[leomontigatti]

T3 update review: 0.1.3-nightly.20261023.3566 → 0.1.3-nightly.20261026.3717 (12 releases)

Needs you

• Decide on pr-auto-link (shared, now ready). T3 links a PR made with gh pr create on its own now (#24813). Tried on this review's own PR docs: T3 update review 0.1.3-nightly.20261026.3717 #28: it was linked (source created) before any call. My pick is to adopt it: a small edit to THREADS.md, implement § 7 and implement-spec § 3 step 3, in its own PR (see Fixes / ready).
• Decide on 1280 px preview images (shared · preview-tools-verification). T3 fixed the blur bug (#24618 closes #20899), but preview_snapshot still caps captures at 1280 px wide. The entry asks you directly: will you accept 1280 px images for a 1440 design width?
  • If yes: the next step is en-escena's dialog probe (/administracion/resultados, Acciones → Descargar, 1440×900). If it passes, the entry becomes ready.
  • If no: the browser CLI stays for wide designs until T3 lifts the cap. No upstream issue asks for that.
• Decide how the shared skills answer T3's new injected instructions (see Breaks / changes). The skills still work, but every T3 session now gets rules that contradict three of them. My pick, in order: (1) a one-line edit to PREVIEW.md now; (2) the delegate_task isolation: worktree trial for implement-spec, which would remove the other two conflicts at their source.

Breaks / changes

T3 rewrote its injected agent instructions and tool descriptions, about 60% shorter (#24850, in 0.1.3-nightly.20261025.3669). No T3 behaviour any active entry relies on changed. What changed is what T3 now tells every agent, and three shared skills say the opposite. An agent that reads both gets two rules and picks one.

• shared · preview-tools-verification (evidence in this PR).
  • T3's instructions (orchestrationInstructions.ts) now say: "Never use preview_* to verify or show your own edit, even a frontend edit or an open tab. … Use them only when the current request asks for browser work or names a page or tab."
  • The preview_snapshot description says the same.
  • The browser section dropped its old exception for "a skill or documented repository workflow that names one".
  • implement step 4 and PREVIEW.md have the agent drive its own edit with exactly those tools. An agent following T3's text skips the drive, or falls back to the browser CLI.
  • Proposed edit to skills/implement/PREVIEW.md (its opening paragraph): "Driving the change is browser work the request asks for: implement step 4 and the project's verification.md require it, so T3's rule against verifying your own edit unasked does not apply here." Keep the closing step ("Close the tab"), which matches T3's "Close a tab you opened before you reply".
• shared · background-subagents-keep-t3-tools (evidence in this PR).
  • The mechanism still holds: isLiveCallerShell and assertLiveCaller only lost comments.
  • T3's instructions now say "Claude Agent always with run_in_background:false" and "Use delegate_task for … a separate checkout (isolation:"worktree"), background work".
  • implement-spec § 2 spawns implementers with run_in_background: true and Agent-tool worktrees. PREVIEW.md step 1 has parallel subagents.
  • Proposed edit to skills/implement-spec/SKILL.md § 2, until the trial below: name the exception where the agent reads it, e.g. "This skill's background implementers are the exception to T3's foreground-Agent and delegate_task-for-background rules: …"
• shared · implementer-worktrees (evidence in this PR).
  • T3 now says "Never make a child's checkout in the shell or in task". agents/implementer.md step 2 runs setup in an Agent-tool worktree by hand.
  • The delegate_task description now reads "Separate checkout or PR per child: isolation:'worktree'".
  • This is the entry's own improvement path. Its trial (does a worktree child commit to a named branch the orchestrator can merge?) is now the way to end the conflict.
  • Proposed: run that trial in a throwaway implement-spec run before changing the skill.

Fixes / ready

• shared · pr-auto-link: watching → ready (in this PR).
  • #24813, in 0.1.3-nightly.20261025.3669: "T3 now links the PR itself … when a settled command runs a host's create command and prints the PR URL on a line of its own", with source created, on the thread that ran it.
  • Gaps (nothing gets linked): a for loop creating several PRs, gh api creates, and output with more PR URLs than create commands.
  • The link_pull_request description dropped "Register every pull request you open" (#24850).
  • Live trial, 2026-10-26: this PR was opened by gh pr create as the last of several shell commands. list_thread_pull_requests showed it linked with source created before any call, and link_pull_request afterwards returned alreadyLinked: true.
  • Proposed edits (own PR):
    • THREADS.md "One thread, one branch, one PR": "T3 links a PR gh pr create opened to the thread; call link_pull_request only for a PR opened another way, or one missing from t3_thread_read's pullRequests."
    • Mirror that in implement § 7 ("and link it to the thread" → "and check it is linked") and in implement-spec § 3 step 3.

Improves

• shared · preview-tools-verification: partly met (in this PR). Captures now render backdrop-filter correctly: full-scale capture, then a canvas resize in the page (shrinkPng). Proposed edit to PREVIEW.md step 3: drop "and at any viewport wider than 640 px draws a blurred backdrop … as mirrored tiles of the page: a capture artifact, not an app bug". The 1280 px sentence stays.
• shared · storage-cleanup-removes-thread-worktrees (evidence in this PR). #23900: a deleted thread now owns a worktree its remaining subagents share, so worktreeOnDelete removes it. Before, the hidden child kept it with "Has commits beyond the default branch".
• shared · pr-watch-is-native (evidence in this PR).
  • Delegated children no longer see the watch tools (#24808).
  • Wake messages say the PR stays linked and watched (#24781). The tool now says "do not call this again on a wake". babysit-pr § 2 already re-watches only "if a wake says it ended", so nothing changes there.
  • A watched PR holds the thread in Working (#24610), so babysit-pr's background gh pr checks --watch wait, which always runs under a watch in T3, doesn't need the new t3_background_wait.
• shared · delegate-task-reviews (evidence in this PR). A result for a running task carries next ("Completion wakes this thread, so end your turn", #24782), and every result carries a link (#24784). summary, target, clientRequestId and the default isolation: shared are unchanged. #24560 now allows t3_thread_send for a follow-up that needs the same child's conversation; independent review rounds stay a new delegate_task, as code-review does.
• Optional edit to THREADS.md (shared, no entry): thread tools now return a ready link, and T3 tells agents to use it verbatim. No shared skill writes t3-thread:// links by hand, so nothing breaks.
• Checked, nothing to change:
  • worktree-per-thread, plus the setup-script-cwd / setup-links-env entries of rubens, enescena and pcp: the setup runner now waits until the shell's line editor is reading before it types the command (#24775). That way an rc-file prompt (fnm, nvm) no longer eats it. cwd, T3CODE_PROJECT_ROOT and the completion wait are unchanged.
  • scheduled-reviews, scheduled-run-status, scheduled-missed-runs, and en-escena · scheduled-run-worktree: ScheduledTaskService.ts only lost comments. scheduledTaskWorkspaceStrategy is unchanged. #17889 is still open. The settings page's taskId link handling moved out of effects; the Base branch field stays.
  • pr-watch-is-native "Breaks if" (inactivity auto-settle): #24617 lets inactivity settle a thread that has an open PR link. Your sidebarAutoSettleAfterDays is null (off), so nothing settles.
  • thread-launch-workspace, own-thread-readable, renamed-branch-in-thread, pr-linked-by-tool: only descriptions got shorter, and a link field was added. Shapes are unchanged. Periodic branch-PR discovery now pauses without a foreground client (#23852). After-turn discovery still runs.
  • en-escena · architecture-review-needs-app: background-service.md gained "Local network access" (LAN and Tailscale), and still says nothing about scheduled tasks.
  • pcp · preview-odoo-login: the preview tools changed only their copy. The same instruction hurdle as PREVIEW.md applies, and the shared PREVIEW.md edit covers it.
  • New, nothing uses them: request_secret with delivery: "file" (#24809); followups steer by default (#24718, a client setting); guest sharing (#24104, off by default).
  • Watched, unchanged: #17813 and its fix PR #22805, #11078, and discussions 18047, 23245, 23305, 24108 are all still open.

Not covered

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

Still ready

• pr-auto-link: T3 links PRs that gh pr create opens; the skills still call link_pull_request after it.

---

Method:

• I triaged the 238 release items by title down to about 30. Dropped: mobile, lint and refactor sweeps, forge refactors, web visuals, and providers we don't use.
• I read all 4,449 lines of watched.diff myself, so no miners.
• I read the bodies of 12 key PRs and followed every Watch issue and discussion.
• 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 (docs: T3 update review 0.1.3-nightly.20261023.3566 #24 to feat(implement): hardening asks in plain words; boundary validation is never speculative #26) add no T3 rule. The project files of en-escena, rubens, pcp and enescena are unaffected beyond the notes above.

Review

Tier: readback, reviewed 746537b

• readback (Luna): no findings

🤖 Generated with Claude Code