Agent-chat ships its first non-prompt verb

Why care

The chat surface has been mechanically working for weeks — Claude Sonnet 4.6 wired via Anthropic SDK in services/prompt-runner/src/chat-turn.ts:21, four-cache-eligible system slabs assembled in services/workspace/src/chat.ts, three response modes locked behind tool_choice: { type: 'any' }, accept- proposal flow rendered in the Svelte transcript. Nothing about the LLM gateway or the chat surface needed building.

What kept it from being useful was a deliberate v0.0.1 scoping decision: the verb roster was hand-written and covered only the three prompt verbs (prompt.draft, prompt.improve, prompt.apply), and the context slab knew only about record_set_id. For the operator who's been spending the last few days inside the Content Reader doing per-funder corpus work, the chat had no idea anything else was happening in the product. The "blank" feeling wasn't about a broken connection — it was about a narrow capability roster meeting a wide actual workflow.

This commit closes the gap with the smallest meaningful proof: ship one new verb that exercises one new context dimension. The verb is /inbox <url> (Vector 2a from yesterday's [[Corpus-Inbox-Capture-and-Triage]] spec). The context dimension is active_client_id. Both are now real. The chat finally has something to do with the funder workflow.

What's new

Backend. services/content-ingest/src/corpus.ts gains addToInbox() + buildInboxFrontmatter() as siblings of the existing addToCorpus() + buildFrontmatter() — same collision- suffix logic, same slug, same YAML quoting, but writes the inbox- extended frontmatter shape per the spec. funder_slug: "inbox", pack_id: "inbox", record_id: null, response_id: null, plus the two new sibling blocks: captured_* (populated at write, immutable) and triaged_* (all null until the future triage layer runs).

services/content-ingest/src/handlers.ts registers the corpus.inbox.add.requested NATS handler. Validates URL parses and is http(s), Jina-fetches (reusing the same per-URL cache the existing manual-add path uses), writes via addToInbox, publishes corpus.inbox.added for downstream subscribers. Gracefully handles Jina fetch failure by writing the file with a stub body and extra_metadata.jina_status: "fetch_failed" — the URL itself is never lost.

Workspace router. services/workspace/src/capabilities.ts wires corpus.inbox.add → corpus.inbox.add.requested with a 30s timeout (single Jina fetch + filesystem write).

Chat verb roster. services/workspace/src/chat.ts:50 V001_CHAT_VERBS gains the corpus.inbox.add entry plus a new VERB RECOGNITION SHORTCUTS block teaching Claude to read /inbox <url> as an explicit verb signal (route via chat_invoke rather than the strict-alignment default chat_propose). Same block covers "save this", "park this", "inbox this", "remember this URL" — natural-language shortcuts the operator can use without the slash prefix.

The chat_propose and chat_invoke tool definitions' enum on capability is widened to include corpus.inbox.add. Without this the model couldn't pick the verb even if it wanted to.

Active-client context. contextSlab() inlines an active_client line:

"The active client is: reach-edu (this is the client_id arg for corpus.inbox.add and any other client-scoped capability)."

This is hardcoded for v0.0.1 — reach-edu is the only client today and this is the line the architecture spec is targeting for proper broker resolution in the next pass. The hardcode is the working seam, not the destination.

Chat-side render. apps/chat/src/ResponseModeRenderer.svelte gains a capability- specific result branch for corpus.inbox.add: instead of the generic "✓ corpus.inbox.add completed" muted line, the bubble shows "✓ Saved to inbox" with the resulting corpus_path rendered in monospace below. CSS adds .bubble.result.inbox and .inbox-path styles.

How it works

End-to-end:

1. Operator types /inbox https://example.com/path in the chat composer and hits cmd-enter.

2. Chat surface's chatTurn() sends the message + thread + context to workspace via the WebSocket.

3. Workspace's dispatchChatTurn() assembles the six-slab system prompt (now including active_client_id: reach-edu in the context slab) + the four chat tool definitions + the thread.

4. Workspace publishes chat.turn.requested to NATS.

5. Prompt-runner's chat-turn handler calls Claude Sonnet 4.6 with tool_choice: { type: 'any' } — the model must pick one of the four tool calls.

6. Claude recognizes the /inbox shortcut + the prefer_invoke_when discipline and returns chat_invoke { capability: 'corpus.inbox.add', args: { client_id: 'reach-edu', url: 'https://example.com/path', captured_from: 'chat-verb' } }.

7. Chat surface receives the invoke turn and immediately runs the tool call via workspace.invoke('corpus.inbox.add', args).

8. Workspace dispatches corpus.inbox.add.requested to NATS.

9. Content-ingest's handler Jina-fetches the URL (reusing the same cache the manual-paste path uses), writes the inbox markdown, publishes corpus.inbox.added.

10. Chat surface gets the capability_result back and renders the inbox-specific bubble — "✓ Saved to inbox" with the corpus_path.

Result file lands at clients/reach-edu/corpus/inbox/2026-06-08_<title-slug>.md with the full extended frontmatter (funder_slug: "inbox", inbox_status: "pending", captured_from: "chat-verb", captured_at: <ISO>, triaged_*: null) and the Jina-fetched markdown body.

What's still loose

• Verb roster is still hand-written. The V001_CHAT_VERBS prose block grew today; each new verb is two edits (prose + enum). Layer 3 of the [[Chat-Context-Awareness-Architecture]] spec is the cleanup — derive the roster from a CHAT_VERBS[] descriptor list with per- verb args schemas + recognition_hints + context_required. Ship the next time a non-trivial verb is added (likely corpus.inbox.list or one of the pack verbs).

• active_client_id is hardcoded reach-edu in contextSlab(). Works because there's one client today; fragile because there will be more. Layer 1 of the architecture spec is the real fix — workspace becomes the source of truth for active client (single top-level field), each microfrontend pushes its state, the slab assembly reads from the broker.

• ActiveView still covers four kinds. Content Reader, Inbox, Pack Runner, Request Reviewer, Response Reviewer, Records Surface, Prompt Template Manager — none of them push their state into the workspace. The chat doesn't know which surface the operator is on; can't tailor proposals to the surface. Same Layer 1 fix; the spec sketches each variant.

• Conversational paste (Vector 2b) NOT shipped. Pasting a bare URL without /inbox doesn't trigger anything yet — /inbox <url> is required. Vector 2b needs the intent classifier (corpus.inbox.classify_intent, Haiku-class) + the inline confirmation card + the per-host snooze heuristic per [[Corpus-Inbox-Capture-and-Triage]] §Vector 2b. Lands when the conversational rhythm is wanted.

• Inbox doesn't have a Browse view yet. Captured items are visible only as files in the per-client repo's filesystem. apps/corpus-inbox/ microfrontend with the Browse mode is step 5 of the inbox spec's migration plan.

• No status-bar badge for inbox count. "N items pending triage" nudge isn't wired; trivial to add once the corpus.inbox.list capability lands.

• Renderer didn't get a "browse inbox" action button. The result bubble shows the path but doesn't offer "open this in the Inbox surface" because the surface doesn't exist yet. Wire it when the surface ships.