Post-cutover gaps: /api/chat/workflow vs open-agents /api/chat

[sweetmantech]

Tracking issue for gaps surfaced by the open-agents → recoup-api chat workflow cutover (recoupable/open-agents#43). The chat-workflow port replicated the request/response wire format but did not port the post-stream side effects from open-agents' app/workflows/chat.ts + chat-post-finish.ts. Filing each gap as its own follow-up PR.

Must-fix (silent data loss / broken UX / silent revenue loss)

• active_stream_id clear after workflow finish — fix(chat): clear active_stream_id when chat workflow finishes #604
• Explicitly close chat workflow writable — fix(chat-workflow): explicitly close the SSE writable so chat.status unsticks #607 (SSE was stuck open ~2m after finish)
• Persist assistant message after workflow finish — fix(chat-workflow): persist the assistant message after a successful run #609 (also bumps last_assistant_message_at so sidebar unread badge fires)
• Charge credits per chat turn — feat(credits): charge credits per chat turn (atomic wallet debit + audit) #612 (consolidated into existing handleChatCredits orchestrator; atomic via deduct_credits_with_audit RPC from feat: deduct_credits_with_audit RPC for atomic wallet + meter writes database#26 — atomicity now applies to research handlers too)
• Persist sandbox state after workflow finish — lifecycle timers stay stale
• Stop monitor — Stop button cancels the workflow but the in-flight model call keeps streaming until natural completion (open-agents polls run.status every 150ms and aborts via AbortController). Also blocks wasAborted gating in auto-commit.
• Persist tool results from client-side messages (ask_user_question style) — tool results lost on device switch
• Chat title generation from first user message (open-agents has isFirstChatMessage → updateChat({title}))
• Sub-agent (task tool) credit attribution — open-agents iterates collectTaskToolUsageEvents(responseMessage) and inserts a usage_events row per sub-agent model with agentType: "subagent". Without this, sub-agent calls aren't billed even after feat(credits): charge credits per chat turn (atomic wallet debit + audit) #612's main-turn fix landed.

Multi-step + metadata (verify whether streamText.stopWhen covers these, then close as appropriate)

• Outer multi-step loop vs single-call streamText.stopWhen — verify multi-tool-call traces complete cleanly on api
• stepFinishReasons[] accumulation across steps in message metadata
• totalMessageUsage / totalMessageCost carryover across steps
• Stable messageId across steps — fix(chat-workflow): persist the assistant message after a successful run #609 (workflow-body generateAssistantMessageId step threaded through every runAgentStep call; resume-aware reuse for in-progress assistant turns)

Lifecycle / observability

• workflow_runs row via recordWorkflowRun — chat workflows have no workflow-run telemetry on api (status, durationMs, stepTimings). Separate from the credits gap above — recordWorkflowUsage in open-agents does BOTH but they're independent concerns
• refreshDiffCache after workflow finish — diff viewer stays stale until next reload
• refreshLifecycleActivity per long step — currently refreshed once before start

Product features

• Auto-commit + push after a natural finish — feat(chat-workflow): auto-commit + push after natural finish #614 (LLM-generated message via gpt-5.4-nano, data-commit UI chunks emit live AND persist onto chat_messages.parts so the GitDataPartCard survives refresh). Three sub-gaps remain blocked on other items: autoCommitEnabled toggle waits on user preferences integration below; wasAborted gating waits on the Stop monitor above; co-author trailer is minor and not blocking.
• User preferences integration — getUserPreferences → filterModelVariantsForSession → sanitizeSelectedModelIdForSession. Also unblocks autoCommitEnabled toggle on the auto-commit flow.
• Subagent model resolution — resolveChatModelSelection currently single-model
• Managed template trial limit (MANAGED_TEMPLATE_TRIAL_MESSAGE_LIMIT enforcement)

Minor

• dedupeMessageReasoning on inbound messages
• pruneMessages({emptyMessages: "remove"})
• reconcileExistingActiveStream — open-agents retries 3× with chat re-fetch; api is single-pass
• upsertChatMessageScoped — open-agents scopes the PK conflict check to the chat; api's global upsertChatMessage could in principle skip on cross-chat id collision (vanishingly unlikely with ai's generateId)
• Conflict logging on duplicate upsert — open-agents console.warn when result.status === "conflict"; api silently returns
• originalMessages threaded into toUIMessageStream — open-agents passes the prior turn for assembly; api doesn't
• sendStart: false, sendFinish: false + explicit sendStart/sendFinish steps — currently relying on toUIMessageStream defaults (wire format verified identical, but interleaved-step architecture diverges)
• Writer per-chunk acquisition — open-agents acquires fresh writer per chunk so other steps can interleave writes; api holds one writer for the whole stream
• streamClosed backup close — open-agents' belt-and-suspenders pattern with a streamClosed flag and a guarded close in finally; api relies on closeChatStream step in finally always running
• Auto-commit co-author trailer — open-agents appends Co-Authored-By: <GitHub App identity> as a second -m; api currently has no GitHub App equivalent
• Auto-commit auth URL helper — open-agents has buildGitHubAuthRemoteUrl({token, owner, repo}); api inlines the template literal in performAutoCommit

Acceptance for full deletion of open-agents /api/chat route

All "Must-fix" boxes checked + multi-step verification + (auto-commit OR explicit decision to deprecate that feature).

Source reference: open-agents/apps/web/app/workflows/chat.ts (~1000 lines) + open-agents/apps/web/app/workflows/chat-post-finish.ts (9 exported step functions, including recordWorkflowUsage at line 230).