[chhoumann]

Why

pstack inside T3 should hand work to children the way T3 manages them. Three things went against that. A writing child got a worktree the parent made by hand, which no T3 thread owns, so T3 never cleans it up and the PR came from a branch outside the thread (#50). Autopilot-full started each PR owner as a top-level thread, which T3 reserves for a thread the user asks for and which task_status and task_cancel cannot reach (#47). The session-start text told every session, children included, "This chat runs the playbook itself", so a child could run a playbook's fan-out (#21).

What changes

• Dispatch on T3 (poteto-mode: inside T3, delegates should use delegate_task isolation instead of git worktree add #50). A writer gets delegate_task's isolation: "worktree". One writer may stay in the parent's checkout while the parent writes nothing. The brief asks the child to end with git branch --show-current and git rev-parse HEAD, because T3 can rename the branch mid-run and task_status does not report it. The parent merges the writer's branch with git merge, and a PR opens only from a T3 thread's branch. When delegate_task refuses with Unable to resolve the parent checkout commit, the rule says to commit first in a repo with no commit. It keeps git worktree add for a thread whose project is not a git checkout, for work in another checkout, and for a child whose worktree gets uncommitted changes copied in, as interrogate's adversaries do. swarm, arena, Orchestrate, Shipping, Feature step 4, and Opening a PR point at it. Isolation always starts at the parent's HEAD, so a child bound for another commit, such as a PR head or fresh trunk, gets its own worktree even when it only reads. It runs git fetch and resets its own branch there first. That covers Shipping verifiers, Autopilot-full owners, and swarm workers on another branch. A child sent into a worktree it does not own, such as a finished Orchestrate worker's, runs shared there by absolute path. Every child's report names pwd, its branch, and its commit, which is how an isolated arena candidate reports its output location.
• Autopilot-full owners (fix(pstack): align delegated task lifecycle with T3 #47). Each PR owner is a delegate_task child with isolation: "worktree" from the root's checkout at trunk. It resets its own branch to fresh trunk before it writes. Its brief names the playbook and the root's thread ID, so it reports heads with t3_thread_send, and it lets the owner delegate. T3 tells a child never to watch a PR, so the owner babysits in drive mode and hands a wait for a person to the root, which watches that PR. This redoes the T3-mechanics part of @pb3975's fix(pstack): align delegated task lifecycle with T3 #47 as substitutions. Its fresh-review-round rule is left out, since Lauren's "Fresh subagents by default" and T3's own instructions already say it.
• A child is a leaf (Recursive subagent explosion: leaf delegates re-read orchestrator skills and delegate further #21). The routing text and Dispatch on T3 say a delegate_task child does the work itself and starts no subagent unless its brief names one. Lauren's runner prompt stays as written. Both hooks files are unchanged.
• Release 2.2.6. VERSION, the CHANGES.md entry, the stamped manifests, and the README's tag example. All three components stay pinned at cursor/plugins c742dd6.

#19 (settle each child thread after its result) is not in this PR. On T3 0.1.8-nightly.20261106 the sidebar leaves out every thread whose lineage has relationshipToParent: "subagent", which every delegate_task child has. So finished children never show in the sidebar, settled or not.

Evidence

All on this Mac with T3 Code (Nightly) 0.1.8-nightly.20261106, in throwaway repos and threads that are deleted since.

#50.

• A thread whose project is not a git checkout called delegate_task with isolation: "worktree" and got invalid_request: Unable to resolve the parent checkout commit.
• A thread in a git repo with no commit got the same error. After one commit, the same call ran the child in ~/.t3/worktrees/<repo>/t3-<hash>.
• A T3 worktree thread made a local commit and left one file uncommitted. Its worktree child started at that commit without the file, committed, and its branch showed in the parent's git worktree list. git merge --ff-only <branch> moved the parent's branch onto it. The child's branch was t3/3098eef5 at its first command and scripted-git-test after its commit. task_status had no branch field. A shared child committed straight onto the parent's branch.
• A thread on a local feature branch, behind an origin whose main and pr-head had moved since it cloned, started two worktree children. Both began at the parent's feature commit. One ran git fetch origin main && git reset --hard origin/main, landed on the trunk commit the parent had never fetched, and committed on top. The other lacked the PR head commit until git fetch origin pr-head, and then git reset --hard <sha> put it on exactly that commit. The parent's checkout stayed on its commit and clean.

#47. A worktree child's t3_thread_send to its parent's thread ID started a new parent run carrying MID-RUN REPORT from child: branch t3/5902a948 head 137ee27, before the child's completion arrived.

#21. From a read-only copy of this Mac's T3 store:

• 488 delegate_task children exist since 2026-10-04. No grandchild ever started a child, so the five-level nesting in Recursive subagent explosion: leaf delegates re-read orchestrator skills and delegate further #21 did not recur here.
• Before 2026-10-30, 6 of 456 children started children their brief did not ask for. One design child, whose brief named no subagent, started 8 delegate_task research children.
• Since 2026-10-30, T3 itself tells each session "If you are a subagent or child, do the work yourself. Start no subagents unless your task says to." All 14 T3 Codex sessions since then carry it, and 0 of 32 children started a child unasked.
• The new line ends the routing text's contradiction of that rule.

#19. In the installed client bundle, the sidebar keeps a thread only when !Df(ref, lineage, set), and Df is true for any lineage with relationshipToParent === "subagent". t3_thread_list counts 856 unsettled threads with subagents and 14 without.

Checks

• bun install --frozen-lockfile and bun tools/generate.mjs --check pass, with 17 generated files current.
• GIT_CONFIG_GLOBAL=/dev/null bun test tests/ --timeout 30000 gives 554 pass, 56 skip, and 0 fail at 5c6b649.
• bun tools/sync.mjs <pstack|cursor-team-kit|thermos> c742dd6 --dry-run exits 0 for each component. Each fork map matches main's path for path, with the same changed-line counts, so each fork carries the same edit as its derived base. No fork is undeclared, and every substitution matches.
• npx markdownlint-cli2@0.18.1 reports 0 errors on the changed Markdown trees.
• The diff and this body contain no workplace name (case-insensitive grep, 0 hits).

Review

Two independent reviewers ran on fresh delegate_task children each round. They were claudeAgent/claude-fable-5-1 at effort xhigh and codex/gpt-6-astra at reasoningEffort xhigh.

• Round 1 on a2f3ae7. Astra found no critical or high issue and raised three mediums. Fable raised one high, interrogate copying uncommitted changes into worktrees that isolation leaves empty, plus two mediums and a low. All are fixed in 86758d2. Dispatch on T3 keeps git worktree add for a worktree that gets copied changes, and the arena line defers to it. A swarm worker on another branch gets its own worktree and resets its own branch. An Orchestrate worker that reuses a worktree runs shared with its path. An owner babysits in drive mode and hands a wait for a person to the root, since a child cannot watch.
• Round 2 on 86758d2. Astra found nothing at any severity. Fable found no critical or high issue and called it mergeable as is. One of its lows corrected a CHANGES claim, fixed in d9f6957.
• Cursor Bugbot on d9f6957. Bugbot raised four mediums with one root: isolation always starts at the parent's HEAD. Shipping verifiers never moved to the PR head, Autopilot-full owners got the root's HEAD rather than trunk, Orchestrate reuse never put the worker in the reused worktree, and an isolated arena candidate had no output path at spawn. All are fixed in 5c6b649, with the base move verified live as above. A fresh Fable review of that fix diff found no critical or high issue and closed all four. It raised two lows. One is that a read-only child's brief now also tells it to reset its own branch. The other is that a PR from a fork needs the brief to name pull/<n>/head for git fetch.

Known follow-up, not fixed here because it is not critical or high:

• The fallback and T3's own rule. T3's prompt says "Never make a child's checkout in the shell". The git worktree add fallback still does that in its three cases. main already did it for every writer, so this PR narrows the conflict. For copied uncommitted changes, an isolated child applying a patch would stay T3-native, but git diff leaves out untracked files.

Closes #50
Closes #21

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 4 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d9f6957. Configure here.}