[leomontigatti]

T3 update review: 0.1.1-nightly.20261017.3238 → 0.1.1-nightly.20261019.3373 (9 releases)

Needs you

• State file was missing. ~/.local/state/t3-update-review/state.json didn't exist, so this run started from 0.1.1-nightly.20261017.3238, the version PR docs: T3 0.1.1-nightly.20261017.3238 update review #6 reviewed. It is marked now. If an earlier run reviewed a later build, some items were looked at twice. Nothing was skipped.
• Decide on two ready entries (shared): setup-runs-in-main-checkout and thread-read-lists-prs (see Fixes / ready). Each one, once adopted, is a small edit to skills/implement/THREADS.md (and contract.md for the first).
• Optional trial, pr-auto-link: open a PR with gh from a thread and don't link it. Then check whether T3's automatic branch association shows its status, can be watched and settles the thread on merge. See Improves.

Fixes / ready

• shared · setup-runs-in-main-checkout → ready. pingdotgg/t3code#18891 (merged 2026-10-18, in 0.1.1-nightly.20261018.3331; closes #18882). "ThreadLaunchService skips the setup action, and its 'Starting setup script' progress step, when the thread has no worktree path." watched.diff: worktreePath === null ? ({ status: "no-script" } as const) : yield* setupScripts.runForThread(...). One case is still open: "an existing_worktree launch whose path is the project root still runs setup."
  • Proposed edit, skills/implement/THREADS.md (Read-only threads): remove "One started in 'local' mode, with no worktree, still runs the runOnWorktreeCreate script, in the main checkout; a setup terminal that fails there is expected, and is no reason to change the main checkout."
  • Proposed edit, contract.md (t3.json row): keep "safe to run in the main checkout", with the new reason "where T3 still runs it for an existing_worktree launch on the project root."
  • Projects: rubens and enescena (setup-script-cwd) and pcp (setup-links-env) say T3 runs the script "the project root for a thread with no worktree". That stopped being true in 0.1.1-nightly.20261018.3331. Proposed edit for each project's docs/agents/t3-dependencies.md: the main-checkout case is now only an existing_worktree launch on the root. Keep the scripts' pwd -P guard.
• shared · new ready entry thread-read-lists-prs. pingdotgg/t3code#17454 (0.1.1-nightly.20261018.3331). t3_thread_list and t3_thread_read now return pullRequests. The tool text says "every linked pull request; linkedPullRequest is a legacy single link and can be null while pullRequests still has entries". Proposed edit, skills/implement/THREADS.md ("A lost context is recovered from T3"): take the PRs from the same t3_thread_read call. Keep list_thread_pull_requests for watch and check detail.

Improves

• shared · pr-auto-link (watching, lead added). watched.diff, ThreadPullRequestService.ts, shows T3 discovering a thread's PR by branch after each turn and saving it as branchPullRequest. pingdotgg/t3code#20223 describes it as an "automatic association" that shows in the sidebar. It is not a link (source has no branch value). Nobody has checked whether it settles the thread or can be watched. Proposed: the trial in Needs you. If it passes, implement § 7 keeps link_pull_request only for PRs from another branch.
• Evidence written into active entries (no action):
  • storage-cleanup-removes-thread-worktrees: #20449 recreates a missing worktree only inside the managed folder (<managed>/<repo>/<folder>). t3_worktree_handoff lost path; no shared file uses it.
  • renamed-branch-in-thread: #20223 drops only automatic associations whose head branch is no longer the thread's. Explicit links are unchanged, so the entry still holds.
  • pr-watch-is-native: #21155. Watching again starts from endedWatch, so a re-watch no longer repeats "all required checks passed". babysit-pr reads the round itself on each wake, so nothing breaks.
  • delegate-task-reviews: #19058 builds summary from the newest assistant message. #21487 makes task_cancel also cancel the task's own queued turn.
  • implementer-worktrees: #20213 refuses isolation: worktree when the project repo lacks the parent's commit.
• Checked, nothing to change:
  • preview-tools-verification: the ServerBrowserPage.ts rewrite (#20446) masks password values and retries a snapshot once. MAX_SCREENSHOT_WIDTH and the scaled clip are unchanged, and #20899 is still open, so the entry stays watching.
  • pcp cleanup-unlinks-env: cleanup now runs git worktree remove -- <path> (still no --force), so the entry holds.
  • scheduled-reviews: bound runs now follow the thread's provider (#18467). Ours are unbound. Webhook values are now framed as untrusted (#20447); we use no webhook.

Watching, moved but not ready

• shared · storage-cleanup-archived-threads. Fix pingdotgg/t3code#16473 merged 2026-10-19T18:37Z (034b506; closes #17811). 3373 was cut from 300a356, and compare 034b506...300a356 returns behind, so the fix isn't installed yet. The Watch line now says ready once the next nightly is installed. Then housekeeping § 1 can drop the archived-thread sweep.
• No change: #17889, #16559/#17813/#11078, #14758, discussions/18047 and #16924, #17703, and #20899 are all still open.

Not covered

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

Still ready

• setup-runs-in-main-checkout: THREADS.md drops the local-thread setup sentence; the contract keeps the safe-script rule for existing_worktree on the root.
• thread-read-lists-prs: recovering a lost context takes the PRs from t3_thread_read.

---

Method: 696 release items were triaged to 382 (web-perf, mobile and marketing dropped). Four gpt-6-luna miners checked them against the dependency files: 36 hits, all evidence or minor. I read watched.diff (20 changed files) in full. Drift check: every shared Used by file and heading exists, and no uncovered T3 rule was found in the shared repo or the four projects.

Projects affected: none by this diff. The rubens, enescena and pcp edits above are proposals for those projects.

🤖 Generated with Claude Code