Skip to content

AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents

35.9 KB
Raw Blame

icon: 💬

Chat

A platform-level AI chat assistant that manages Activepieces projects via natural language. Streams LLM responses over WebSocket and exposes project resources (flows, tables, connections, runs) as callable tools through the project's MCP server. Conversations persist per-user with cross-session memory (personal instructions + remembered facts injected into every turn), compaction, attachments, multi-project context, and two-phase tool gating. EE/Cloud only (not registered in CE).

Execution model (read first)

The chat LLM loop runs in the worker, not the API. Send path: agent-conversation-controller.ts (POST /conversations/:id/messages) enqueues a WorkerJobType.EXECUTE_AGENT_RUN job → worker execute-agent-run.ts calls getAgentConfig RPC, assembles tools, runs run-agent-turn.ts (shared streamText() DI loop) → chunks stream back via sendAgentEvent RPC → websocket CHAT_MESSAGE_CHUNK (filtered by runId) → frontend reducer. agent-conversation-service.ts only does conversation CRUD + persistence.

Entities & services

  • ChatPersonalization (chat_personalization) — first-run onboarding: role + company, background research, researched empty-state cards. See chat personalization.
  • AgentConversation (agent_conversation) — per-user, per-platform, optionally per-project; status STREAMING/IDLE/ERROR, activeRunId, messages (ModelMessage[] JSONB), uiMessages, summary/summarizedUpToIndex for compaction.
  • UserMemory (user_memory) — one row per (platformId, userId): instructions (nullable text) + memories (jsonb string[]); capped at 50 facts × 280 chars and 4000 chars of instructions (agentHelpers.capMemories). Every write to memories is { next, base } and goes through a three-way merge under a row lock (agentHelpers.saveUserMemory); there is no blind overwrite, so a writer working from a stale list can't resurrect removed facts or wipe ones saved meanwhile.
  • Tool logic in ee/agent/; shared tool phase/classification in core/shared/.../ee/agent/.

How it works

  • Tools: local (ap_execute_action, ap_select_project, ap_fetch_url…), display cards (ap_show_connection_picker, ap_show_questions, ap_show_quick_replies…), and project-scoped MCP tools. Each tool is wired across up to four files — see gotcha: a chat tool lives in four files.
  • Two phases — every message starts in discovery and moves to build when a build skill loads (entersBuildPhase) or a build-only tool runs, including through ap_lazy_tool. It never goes back within a message. In skills mode the phase only turns thinking on; it does not hide tools and does not change the model (the fast model is step one only).
  • Gates (Redis pub/sub, 5-min timeout): display-tool cards, the action-run action preview, and the test-flow write gate. Flow build + publish are NOT gated.
  • Web access: ap_web_search is Tavily when configured, otherwise a sub-call to the provider's own search on the configured LLM credential (Anthropic web_search_20250305, Google grounding, OpenRouter web plugin); ap_fetch_url works everywhere.
  • Cross-session memory: instructions + facts injected into every turn (buildMemoryNote in agent-rpc-handlers.ts). Writes go through agentMemoryAi.applyInstruction — an LLM reconcile on the fast-tier model (add/forget, dedupe, supersede contradictions; non-AI fallbacks so it never hard-fails) — used by both the ap_remember tool and the /v1/chat/memory[/import|/instruct] endpoints; concurrent saves are merged under a pessimistic_write lock. UI lives in the settings hub (packages/web/src/app/components/settings-hub/).
  • Billing & credit gating: POST /conversations/:id/messages gates pre-enqueue — assertCreditsAndAppSumoNotExceeded blocks ALL chat (any provider) on the platform's credit/AppSumo balance (QUOTA_EXCEEDED), next to a per-user rate limit (40 messages / 10 min, HTTP 429). After each turn chatUsageTracker.track meters Autumn credits with creditValue = creditWeight + billableToolCalls (tier's weight for the managed ACTIVEPIECES provider, default 2; 1 for BYO), idempotency key {conversationId}:chat:{turnIndex} (CreditUsageSource.CHAT), plus the AppSumo meter on AppSumo plans; it then emits the PostHog chat_message billing event (skipped when the platform has no license key — the Autumn tracking always runs). chat-tool-billing.ts decides which tool calls bill: every mcp__ tool plus a fixed set (ap_web_search, ap_scrape_url, ap_generate_image, ap_execute_action, ap_explore_data, ap_run_code).

Turn liveness — three independent timers (get this right)

A turn is kept alive / reclaimed by three separate mechanisms in execute-agent-run.ts; confusing them causes "chat randomly stops" bugs:

  • Heartbeat (HEARTBEAT_INTERVAL_MS 15s): a setInterval that bumps conversation.updated (via heartbeatAgentConversation RPC) + sends an empty keepalive chunk, so a live-but-slow turn is never reclaimed as stale.
  • DB stale-recovery (STREAMING_STALENESS_TIMEOUT_MS 90s, agent-helpers.ts): on-read (getConversationOrThrow) + a per-minute sweep flip any STREAMING conversation whose updated is >90s old back to IDLE. The heartbeat is what holds this off.
  • Stream idle watchdog (STREAM_IDLE_TIMEOUT_MS 90s, in streamChunksToClient): aborts the turn if the drain-stream reader is silent 90s. It must be SUSPENDED while legitimate silent work is in flight — pending tool calls AND in-flight reasoning (reasoning-start→reasoning-end). Reasoning-awareness was missing and caused the bug where long "thinking" on the Expert tier randomly aborted a healthy turn (a >90s gap between reasoning deltas looked like a wedge). Backstop for a genuine mid-reasoning wedge is MAX_TURN_WALL_CLOCK_MS (20 min).

Gotchas

  • A piece's declared classification beats its action name. agentToolClassification still guesses read or write from words in the action name, but a piece that declares READ, SEARCH, WRITE or DESTRUCTIVE wins: mark_as_read and add_member_to_list are writes. The API refuses a declared write in ap_explore_data (runAgentAction, readOnlyOnly), and the worker's approval card asks the API through __action_read_only before deciding. If that lookup fails, the gate falls back to the name rule. custom_api_call keeps following its HTTP method (AI-106).

  • Compaction is not metered. getAgentConfig builds the compaction model with aiUtils.createModel and no billing, so the summarizing call on the main model never reaches credits or usage reporting. Count it by hand when measuring cost per turn.

  • Bedrock ignores our cache and thinking options. agentProviderOptions sends anthropic.cacheControl and anthropic.thinking for BEDROCK, but @ai-sdk/amazon-bedrock reads only bedrock/amazonBedrock keys (cachePoint, reasoningConfig). Bedrock chat therefore runs with no prompt caching and default thinking.

  • Tiers only resolve on Anthropic-shaped providers. PublishedTier.nativeModelId is a single Anthropic id. On OpenAI and Google both fast and smart fall to the first curated model; on Azure and Bedrock the fast model is passed fallbackModelId: resolvedModelId, so it equals the main model.

  • Managed chat has no system-prompt cache breakpoint. buildSystemPromptWithCaching sets one only for ANTHROPIC/BEDROCK; managed/OpenRouter rely on automatic caching at the end of the request. So any change to the tool list or thinking settings (the discovery→build flip swaps activeTools) rewrites the whole ~22k-token prefix. @openrouter/ai-sdk-provider does honor cacheControl on system messages.

  • Refusals aren't handled. Nothing branches on content-filter. A refused step with no visible output hits decideLoopAction → continue_empty and re-sends the full context with "Continue".

  • The per-user tail breaks the prompt cache. The date, memory, connection inventory and {{PROJECT_LIST}} sit inside the same single system block as the ~22k-token static prompt. Any change re-writes the whole cached prefix, and the prompt is effectively unique per user.

  • A step's response.messages holds only that step's messages (AI SDK v7). The older SDK made it cumulative, so collectStepMessages once took the last step; under v7 that saved only a turn's closing text and dropped every tool call and result. The next turn then saw replies like "New agent is updated" with no tool call behind them, and repeated the pattern by claiming changes it never made. It now joins every step, and run-agent-turn-history.test.ts runs a real streamText turn so a future SDK change to these semantics fails loudly.

  • Chat has no fixed step cap. A chat turn is bounded by credits (checked after every step), the runaway-token ceiling (RUNAWAY_TURN_CONTEXT_MULTIPLE × the context window), the identical-failure guard and the 2-hour MAX_AGENT_TURN_WALL_CLOCK_MS, so it finishes long jobs. A saved agent's author-set maxSteps is still honored on its last step: an agent with structured output is forced to call TASK_COMPLETION_TOOL_NAME so the flow gets a result, and any other agent runs with toolChoice: 'none' plus a wrap-up note, so it replies with what's done and what's left. Never a silent stop or an error.

  • A turn limit only works as a stop condition. One streamText call runs every step of a turn, so a check after it returns can't stop a runaway. The token ceiling used to be checked only there, and a turn on 24 Oct 2026 ran 181 steps and 50.8M input tokens (AI-86). Put every per-turn bound in loopStopCondition. Test mocks must report usage in the V3 shape (inputTokens: { total, ... }); flat numbers read as 0 tokens.

  • The streaming lock is per run, not per conversation. A re-send can claim activeRunId before the first job reaches getAgentConfig. agentHelpers.acquireStreamingLock only lets the owning run set STREAMING; an older run gets AGENT_RUN_SUPERSEDED and the worker exits quietly, with no ERROR save and no events. A lock that ignored runId let the old run stream unowned while the new one was rejected as busy and marked ERROR, which is how users saw nothing after tapping a starter card twice.

  • A redelivered run is a duplicate, not busy. A worker socket blip makes the app requeue a chat job whose worker is still running it (releaseConnectionJobs). The second delivery finds its own activeRunId STREAMING, so acquireStreamingLock returns duplicate and it exits quietly like a superseded run. When this came back as busy it saved ERROR, and users saw "An agent is already running for this conversation" in the middle of a working turn (prod, 2026-10-13). If the first worker really died, the 90s CHAT_STALE_SWEEP frees the conversation. This only works when the conversation carries the run's activeRunId: chat claims it in the controller, and a flow-step run claims it when getAgentConfig creates the conversation. Without that claim the redelivery read as busy and failed the job (prod, 2026-10-21). A copy that arrives after its run already finished finds the row IDLE, so the lock alone would let the agent run again and its tools act twice (AI-153). acquireStreamingLock therefore also sets a Redis marker per conversation and run (agent-run-started:<conversationId>:<runId>, 24h) with putIfAbsent, and any later delivery of that run is a duplicate. That's safe because every enqueue mints a fresh runId and agent jobs run with attempts: 1. If the first worker really died, the redelivery is still dropped, so CHAT_STALE_SWEEP fails the step itself: a flow-step conversation keeps its waitpointId, and the sweep resumes that waitpoint with a failure result, so the flow doesn't sit paused until the 3h step timeout. It never re-runs the agent, because its tools may already have acted.

  • A flow-step run that loses its getAgentConfig response exits quietly, and the redelivered copy finishes it. A socket drop mid-RPC surfaces on the worker as RpcTimeoutError.disconnected. The API handler doesn't stop when the socket drops: it still takes the streaming lock, so the copy used to get duplicate and the step failed, either right away as "provider temporarily unavailable" or 90s later through CHAT_STALE_SWEEP. Now the worker exits quietly on a disconnect, and getAgentConfig in worker-rpc-service.ts calls agentHelpers.handLockToRedelivery when the worker socket is gone by the time it replies. That writes a one-time agent-run-handed-off token (60s), and the next delivery of that run consumes it in acquireStreamingLock instead of being dropped ('handed-off'), so the copy runs the step (AI-156). The first delivery already saved the user message, so a handed-off copy re-reads the history after the lock and does not append it again. The hand-off never touches the row: setting it IDLE would hide the run from CHAT_STALE_SWEEP when the copy was already turned away, and the flow would sit paused for 3h. The token's TTL is shorter than the sweep window, so a late copy can't revive a step the sweep already failed. A plain timeout on a live socket is never redelivered, so it still fails the step. Chat still reports the error, because its buffered emits would mark ERROR over a copy that is streaming.

  • A failed turn used to leave status ERROR with an empty reply. The error text only travelled as a websocket ERROR event, so a reload, another tab, or a user who was away saw a dead conversation (and often no user message either when the job died inside getAgentConfig, before it persists the turn). The execute-agent-run.ts catch now sends failure: { message, userMessage? } on the empty saveAgentMessages call and conversation-rpc.ts appends it to uiMessages as an assistant text reply; use-chat.ts shows the Retry banner when a loaded conversation is ERROR. Stale STREAMING to IDLE recovery, a worker that never picks the job up, and assertAgentMessageRateLimitNotExceeded still leave no reply of their own.

  • A built flow ends with a "Turn it on?" card, never an auto-publish: an audit found 40% of built-but-unpublished flows ended on "open it to review", some claiming LIVE. Chat never publishes on its own; only a yes to the card calls ap_lock_and_publish (build_flow.md "Turn it on?"). Pinned by the build-then-ask-turn-it-on fixture (askedToTurnItOn, noLiveClaimWithoutPublish). The live-claim check is a regex plus "was ap_lock_and_publish called", so it cannot see whether publish succeeded.

  • Server-managed connections: the LLM never sees connection externalIds; ap_execute_action auto-fills them from a Redis store.

  • Prompt-injection taint: a per-turn taintState flips to tainted once the turn reads outside or user data (MCP tools, web/search/scrape, ap_explore_data, ap_execute_action, ap_run_code, ap_list_across_projects, configured tools). It forces the action-preview gate on non-read-only actions, makes saved-agent edits (ap_update_agent, ap_add_agent_tool, ap_remove_agent_tool, ap_create_agent) ask for approval on the action card instead of refusing, and makes ap_remember ask before saving (memory reaches every future conversation). It carries one message: each saved reply stores tainted (this reply's own reads, via createTaintState), and the next message starts from it. Deletes are confirmed always, not by taint.

  • Write-check gate: before a live ap_test_flow, __flow_write_check RPC flags write/destructive PIECE steps; read-only flows run ungated; if the check itself errors, the approval card still shows (fails closed).

  • Chat follows the plan on Cloud and Enterprise. chatVisibility returns plan.chatEnabled on both, and Community and embedded sessions never see chat. Chat reaches a Cloud user only through CHAT_ENABLED on their Autumn plan, which toPlatformPlanFlags writes to chatEnabled; the code has no Cloud override.

  • Flow correctness is 100% prompt/guide-driven — nothing in code enforces it. The "#1 silent bug" ("Class A"): the agent frames a recurring automation as a one-time task and omits any anti-reprocessing step, so run N+1 redoes run N's work (re-pays, re-sends). It's a design-time reasoning gap, not a testing gap — ap_test_flow runs ONCE, so a single test looks perfect; the bug only shows on the 2nd run. Fix lives in the prompt (chat-system-prompt.md <decision_framework> + build_flow.md "Recurring flows must not reprocess") + capability eval fixtures with a recurring_avoids_reprocessing judge dimension. The platform already has every primitive (Tables New-Record webhook, polling DedupeStrategy, _dedupe_key, Store, update/delete-record); the agent just wasn't reaching for them. Watch the build_flow.md "don't over-build" bias — it once actively discouraged the fix.

  • Compaction runs inside the 60s getAgentConfig RPC, so it must stay bounded. The summarizer call is capped at SUMMARY_OUTPUT_RESERVE_TOKENS and aborted after COMPACTION_TIMEOUT_MS (35s). On failure the previous summary is kept and buildCompactedPayload still fits the history to the budget, flagging trimmed messages to the model. An unbounded summary once ran for 2.5 min and failed the turn as RPC [getAgentConfig] failed (timeout: 60000ms).

  • The eval replay harness drifts silently when a worker tool's signature changes. buildEvalToolSet (agent-eval/core/runner.ts) hand-builds the tool set with stubs, and nothing type-checks test/lib/agent-eval/ in CI, so a new required param never fails a build. Until Oct 2026 the eval passed no taintState to createCrossProjectTools (every data tool threw, transcripts showed TOOL_RESULT: null, and fixtures like the injection one never saw their data), onSetProjectContext returned void, and approvals lacked outcome so all counted as declined. When you change a tool factory's params, update the eval stubs too, and treat TOOL_RESULT: null in an eval transcript as a harness bug, not agent behaviour. Check with npx tsc -p packages/server/worker/test/lib/agent-eval/cli/tsconfig.json --noEmit.

  • An eval verdict is a majority over repeats, graded by a different model. LLM output is random, so one run proves little (two full runs of the same commit 13 seconds apart, Oct 2026, scored 55/78 vs 54/78 repeats but disagreed on 7 of 26 fixtures): CHAT_EVAL_REPEATS (nightly 3) runs each fixture N times; judge dimensions pass on a majority, but a deterministic assertion must hold on every run (one run that emails the attacker is a real failure, not noise). The report shows passes/runs and the first failing run's checks; runVerdicts in the JSON keeps every run. The judge is CHAT_EVAL_JUDGE_MODEL (default an Opus slug), never the agent's own model, which went easy on itself. The judge's transcript carries each tool call's input and result, not just its name, so it can see where an email went. The report's "expected-label match" only compares the judge to each fixture's expectedLabel, so an agent miss counts as a judge error; judge accuracy is the separate "judge vs human labels" line, from person-reviewed cases in agent-eval/calibration/. A model-written label measures Claude against Claude, so labels carrying "labelledBy": "claude" are excluded from it and only feed a separate "draft estimate" line until a person reviews them. Each run is published (public) to cdn.activepieces.com/ai/evals/runs/. Nightly runs cover gate fixtures only; a Sunday 07:00 UTC run covers all fixtures, and only those full runs feed the Config Console's Agent score (share of repeats passed), whose "+N pts" compares just the fixtures present in both runs so adding a fixture never fakes a gain. Prefer a deterministic assertion (noToolArgMatches, neverCalledTool) over a judge dimension wherever code can see the behaviour.

  • A stopped npm run agent-evals keeps nothing and leaks its key. The CLI writes .agent-eval/last-run.json only after every fixture, and Ctrl-C skips cleanupAuth, so the minted "activepieces-agent-eval (ephemeral)" OpenRouter key ($25 cap) stays alive until deleted by hand. Try a model with --limit 3 or --only <ids> first. The cost it shows is what OpenRouter reported for the agent's own calls, and — means some call reported none; judge cost is a separate line.

  • The context budget subtracts what the request reserves. Anthropic/OpenRouter count tool schemas and the reserved max_tokens against the 200k window, so getAgentConfig passes reservedTokens (the clamped output window from agentAiUtils.affordableOutputTokens, the same helper the worker uses for maxOutputTokens, plus TOOL_SCHEMA_TOKEN_ESTIMATE = 12_000) into agentCompaction. The threshold, the fit check and the recent window are all measured against maxContext - reservedTokens, and the recent window is sized by estimated tokens, not message count. The tool-schema figure is a constant because the tools are built in the worker and the API never sees their definitions. The worker sets maxOutputTokens per step in prepareStep, so a thinking-disabled step (round one, discovery phase) does not reserve thinkingBudget. The system prompt's size is added to the reserve for the window and fit check, and each message is capped at 20k chars in the summary request so the summarizer itself cannot overflow.

  • ap_set_phase and ap_load_guide are gone. Skills replaced both: ap_load_skill returns the guide and its tools' schemas, and the build phase flips on its own. BUILD_ONLY_TOOL_NAMES now only decides when thinking turns on, and only hides tools on surfaces outside skills mode. The web app keeps both names so old conversations still render.

  • Read-only checks don't belong in BUILD_ONLY_TOOL_NAMES. ap_validate_flow was in it, so "check that this solution fits together" never saw the tool. The agent inspected flows by hand instead and misread a Call Flow flowId (an externalId) as a mismatch with the flow's id. It now stays visible in discovery, like ap_flow_structure.

  • A dynamic property only runs with a saved schema. The engine processes a DYNAMIC prop's sub-fields (JSON parse, numbers, arrays) only from propertySettings[prop].schema, which the builder saves when it resolves the prop. Chat tools used to save propertySettings: {}, so an advanced-mode Call Flow payload reached the subflow as one string and a live run saved empty fields. mcpUtils.resolveDynamicPropertySettings now resolves and saves it on build, add and update.

  • Focused goals run as task sub-agents. ap_run_task (task-subagent-tools.ts) runs a second runAgentTurn with a fresh context and the main's tools, minus anything that talks to the user, nests, tests, publishes or flips flow status. The prompt is task-subagent-prompt.md. Tasks start in the build phase with their own ap_set_phase: any tool whose closure holds run state must be rebuilt per task, because reusing the main's copy flipped the main's phase. A task must end with updateTaskStatus. If it's cut off it continues (up to 2 times), and if it forgets to report it gets one forced report step, so failed is rare and the main retries it. Histories live in agent_task, so taskId resumes the same sub-agent. Approvals inside a task carry taskTitle (via AsyncLocalStorage). Flat-billed calls come back as billedToolCalls, and toModelOutput keeps the card data out of the main's context. Delegation happens only when it buys something: multi-flow solutions build in dependency waves, one task per flow; single flows stay with the main.

  • Research goes through ap_deep_research, not prompt rules. Prompt rules alone got research delegated in 3 of 8 runs, because the model reaches for the quick ap_web_search. A dedicated tool whose name matches the intent fixed it. It runs one task per subject under progress ids toolCallId:index, and the web expands it into one row per subject. Its timeline (statuses, searches, page reads) streams into the right-side task panel (task-panel.tsx), which opens on its own on the first search.

  • Capability notes are built where discoveryOnly is not known, so a prompt can promise tools the worker has stripped. getAgentConfig composes the system prompt in the api, while discoveryOnly rides on the job data; before Aug 2026 it never crossed that boundary. Meanwhile the worker strips image tools, email tools and the agent tools on such a run, so the notes claimed all three. Two of the three had been wrong since long before anyone noticed, because each note computed its own availability term. The flag now travels with the config request and the three notes read one shared actingRun = !dryRun && !discoveryOnly. Whenever a tool group is gated on a run mode in the worker, the note that advertises it has to be gated on the same term, in one place.

  • Tool output can't be trusted to waive a charge. A user's own MCP server or ap_run_code controls its output's top-level keys, and they reach the persisted part unchanged. Chat billing only honours billedAtCost from ap_web_search and ap_generate_image; a new built-in tool that bills at cost must be added to TOOLS_THAT_BILL_AT_COST in chatBilling.

  • Credits are checked after every step, not just before the turn. runAgentTurn asks the API (agentHasCredits) after each tool step, sending what the turn has used so far (chatBilling.creditsForTurn, the same rule end-of-turn billing uses), and stops once the remaining balance no longer covers it. Flat credits are only charged when the turn ends, so without that pending count the balance never moves mid-turn. The check fails open if the RPC errors.

  • Chat files belong to a conversation, not just a project. Conversations are private to their user, so a project check is not enough. saveAgentFile and persistAgentAttachments stamp conversationId into the file metadata, and readAgentFile refuses a file whose stamp doesn't match. Files saved before the stamp existed can't be read back (image editing can't use them). Any tool that reads chat files by a model-supplied ID must go through readConversationFile (agent-file-utils.ts), and any new save that hands the model a file ID must stamp conversationId.

  • Local dev needs AP_DB_TYPE=POSTGRES + Redis; refuses PGLite. Prefer AP_EDITION=cloud over ee for chat work. Cloud boots locally against plain Postgres and Redis with no Autumn, Stripe or license-key config (verified Aug 2026: API healthy, migrations applied, zero billing or license errors), and on Cloud chat follows plan.chatEnabled, so set it on the local platform's plan. On ee it is gated behind plan.chatEnabled and you have to get a plan onto the platform first. Note SMTP is usually unset locally, which makes the auth card open on the password form rather than the email-code step. Debug a run with npm run chat:logs -- <conversationId> [runId] (needs LOG_FILE=true/AP_LOG_FILE=true set when the turn ran — otherwise .evlog/logs is empty).

  • Chat was renamed to agent in code and DB, but only the storage half. As of release 0.87.1 (1823000000000-AddRenamedChatTableCompatViews) chat_conversation → agent_conversation and user_chat_memory → user_memory, the server module moved ee/chat/ → ee/agent/ (entry point agentModule), the worker dir moved jobs/ee/chat/ → jobs/ee/agent/, shared types moved core/shared/.../ee/chat/ → .../ee/agent/, and server/utils/src/chat-ai-utils.ts → agent-ai-utils.ts. The rename is not uniform, and the split is the thing to learn: files describing chat as a user-facing surface deliberately kept their chat- names inside ee/agent/ — chat-visibility.ts, chat-analytics-sync.ts, chat-tool-billing.ts, chat-usage-tracker.ts, chat-plan-grant.ts. So a new chat-surface concern keeps the chat- prefix; a new stored entity takes agent_. The migration also leaves CREATE OR REPLACE VIEW compat views at both old table names, so raw SQL against chat_conversation still reads fine and will NOT tell you the rename happened — grep the entity, not the database.

  • An AI SDK major bump can typecheck clean while a callback payload silently changed shape. v7 keeps most v6 option names as working deprecated aliases (system, onStepFinish, experimental_repairToolCall, stepCountIs, result.toUIMessageStream), so the option compiles but the data underneath can differ: experimental_onToolCallFinish survived as an alias for onToolExecutionEnd while its event lost durationMs/success/error (now toolExecutionMs plus a toolOutput.type === 'tool-result' discriminator). A type-probe that only names the option passes; you have to exercise each callback's property access. onStepEnd's content is also cast to a structural ContentPartLike with an args ?? input fallback (agent-ai-utils.ts), which means a shape change there fails at runtime, not compile time — always smoke a real turn after a provider/SDK major.

  • ai and evlog are version-coupled. evlog ≤2.18.1 imports TelemetryIntegration from ai, which v7 renamed to Telemetry, so bumping ai to 7 without bumping evlog (≥2.22.4, which peers ai >=6.0.168 <8.0.0 and supports both v6 and v7 hooks) will not compile. That evlog bump in turn changes DefinedAuditAction from <TargetType> to <Action, Options> and breaks helper/audit-events.ts — drop the explicit annotation and let defineAuditAction's inference supply it.

  • AI SDK v7 is ESM-only, and that is NOT a reason to convert the server to ESM. ai@7 ships type: module with no require condition, but the CJS server consumes it fine through Node's require(esm) (Node 22.12+/24, verified), and TS 5.5.4 resolves its types under module: CommonJS + moduleResolution: node because a root main and an adjacent index.d.ts still exist and skipLibCheck is on. No ESM migration, no TypeScript upgrade. Mixed ai majors across workspaces are also safe and intentional — bunfig.toml sets linker = "isolated", so pieces/framework/engine can stay on v6 while the agent path runs v7.

  • ap_show_connection_required is an alias of ap_show_connection_picker, not a smaller capability. Both names resolve to the same ConnectionPickerCard, which lists every account the caller has for that piece and offers "Use a different account"; the only schema difference is an optional status: 'missing' | 'error' hint. So an allow-list that grants one name and asserts the other is absent proves nothing: verified live on the agent surface, granting only ap_show_connection_required renders "Which account should I use?". The card also fetches the account list itself from the frontend, keyed by conversationId, so the tool payload cannot constrain what it offers. A repair-only variant therefore lives in the endpoint feeding the card, not in the tool set.

  • On a saved-agent run, choosing a different account in the connection card does nothing. onConnectionSelected writes into selectedConnectionByPiece (execute-agent-run.ts), which is read only through getSelectedAuth, passed only to the MCP tool set — and AgentRunSource.AGENT is not granted groups.mcp at all. Configured piece tools carry the agent's stored pieceMetadata auth instead. So the card reports the account as connected while the tool keeps calling on the pinned one. Only the in-place Reconnect actually repairs an agent run, because it re-authorizes the same connection row the agent is pinned to. So /v1/agents/conversations/:id/connections answers { connections, reconnectOnly }, and for an AGENT-source conversation returns only the accounts that agent's tools pin. Three things that branch has to get right, each of which was a live bug first: match on (projectId, externalId), because externalId is caller-supplied and its index is not unique, so a same-id row in another project can pose as the pinned one; read the pin through published ?? draft, the same as the run; and unwrap the pre-0.87 {{connections['id']}} template form, or an older agent reads as having no pinned account and the card tells the user their live account is gone. The card must also carry the row's own projectId into the reconnect dialog, which otherwise falls back to the session project and repairs the wrong one.

  • A task agent's cost is steps × context, and a cache read isn't free. A cache read still costs 10% of the normal input price, so a 30-step task over a 30k context costs about $0.60 on Sonnet even with 92% cache hits. Three things broke task cost (measured 2026-10-11): a status-only step every other step (tasks get no ap_update_thinking_status; the card falls back to the latest tool title), a report turn that swapped the tool set (any change to the tools, or activeTools, re-writes the whole cached context: 25–44k at write price), and researchers on Sonnet with full 128KB pages. After the fixes a researcher costs 3–5¢ on Haiku. Builds are the opposite: a task per flow costs $0.21–0.38 because it repeats the chat's setup, while the main chat builds four flows with parallel ap_build_flow calls for about $0.51, so builds stay in the main chat.

  • The chat's internal MCP token lives as long as a turn can (2 hours), not 15 minutes. It's issued once per turn by issueChatTurnAccessToken (called from agentMcp.getCredentials) and kept in worker memory. issueInternalAccessToken, which the /mcp-server/token route hands to callers, keeps 15 minutes. With the 15-minute outside-client lifetime, any turn that waited on a card long enough lost every ap_* tool with "authentication failed" and stopped half done. Outside MCP clients keep their 15-minute tokens.

Key files

Entry point: agentModule, the Fastify plugin registered in packages/server/api/src/app/app.ts.

  • packages/server/api/src/app/ee/agent/ — the API module: controllers, service, helpers, approval gate, compaction, console sync, billing (chat-usage-tracker.ts, chat-tool-billing.ts), memory (agent-memory-ai.ts, user-memory-entity.ts), entities, plus tools/, mcp/, prompt/, history/ subdirs
  • packages/server/worker/src/lib/execute/jobs/ee/agent/ — where the LLM loop actually runs: execute-agent-run.ts job handler (+ the three liveness timers + streamChunksToClient idle watchdog), run-agent-turn.ts DI streaming loop, agent-worker-tools.ts tool defs
  • packages/server/utils/src/ai-utils.ts — the provider-agnostic half: createModel per provider, createEmbeddingModel/toStorageEmbedding, supportsWebSearch/buildWebSearchTools
  • packages/server/utils/src/agent-ai-utils.ts — what is genuinely agent-shaped: collapseStaleToolOutputs history hygiene, the onStepEnd content handling
  • packages/core/shared/src/lib/ee/agent/ — shared zod schemas and types, tool-phases.ts gating, tool-classification.ts, chat-visibility.ts
  • packages/server/api/src/assets/prompts/ — system prompt + project-context markdown and the on-demand guides/; agent-eval fixtures live in packages/server/worker/test/lib/agent-eval/
  • packages/web/src/app/routes/chat-with-ai/ — the chat page, chat box, conversation list, and components/ cards
  • packages/web/src/features/chat/ — API client, Zustand store, use-chat.ts, chunk-reducer.ts, streaming and voice hooks

Paths verified 2026-08-19 against main. An earlier version pointed at ee/chat/chat-model-factory.ts and ee/chat/chat-history-hygiene.ts; both were folded into packages/server/utils/src/agent-ai-utils.ts. Every ee/chat/ path on this page before that date is dead — see the chat-to-agent rename gotcha above.

  • setConversationId is a reload, not a setter. It calls stopStream(), resets the interaction stores and refetches history, so handing it the id of a conversation the hook is already in destroys the turn in flight. AIChatBox seeds it from the conversationId prop in an effect, which makes the obvious wiring — feed onConversationCreated back into that prop — kill the very turn that created the conversation: the pane goes blank while the reply completes fine on the server. It now early-returns when the id is unchanged, so re-seeding is a no-op, but the shape is worth knowing before adding another caller.