refactor(decopilot): drop /attach orphan-resume; collapse to pure tail

[viktormarinho]

What is this contribution about?

Pre-PR-#3376, POST /messages dispatched runs directly, so a pod death mid-run left orphan threads that only the next /attach could resurrect — hence the orphan-resume code path in /attach (routes.ts:442-537). Now every user message lives inside a threadGateWorkflow DBOS step, and the recovery executor replays it on a healthy pod with streamBuffer wired in — chunks land back on the per-thread JetStream subject and the existing /attach tail picks them up. The heartbeat watcher in app.ts stays as a backstop. This PR strips the now-vestigial client-triggered resume, removes the dead fire-and-forget dispatchRun export (only the orphan-resume called it), and drops threadStorage from DecopilotDeps. Net -150 lines; no behavior change for the happy path.

How to Test

1. bun run dev, open a thread, send a message, confirm the stream renders normally and reconnecting via the page-refresh /attach tail works.
2. Run bun test apps/mesh/src/api/routes/decopilot/ — 350/350 pass.
3. Pod-death scenario (optional, requires multi-pod setup against shared Postgres + NATS): kill the pod owning an in-flight run mid-stream and confirm a client tailing /attach continues to receive chunks (DBOS replays the workflow step on another pod, which re-pumps into the same JetStream subject).

Migration Notes

None.

Review Checklist

• PR title is clear and descriptive
• Changes are tested and working
• Documentation is updated (if needed)
• No breaking changes

---

Summary by cubic

Removed the client-triggered orphan-resume from /attach and collapsed the endpoint to a pure JetStream tail. All runs now go through the per-thread DBOS threadGateWorkflow via dispatchRunAndWait, which pumps to JetStream when streamBuffer is present.

• Refactors

  • Dropped /attach orphan-resume; endpoint now only tails the per-thread JetStream subject and serves SSE.
  • Removed fire-and-forget dispatchRun; all producers use dispatchRunAndWait (drains and pumps when configured).
  • Recovery relies on DBOS replay and the heartbeat watcher; chunks return to JetStream and existing /attach tails pick them up.
  • Removed threadStorage from DecopilotDeps and cleaned up docs/comments to reference dispatchRunAndWait.

• Bug Fixes

  • /attach now sets deliverPolicy from DB thread.status (cluster-wide) instead of pod-local isRunning, preventing missed chunks in multi-pod setups.
  • Corrected a stale dispatchRun reference in prepareRun error messaging to dispatchRunAndWait.

^{Written for commit 0cf93de. Summary will update on new commits. Review in cubic}

[github-actions]

🧪 Benchmark

Should we run the Virtual MCP strategy benchmark for this PR?

React with 👍 to run the benchmark.

Reaction	Action
👍	Run quick benchmark (10 & 128 tools)

Benchmark will run on the next push after you react.

[github-actions]

Release Options

Suggested: Patch (2.330.2) — based on refactor: prefix

React with an emoji to override the release type:

Reaction	Type	Next Version
👍	Prerelease	2.330.2-alpha.1
🎉	Patch	2.330.2
❤️	Minor	2.331.0
🚀	Major	3.0.0

Current version: 2.330.1

Note: If multiple reactions exist, the smallest bump wins. If no reactions, the suggested bump is used (default: patch).

[cubic-dev-ai]

1 issue found across 7 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="apps/mesh/src/api/routes/decopilot/routes.ts">

<violation number="1" location="apps/mesh/src/api/routes/decopilot/routes.ts:403">
P2: `/attach` chooses `deliverPolicy` from pod-local `runRegistry`, so attaches handled by a different pod can miss buffered chunks of an in-flight run. Use thread state from `validateThreadAccess` (global DB view) when deciding `"all"` vs `"new"`.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
Re-trigger cubic}