fix(cursor,hub): defer ACP reopen merge until ACP session/load ready

[heavygee]

Fixes #917

Summary

• CLI emits a { type: 'ready' } event immediately after successful Cursor ACP bootstrap so the hub can detect load success before merging sessions.
• Hub defers mergeSessions on Cursor ACP resume/reopen until the spawned session is active and has emitted ready (30s timeout).
• Hub keeps archived metadata intact during reopen until resume succeeds; failed handshakes restore the archived row instead of silently merging it away.
• Dedup skips merging archived Cursor ACP duplicates until the active target has emitted ready.

Problem

Investigation on the reported repro showed the archived session was merged away in ~2s while session/load was still failing — not after a ~90s idle timeout. Premature merge destroyed the recovery path.

Testing

• bun typecheck
• Hub: sessionModel.test.ts — three new POST /reopen on archived cursor ACP session: spawned cursor-agent dies ~90s after successful ACP load, merges away the original (no recovery) #917 handshake tests (all pass)
• CLI: cursorAcpRemoteLauncher.test.ts — ready emission + load failure logging

Notes

Does not fix sessions whose ACP store genuinely cannot load (missing/corrupt store.db); those now fail with an error and preserved archive row instead of silent merge-away.

Made with Cursor

[github-actions]

Findings

• [Major] Legacy Cursor rows without cursorSessionProtocol now wait for an ACP-only ready handshake — resumeSession treats missing protocol as ACP via (metadata.cursorSessionProtocol ?? 'acp') === 'acp', but the CLI still routes the same metadata shape (flavor: cursor, cursorSessionId present, protocol missing) to the legacy stream-json launcher. Legacy stream-json does not emit the new ACP bootstrap ready event on the spawned new session, so reopen/resume waits 30s and returns resume_failed even though the legacy runner started. Evidence: hub/src/sync/syncEngine.ts:1162, cli/src/cursor/utils/cursorProtocol.ts:15.

Summary

• Review mode: initial
• One Major issue found: hub-side ACP detection diverges from CLI Cursor protocol routing for pre-#799 legacy rows.

Testing

• Not run (automation). Suggested: add a hub reopen/resume test for a Cursor archived row with cursorSessionId and no cursorSessionProtocol, asserting it does not call waitForSessionResumeReady and can complete through legacy routing.

HAPI Bot

[heavygee]

Addressed the [Major] in 62348c5:

• Moved `isLegacyCursorSession` / `resolveCursorRemoteProtocol` to `shared/src/cursorProtocol.ts` (same rules as CLI).
• Hub resume + dedup now call `resolveCursorRemoteProtocol(metadata) === 'acp'` instead of defaulting missing protocol to ACP.
• Added tests: legacy archived reopen skips `waitForSessionResumeReady`; legacy dedup merges without ready event.

CI re-running on push.

[github-actions]

Findings

• No issues found.

Summary

• Review mode: follow-up after new commits
• Prior bot finding about legacy Cursor rows without cursorSessionProtocol was addressed by sharing resolveCursorRemoteProtocol between CLI and hub and covering the legacy reopen/dedup cases. Residual risk: this review did not execute the test suite.

Testing

• Not run (automation).

HAPI Bot

[github-actions]

Findings

• No issues found.

Summary

• Review mode: follow-up after new commits
• No issues found in the current diff. Re-checked Cursor ACP ready gating, archived-row preservation on failed reopen, legacy protocol routing, and the added regression tests. Residual risk: this review did not execute the test suite.

Testing

• Not run (automation).

HAPI Bot

[github-actions]

Findings

• [Major] Preserve legacy Cursor protocol before spawning — reopenSession now calls resumeSession before clearSessionArchiveMetadata, so pre-#799 Cursor rows with cursorSessionId but no cursorSessionProtocol are spawned without ever stamping stream-json. The hub-side resolveCursorRemoteProtocol(metadata) correctly skips the ACP ready wait, but the new CLI process only receives --resume plus existingSessionId; bootstrapExistingSession preserves no protocol when the old row has none, so cursorRemoteLauncher defaults to ACP and tries session/load on a legacy stream-json id. This breaks the legacy fallback path covered by the new test because the test pre-creates the spawned session with no protocol instead of exercising CLI metadata bootstrap. Evidence: hub/src/sync/syncEngine.ts:1291, related CLI routing at cli/src/cursor/runCursor.ts:45, cli/src/cursor/loop.ts:59, cli/src/cursor/cursorRemoteLauncher.ts:11.

Questions

• None.

Summary

• Review mode: follow-up after new commits
• One Major regression found: archived legacy Cursor sessions can be reopened through the ACP launcher instead of stream-json when migration is disabled/skipped.

Testing

• Not run (automation).

HAPI Bot

[github-actions]

Findings

• No findings.

Questions

• None.

Summary

• Review mode: follow-up after new commits
• The prior Major finding around legacy Cursor protocol stamping before spawn is addressed in this head. I did not find any new Blocker/Major/Minor issues in the changed lines. Residual risk is mainly the Cursor ACP reopen/dedup timing path, but the PR adds focused hub and CLI tests for the ready handshake, rollback, legacy skip, and dedup behavior.

Testing

• Not run (automation).

HAPI Bot