aliases plus the Executive's own address, always, and — for a teammate — any address on one of the company's own domains whose local part (`+tag` dropped) matches theirs, the same rule inbound mail is matched by; a **contact's** email joins it only when `people_tools.contacts_reachable_now()` is true — a turn the principal started on a verified surface (web chat, their own Slack or Discord, a verified private Telegram chat), or inside `grant_contact_egress()` (the decisions approve route, when the approver is the principal). On any other turn — an inbound email, a teammate, the scheduler, a workflow, the alert review — a contact's address is refused exactly like a stranger's, so the refusal cannot reveal that it belongs to a contact (contacts are private to the principal). The single exception to the roster rule is the fixed roster-request acknowledgement above, under its one-shot `roster_ack_grant`. A turn about the principal's private mail (`Session.private_to_principal`) narrows the list to the principal's own address and the Executive's. Such a turn may call only the MCP tools in `schedule_tools.PRIVATE_TURN_MCP_TOOLS`: Google Workspace reads (`search_gmail_messages`, `get_gmail_message_content`, `get_events`, `list_calendars`, `query_freebusy`, and the Drive, Docs and Sheets reads such as `search_drive_files`, `get_drive_file_content`, `get_doc_content` and `read_sheet_values`) and the two Gmail tools this gate narrows to the principal (`send_gmail_message`, `draft_gmail_message`). It is an exact allow-list, not a server: workspace-mcp runs at the tier `WORKSPACE_MCP_TOOL_TIER` names, `complete` by default, where a Google Chat message to a space, a Docs, Sheets or Drive write into a file already shared with others, `manage_event` on a shared calendar (its gate checks attendees and pins notifications, nothing else) reaches people with no recipient check; and every other server has none at all (a Slack or Notion server, the default `fetch` server, whose URL can carry the turn's content, a server loaded at runtime). Its `search_tools` results show only the allowed tools (`workflows.tool_catalog.filter_search_results`; output it cannot parse reads as no match), and a `call_tool` naming anything else is refused with an error tool_result and an audit row private to the principal (`schedule_tools.private_turn_withholds`), a missing name included. `search_tools` and `call_tool` stay offered, so the tool list and its cached prefix do not change. The typed calendar tool (`_resolve_attendees`) and the chat DM tools (`message_person`, `send_telegram_message`, `send_discord_dm` and their person-id recovery) apply the same rule — a contact's id reads as an unknown one off the principal's turn — and a private turn may DM or invite only the principal (`PRIVATE_TURN_REFUSAL` otherwise); `message_person` never falls back to a \"Needs you\" alert for a contact, who cannot sign in to see it. The Google Workspace server (`workspace-mcp`) behind those Gmail/Calendar/Drive tools runs **co-located inside the API** as a stdio child of the gateway — fitting the single-tenant model (one install = one company = one exec Google account), so there's no separate service to operate. It's baked into the API image and launched by `docker/workspace-mcp-launch.sh`, which selects single-user OAuth (token persisted at `/data/google_credentials`) or service-account domain-wide delegation from `GWORKSPACE_AUTH_MODE`. The optional `WORKSPACE_MCP_TOOLS` (comma-separated service names, default `all`) limits which Google services it loads. The gateway forwards the Google secrets to the child via `_FORWARDED_ENV_VARS`. Every `google_workspace__*` call acts as the Executive's own account. workspace-mcp takes the acting account per call (`user_google_email`) and nothing used to check it, so under domain-wide delegation a model-chosen address could open anyone's mailbox in the domain; `mcp_gateway._check_acting_account` now refuses a value naming any other address — before every other gate, with an `integration_outbound_blocked` audit row. An absent value stays allowed (the server's own account).\n\n**What the model is told is connected.** `MCPGateway.start()` records the configured server names (`server_names`), and the Executive renders them with the channel settings into block 0's *Connected Systems* section (`prompts.connected_systems`; see the caching section). While Google is connected that section pins the Google tool names the Executive uses constantly (`GOOGLE_TOOL_MANIFEST`, narrowed by `connected_systems.google_tool_manifest` to the services and tier actually loaded) and tells it to call them with `call_tool` directly, with no `search_tools` round trip. The narrowing is the point: a name pinned for a service workspace-mcp never imported is not in the index, so the model calls it, receives extensible-mcp's “has not been discovered via search_tools” refusal, and has no way to square that with its own instructions. extensible-mcp only runs a tool one of the session's searches has returned, and remembers it for the session's life, so the gateway does that discovery itself: the lifespan starts `MCPGateway.prime_pinned_tools()` in the background once the gateway is up with `google_workspace` configured (an exact-name search per pinned tool, the way `workflows.tool_catalog.resolve` does it, over the same narrowed set the prompt advertises — `connected_systems.pinned_google_tools`, so priming never probes a service that was not loaded). A name neither the prompt nor priming can account for is logged once, naming both causes: excluded by `WORKSPACE_MCP_TOOLS` / `WORKSPACE_MCP_TOOL_TIER`, or renamed by workspace-mcp since the pin. A pinned call that still comes back with extensible-mcp's \"has not been discovered via search_tools\" refusal, for example before priming finished, is discovered and retried once. That happens after every gate has passed the same arguments, and only for pinned names: any other tool still needs the model's own search, and nothing here widens what a call may do, including the private-turn allow-list.\n\n**Microsoft 365** is the Outlook twin: `@softeria/ms-365-mcp-server` (a Node package, pinned) is installed in a throwaway Node build stage and only the `node` binary plus the package are copied into the API image, launched by `docker/ms365-mcp-launch.sh` as the `microsoft_365` entry with an **explicit tool allow-list** (`--enabled-tools`, an anchored regex) rather than the upstream `mail,calendar` preset — that preset also registers inbox forwarding rules, mailbox settings (external auto-reply), calendar sharing and event forwarding, which can address people the gateway never sees. Exactly the Outlook mail + calendar tools and the OneDrive tools are exposed (`microsoft_365__send-mail`, `microsoft_365__create-calendar-event`, `microsoft_365__search-onedrive-files`, …; hyphenated Graph endpoint aliases, proxied verbatim), and every admitted tool is either roster-gated or carries no recipient; the gate also covers the excluded writers in case an operator widens the list. Both servers can be configured at once. Auth is the MSAL device-code flow against the operator's own Entra app registration (`MS365_MCP_CLIENT_ID` is required; the server's built-in public client is refused); the refresh token persists at `/data/ms365_credentials` and is seeded once with `ms365-mcp-launch.sh --login` from inside the container — the server has a native device-code CLI, so there is no mint script and nothing to copy. The launcher scrubs the literal `$VAR` placeholders extensible-mcp leaves for unset variables (otherwise MSAL would be handed `$MS365_MCP_TENANT_ID` as an authority), and `MS365_MCP_EXPECTED_USERNAME` pins the bound mailbox. The egress gate for this family walks Graph's nested shapes (`body.Message.toRecipients[].emailAddress.address`, `body.attendees[]…`) — every email-shaped string outside the free-text fields must be on the roster, control characters and non-ASCII addresses are refused, unforeseen keys are still scanned (fail-closed) — and covers the whole write surface: send, reply, reply-all, forward, every draft tool, `update-mail-message` (it can PATCH recipients onto a draft) and create-/update-calendar-event; reads and `delete-calendar-event` (the only cancel path) pass through. Graph's reply, reply-all and send-draft actions carry no recipient in their arguments (the recipient is whoever the referenced `messageId` names), so for those the gate first reads the referenced message through the server and roster-checks the recipients Graph will derive (`replyTo` else `from` for a reply; `replyTo` and `from` together, plus the original to/cc, for reply-all; the draft's own to/cc/bcc for send-draft), refusing on any lookup failure; RSVP and cancel actions get the same treatment through an event lookup, because their comment is emailed to the organizer or every attendee; `download-bytes` is pinned to mail attachment bytes and OneDrive file content (`/drives/{id}/items/{id}/content`, optionally Graph's `?format=pdf` rendering), and a OneDrive result comes back as the file's extracted text rather than base64 — the attachment pin is what makes the poller's promise that an unrostered sender cannot be auto-replied to hold on Outlook as it does on Gmail; and the rendered inbound text quotes any forged `--- REPLY ---` line inside a body so the persona only ever reads the real block. Every mail write's `from`, `sender` and `replyTo` may name only the Executive's own address (the Outlook twin of the Gmail acting-account pin), so the model cannot send as a rostered colleague's mailbox or steer replies elsewhere; and `move-mail-message` into Deleted Items, Junk Email or Recoverable Items is refused, so a move cannot trash mail ungated. The server registers its account tools whatever its tool allow-list says, so the gateway itself refuses `login`, `logout`, `select-account` and `remove-account` (audited like any other refusal) and drops them from `search_tools` results — a chat turn cannot sign the mailbox out, switch accounts or start a sign-in; setup is only the launcher's `--login`, and the read-only `verify-login` / `list-accounts` stay callable. The example config's hyphenated patterns still deny `delete-mail-*`, `logout`, `remove-account` and `download-bytes-to-file` as defense in depth. A successful `send-mail` records the same outbound-context linkage Gmail does.\n\n**OneDrive in chat** (opt-in with `MS365_MCP_ONEDRIVE=true` and a fresh `--login`, since the server asks for its tools' permissions on every token refresh and a sign-in without file consent would otherwise fail mail and calendar too) matches what the Executive can do on Google Drive: search, list a folder, read a file, upload and edit, create folders, move, rename, copy and share; there is no delete tool. The launcher adds `Files.Read.All` at sign-in (`--extra-scopes`), so folders shared *with* the Executive can be read too; writes (upload, new folder, move, copy) land only in the Executive's own drives, checked against `list-drives` at call time, so a planted instruction cannot save company files into a stranger's shared folder. Sharing is gated like Drive sharing: `share-drive-item` invitees must be People on the roster and named by email (an `objectId` or `alias` grantee is refused, since nothing can check it), no `owner` role, and `create-drive-item-share-link` must be scoped to specific people (`scope: \"users\"`): an anonymous or whole-organization link, or one with no scope (the tenant default), is refused. An invite must require sign-in and grant only view or edit, a link must be a view or edit link, and `search-onedrive-files` takes plain words only, since its query goes into the request path unencoded.\n\nAs a *server*, `openexecutive.mcp_server` exposes OE's own company context and specialist council to external MCP clients at `/mcp` — the inverse direction; see the dedicated **MCP Server** section. The gateway launches extensible-mcp as a `uvx` subprocess whose fastembed tool-search index uses an embedding model baked into the API image and loaded offline, so it no longer re-downloads that model from the Hugging Face Hub on every cold start. extensible-mcp is pinned twice over: to a commit, which fixes its own code, and with `--exclude-newer`, which fixes the resolution of its dependencies — without the cutoff uvx re-resolves them against PyPI on every networked start, so a new release of any of them would reach running deployments on restart. Both live in `_EXTENSIBLE_MCP_LAUNCH_ARGS` in `orchestrator/mcp_gateway.py`, the Dockerfile's pre-warm uses the same arguments (a unit test enforces the match), so an image whose pre-warm succeeded starts the gateway with no network. The pre-warm is best-effort (`|| true`): an image built while GitHub was unreachable fetches the pinned commit on its first start instead. Updating the gateway is a deliberate bump of the commit and cutoff; `uv` itself is not pinned. Because the MCP stdio client does not inherit the parent environment, `MCPGateway.start()` explicitly forwards the embedding-cache/offline env vars (`FASTEMBED_CACHE_PATH`, `HF_HUB_OFFLINE`, …) into the child.\n\n**Emailing an artifact.** A Gmail send or draft may carry `attachments=[{\"artifact_id\": \"alert:12\", \"as\"?: \"docx\"}]`. `call_tool` runs the recipient gate first; only if it passes does `_expand_artifact_attachments` replace each such entry with the rendered file (`content` base64 + `filename` + `mime_type`, workspace-mcp's own attachment shape), produced by `artifact_records.render_artifact_file` — the same renderer as `/artifacts/{id}/download`. Only artifact rows resolve (never an arbitrary alert), links have no file, and an entry mixing `artifact_id` with other keys is refused. A stringified or single-object `attachments` is normalised first, so an artifact entry can never reach workspace-mcp unexpanded (it would silently drop it). At most 5 distinct artifacts per email (repeats are dropped) and 15 MB rendered in total; rendering runs off the event loop via `asyncio.to_thread`. A refusal writes `artifact_attachment_refused`; a successful send writes `artifact_attached` — both carry the recipients. Other attachment entries pass through unchanged. The same expansion serves the Microsoft 365 mail writes: the entry becomes a Graph `#microsoft.graph.fileAttachment` (`name`, `contentType`, `contentBytes`) placed where Graph expects it for that tool (`body.Message.attachments` for `send-mail`, reply, forward and the createReply drafts; `body.attachments` for `create-draft-email`), whether the model put it at the top level or inside the message; an entry anywhere else, one on `update-mail-message` or `send-draft-message` (their requests cannot carry attachments), or one behind a body that is not an object is refused rather than forwarded unexpanded.\n\n**Enablement, and why it degrades instead of dying.** `MCP_ENABLED` left unset is inferred from the presence of `mcp_servers_config_path` — dropping an `mcp_servers.json` next to `profile.yaml` turns MCP on with no further configuration. An **explicit** setting always wins in both directions: `MCP_ENABLED=false` keeps MCP off with the file still in place. It did not: `_resolve_mcp` tested `not self.mcp_enabled`, which cannot tell \"never set\" from \"set to false\", so the file's presence silently overrode the operator. The file is probed through `config.mcp_config_file_present`, never `Path.exists` directly — that re-raises an EACCES on a parent directory, and from inside a validator that comes out of `Settings()` itself, which `api/main.py` calls at module level. The check now reads `model_fields_set` (the fields the env actually supplied) and stores that on a private attr behind the read-only `mcp_auto_enabled` property, because the validator's own assignment joins that set and would otherwise make every auto-enable look explicit downstream.\n\nStartup is **best-effort**, in `main._start_mcp_gateway`. Two checks come before the subprocess: MCP off is a no-op, and a config defining no servers under `mcpServers` is logged and skipped — extensible-mcp's own loader raises on that and exits, and because the child is already gone the only thing that surfaces is anyio's \"Attempted to exit cancel scope in a different task\" from `stdio_client` unwinding, a traceback naming nothing about configuration. The start is also bounded (120s, generous because that child may still clone extensible-mcp on a cold start if the image's pre-warm failed): `ClientSession` carries no `read_timeout_seconds`, so `initialize()` waits forever on a child that is *alive but silent*, and an unbounded wait is a lifespan that never yields — the same healthcheck-never-passes symptom, which no `except` can see. A gateway that fails anyway is logged at ERROR with the config path, the servers it named, and whether the config file or an explicit `MCP_ENABLED` turned MCP on, then torn down under its own timeout; the API boots **without MCP tools and without the email poller** rather than refusing to boot. The guard catches `BaseException`, not `Exception`, re-raising a bare `CancelledError` first: the original failure came from anyio, which wraps a task group's failures in `BaseExceptionGroup` as soon as one is a `BaseException` — and that group is not an `Exception`, so `except Exception` would miss the very shape this contains. The teardown is contained too, since it runs from inside that `except` and would otherwise replace the actionable log with the crash it prevents. `configured_server_names` never raises either, and deliberately catches broadly: it reads an operator-owned path during startup, where an exception is the crash loop rather than a bug report — `Path.is_file` re-raises an EACCES on a parent directory, and `json.loads` answers deeply-nested input with `RecursionError`, which is not a `ValueError`. It tests `is_file()` so a directory or a symlink to a FIFO is never opened, and caps the read at 1 MiB. That exception used to propagate out of the lifespan and kill the container — under `restart: unless-stopped`, a crash loop whose only outward symptom was a healthcheck that never passed. Outbound tools: per-Person DMs (`send_slack_dm`, `send_discord_dm`, `send_telegram_message`) plus room-targeted `send_department_message` and `send_company_broadcast`. All three per-Person DMs refuse a recipient id that matches no non-archived Person row (`schedule_tools._dm_recipient_on_roster`), which is what stops untrusted content in a turn — inbound mail, a chat message, a knowledge document — from directing a DM at an arbitrary workspace member. `send_slack_dm` previously consulted the roster only when the turn was already private to the principal, so on an ordinary turn it applied no roster check at all while its Discord and Telegram siblings did; the room-targeted tools are deliberately ungated, since those channels are team-visible. The per-Person DMs are additionally rate-limited and de-duplicated by the outbound anti-spam guard (below).\n\n### Outbound anti-spam guard\n\nEvery per-Person DM passes a deterministic guard (`orchestrator.outbound_guard.check_outbound_allowed`) after the roster gate and recipient-id recovery but before the network send — the single chokepoint **all** proactive DM paths reach (ad-hoc chat sends, the scheduler's followup/nudge dispatch, principal-brief delivery, the reflection turn, and research-finding routing). On a violation it returns a human-readable reason and the handler **suppresses** the send (`{\"status\":\"suppressed\",\"reason\":…}`), writes a `tool_invocation` audit row with `suppressed=true`, and writes **no** `done` activity row — so a blocked attempt never counts toward its own rate cap, and the Executive can read the reason and hold or `schedule_followup` for the recipient's next window. Three controls: **content dedup** (a near-identical message — whitespace/case-normalized, first 160 chars — already sent to the same recipient within `OUTBOUND_DEDUP_WINDOW_MINUTES`, default 360); a **per-recipient rate cap** (`OUTBOUND_MAX_PER_RECIPIENT_PER_WINDOW`, default 5, over `OUTBOUND_RATE_WINDOW_MINUTES`, default 60); and **quiet hours / availability** (`OUTBOUND_RESPECT_QUIET_HOURS`, default on — refuse while the recipient is on leave or outside their `people.channel` availability windows, surfacing the next open window). Delivery history is read from the existing `status='done'` `scheduled_actions` rows (`episodic.recent_sends_for_channel_ref`), so there is no new table. The guard is **fail-open**: any internal lookup error returns \"allow\", so it can never turn a working send into a crash or a silently-dropped message. Principal briefs and the internal heartbeat kinds use the `__internal__` channel and never reach these handlers, so they are unaffected.\n\n### Calendar & video links\n\nThe Executive books meetings through two typed tools over the configured calendar backend (`CALENDAR_PROVIDER`: Google Calendar via `manage_event`, or Outlook via `create-calendar-event` — see *Workspace providers*). `create_calendar_event` schedules a **future** meeting (propose→approve by default, or auto-execute once that decision class is promoted) and may be used proactively when a stalled goal warrants a sync. `create_instant_meeting` spins up an **impromptu call starting now**, auto-executing without the approval step and skipping the business-hours/horizon caps, so the Executive can hand back a join link immediately. Both attach a video-meeting link by default (`add_video_link`; opt out for in-person) — a **Google Meet** link parsed from the created event's `conferenceData`, or a **Microsoft Teams** link from `onlineMeeting.joinUrl` — surfaced to the caller plus persisted on the decision ledger as `meet_link`. The tool descriptions are static, backend-neutral text, so the cached tool block never depends on settings; the pre-rename `add_google_meet` key is still honoured for old ledger rows. Either way the caps that matter still hold — roster membership, principal-protection, max attendees, and a daily booking ceiling.\n\n**Around-the-meeting participation.** The Executive doesn't join the call as a live voice; instead, on every successful booking it schedules a best-effort post-meeting follow-up (~5 min after the end) that DMs a human attendee asking for a quick recap — key decisions and action items. The reply flows back through the same cross-session DM context linkage described below, so the Executive can act on it (summarize, follow up) in a later turn. No meeting-transcript or Drive integration is required.\n\n### Workspace providers\n\nThe Executive can *discover* any mail or calendar tool the gateway exposes, but three code paths call a fixed backend with no model in the loop and used to spell Gmail tool names inline: the inbound mailbox poller (list unread → fetch → mark read), alert email dispatch plus the scheduler's email send-tool hint, and the typed booking tools plus the approval-time conflict check. `integrations.workspace` puts those behind two Protocols (`MailProvider`, `CalendarProvider`) with one module per backend (`google.py`, `microsoft.py`) and a registry keyed on `EMAIL_PROVIDER` / `CALENDAR_PROVIDER` — separate switches (Outlook mail with a Google calendar is legitimate), both defaulting to `google`, so an existing install changes nothing until an operator opts in. The Google backend is the old inline code moved verbatim; tests pin its argument dicts byte-for-byte. The Microsoft backend polls only the Inbox folder (`list-mail-folder-messages` with `isRead eq false`, never the all-folders listing that includes Junk), converts Graph's HTML bodies to text with a real parser, lists a message's attachments so the names appear in the same `N. name (mime, size KB)` shape Gmail's text uses (and reach peer memory), marks read with `update-mail-message`, sends through `send-mail`'s nested shape, books with `isOnlineMeeting` + `onlineMeetingProvider=teamsForBusiness`, cancels with `delete-calendar-event`, and answers the advisory conflict check from the Executive's **own** calendar view (delegated permissions cannot read attendees' free/busy — narrower than Google's `query_freebusy`).\n\nEvery backend renders an inbound message into the same `Subject:` / `From:` / `To:` / `Cc:` + `--- BODY ---` text the persona already expects, and the poller adds a `--- REPLY ---` block naming the exact reply tool and threading ids for *that* message (Gmail: `send_gmail_message` + `thread_id`; Outlook: `reply-mail-message` + `messageId`) after the mail's `<untrusted_content>` block, never inside it. The block rides in the user turn — which is what lets the cached persona say \"use the reply tool in the REPLY block\" instead of naming a vendor tool — and the scheduler's proactive-email trigger names the backend's send tool the same way. Providers never raise on a malformed backend response (they return empty / `None` / `{\"error\": …}`), so one bad cycle never kills a loop, and a chosen backend whose MCP server is missing from `mcp_servers.json` is logged at startup and skipped by the poller: fail-soft, never a boot failure. The poller's newer checks run on that rendered text for both backends (contacts and private turns, roster requests and their emailed answers, the audit row), with three Outlook differences that fail closed: no sender is ever authenticated (the `dmarc=pass` read is Gmail's), so there is no acknowledgement to an unknown sender, no emailed fact confirmation and no principal-authenticated mail; attachments stay a list (the download is Gmail's); and the brief's calendar block is skipped (its listing is Google Calendar's). The code's own emails to the principal (roster prompts, fact confirmations, scheduled deliveries) go through `workspace.registry.send_from_executive`, and `runner.email_ready` follows `EMAIL_PROVIDER`. A private turn's allow-list (`schedule_tools.PRIVATE_TURN_MCP_TOOLS`) and Act as me's mail-touched reads include the Outlook twins of the Gmail and Calendar tools, and roster answer tokens are hidden from Microsoft 365 results as from Google's.\n\n### Cross-session DM context\n\nWhen the principal says \"DM Alex about X\" in one session, OE sends the DM and — if a live conversation is active and the recipient is not one of the principal's contacts (a contact cannot DM the bot, and quoting the principal's conversation into their email reply's unattended turn would put it in the audit log) — records an `outbound_context` linkage row (recipient, full outbound text, originating session id). Email is covered too but recorded one layer deeper — there is no `send_email` tool, so the linkage is written in `orchestrator.mcp_gateway` after a successful `google_workspace__send_gmail_message` or `microsoft_365__send-mail`, one row per `to`/`cc` recipient keyed by the bare lowercased address (OE's own address skipped). When Alex later replies on the 1:1 DM path — or by email — the shared helper `integrations.inbound_hydration.hydrate_user_message` (called by the Discord, Telegram, and Slack handlers, and by the email poller keyed on the sender address) looks up the most-recent `open` linkage for that sender within 72h and prepends an `<outbound_reply_context>` block — the outbound text plus a backstory excerpt from the originating session — to the message handed to `Executive.chat()`. So OE knows what the reply is about instead of starting from an empty per-person session. The block is context for the LLM turn only: peer memory strips it before recording what the person said (`strip_outbound_reply_context`, applied in `sync_turn`, `sync_department_turn` and to the prefetch query), so the Executive's own DM is never attributed to the person who replied to it. The block lands in the user turn (caching rule) and is injected into a copy only — the pristine reply is persisted to history. The linkage is one-shot (consumed on first matched reply, race-safe) so an unrelated later DM isn't re-hydrated. This is deterministic and independent of the per-person `peer_memory` (Honcho) layer.\n\nEvery adapter that persists a turn (Slack, Discord, Telegram, and the web route) records the message's **resolved sender** on `chat_messages.sender_person_id` — the rostered Person the adapter matched, or NULL for an unrostered sender — independently of the session owner, which falls back to the principal on Discord and Telegram so those threads stay visible. Attunement's open loops and feedback rely on this so an outsider's words are never read as someone on the roster's.\n\n### OneDrive folder sync\n\nThe Microsoft twin of the Drive folder sync below (`knowledge.onedrive_sync`, `ONEDRIVE_SYNC_ENABLED`; setup in `docs/onedrive_sync_setup.md`). It copies the files in `ONEDRIVE_SYNC_FOLDERS` (`<drive id>/<item id>` pairs; `openexecutive onedrive-folder <link>` resolves a sharing link) into the isolated `onedrive_docs` collection on its own heartbeat, and shows them on the knowledge page as \"OneDrive\" rows with a **Sync now** button. Microsoft has no service account that sees only what is shared with it (app-only Graph access covers the whole tenant), so it reads as the Executive's own Microsoft 365 sign-in, and the folder list is the scope. It never stores a credential: each tick asks the launcher for a token (`ms365-mcp-launch.sh --access-token`, which runs `docker/ms365-access-token.mjs` through the server's own token cache and will only mint file-read tokens), keeps it in memory, and sends it only to Graph. `knowledge.onedrive_client` issues GETs only, follows a paging link only back to Graph, and fetches a file's pre-authenticated download link without the token; errors never carry that link. PowerPoint and older Office formats are read through Graph's PDF rendering.\n\n### Google Drive folder sync\n\nThe synced files show on the knowledge page's Company documents list as \"Google Drive\" rows (sync time, last edit in Drive, an Open in Drive link and a text view), and the page's source card shows the last run, any sign-in error or count of files that failed to sync, and a **Sync now** button that runs one tick on demand.\n\n\nSeparate from the Drive tools the gateway proxies. The optional folder sync (`knowledge.drive_sync`, `DRIVE_SYNC_ENABLED`; setup in `docs/drive_sync_setup.md`) copies the files in `DRIVE_SYNC_FOLDER_IDS` into the isolated `drive_docs` knowledge collection on a scheduler heartbeat, so the Executive knows those folders without searching Drive live (see RAG). It calls the Drive v3 API directly (`knowledge.drive_client`, `httpx`) as a **service account** with the `drive.readonly` scope (`DRIVE_SYNC_SERVICE_ACCOUNT_FILE`), not as the Executive's own account through workspace-mcp. Google enforces the folder scope, because the account sees only what is shared with it, and the sync can change nothing. Structured JSON means no display text to parse, and every file or folder id is checked against `[A-Za-z0-9_-]{1,128}` before it goes into a URL or a query. The two paths meet in two places. A synced chunk carries its Drive file id and sync time, so the Executive can open the live file with its own `get_drive_file_content`. The setup doc says to share each synced folder with the Executive's account too, so that live open works.\n\n### Confluence space sync\n\nThe optional Confluence sync (`knowledge.confluence_sync`, `CONFLUENCE_SYNC_ENABLED`; setup in `docs/confluence_sync_setup.md`) copies the pages of the spaces in `CONFLUENCE_SYNC_SPACE_KEYS` into the isolated `confluence_wiki` knowledge collection on a scheduler heartbeat, so the Executive knows the company wiki without a tool call (see RAG). It works with Confluence Cloud and Server/Data Center: `knowledge.confluence_client` calls the v1 REST API both serve, directly with `httpx`, signing in with a personal access token (Server/DC) or a username and API token (Cloud). It reads only, never follows a redirect (so the token stays on the configured site), streams responses with a 20 MB cap, and checks page ids and space keys before they reach a URL or a CQL query. The URL must be `https://` unless `CONFLUENCE_SYNC_ALLOW_HTTP` is set, and `CONFLUENCE_SSL_VERIFY` takes `true`, `false` or a CA bundle path for a private certificate authority. With `CONFLUENCE_SYNC_PUBLIC_HOSTS_ONLY` (off by default, for an install where whoever sets the URL must not reach the backend's own network) every request resolves the site's host, refuses it when any address is not public (loopback, private, link-local, shared), and connects to the checked address with the site's name kept for the Host header and the TLS certificate check, directly rather than through a proxy. `confluence_client.config_problem` is the same settings check the app runs at start, for a caller that writes these settings on someone's behalf.\n\nIt is separate from a Confluence MCP server such as mcp-atlassian, which a deployment can add to `mcp_servers.json` for live reads and edits in chat. The two share the env names `CONFLUENCE_URL`, `CONFLUENCE_PERSONAL_TOKEN`, `CONFLUENCE_USERNAME`, `CONFLUENCE_API_TOKEN` and `CONFLUENCE_SSL_VERIFY`, so one set of values serves both, and a synced chunk carries its page id so the Executive can open the live page with such a tool. A Confluence token is a user's, and what the sync reads becomes retrievable by everyone who talks to the Executive, so the space list is required and pages with a read restriction (their own or a parent's) are skipped unless `CONFLUENCE_SYNC_SKIP_RESTRICTED=false`. That keeps any existing member's token safe to use, without a dedicated licence. Manual runs: `openexecutive sync-confluence` / `purge-confluence`.\n\n### Your own Gmail or Outlook (Act as me)\n\nA second, separate mail connection, for **Act as me** (see **Act as Me (Delegation)**): a person's *own* mailbox, which `delegation.gmail.DelegateGmail` reads and drafts into over the Gmail REST API. It is never registered with the MCP gateway, so the model can reach it only through the typed `ghostwrite_email` handler — never by tool name through `call_tool`. The credential is one file per person in `DELEGATION_GOOGLE_CREDENTIALS_DIR` (default `company/delegation_google/`, `/data/delegation_google` in Docker), minted on a computer with a browser by `scripts/connect-own-gmail.py --email <you>` with the Executive's own OAuth client and the scopes `gmail.readonly` and `gmail.compose` — the same seed-it-yourself flow as the Executive's token, since the API serves no OAuth callback. It never lives in `WORKSPACE_MCP_CREDENTIALS_DIR`, where workspace-mcp picks a credential by address.\n\nEvery use re-checks with Google that the credential opens the mailbox on the person's People entry, and refuses outright when that address is the Executive's own: a shared mailbox would make everything the Executive writes look like theirs. Drafts are threaded (`In-Reply-To`, `References` of whole ids ending with the parent, `threadId`) and stamped `X-OE-Ghostwritten`; the only send is `drafts.send` of an existing draft by id, when the person taps Send on its reply card (`delegation.reply_send`). A Gmail rate limit, even as a 403, is an error to retry rather than a reconnect. The voice learner and the thread shown to the ghostwriter keep only each sender's own words through `email_poller.sender_new_text`, the poller's quoted-reply stripper made public. Setup status shows it as **Your own mailbox (Act as me)**.\n\nThe same file can instead hold an Outlook sign-in (`\"provider\": \"microsoft\"`), minted by `scripts/connect-own-outlook.py --email <you>` with a device-code sign-in on the Executive's own Entra app (`MS365_MCP_CLIENT_ID`, `MS365_MCP_TENANT_ID`, default `common`) and the delegated scopes `User.Read`, `Mail.ReadWrite`, `Mail.Send` and `offline_access`. `delegation.gmail.gmail_for` then hands back `delegation.outlook.DelegateOutlook`, which talks to Microsoft Graph directly — never the Microsoft 365 MCP server the Executive's own Outlook uses — with the same methods and the same one send: Graph's `send` on an existing draft, with no body. Microsoft rotates the refresh token on use and the new one is written back into the file. A sender counts as verified only from the topmost `Authentication-Results`, the one Exchange wrote, and only mail in their Sent Items counts as theirs.",
"mermaid": "flowchart TD\n Channels[\"Slack / Discord<br/>Telegram / GChat / Email\"]\n Resolve[\"find_person_by_*_id\"]\n Gate[\"roster gate\"]\n Drop[\"drop\"]\n Hold[\"roster request<br/>(hold + fixed ack,<br/>ask the principal)\"]\n Exec[\"Executive.chat()\"]\n Attach[\"attachments<br/>(vision / inbound_attachments)\"]\n MCP[\"MCP gateway\"]\n GW[\"Google Workspace<br/>(workspace-mcp child)\"]\n M365[\"Microsoft 365<br/>(ms-365-mcp-server child:<br/>Outlook + OneDrive)\"]\n WP[\"workspace providers<br/>(poller / alerts / booking)\"]\n Guard[\"outbound anti-spam guard<br/>(dedup / rate cap / quiet hours)\"]\n Out[\"DM / department<br/>/ broadcast tools\"]\n Channels --> Resolve\n Resolve --> Gate\n Gate -- \"unknown sender\" --> Hold\n Hold -- \"approved: replay\" --> Exec\n Hold -- \"contact / declined\" --> Drop\n Gate -- \"Person\" --> Exec\n Gate --> Attach\n Attach --> Exec\n Exec --> MCP\n MCP -- \"roster egress gate\" --> GW\n MCP -- \"roster egress gate\" --> M365\n WP -- \"EMAIL_PROVIDER / CALENDAR_PROVIDER\" --> MCP\n WP --> Exec\n Exec --> Guard\n Guard -- \"suppress\" --> Drop\n Guard -- \"allow\" --> Out\n Out --> Channels\n Own[(\"Principal's own Gmail<br/>(Act as me: REST, drafts only)\")]\n Exec -- \"ghostwrite_email\" --> Own\n class Channels entry\n class Resolve compute\n class Gate compute\n class Drop external\n class Hold compute\n class Exec hot\n class Attach compute\n class MCP external\n class GW external\n class M365 external\n class WP compute\n class Guard compute\n class Out compute\n class Own external",
"generated_at": "2026-10-09T18:00:00Z"
}
diff --git a/packages/core/openexecutive/config.py b/packages/core/openexecutive/config.py
index c953b9c3..28a02d10 100644
--- a/packages/core/openexecutive/config.py
+++ b/packages/core/openexecutive/config.py
@@ -478,6 +478,63 @@ class Settings(BaseSettings):
exec_display_name: str = Field("Open Executive", alias="EXEC_DISPLAY_NAME")
email_poll_interval_seconds: int = Field(60, alias="EMAIL_POLL_INTERVAL_SECONDS")
+ # Which Google services the co-located workspace-mcp child loads, and at
+ # which tier. These are the same two variables `docker/workspace-mcp-launch.sh`
+ # turns into `--tools` / `--tool-tier`; they are read here so the prompt can
+ # describe what actually loaded instead of the whole manifest. Both are
+ # process-constant, which is what keeps cached block 0 warm.
+ workspace_mcp_tools: str = Field("all", alias="WORKSPACE_MCP_TOOLS")
+ workspace_mcp_tool_tier: str = Field("complete", alias="WORKSPACE_MCP_TOOL_TIER")
+
+ @property
+ def workspace_mcp_service_list(self) -> list[str] | None:
+ """The Google services loaded, or ``None`` meaning "all of them".
+
+ Mirrors `workspace-mcp-launch.sh`'s normalisation exactly, including the
+ cases that are easy to miss — and every one of them resolves to "all",
+ so a Python parser that handled only `split(",")` would answer with a
+ NARROWER set than the child actually loaded. That direction matters:
+ this value gates what the prompt advertises, and under-reporting hides a
+ tool the model can really call, which is the mirror image of the bug
+ this property exists to fix.
+
+ The four rules, in the launcher's order:
+
+ 1. unset or empty -> ``all``
+ 2. a value still reading ``$SOMETHING`` -> ``all``. extensible-mcp
+ leaves an unresolved `$VAR` in a config `env` block as that literal
+ text, so the child can receive the string "$WORKSPACE_MCP_TOOLS".
+ 3. commas become spaces, then a value with no non-whitespace left
+ (``","``) -> ``all``
+ 4. ``all`` only means every service when it is the WHOLE value, after
+ the launcher collapses whitespace. ``"gmail,all"`` becomes
+ ``--tools gmail all``, and since ``--tools`` is declared
+ ``nargs="*" choices=VALID_SERVICES`` and ``all`` is not a service
+ name, **argparse exits 2 and the child never starts** — so nothing
+ loads at all. This property reports ``["gmail", "all"]`` there,
+ matching what the launcher passes rather than what survives it;
+ neither answer describes a dead child, which is a gap the prompt
+ cannot express (``server_names`` comes from the config file, not
+ from child health). Treating ``all`` as a sentinel wherever it
+ appeared made this answer "every service" instead, the widening
+ direction this property exists to prevent. A parity test against
+ the real script catches both.
+ """
+ raw = (self.workspace_mcp_tools or "").strip()
+ if not raw or raw.startswith("$"):
+ return None
+ names = [p for p in raw.replace(",", " ").split() if p]
+ if not names or names == ["all"]:
+ return None
+ return list(dict.fromkeys(names))
+
+ @property
+ def workspace_mcp_tier(self) -> str:
+ """The tier the child runs at. An unresolved ``$VAR`` reads as the
+ launcher's default, for the same reason as above."""
+ raw = (self.workspace_mcp_tool_tier or "").strip()
+ return "complete" if not raw or raw.startswith("$") else raw
+
# Roster requests (people.roster_requests): someone off the roster who
# writes in is held for the principal to confirm, and told so — at most
# once per sender per ROSTER_ACK_WINDOW_DAYS, and at most
diff --git a/packages/core/openexecutive/orchestrator/mcp_gateway.py b/packages/core/openexecutive/orchestrator/mcp_gateway.py
index 66c079dc..fcf8fb43 100644
--- a/packages/core/openexecutive/orchestrator/mcp_gateway.py
+++ b/packages/core/openexecutive/orchestrator/mcp_gateway.py
@@ -2779,24 +2779,37 @@ class MCPGateway:
return False
if any(info.name == tool_name for info in parse_search_results(text)):
return True
+ # Both causes, because the old text named only the rename and sent the
+ # operator to the manifest for what was really a service-list setting.
+ settings = get_settings()
logger.warning(
- "MCPGateway: pinned tool %s not found by search — has workspace-mcp "
- "renamed it? (prompts/connected_systems.GOOGLE_TOOL_MANIFEST)",
+ "MCPGateway: pinned tool %s not found by search — not loaded under "
+ "WORKSPACE_MCP_TOOLS=%s / WORKSPACE_MCP_TOOL_TIER=%s, or renamed by "
+ "workspace-mcp? (prompts/connected_systems.GOOGLE_TOOL_MANIFEST)",
tool_name,
+ settings.workspace_mcp_tools,
+ settings.workspace_mcp_tool_tier,
)
return False
async def prime_pinned_tools(self) -> list[str]:
- """Discover every pinned Google tool (PINNED_GOOGLE_TOOLS) up front, so
- the model's direct ``call_tool`` on one works without a search of its
- own. Returns the names not found, a drift signal. Run once at startup
- when Google is configured."""
- from openexecutive.prompts.connected_systems import PINNED_GOOGLE_TOOLS
-
- missing = [n for n in sorted(PINNED_GOOGLE_TOOLS) if not await self._discover(n)]
+ """Discover the pinned Google tools this install actually loaded, so the
+ model's direct ``call_tool`` on one works without a search of its own.
+ Returns the names not found, a drift signal. Run once at startup when
+ Google is configured.
+
+ Probes the same filtered set the prompt advertises
+ (``connected_systems.pinned_google_tools``), not the whole manifest: with
+ a ``WORKSPACE_MCP_TOOLS`` subset or the ``core`` tier, the rest were
+ never registered, so probing them produced a warning per tool every boot
+ and told the operator to look for a rename that had not happened."""
+ from openexecutive.prompts.connected_systems import pinned_google_tools
+
+ pinned = pinned_google_tools(get_settings())
+ missing = [n for n in sorted(pinned) if not await self._discover(n)]
logger.info(
"MCPGateway: primed %d pinned Google tools (%d missing)",
- len(PINNED_GOOGLE_TOOLS) - len(missing), len(missing),
+ len(pinned) - len(missing), len(missing),
)
return missing
diff --git a/packages/core/openexecutive/prompts/connected_systems.py b/packages/core/openexecutive/prompts/connected_systems.py
index 0080d507..2d2ce9ec 100644
--- a/packages/core/openexecutive/prompts/connected_systems.py
+++ b/packages/core/openexecutive/prompts/connected_systems.py
@@ -73,6 +73,71 @@ PINNED_GOOGLE_TOOLS: frozenset[str] = frozenset(
name for names in GOOGLE_TOOL_MANIFEST.values() for name in names
)
+# Which workspace-mcp service module each manifest heading comes from, so a
+# `WORKSPACE_MCP_TOOLS` subset can be matched against it. These are the names
+# `--tools` accepts (workspace-mcp's `SERVICE_MODULES`), not display names.
+GOOGLE_SERVICE_OF: dict[str, str] = {
+ "Gmail": "gmail",
+ "Calendar": "calendar",
+ "Drive": "drive",
+ "Docs": "docs",
+ "Sheets": "sheets",
+}
+
+# Pinned tools that workspace-mcp's `core` tier does NOT include, read from
+# `core/tool_tiers.yaml` in the pinned 1.29.0. Tiers are cumulative, so
+# `extended` and `complete` carry all 24 and only `core` needs subtracting.
+# Re-check when bumping the pin, exactly as for the manifest above.
+GOOGLE_TOOLS_ABOVE_CORE: frozenset[str] = frozenset({
+ "google_workspace__draft_gmail_message",
+ "google_workspace__get_gmail_attachment_content",
+ "google_workspace__get_gmail_thread_content",
+ "google_workspace__get_spreadsheet_info",
+ "google_workspace__list_docs_in_folder",
+ "google_workspace__list_drive_items",
+ "google_workspace__list_sheet_tables",
+ "google_workspace__list_spreadsheets",
+ "google_workspace__query_freebusy",
+ "google_workspace__search_docs",
+})
+
+
+def google_tool_manifest(settings: Settings) -> dict[str, tuple[str, ...]]:
+ """``GOOGLE_TOOL_MANIFEST`` narrowed to what workspace-mcp actually loaded.
+
+ The manifest is a static list of 24 names, but the child imports only the
+ service modules `WORKSPACE_MCP_TOOLS` names and registers only the tools in
+ `WORKSPACE_MCP_TOOL_TIER`. Advertising the unfiltered list tells the model a
+ tool is "already available, no `search_tools` needed" when it is not in the
+ index at all — it then calls it, gets extensible-mcp's "has not been
+ discovered via search_tools" refusal, and has no way to reconcile that with
+ its own instructions.
+
+ Both inputs are process-constant, so the result is too and block 0 stays
+ cacheable. A heading with nothing left is dropped rather than rendered empty.
+ """
+ services = settings.workspace_mcp_service_list
+ core_only = settings.workspace_mcp_tier == "core"
+ out: dict[str, tuple[str, ...]] = {}
+ for heading, names in GOOGLE_TOOL_MANIFEST.items():
+ if services is not None and GOOGLE_SERVICE_OF[heading] not in services:
+ continue
+ kept = tuple(n for n in names if not (core_only and n in GOOGLE_TOOLS_ABOVE_CORE))
+ if kept:
+ out[heading] = kept
+ return out
+
+
+def pinned_google_tools(settings: Settings) -> frozenset[str]:
+ """The pinned names worth probing and advertising on this install.
+
+ One source of truth shared by the prompt and `prime_pinned_tools`, so the
+ gateway never probes for a tool the prompt does not claim, and vice versa.
+ """
+ return frozenset(
+ name for names in google_tool_manifest(settings).values() for name in names
+ )
+
def _on(flag: bool) -> str:
return "on" if flag else "off"
@@ -104,9 +169,18 @@ def render_connected_systems(
]
if google:
+ manifest = google_tool_manifest(settings)
+ # Derived, never hardcoded: naming all five while a subset is loaded is
+ # the whole bug. Ordered as the manifest is, so the text is stable.
+ loaded = [s for s in GOOGLE_TOOL_MANIFEST if s in manifest]
+ services = (
+ " and ".join(filter(None, [", ".join(loaded[:-1]), loaded[-1]]))
+ if loaded
+ else "no services"
+ )
lines.append(
f"- **Google Workspace: connected** as your own account ({email}) — "
- "Gmail, Calendar, Drive, Docs and Sheets. Act from your own account: "
+ f"{services}. Act from your own account: "
"never ask the user which address to send from — always send, create "
"events, and own documents from your own account. If you need the "
"user's email or a third party's, ask for that specifically by name. "
@@ -118,8 +192,8 @@ def render_connected_systems(
"available, no `search_tools` needed. Use `search_tools` only for "
"something not listed here."
)
- for service in sorted(GOOGLE_TOOL_MANIFEST):
- names = ", ".join(f"`{n}`" for n in GOOGLE_TOOL_MANIFEST[service])
+ for service in sorted(manifest):
+ names = ", ".join(f"`{n}`" for n in manifest[service])
lines.append(f" - {service}: {names}")
else:
lines.append(
diff --git a/packages/core/tests/unit/test_connected_systems.py b/packages/core/tests/unit/test_connected_systems.py
index d11ad36c..b5e9e1c7 100644
--- a/packages/core/tests/unit/test_connected_systems.py
+++ b/packages/core/tests/unit/test_connected_systems.py
@@ -16,7 +16,10 @@ from openexecutive.orchestrator.mcp_gateway import MCPGateway, gateway_server_na
from openexecutive.prompts.cache_manager import build_system_blocks
from openexecutive.prompts.connected_systems import (
GOOGLE_TOOL_MANIFEST,
+ GOOGLE_TOOLS_ABOVE_CORE,
PINNED_GOOGLE_TOOLS,
+ google_tool_manifest,
+ pinned_google_tools,
render_connected_systems,
)
@@ -65,6 +68,81 @@ def test_google_connected_lists_every_pinned_name(channel_env: Any) -> None:
assert "no `search_tools` needed" in text
+def test_a_service_subset_is_neither_named_nor_pinned(channel_env: Any) -> None:
+ """The bug this fix is for: `WORKSPACE_MCP_TOOLS` decides which service
+ modules workspace-mcp imports, so naming all five and pinning all 24 tells
+ the model a tool is "already available" when it is not in the index at all.
+ It then calls it, gets extensible-mcp's "has not been discovered via
+ search_tools" refusal, and cannot reconcile that with its instructions.
+ """
+ channel_env.setenv("WORKSPACE_MCP_TOOLS", "gmail,calendar,drive")
+ text = _render(("google_workspace",))
+
+ assert "Gmail, Calendar and Drive" in text
+ assert "Docs" not in text and "Sheets" not in text
+
+ for name in GOOGLE_TOOL_MANIFEST["Gmail"]:
+ assert f"`{name}`" in text
+ for dropped in GOOGLE_TOOL_MANIFEST["Docs"] + GOOGLE_TOOL_MANIFEST["Sheets"]:
+ assert dropped not in text
+
+
+def test_the_core_tier_drops_the_tools_it_does_not_register(channel_env: Any) -> None:
+ """The second axis. Tiers are cumulative and `extended`/`complete` carry all
+ 24, so only `core` subtracts anything — the ten names in
+ `GOOGLE_TOOLS_ABOVE_CORE`, read from workspace-mcp 1.29.0's
+ `core/tool_tiers.yaml`."""
+ channel_env.setenv("WORKSPACE_MCP_TOOL_TIER", "core")
+ text = _render(("google_workspace",))
+
+ # A service whose core tools survive is still named...
+ assert "Gmail" in text
+ assert "`google_workspace__search_gmail_messages`" in text
+ # ...while the above-core names are gone.
+ for name in GOOGLE_TOOLS_ABOVE_CORE:
+ assert name not in text
+
+
+def test_the_default_install_is_unchanged(channel_env: Any) -> None:
+ """Backward compatibility, stated as a test rather than assumed: with the
+ Dockerfile's `all` / `complete` the filter is a no-op and every pinned name
+ is still advertised."""
+ channel_env.setenv("WORKSPACE_MCP_TOOLS", "all")
+ channel_env.setenv("WORKSPACE_MCP_TOOL_TIER", "complete")
+ text = _render(("google_workspace",))
+ assert "Gmail, Calendar, Drive, Docs and Sheets" in text
+ for name in PINNED_GOOGLE_TOOLS:
+ assert f"`{name}`" in text
+ assert pinned_google_tools(get_settings()) == PINNED_GOOGLE_TOOLS
+
+
+def test_priming_probes_exactly_what_the_prompt_advertises(channel_env: Any) -> None:
+ """One source of truth. If priming probed the full manifest while the
+ prompt advertised a subset, the boot log would warn once per unloaded tool
+ and send the operator looking for a rename that never happened — which is
+ what it did before this change."""
+ channel_env.setenv("WORKSPACE_MCP_TOOLS", "gmail")
+ settings = get_settings()
+
+ advertised = {
+ n for names in google_tool_manifest(settings).values() for n in names
+ }
+ assert pinned_google_tools(settings) == advertised
+ assert advertised == set(GOOGLE_TOOL_MANIFEST["Gmail"])
+ assert advertised < PINNED_GOOGLE_TOOLS
+
+
+def test_block_0_stays_one_stable_cached_block_under_a_subset(channel_env: Any) -> None:
+ """The expensive failure mode would be a block 0 that varies. Both inputs
+ are process-constant, so a subset changes the text once and then holds."""
+ channel_env.setenv("WORKSPACE_MCP_TOOLS", "gmail,drive")
+ a = build_system_blocks(mcp_servers=("google_workspace",))
+ b = build_system_blocks(mcp_servers=("google_workspace",))
+ assert a == b
+ assert a[0]["cache_control"] == {"type": "ephemeral", "ttl": "1h"}
+ assert a[0]["text"].count("## Connected Systems") == 1
+
+
def test_without_google_nothing_claims_gmail_or_drive(channel_env: Any) -> None:
text = _render(("fetch",))
assert "Google Workspace: not connected" in text
diff --git a/packages/core/tests/unit/test_workspace_mcp_services.py b/packages/core/tests/unit/test_workspace_mcp_services.py
index 3aa6fb28..396372e0 100644
--- a/packages/core/tests/unit/test_workspace_mcp_services.py
+++ b/packages/core/tests/unit/test_workspace_mcp_services.py
@@ -99,3 +99,77 @@ def test_dockerfile_defaults_match_the_launcher() -> None:
def test_gateway_forwards_the_service_list() -> None:
# Without it an operator's WORKSPACE_MCP_TOOLS never reaches the child.
assert "WORKSPACE_MCP_TOOLS" in mcp_gateway._FORWARDED_ENV_VARS
+
+
+# --- the Python side must agree with the shell ------------------------------
+#
+# `Settings.workspace_mcp_service_list` exists so the prompt can describe what
+# the child loaded. If it disagrees with the launcher, the prompt is wrong again
+# — and the dangerous direction is the Python side answering NARROWER, which
+# hides a tool the model really can call.
+
+
+ "value",
+ [
+ None, # unset
+ "all",
+ "gmail,calendar,drive",
+ "gmail, calendar ,drive", # the launcher splits on commas AND whitespace
+ "gmail",
+ ",", # no service names at all -> all
+ " ", # whitespace only -> all
+ "", # empty -> all
+ "$WORKSPACE_MCP_TOOLS", # extensible-mcp left the literal in place
+ "all,", # a decorated "all" still means every service
+ " all ",
+ ",all",
+ "gmail,all", # "all" beside real names is an operator error
+ "gmail,gmail,drive", # duplicates
+ ],
+)
+def test_the_python_service_list_matches_the_real_launcher(
+ tmp_path: Path, monkeypatch: pytest.MonkeyPatch, value: str | None
+) -> None:
+ """Parity against the shell, not against a second copy of the same reasoning.
+
+ The expectation comes from executing `workspace-mcp-launch.sh` and reading
+ its argv, so this cannot pass by both sides sharing a mistake — which is
+ how it caught two real divergences while this was being written. Both had
+ Python answering WIDER than the launcher, hiding the very mismatch the
+ filter exists to prevent.
+ """
+ from openexecutive.config import get_settings
+
+ env = {} if value is None else {"WORKSPACE_MCP_TOOLS": value}
+ from_shell = _services(_launch_args(tmp_path, **env))
+
+ if value is None:
+ monkeypatch.delenv("WORKSPACE_MCP_TOOLS", raising=False)
+ else:
+ monkeypatch.setenv("WORKSPACE_MCP_TOOLS", value)
+ from_python = get_settings().workspace_mcp_service_list
+
+ # `None` on both sides means "every service"; the launcher omits --tools.
+ if from_shell is None:
+ assert from_python is None, f"{value!r}: shell loads all, python says {from_python}"
+ else:
+ # The launcher keeps duplicates in argv; workspace-mcp imports a module
+ # once either way, so compare as sets with order preserved on our side.
+ assert from_python is not None, f"{value!r}: shell narrowed, python says all"
+ assert set(from_python) == set(from_shell), f"{value!r}"
+
+
+def test_the_python_tier_matches_the_real_launcher(
+ tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+ from openexecutive.config import get_settings
+
+ for value in (None, "core", "extended", "complete", "$WORKSPACE_MCP_TOOL_TIER"):
+ env = {} if value is None else {"WORKSPACE_MCP_TOOL_TIER": value}
+ from_shell = _tier(_launch_args(tmp_path, **env))
+ if value is None:
+ monkeypatch.delenv("WORKSPACE_MCP_TOOL_TIER", raising=False)
+ else:
+ monkeypatch.setenv("WORKSPACE_MCP_TOOL_TIER", value)
+ assert get_settings().workspace_mcp_tier == from_shell, f"{value!r}"