connect-aidream — outbound calls into the AI Dream backend

The matrx-extend extension talks to aidream over a single SSE-streamed chat endpoint. The contract that defines what tools the LLM gets to see, and what state it gets to read, is the capability envelope. This skill is the entry point when you're working in this repo and need to extend or debug what flows out to aidream.

When to use

Adding a new key to client.state["browser-dom"] (e.g. a new flag the server needs to gate tool advertisement on).
Wiring a new tool category that the model can call load_chrome_tools with (extension side: handler + alias; server side: routing).
Debugging "the model never sees my tool" — usually a discovery / merge bug, not a handler bug.
Reading or extending the RESOURCE_CHANGED kind=active_tools event family that the extension subscribes to for live tool-set updates.

When NOT to use

Changes to aidream's Python source. Switch to that repo and use its connect-matrx-extend skill.
Changes to a tool handler (the code that actually runs in the service worker when a tool fires). That's local to this repo and lives in src/lib/tools/handlers/.
Changes to how the extension renders a tool in the chat surface — that's the matrx-extend-tool-display skill.

30-second mental model

extension → POST /ai/agent/{agent_id} ─┐
  client.capabilities = ["browser-dom"]│
  client.state["browser-dom"] = {…}    │
                                       ▼
                                aidream apply_unified_tools
                                  (tool_merge.py)
                                       │
                                advertises load_chrome_tools to LLM
                                       │
                              LLM calls load_chrome_tools(category="forms")
                                       │
                                browser_discovery.py reads state,
                                queues tool changes (add/remove)
                                       │
                                  RESOURCE_CHANGED kind=active_tools
                                       ▼
                              extension updates Tools-tab badge,
                              records loaded category

Quick start

When you need the server to gate a tool category on a new flag:

Extension side — add the flag to the browser-dom state bundle that ships with every chat request. The bundle is the small orchestration object (~12 keys today) sent under client.state["browser-dom"].
Server side — extend browser_discovery.py to read the new flag and filter the category-routing table. (This is server work; switch repos to do it.)
Verify — admin → Tools tab → trigger load_chrome_tools for the gated category; confirm the tools appear iff the flag is true.

The server-to-extension event you watch for is RESOURCE_CHANGED kind=active_tools. The extension already subscribes to this and updates useActiveToolsStore so the next request hints the loaded categories back. No new event plumbing needed.

Client-tool resume contract (load-bearing — read before touching tool results)

The aidream backend HARD-SUSPENDS the loop when a client-delegated tool is pending: it persists the turn, emits a complete phase, and ends the SSE stream. The originating stream is GONE the moment the model delegates a client tool. To continue, the extension must:

POST the tool result through postToolResults (src/lib/api/routes/tool-results.ts) — the ONE funnel.
Read the response. When data.continuation_needed === true, broadcast CHANNELS.STREAM_CONTINUE with {conversationId, userRequestId}. The SW dispatcher already does this in dispatch.ts::postResult — do NOT add a second tool-result POST path that bypasses it.
The sidepanel's useChatStream subscribes to STREAM_CONTINUE and runs resumeRun(conversationId, userRequestId): pushes a fresh assistant bubble, allocates a new runId, rebuilds the client.state["browser-dom"] envelope against the current active tab, and STREAM_STARTs against conversationResumePath(conversationId). The existing STREAM_CHUNK consumer routes the continuation chunks into the new bubble.

Invariant. Never wait on the original stream after a tool_delegated event. It has ended. Adding a "listen for more events on the closed stream" pattern reintroduces the entire bug class this protocol exists to kill.

Disambiguation — src/lib/stream/resume.ts::attemptResume is NOT this. That file is the SEPARATE, still-unbuilt stall-recovery cursor-replay feature (keyed by request_id + a cursor of events seen; triggered by the watchdog on a stall). It points at GET /ai/agent/runs/{request_id}/resume?cursor= which the backend has not built. Do not conflate. The canonical doc that covers both, with the differences laid out, is matrx-frontend/features/agents/docs/CLIENT_TOOL_SUSPEND_RESUME.md.

Wire shape

SW → server:  POST /ai/conversations/{id}/tool_results
              body: { results: [{call_id, tool_name, output, is_error, error_message}] }

server → SW:  200 { resolved, already_resolved, not_found,
                     continuation_needed: bool,
                     user_request_id: string|null,
                     conversation_id: string }

(if continuation_needed === true)

SW → sidepanel (broadcast):  CHANNELS.STREAM_CONTINUE
                              { conversationId, userRequestId }

sidepanel → server:  POST /ai/conversations/{id}/resume
                     body: { user_request_id, client: { capabilities: [...], state: {...} } }
                     → 200 NDJSON stream of the continuation
                     → 409 outstanding_delegated_calls (more answers still pending)
                     → 404 conversation not found

The server reconstructs the conversation from the DB on every /resume call; the answer the SW just POSTed is already embedded — do NOT include the tool result in the resume body.

File index (extension side)

File	Role
`src/lib/chat/context/v2-bundled.ts`	The canonical context shape; `browser-dom` state lives in the orchestration bundle attached to every chat request
`src/lib/tools/aliases.ts`	Wire-format aliasing — strips `matrx-extend__` and bundle prefixes (e.g. `forms__fill_form`) before dispatch
`src/lib/tools/registry.ts`	`lookup(name)`, `assistantToolNames`, `pilotToolNames` — local handler library
`src/lib/tools/dispatch.ts`	SW dispatcher; subscribes to `STREAM_OPENED` + `STREAM_CHUNK`, runs handlers, posts results
`src/lib/tools/handlers/*.ts`	One file per tool domain; signatures and Zod schemas
common-docs/systems/agents/agent-tools/STATE.md	Full wire-format playbook (canonical `:` vs wire `__`)

Server-side reference (read-only from this repo)

aidream/api/utils/tool_merge.py::apply_unified_tools — the unified merge that decides which tools the LLM sees on each turn.
packages/matrx-ai/matrx_ai/capabilities/browser_dom.py — capability definition; the JSON metadata sibling lists every always-on tool.
packages/matrx-ai/matrx_ai/tools/implementations/browser_discovery.py — server-side handler for load_chrome_tools; reads client.state["browser-dom"] and decides the add/remove list.

Extending an existing capability vs. injecting inline

Extend the existing browser-dom capability when the new tool belongs to a category and follows the discovery loop (cheap; reuses the always-on load_chrome_tools).
Inject inline via the unified tool merge (apply_unified_tools) only when the tool is request-specific or shouldn't be discoverable by category. Inline tools cost the full schema in the prompt every turn — use sparingly.

Cross-turn limitation (current state, May 2026)

Tool mutations are per-request only. Each new user message restarts with [load_chrome_tools] — discovery re-runs. Discovery is cheap (server-side lookup, no LLM round-trip), so this is acceptable. Cross-request persistence (cx_conversation.dynamic_tool_state) is the Phase D-persist roadmap item in aidream's TOOL_INJECTION_REFACTOR.md. No extension changes are needed when it lands. Do not work around this limitation in extension code; let the server-side fix arrive.

Failure modes

Silent: tool never advertised. Usually means client.state["browser-dom"] is missing a flag the server gates on, or a category-routing rule excludes it. Check the request payload in DevTools Network tab; confirm state keys; then look at server logs.
Loud: schema rejection at provider. Tool name has : instead of __. The wire format is __; canonical is :. See common-docs/systems/agents/agent-tools/STATE.md §1.
Silent: dispatch falls through to default. Handler isn't keyed on the bare local name. Aliasing in src/lib/tools/aliases.ts must strip both matrx-extend__ and bundle-style <bundle>__ prefixes.

Pointer

For the full topology and the parallel inbound channels, see /Users/armanisadeghi/code/common-docs/systems/apps/extension/CHANNELS.md.

connect-aidream

SKILL.md.claude/skills/connect-aidream/SKILL.md