[leomontigatti]

T3 update review: 0.1.2-nightly.20261021.3473 → 0.1.3-nightly.20261023.3566 (7 releases)

Needs you

• The ready entry storage-cleanup-archived-threads is still waiting on your decision (shared). It was raised in the last review and hasn't been adopted: skills/housekeeping/SKILL.md:19 still has the row "this sweep, until T3 fixes it". My pick is still to adopt it (low risk). It's a small housekeeping edit, in its own PR.
• Optional: post a T3 Ideas discussion for stack-layer cleanup. #14758 (cleanup for PRs merged into a stack parent) was closed not_planned on 2026-10-21. The maintainers say feature requests now belong in Ideas discussions, and I found no discussion carrying it over. storage-cleanup-stack-layers now has nothing upstream to watch. Posting one is a public action, so it's your call. If you don't, the housekeeping workaround stays as it is.

Improves

• shared · pr-watch-unreported-checks: partly met, still watching (evidence in this PR). Both fixes below are in the installed build (compare → ahead for both).
  • #23904 ("wake PR watches when remaining CI checks finish", 09bb9f1): when the required checks pass early, the other checks still running are recorded in owedChecks, and "once all checks finish successfully, send one final all-checks notification".
  • #23863 ("wake PR watches when no checks report", b3c0109): GitHub only. When a head has had no checks for at least 15 minutes (usually about 30 in practice), T3 wakes the thread once with "No checks have reported". It never claims CI passed.
  • Still unmet: a wake while a required check hasn't reported but others have (discussions/18047), and a wake when the rest finish after a red check woke the thread. So the background gh pr checks --watch wait stays, and the entry stays watching.
  • The Watch line moves: #16924 was converted to discussions/23245 on 2026-10-22, not fixed.
  • Proposed edit to skills/babysit-pr/SKILL.md § 2 (optional, its own PR): the T3 bullet's wake list gains "when the other checks finish after the required ones passed, and once when a GitHub head has no checks after about 15 minutes". The "A check still running" paragraph could then say that in T3, after a required-checks-passed wake with others still running, the session ends its turn on the watch instead of waiting in the background. The wait stays after a red-check wake and after a fresh push.
• shared · pr-watch-is-native (evidence in this PR). The Relies on line gains the two new wake events above. The rest is unchanged.
• shared · delegate-task-reviews (evidence in this PR). #22889: orchestrator_capabilities now returns a short list, and takes search. From OrchestratorMcpService.ts: FULL_CATALOG_DRIVERS = new Set(["codex", "claudeAgent", "muse", "grok"]). That means every current Codex and Claude model is still listed. Only legacy models need search. The tool also says "delegate_task accepts any catalog model, listed or not". Model options are now compact, {id, values, default}, instead of full descriptors, and label is left out when it just repeats the id.
  • I checked the live call: every id in agents/models.md (gpt-6.1-sol, gpt-6-luna, claude-opus-5-5, claude-sonnet-5-5) is listed. claude-sonnet-5-5 now shows up, so it's no longer hidden.
  • Nothing here reads descriptor fields or label, so nothing breaks.
  • Proposed edit to agents/models.md § Reaching each model (optional): "orchestrator_capabilities lists the live ones" → "lists the live ones (pass search for a legacy id)".
  • Related fixes, all consistent with the entry: #22850 (answering an already-answered question says "already answered"), #23657 (approval feedback goes to the agent that asked), #22841 (delegated Claude children follow the subagent cache TTL setting).
• shared · storage-cleanup-renamed-branches: Watch line updated (in this PR). #16559 was closed on 2026-10-22 as a duplicate of #17813, which is still open. The fix is #22805, still an open PR ("threads follow branches the agent renames in their worktree"): after a run, finalization updates thread.branch to the branch that's actually checked out. Once that merges and ships, this entry becomes ready.
• shared · pr-watch-unresolved-threads: Watch line updated (in this PR). #17703 was converted to discussions/23305 on 2026-10-22, not fixed.
• shared · storage-cleanup-stack-layers: Watch line updated (in this PR). #14758 was closed not_planned and nothing upstream replaces it (see Needs you).
• Checked, nothing to change:
  • storage-cleanup-removes-thread-worktrees, pcp · cleanup-unlinks-env: cleanup now removes worktrees through removeLinkedWorktree (#21099). That changes behaviour only on Windows. On Linux it still runs git worktree remove (with --force only for submodules), and the branch is kept ("Preserve branch and path").
  • scheduled-run-status, scheduled-missed-runs, en-escena · scheduled-run-worktree: ScheduledTaskService.ts only gained read-only SQL and sorts by created_at (#22899). #17889 is still open. Webhook [] filters (#23282) don't apply: we have no webhook tasks.
  • own-thread-readable: t3_thread_read now clamps a limit that's too large instead of failing (#22901). A bare thread uuid now gets a "Did you mean mcp:<uuid>?" hint (#22452). Both are only friendlier.
  • thread-launch-workspace: the orchestration instructions changed only the capabilities sentence.
  • background-subagents-keep-t3-tools: #23389 (a Claude agent stops when its server dies) and #23783 don't change what this entry relies on.
  • preview-tools-verification: #20899 is still open.

Not covered

• infra (~/.t3/projects/infra): no docs/agents/t3-dependencies.md on any branch.

Still ready

• storage-cleanup-archived-threads: housekeeping hands archived threads' worktrees to T3's cleanup.

---

Method: I triaged the 174 release items by title down to about 30 (mobile, the GitCafe/Forgejo forges, providers we don't use and web refactors were dropped), and read all 824 lines of watched.diff myself, so no miners. I followed every Watch issue's timeline. #16924 and #17703 were converted to discussions, #16559 was closed as a duplicate, and #14758 was closed not_planned; none of them was fixed. Still open: #17889, #17813, #11078, #20899, discussions/18047. WATCHED_PATHS needs no change: the dependency files cite no T3 path it doesn't already watch. Drift check: every Used by file and heading in the shared file exists. The commits since the last review (#19 to #23) add no T3 rule. The project files (en-escena, rubens, pcp, enescena) are unaffected.

Projects affected: none. This PR updates only the shared t3-dependencies.md with evidence and Watch lines; no project needs to add or rename anything. The skill edits above are proposals, not part of this diff.

🤖 Generated with Claude Code

[leomontigatti]

Update: the user posted the Ideas discussion, pingdotgg/t3code discussions/24108, so storage-cleanup-stack-layers now watches it. The storage-cleanup-archived-threads adoption has its own thread and PR.