chore(orchestrator): reconcile current main with V2

[saphid]

What changed

Reconcile upstream main through 5b68b2c8e1e971e4ee6956157a1698646ef52a55 with Orchestrator V2 at 415ed0f73b97f1655b6282492f81d0b2bba3a9cc. The resulting head is 61c87ae14483f2c50a686c0e2af8b402460e88b2 (including the subsequent CI reconciliation). This remains a reconciliation PR targeting t3code/codex-turn-mapping; the large diff includes upstream changes.

The merge keeps V2 run, provider-session, checkpoint, subscription, and client projection identities while carrying forward current main behavior. In particular:

• Preserve main's canonical database migrations 48–49, move V2 migrations to 50–62, and reconcile previously shipped V2 migration ledgers transactionally.
• Preserve provider and checkpoint execution, setup-script failure propagation, Git lock handling, background draft intent, and branch PR state without restoring retired V1 reactors.
• Preserve completed plan-step durations across reconnects and server restarts, including repeated labels. Timing is stored in V2 plan artifacts, with backward-compatible optional contract fields. Queued/preparing timers use the durable request timestamp when no provider start time exists yet.
• Carry React Native's outbox-first task creation and immediate navigation into V2. A pending shell and prompt cover creation; failure restores the draft and exposes recovery. Existing V2 queue controls and runtime requests remain connected.

Main changes after the frozen 5b68b2c8 snapshot are outside this reconciliation. Separate V2 leaf fixes still need their own landing order; this PR does not make those leaves obsolete.

Verification

Focused checks passed during reconciliation; these are separate suites, not an aggregate count:

• Migration history and migration suites: 38 tests across 21 files, exercising historical ledgers, canonical IDs, upgrades, and rollback with real in-memory SQLite.
• Modified non-adapter V2 runtime suites: 145 tests across 17 files; provider adapter suites: 272 tests across ACP, Claude, Codex, Cursor, and OpenCode.
• Composer/draft/pending-input compatibility: 281 web tests across four files. Restored setup-script and Git-lock behavior checks: 78 tests across two files.
• Plan artifacts: vp test run src/orchestrationV2.test.ts in contracts, 20 passed; vp test run src/orchestration-v2/ProviderEventIngestor.test.ts in server, 14 passed; vp test run src/session-logic.test.ts in web, 37 passed. The web suite was repeated on final c799952f and passed.
• Final React Native creation/outbox suites: 39 tests; related status/support suites: 9 tests. Shared orchestration timing: 29 tests.
• Contracts, server, client-runtime, web, and React Native package typechecks passed at their respective frozen reconciliation stages. The final mobile/shared typechecks and scoped lint/format checks passed after the last 21-file main port. No repository-wide checks were run locally; GitHub CI owns those checks.

The CI repair at 71deee98 also passed the exact export scan, 85 focused tests (desktop settings, server settlement/migration indexing, native archive grouping, and web terminal), and six affected package typechecks. It restores manual dismissal of asynchronous questions while retaining automatic-settlement, native callback, and approval blockers; corrects the migration-index test boundary; and removes unused exports left by reconciliation. A subsequent one-line relay export restoration at 54b1c1c0 passed 13 relay API tests, infra/contracts typechecks and the exact export scan. The final palette-status repair at 61c87ae1 restores retained/snapshot PR badges and leases fallback VCS queries only for visible rows without a persisted PR link. Its exact CI regression was reproduced failing, then all 43 focused status tests and web typecheck passed. Fresh GitHub CI is running on the latest pushed repair.

Real-client execution

In the full running web client at 1280×800, a disposable V1 fixture was copied consistently into an isolated T3 home. V2 imported its thread and user/assistant transcript. Legacy work activities are not part of that transcript import.

On 2eb2424c0098955e5924846dc6f56d8247865a76, an actual Codex / GPT-5.6-Sol turn ran pwd and a harmless marker command. The client displayed T3_V2_BRIDGE_OK; the persisted run completed, and the checkpoint effect succeeded. Browser reload retained the messages and completed run. Actual ArrowUp recalled the latest prompt and ArrowDown restored the empty composer with Submit disabled. No response text or provider result was injected.

Clean recording · Annotated recording

The GIF samples the actual recording at 15 fps; the roughly 18-second execution segment retains elapsed time. This is candidate execution evidence, not a fabricated before/after comparison. The later c799952f commit changes React Native creation and shared timing; after restarting at that final revision, the existing transcript and completed run remained present. A second actual Codex turn then resumed successfully, displayed T3_V2_RESUME_OK, completed its persisted run, and saved another successful checkpoint.

On final repaired head 61c87ae14483f2c50a686c0e2af8b402460e88b2, a third actual Codex turn ran the harmless marker command and displayed T3_V2_FINAL_OK. Run ordinal3 completed at 09:10:59 UTC; both the provider-start and checkpoint-capture receipts succeeded without errors. The 15-second clip retains the full execution interval and ends after the result.

Final clean video · Final annotated video

Limits and review

The development watcher initially stalled while dependency files changed. Only the captured disposable process group was restarted; the isolated server then started successfully. Fresh native mobile, Electron, relay/tunnel, and full production-database migration runs were not performed in this pass. The focused native tests establish the mapped behavior but do not replace an integrated device journey.

GPT-6 reviewed the reconciliation in the Codex harness. Bounded reviews were also performed by Codex workers; exact worker model metadata was not exposed. The direct claude --model claude-opus-5 --effort high --print --max-turns 1 read-only review attempt exited 1 because OAuth had expired before inference; no Claude review occurred.

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.