Browser-based graphical shell and control plane for a Linux server you own — real terminal, file manager, live metrics and BYOK AI in one mobile-first workspace. Next.js 16 · React 19 · node-pty · Tailwind v4 · shadcn/ui. Public alpha, self-hosted.
TypeScript browser-shellcontrol-planefile-managerhomelabmobile-firstnextjsptyreactself-hostedserver-managementshadcn-uisysadmintailwindcssterminaltypescriptvpsweb-desktopweb-terminal
docs: reconcile RASMIC with latest agent docs
Rahmane2b8dbc57ee2
diff --git a/CLAUDE.md b/CLAUDE.mdindex b1cea758..32773cb9 100644--- a/CLAUDE.md+++ b/CLAUDE.md@@ -349,8 +349,8 @@ the real runtime and working analogue before changing a tool or skill contract.- **Tool names are a cross-repo contract.** `rahmanef63/connectors-gateway` registeredmso as a connector on 2026-08-17 and historically pinned a subset of the MCP names asliteral `x-upstream` strings.- <!-- mcp-toolset: server=1.6.0 version=2026.09.02.7 tools=66 read=32 write=22 exec=12 -->- MSO currently exposes **67 transport tools**: **66 model/operator tools** (32 read, 22 write,+ <!-- mcp-toolset: server=1.6.0 version=2026.09.02.9 tools=69 read=33 write=24 exec=12 -->+ MSO currently exposes **70 transport tools**: **69 model/operator tools** (33 read, 24 write,12 exec) plus the app-only `workflow_status` bridge; `GET /mcp` plus `_meta.toolset` exposethe current version/hash/name manifest.Treat the gateway mapping as a separate cross-repo contract and verify it before renaming. Renaming or removingdiff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.mdindex 5641d9af..a83e3d2f 100644--- a/docs/ARCHITECTURE.md+++ b/docs/ARCHITECTURE.md@@ -160,8 +160,8 @@ park a visible Approve/Deny card. The complete semantic contract is**MCP** is for external clients such as ChatGPT, Claude.ai and Cursor. Its catalog lives in`lib/mcp/`; access is controlled by the OAuth token scope rather than Alfa approval cards.-<!-- mcp-toolset: server=1.6.0 version=2026.09.02.7 tools=66 read=32 write=22 exec=12 -->-At the current toolset it exposes **67 transport tools**: **66 model/operator tools** (32 read, 22 write, 12 exec) plus the app-only `workflow_status` progress bridge. Project-specific+<!-- mcp-toolset: server=1.6.0 version=2026.09.02.9 tools=69 read=33 write=24 exec=12 -->+At the current toolset it exposes **70 transport tools**: **69 model/operator tools** (33 read, 24 write, 12 exec) plus the app-only `workflow_status` progress bridge. Project-specificfunction names remain data behind `project_capabilities` / `project_function_call`, so oneproject cannot dynamically rewrite the global MCP tool prefix. `GET /mcp` remains the live count/hash authority.diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.mdindex a3d8f75d..2f8f58e4 100644--- a/docs/CHANGELOG.md+++ b/docs/CHANGELOG.md@@ -8,6 +8,7 @@ this is the *what*, and it is what Settings → About shows as “What's new”.**Added**+- add RASMIC orchestration memory automation- `agent` section terminal and recover API errors- `agent` name sessions and wrap composer- `agent` correlate local replies and add subagents@@ -25,14 +26,30 @@ this is the *what*, and it is what Settings → About shows as “What's new”.- `agent` make restart process-safe- `a2a` harden remote v1 delegation+**Faster**++- `agent` route capabilities before model reasoning++**Tests**++- follow RASMIC storage split+**Docs**+- reconcile RASMIC with latest agent docs+- publish catalog-first RASMIC runtime+- publish RASMIC MCP contract- `testing` record 1.12 live verification**Chores**- `agent` sync docs and trim internal exports+**Other**++- pin ZAP Docker Hub mirror+- cover RASMIC orchestration+## 2026-09-01**Added**diff --git a/docs/CHATGPT-PLUGIN.md b/docs/CHATGPT-PLUGIN.mdindex 5d17e2b2..1c21b41e 100644--- a/docs/CHATGPT-PLUGIN.md+++ b/docs/CHATGPT-PLUGIN.md@@ -12,10 +12,10 @@> <https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt-beta>> and <https://help.openai.com/en/articles/20001256-plugins-in-chatgpt-and-codex>.-<!-- mcp-toolset: server=1.6.0 version=2026.09.02.7 tools=66 read=32 write=22 exec=12 -->+<!-- mcp-toolset: server=1.6.0 version=2026.09.02.9 tools=69 read=33 write=24 exec=12 -->-MSO currently exposes MCP server **1.6.0**, toolset **2026.09.02.7**: **67 transport tools** total; **66 model/operator tools**-(32 read, 22 write, 12 exec) plus app-only `workflow_status` for the progress widget. Use `GET /mcp` or Settings → MCP as the live authority if+MSO currently exposes MCP server **1.6.0**, toolset **2026.09.02.9**: **70 transport tools** total; **69 model/operator tools**+(33 read, 24 write, 12 exec) plus app-only `workflow_status` for the progress widget. Use `GET /mcp` or Settings → MCP as the live authority ifthis document and a deployed instance ever disagree.### Local-agent and subagent tools@@ -145,7 +145,7 @@ A token sees a scope prefix; there is no per-project or per-agent hidden tool fi`tools/list` filters the catalog and `tools/call` independently re-checks the requiredscope.-### `read` — 32 model/operator tools+### `read` — 33 model/operator tools- `a2a_agent_discover`- `a2a_agents_list`@@ -171,6 +171,7 @@ scope.- `local_agent_request_wait`- `local_agents_list`- `project_capabilities`+- `project_memory_search`- `projects_list`- `screen_capture`- `skills_list`@@ -180,7 +181,7 @@ scope.- `sys_stats`- `tool_forge_candidates`-### `write` — read + 22 tools (54 visible at write scope)+### `write` — read + 24 tools- `a2a_agent_register`- `a2a_agent_remove`@@ -200,12 +201,14 @@ scope.- `hostinger_dns_upsert`- `local_agent_message_send`- `local_agent_reply`+- `project_memory_upsert`+- `project_script_run`- `tool_forge_propose`- `workflow_cancel`- `workflow_finish`- `workflow_start`-### `exec` — write + 12 tools (66 visible at exec scope)+### `exec` — write + 12 tools- `a2a_handoff`- `a2a_message_send`@@ -220,6 +223,8 @@ scope.- `tool_forge_evaluate`- `tool_forge_promote`+RASMIC's `project_memory_search`, `project_memory_upsert`, and `project_script_run` keep compact redacted orchestration memory/evidence and deterministic automation portable under a managed repo's `.agent/` directory. `project_memory_search` uses one schema for ranked `search`, deterministic `related`, and compact `timeline` views instead of adding more tool names. They do not grant access to another repo or external memory service.+`exec_run` is full host shell power as the MSO service user. The command filter is anaccident tripwire, not a sandbox. Grant `exec` only to a ChatGPT app/workspace you trustwith the same care as a remote shell.@@ -267,7 +272,7 @@ Settings → MCP:server: 1.6.0toolset: 2026.09.02.7hash: <live schema hash>-count: 66 model/operator tools (+ 1 app-only bridge)+count: 69 model/operator tools (+ 1 app-only bridge)```Use this sequence after an MSO MCP change:diff --git a/docs/COGNITIVE-RUNTIME.md b/docs/COGNITIVE-RUNTIME.mdindex 01f5373e..01313bf8 100644--- a/docs/COGNITIVE-RUNTIME.md+++ b/docs/COGNITIVE-RUNTIME.md@@ -70,15 +70,9 @@ The external MCP server keeps a stable, complete scope-filtered catalog for stanMSO's own terminal model harness uses a provider-neutral **per-turn capability router** instead of sendingall schemas on every call.-The router always retains a small control/discovery core (workflow lifecycle, `skills_search`, project and-session discovery), scores the current request plus recent evidence, preserves recently used tools, and-adds required companions such as `exec_job_status`/`exec_job_cancel` when a long job is selected.-The normal target is at most 14 selected schemas before mandatory companions.+The router is **catalog-first, not model-first**. It classifies the latest user intent through a deterministic bilingual capability catalog before the model call, loads only that capability pack plus proven dependencies, and falls back to `skills_search`/lexical discovery only when the catalog cannot classify the request. There is no always-on schema core. Short continuation prompts inherit only a bounded prior user intent; routing never needs the full projected conversation.-If a missing capability is needed, `skills_search` searches the unified catalog of tools, Skills, and-learned recipes. A tool named by discovery/tool output becomes eligible on the next round. The full local-catalog still performs invocation validation, so hiding a schema from one turn neither deletes the tool nor-changes its permission.+Capability packs are phase-aware. A repository change starts with `workflow_start` alone; after that call, the next round exposes the bounded execution/finish pack instead of sending `workflow_start` again. Long jobs add status/cancel companions, writes add their bounded read prerequisite, and local-agent/A2A/Forge routes add only the lifecycle tools needed for that class. Hiding a schema from one turn neither deletes the tool nor changes its permission because invocation is still validated against the complete local catalog.## Tool-result budget@@ -117,15 +111,16 @@ bun run bench:agents -- --model gpt-5.6-terra # dry-run cross-agent planbun run bench:agents -- --model gpt-5.6-terra --run # scratch-only external smoke```-The core benchmark is deterministic and provider-neutral. Its checked scenarios cover application logs,-server health, safe file editing, long/short execution, project functions, Cloudflare/Hostinger/Dokploy,-browser/screenshots, memory, session resume, and Skills.+The core benchmark is deterministic and provider-neutral. Its checked scenarios cover application logs, server health, safe file editing, long/short execution, project functions, Cloudflare/Hostinger/Dokploy, browser/screenshots, memory, session resume, Skills, manual regression evidence, Indonesian repository changes, local-agent requests, subagents, A2A and Tool Forge. It also verifies the second phase after `workflow_start`.Release gates require:- **100% required-tool recall** on the checked scenarios;- **100% typed-memory retrieval, temporal-resolution and conflict-resolution accuracy** on the deterministic fixture suite;- deterministic selection for identical input;+- average active tools ≤ 4 and active-schema reduction ≥ 95%;+- catalog hit rate ≥ 90%, average routing text ≤ 12 KiB, and average route-requested history budget ≤ 16k tokens;+- phase-aware repository execution after `workflow_start`;- lower active schema bytes than the full MSO catalog;- when Hermes' installed CLI exposes `hermes prompt-size`, lower active tool-schema bytes than that exactcomparable metric.@@ -138,20 +133,21 @@ scenarios before making that claim.## Current P2 + A2A baseline-Benchmark snapshot captured on the development host after P2 + outbound A2A integration (before later Local Agent/subagent catalog growth):+On the development host after P2 + outbound A2A integration:| Metric | MSO P2 + A2A | Hermes local baseline ||---|---:|---:|-| Full MSO transport catalog | 60 tools | — |-| Full MSO schema bytes | 41,519 | — |-| Average MSO active tools / turn | 15.4 | — |-| Average active schema bytes / turn | 11,614 | 44,758 tool-schema bytes |-| MSO schema reduction vs full catalog | 72.0% | — |-| Required-tool routing recall | 100% | not the same benchmark |+| Full MSO transport catalog | 70 tools | — |+| Full model-tool schema bytes | 53,108 | — |+| Average MSO active tools / turn | 2.7 | — |+| Average active schema bytes / turn | 2,252 | 44,758 tool-schema bytes |+| MSO schema reduction vs full catalog | 95.8% | — |+| Required-tool routing recall / catalog hit | 100% / 100% | not the same benchmark |+| Average routing text / requested history budget | 76 bytes / 10,571 tokens | — || Typed-memory fixture accuracy | 100% (8/8) | not implemented in this harness || Deterministic routing / memory resolution | yes / yes | not the same benchmark |-At that P1/P2 rollout snapshot, the learned-recipe store reported **179 successful / 186 total workflows (96.2%)**. Those recipes predated the P1 quality schema, so step-level telemetry coverage correctly started at **0%** rather than fabricating historical tool-error/retry data. New or re-learned recipes carry `qualityVersion: 1`; coverage grows naturally from real post-P1 executions.+The existing learned-recipe store currently reports **179 successful / 186 total workflows (96.2%)**, but those recipes predate the P1 quality schema, so step-level telemetry coverage correctly starts at **0%** rather than fabricating historical tool-error/retry data. New or re-learned recipes carry `qualityVersion: 1`; coverage grows naturally from real post-P1 executions.A non-interactive full MSO tool loop is now available as `mso agent --oneshot <prompt> --json`. It defaults to read-only autonomous approval; write/exec require explicit `--approve-scope write|exec`. The cross-agent harness uses a private scratch read task whose nonce and values exist only in a file. A first smoke proved MSO on `openai-codex/gpt-5.6-terra`; Hermes failed because its local configuration does not expose an `openai` provider and OpenClaw rejected that model override for agent `main`, so the harness correctly returned **`comparable=false`**. That is connectivity/configuration evidence, not a ranking.@@ -163,14 +159,14 @@ P1 supplies typed/provenance-aware temporal memory, deterministic retrieval evalForge only creates inert candidates from repeated high-quality recipes. Skill candidates are deterministic redacted guidance; executable project-function candidates require an exec-scope verified recipe and may reference only existing project-owned Node code. Fixtures run in the dedicated cached Docker sandbox with no network and a read-only project mount. Evaluation records current toolset, candidate/target state, source SHA-256 and exact sandbox image evidence; promotion requires literal confirmation and immediately repeats the checks. A candidate cannot silently raise scope, overwrite an existing capability, or turn generated text into host-executable code.-Outbound A2A v1 remains orthogonal: it adds remote-agent delegation, while Forge turns repeated **local verified procedures** into reviewed project capabilities. That recorded combined routing benchmark covered both surfaces with 100% required-tool recall while selecting 15.4 of the then-current 60 tools on average; current catalog size is intentionally read from the live MCP descriptor instead of rewriting this historical benchmark.+Outbound A2A v1 remains orthogonal: it adds remote-agent delegation, while Forge turns repeated **local verified procedures** into reviewed project capabilities. The combined routing benchmark covers these surfaces with 100% required-tool recall while selecting only 2.7 of 70 tools on average; it also requires the catalog-first routing and history-budget gates above.The next phase should continue from measured primitives:-1. **Programmatic read-only orchestration** — bulk filtering/aggregation may reduce model round-trips while remaining constrained by caller scope and normal host guards.-2. **Benchmark corpus expansion** — add multiple read/write/recovery/security task classes and configure Hermes/OpenClaw with an actually equivalent model/provider path before any overall comparison.-3. **Memory calibration** — grow post-P1 quality telemetry and test confidence/authority policy against real corrections before introducing graph-memory complexity.-4. **Cost-per-success accounting** — normalize provider usage reports only where token/cost semantics are comparable; missing usage must remain unknown, never zero.+1. **Benchmark corpus expansion** — keep adding read/write/recovery/security task classes and configure Hermes/OpenClaw with an actually equivalent model/provider path before any overall comparison.+2. **Real token-per-success telemetry** — use provider-reported usage plus the non-prompt routing telemetry; missing provider usage remains unknown, never zero.+3. **Memory calibration** — grow post-P1 quality telemetry and validate deterministic relation/conflict policy against real corrections.+4. **Optional external memory transport** — keep repo-local `.agent/` canonical; any sync service consumes/produces the portable bundle and never becomes a runtime dependency.## Security notesdiff --git a/docs/CONNECTORS-GATEWAY-INTEGRATION.md b/docs/CONNECTORS-GATEWAY-INTEGRATION.mdindex 7f520bb1..6aace2a0 100644--- a/docs/CONNECTORS-GATEWAY-INTEGRATION.md+++ b/docs/CONNECTORS-GATEWAY-INTEGRATION.md@@ -4,16 +4,18 @@> can evolve independently, so do not claim its mapping count is current without checking> that repository at the time of the change.-<!-- mcp-toolset: server=1.6.0 version=2026.09.02.7 tools=66 read=32 write=22 exec=12 -->+<!-- mcp-toolset: server=1.6.0 version=2026.09.02.9 tools=69 read=33 write=24 exec=12 -->-MSO currently exposes MCP server **1.6.0**, toolset **2026.09.02.7**: **67 transport tools** total; **66 model/operator tools**-(32 read, 22 write, 12 exec) plus app-only `workflow_status`. `GET /mcp` exposes the live names, version/hash and scoped+MSO currently exposes MCP server **1.6.0**, toolset **2026.09.02.9**: **70 transport tools** total; **69 model/operator tools**+(33 read, 24 write, 12 exec) plus app-only `workflow_status`. `GET /mcp` exposes the live names, version/hash and scopedmanifest and is the machine-readable parity source.## Local collaboration additionsToolset `2026.09.02.7` tracks `local_agents_list`, `local_agent_message_send`, `local_agent_reply`, `local_agent_request_wait`, `local_agent_inbox`, `local_agent_request`, and `agent_subagent_run`. A gateway should preserve these literal names if it chooses to expose them. The six `local_agent_*`/`local_agents_list` tools address durable same-principal MSO sessions; `agent_subagent_run` is a foreground same-session worker and must not be modeled as a remote A2A registration or long-lived peer.+Toolset `2026.09.02.8` introduced RASMIC's `project_memory_search`, `project_memory_upsert`, and `project_script_run`; toolset `2026.09.02.9` keeps those three names stable while extending `project_memory_search` with `search | related | timeline` views and the catalog-first runtime. A gateway may expose these literal names, but must preserve their MSO scopes and must not reinterpret repo-local `.agent` records as gateway-owned/global memory.+## Why this contract mattersConnector/gateway manifests commonly bind an action to an MSO tool with a literal upstream@@ -37,13 +39,13 @@ Before changing an MCP tool name:## Current MSO catalog-### Read (32 model/operator)+### Read (33 model/operator)-`a2a_agent_discover`, `a2a_agents_list`, `a2a_task_get`, `agent_memory_read`, `agent_memory_search`, `agent_session_current`, `agent_session_resume`, `agent_sessions_list`, `apps_list`, `apps_logs`, `browser_status`, `cloudflare_zones_list`, `dokploy_projects_list`, `exec_job_status`, `fs_list`, `fs_read`, `fs_search`, `fs_usage`, `infra_provider_doctor`, `infra_providers_list`, `local_agent_inbox`, `local_agent_request_wait`, `local_agents_list`, `project_capabilities`, `projects_list`, `screen_capture`, `skills_list`, `skills_read`, `skills_search`, `sys_processes`, `sys_stats`, `tool_forge_candidates`.+`a2a_agent_discover`, `a2a_agents_list`, `a2a_task_get`, `agent_memory_read`, `agent_memory_search`, `agent_session_current`, `agent_session_resume`, `agent_sessions_list`, `apps_list`, `apps_logs`, `browser_status`, `cloudflare_zones_list`, `dokploy_projects_list`, `exec_job_status`, `fs_list`, `fs_read`, `fs_search`, `fs_usage`, `infra_provider_doctor`, `infra_providers_list`, `local_agent_inbox`, `local_agent_request_wait`, `local_agents_list`, `project_capabilities`, `project_memory_search`, `projects_list`, `screen_capture`, `skills_list`, `skills_read`, `skills_search`, `sys_processes`, `sys_stats`, `tool_forge_candidates`.-### Write (22 beyond read)+### Write (24 beyond read)-`a2a_agent_register`, `a2a_agent_remove`, `agent_memory_forget`, `agent_memory_remember`, `agent_session_note`, `agent_session_rename`, `apps_power`, `cloudflare_dns_upsert`, `dokploy_project_ensure`, `fs_copy`, `fs_delete`, `fs_mkdir`, `fs_move`, `fs_upload_file`, `fs_write`, `hostinger_dns_upsert`, `local_agent_message_send`, `local_agent_reply`, `tool_forge_propose`, `workflow_cancel`, `workflow_finish`, `workflow_start`.+`a2a_agent_register`, `a2a_agent_remove`, `agent_memory_forget`, `agent_memory_remember`, `agent_session_note`, `agent_session_rename`, `apps_power`, `cloudflare_dns_upsert`, `dokploy_project_ensure`, `fs_copy`, `fs_delete`, `fs_mkdir`, `fs_move`, `fs_upload_file`, `fs_write`, `hostinger_dns_upsert`, `local_agent_message_send`, `local_agent_reply`, `project_memory_upsert`, `project_script_run`, `tool_forge_propose`, `workflow_cancel`, `workflow_finish`, `workflow_start`.### Exec (12 beyond write)diff --git a/docs/MCP.md b/docs/MCP.mdindex ea4c5ff0..cec04502 100644--- a/docs/MCP.md+++ b/docs/MCP.md@@ -75,8 +75,8 @@ Picked per token, on the consent screen, capped by `OS_MCP_MAX_SCOPE`. The highe| Scope | Tools ||---|---|-| `read` | `a2a_agent_discover` `a2a_agents_list` `a2a_task_get` `agent_memory_read` `agent_memory_search` `agent_session_current` `agent_session_resume` `agent_sessions_list` `apps_list` `apps_logs` `browser_status` `cloudflare_zones_list` `dokploy_projects_list` `exec_job_status` `fs_list` `fs_read` `fs_search` `fs_usage` `infra_provider_doctor` `infra_providers_list` `local_agent_inbox` `local_agent_request_wait` `local_agents_list` `project_capabilities` `projects_list` `screen_capture` `skills_list` `skills_read` `skills_search` `sys_processes` `sys_stats` `tool_forge_candidates` |-| `write` | + `a2a_agent_register` `a2a_agent_remove` `agent_memory_forget` `agent_memory_remember` `agent_session_note` `agent_session_rename` `apps_power` `cloudflare_dns_upsert` `dokploy_project_ensure` `fs_copy` `fs_delete` `fs_mkdir` `fs_move` `fs_upload_file` `fs_write` `hostinger_dns_upsert` `local_agent_message_send` `local_agent_reply` `tool_forge_propose` `workflow_cancel` `workflow_finish` `workflow_start` |+| `read` | `a2a_agent_discover` `a2a_agents_list` `a2a_task_get` `agent_memory_read` `agent_memory_search` `agent_session_current` `agent_session_resume` `agent_sessions_list` `apps_list` `apps_logs` `browser_status` `cloudflare_zones_list` `dokploy_projects_list` `exec_job_status` `fs_list` `fs_read` `fs_search` `fs_usage` `infra_provider_doctor` `infra_providers_list` `local_agent_inbox` `local_agent_request_wait` `local_agents_list` `project_capabilities` `project_memory_search` `projects_list` `screen_capture` `skills_list` `skills_read` `skills_search` `sys_processes` `sys_stats` `tool_forge_candidates` |+| `write` | + `a2a_agent_register` `a2a_agent_remove` `agent_memory_forget` `agent_memory_remember` `agent_session_note` `agent_session_rename` `apps_power` `cloudflare_dns_upsert` `dokploy_project_ensure` `fs_copy` `fs_delete` `fs_mkdir` `fs_move` `fs_upload_file` `fs_write` `hostinger_dns_upsert` `local_agent_message_send` `local_agent_reply` `project_memory_upsert` `project_script_run` `tool_forge_propose` `workflow_cancel` `workflow_finish` `workflow_start` || `exec` | + `a2a_handoff` `a2a_message_send` `a2a_task_cancel` `agent_subagent_run` `browser_power` `exec_job_cancel` `exec_job_start` `exec_run` `local_agent_request` `project_function_call` `tool_forge_evaluate` `tool_forge_promote` |Alfa — the in-app assistant — overlaps the same host capabilities under dot.case names,@@ -85,7 +85,7 @@ without a written reason. `skills_search` maps to Alfa's `skills.search`. The inprovider tools are intentionally MCP-only: the terminal MSO Agent consumes this canonicalcatalog through the owner-only agent-tools bridge, while in-shell Alfa has dedicatedDokploy/Cloudflare feature surfaces instead of receiving public-DNS credentials/capabilities.-`screen_capture`, `projects_list`, `project_capabilities`, `project_function_call`, `fs_upload_file`, `workflow_start`, `workflow_cancel`,+`screen_capture`, `projects_list`, `project_capabilities`, `project_function_call`, `project_memory_search`, `project_memory_upsert`, `project_script_run`, `fs_upload_file`, `workflow_start`, `workflow_cancel`,`workflow_finish`, the async `exec_job_*` lifecycle and the `infra_*` / Dokploy / Cloudflare /Hostinger provider tools are explicitly MCP-only: the external connector needs visual proof,an explicit project enumeration and an actor-scoped task boundary, while Alfa already@@ -126,10 +126,10 @@ The catalog has a stable server version plus a schema-derived toolset signature.Settings → MCP shows the current version/hash/count and stores a browser-local acknowledgement when the operator marks ChatGPT refreshed. A later signature change becomes an explicit stale-snapshot warning. This does not mutate ChatGPT remotely; it makes the required refresh visible instead of relying on memory.-<!-- mcp-toolset: server=1.6.0 version=2026.09.02.7 tools=66 read=32 write=22 exec=12 -->+<!-- mcp-toolset: server=1.6.0 version=2026.09.02.9 tools=69 read=33 write=24 exec=12 -->-Current transport catalog: **67 tools**, server `1.6.0` / toolset `2026.09.02.7`.-The model/operator catalog is **66 tools** (32 read, 22 write, 12 exec); `workflow_status` is the+Current transport catalog: **70 tools**, server `1.6.0` / toolset `2026.09.02.9`.+The model/operator catalog is **69 tools** (33 read, 24 write, 12 exec); `workflow_status` is theone app-only MCP Apps bridge used by the progress widget and is documented separately.`agent_memory_search` is the typed-memory retrieval surface. It resolves semantic/episodic/procedural claims at an optional point in time, returns confidence/provenance and competing effective claims, and can expose superseded/retracted history when explicitly requested. `agent_memory_remember` remains the write surface and now accepts typed metadata; raw ChatGPT conversation ids are never stored as provenance.@@ -137,6 +137,10 @@ Session/memory tools add durable conversation context without creatingdynamic per-project tool names; `project_capabilities` and `project_function_call` remain thestable project-automation seam.+### RASMIC repo-local orchestration memory++RASMIC adds three stable generic MCP tools without coupling MSO to another repository or service: `project_memory_search` reads compact `.agent` task/debug/test/decision/failure memory through `search`, deterministic `related`, or chronological `timeline` views; `project_memory_upsert` writes one redacted structured record, including `source=user-manual` test evidence; and `project_script_run` replays only bounded read-only script manifests and promotes a candidate to `tested` only after a real successful replay. The script runner is write scope because that candidate→tested manifest transition is a mutation; every nested tool is revalidated as read-only at runtime.+### Local Agent collaborationLocal same-owner session collaboration is intentionally separate from public A2A. `local_agents_list` reports each durable session's short public `name`, lease status, and live receiver observability (`consumerConnected` / `consumerCount`). `local_agent_message_send` + `local_agent_reply` provide durable correlated mailbox semantics; `local_agent_request_wait` gives a bounded foreground request outcome without polling/resending; `local_agent_request` is an explicit exec-gated fresh worker from another durable session context and never claims to wake/control its original terminal/ChatGPT process. Human `@name` mentions resolve active public names only.@@ -360,13 +364,7 @@ its `project`, and the bootstrap carries a `discovery.complete` flag plus a## MCP-first cognitive runtime-MSO keeps the public MCP catalog capability-complete and scope-stable, but its own terminal-agent does not stuff every schema into every model turn. A provider-neutral router keeps a-small core (`workflow_*`, `skills_search`, project/session discovery), scores the current user-intent plus recent tool evidence, and normally loads at most 14 relevant schemas plus required-companions. `skills_search` already searches tools as well as skills/recipes, so a tool named by-discovery becomes available on the next model round without changing permissions. Public MCP-`tools/list` remains the full scope-filtered catalog for interoperability.+MSO keeps the public MCP catalog capability-complete and scope-stable, but its own terminal agent does not stuff every schema into every model turn. A shared deterministic bilingual capability catalog routes the latest intent **before the model call** and loads only the matching capability pack plus mandatory dependencies. There is no always-on schema core. `skills_search`/lexical discovery is a fallback for genuinely unknown intent, and short continuation prompts use only bounded prior user intent. Repository work is phase-aware: the bootstrap round exposes `workflow_start`; after it runs, the execution round exposes the bounded edit/test/finish pack instead of paying for the bootstrap schema again. Public MCP `tools/list` remains the full scope-filtered catalog for interoperability.Generic MCP text results are also context-budgeted: 32 KiB by default, with explicit boundedexceptions for file reads and long-job output. Oversize results return a parseable@@ -375,14 +373,10 @@ the model context. Explicit MCP Apps structured projections remain separate. Theapplies regardless of whether the selected provider is OpenAI, Anthropic, Google, Qwen, GLM,DeepSeek, or another compatible backend.-The CLI persists much more session history than it shows the model. Before a model call, recent-history is projected to roughly 55% of the model's context window (capped at 120k history-tokens) so system instructions, memory, skill/tool schemas, reasoning and output retain-headroom. At the durable 700k session threshold, structured compaction + sanitized archival+The CLI persists much more session history than it shows the model. The route is chosen first; then history is projected using the smaller of the provider-derived budget and the capability pack's requested budget. Without an override the old roughly-55%-of-context / 120k-cap behavior remains the ceiling, so system instructions, memory, skill/tool schemas, reasoning and output retain headroom. At the durable 700k session threshold, structured compaction + sanitized archivalreduces storage context and feeds the compact summary back into later turns.-Run `bun run bench:cognitive` for the reproducible provider-neutral routing/footprint gate. It-requires 100% required-tool recall and deterministic routing across the checked scenarios.+Run `bun run bench:cognitive` for the reproducible provider-neutral routing/footprint gate. It requires 100% required-tool recall and deterministic routing, ≥95% schema reduction, ≥90% catalog hit, ≤4 average active tools, bounded routing text/history budget, and the phase-aware repository transition. The current checked corpus reports 2.7 active tools and 2,252 schema bytes on average (95.8% reduction), with 100% recall/catalog hit.Hermes is compared only where an equivalent offline `prompt-size` metric exists; OpenClaw isreported but is not declared beaten on a non-comparable metric.