Workflow Chat → production parity + retire rooms/memories

[sweetmantech]

Tracking issue for bringing the new workflow chat to production parity and decommissioning the legacy substrate, following the /api/chat/workflow → /api/chat cutover (chat#1767, shipped to prod 2026-06-05). The cutover is fully live; this issue captures the capabilities still stubbed/deferred and the legacy teardown. Inherited-gap detail lives in api#605.

Goal

/api/chat (durable Vercel Workflow + per-session sandbox) is the only chat path in prod. Make it feature-complete and production-grade: full streaming control, faithful tool-call UX, restored sandbox tools + artist context, correct telemetry/billing, docs in lockstep — then drop rooms/memories once no reader remains. Done = a user on chat.recoupable.com has everything they had pre-cutover, and the legacy tables are gone.

What shipped (context)

• Cutover live: api.recoupable.com/api/chat streams; /api/chat/workflow → 404. chat + open-agents repointed. (Post-cutover cleanup + follow-up items #1767, api#645/#648, chat#1783/#1784, open-agents#47, docs#235)
• Model-selector billing fixed: chat persists the picked model to chats.model_id before send (chat#1781).
• open-agents inherited gaps partially closed: active_stream_id clear (api#604), close-writable (api#607), persist assistant message (api#609), per-turn credits (api#612).

Done

• docs#229 — align the Sessions OpenAPI spec with the live api.recoupable.com contract.
  ✅ Shipped 2026-06-05 (squash-merged to main, commit 2b683d5). Closes the former "Docs in lockstep" item.
  Renamed the PATCH request schema PatchSessionRequest→PatchSessionBody (matches api's validatePatchSessionBody.ts); added artistId to Session.required typed OAS 3.1 ["string","null"] (was deprecated nullable: true); clarified isNewBranch as always-present on responses; hyperlinked the POST/GET/PATCH endpoints in the Session description and trimmed redundant schema descriptions (KISS, mirroring the CreateSessionRequest sibling).
  Verified: the "required" claims are accurate against api/lib/sessions/toSessionResponse.ts, which always emits isNewBranch and artistId — so the original "marked required but missing from responses" framing was a doc-clarity gap, not an API bug. An earlier patch.mdx frontmatter regression (dropped the OpenAPI file-path prefix → would break the PATCH page render) was caught in review and fixed in 05cd6a18. JSON valid; CI green (Mintlify/CodeRabbit/cubic); merged CLEAN.

Open — Streaming control & resilience

• Stop actually halts the model. (server api#590 ✅ shipped; client chat#1770 pending)
  • Server — ✅ shipped 2026-06-06 (squash → test, then test → main via api#652; both synced). POST /api/chat/{chatId}/stop cancels the run, polls getRun().status every 150ms to trip an AbortController on the in-flight model loop, waits for terminal status (≤8s, else 504 with the slot held), then releases active_stream_id via CAS; runAgentWorkflow gates auto-commit on !result.aborted. Verified on preview: stop returned 200 {stopped:true} sub-second (~0.6–0.8s), the in-flight bash tool emitted tool-output-error: "Cancelled" (cancelled mid-execution), and active_stream_id cleared to null.
  • Remaining — client chat#1770: wire the Stop button to POST /api/chat/{chatId}/stop so the UI "clicking Stop" path is exercised end-to-end.
  • Done when: clicking Stop in the UI ends the stream within ~1s and wasAborted gates auto-commit.
• Resume in-flight stream on reopen. (api#596 GET /api/chat/[chatId]/stream + chat#1771 + docs#225)
  • Done when: reopening a chat mid-generation reconnects to the live stream instead of showing a frozen partial turn.
  • Supporting refactor: chat#1776 (transportChatId).
  • Merge order: api#596 → chat#1776/feat(chat): wire stop button to POST /api/chat/{chatId}/stop #1770/feat(chat): resume in-flight workflow stream on chat mount #1771 → docs#225. (api#590 ✅ merged.)

Open — Agent UX parity

• Interactive tool-call traces. (chat#1780) — collapsible per-tool rendering incl. nested sub-agent calls; fixes broken multi-tool client traces.
• Auto chat-title from first message. (net-new; api#605) — pairs with chat#1662 (new title not shown in sidebar). Done when: the first turn sets a generated (non-verbatim) title that shows in the sidebar.
• Model picker polish. (chat#1682)
• Edit / regenerate-from-message deletes from the wrong store (missed rooms/memories reader). (api#636 + chat#1776)
  • Why: DELETE /api/chats/{id}/messages/trailing still resolves and deletes via the legacy memories substrate (selectMemories → fromTimestamp → deleteMemories) — an untracked rooms/memories reader the cutover missed. Post-cutover history lives in chat_messages, so editing or regenerating from a prior turn in a workflow chat doesn't remove the trailing messages the user actually sees.
  • Fix: api#636 migrates validation + deletion to chat_messages — boundary {id, createdAt} deleted via the stable (created_at, id) ordering, mirroring open-agents' deleteChatMessageAndFollowing. chat#1776 threads transportChatId so the editor calls the endpoint with the session-scoped chat id. Confirm two open-agents parity behaviors are kept or consciously deferred: restrict deletion to role === "user", and bump last_assistant_message_at after delete.
  • Done when: editing/regenerating from a prior user message in a workflow chat removes that message + all newer from chat_messages (verified in DB), the sidebar's last message updates, and no memories reader remains on this path.

Open — Restore sandbox tools & context (net-new — biggest gap; "accepted regressions")

• MCP tools in the workflow sandbox — artist data, music-industry APIs, send_email.
• Composio tools — Sheets / Drive / Docs / TikTok.
• Artist context in the system prompt.
• Telegram new-conversation notifications.
  • Done when: a workflow chat can invoke these tools and the system prompt carries artist context, matching pre-cutover behavior.

Open — Telemetry & billing correctness (net-new; api#605)

• Persist sandbox state after workflow finish — lifecycle timers stay stale otherwise.
• Sub-agent (task tool) credit attribution — one usage_events row per sub-agent model (collectTaskToolUsageEvents).
• workflow_runs telemetry.
• Persist ask_user_question tool results across device switch.

Open — Phase 4 decommission (LAST — only after the workflow tools above are restored and stable)

• Remove remaining rooms/memories readers — getGeneralAgent / setupChatRequest / setupToolsForRequest / MCP-Composio plumbing, still reachable via /api/chat/generate + Slack + email-inbound. (Note: the legacy handleChatStream/handleChatCompletion/app/api/chat/route.ts were already deleted in api#632; the current app/api/chat/route.ts is the canonical endpoint — do not delete it.)
• Drop legacy rooms + memories tables — migration in recoupable/database (+ FK-dependents room_reports / segment_rooms / memory_emails). Destructive + irreversible — snapshot first; confirm zero readers.

Architecture decisions

• Restore-before-remove. Workflow MCP/Composio (Restore-sandbox-tools bucket) must land before Phase 4 deletes the legacy tool plumbing, or capability is lost in the gap.
• /api/chat is the single canonical chat path. /api/chat/generate + Slack + email-inbound are the remaining legacy rooms/memories readers Phase 4 targets.

Out of scope (separate trackers)

• API consolidation (chat → recoup-api): upload (chat#1740/api#540/docs#196), agent templates (chat#1745), message persistence (api#572/#556/docs#215), task runs (chat#1735/api#523), subscriptions (chat#1724/api#505/chat#1723).
• Research / SongStats refactors (api#641–647, Sweetmantech/myc 1868 p1 tool twitter search twitter #575).
• Model providers (chat#1677 / api#442 / api#445).
• Other — predictions (api#425 / database#17 / docs#126), i18n (chat#1670), ElevenLabs (api#391), Next.js upgrade (chat#1590/feat: update Next.js to 16.2.1 and eslint-config-next #1591), dead-code (chat#1741/refactor: remove unused hooks from the project #1742/chore: remove unused UI components and utility functions #1743), stale agent:@U0AJM… auto-PRs (triage/close).