gabimoncha/oh-my-t3/omt-swarm

Swarm

Each track is a top-level T3 thread on its own worktree and branch, which builds its units with omt-ulw and commits. You split, launch, wait, review, merge, test, and open the PR. You do not build.

Invoking this skill is the user request that t3_thread_launch needs (Top-level threads).

Rules

• The input is a plan from omt-plan. Ask no question batches, and run no plan review.
• Wait inside your turn: t3_thread_wait for tracks, mode: "wait" for children. Use no schedule_task and no heartbeat.
• End your turn only for a blocker, a check that fails after the last QA re-run, or the report.
• Tracks never push. Children never commit. You commit their accepted fixes and own the PR.
• Dispatch review and QA children by Dispatching a child. A track is not a delegate_task child: record it with omt ledger add --plan <slug> --unit <track> --role swarm --round 1 --attempt 1 --status <status> --note <note>, and never settle it.
• Tick plan items only with omt plan mark <slug> <id> <done|open|blocked>.
• After a context reset, read omt plan status <slug> and omt ledger list --plan <slug> first.

Steps

0. Check the modes

Call t3_thread_configure with no arguments. t3_thread_launch needs full-access runtime and default interaction mode. Keep the modelSelection driver as <your family>.

On the default branch, ask the owner for a feature branch and stop. Then run omt git behind --ff (Branch check). On the first run in a project, or after a WARN worktree line, follow worktrees.md.

Done when: both modes match and you are on an up-to-date feature branch, or you stopped and told the user what to change.

1. Read the plan

Find the plan slug in the user's message or omt plan list. With no plan, tell the user to run omt-plan first, and stop. Keep the output of omt plan status <slug> and omt profile show --plan <slug> --json.

When git status --porcelain lists changes outside .omt3/ or .omt/, tell the user that uncommitted edits do not reach the tracks, and stop. Then run omt git start --plan <slug>.

Done when: the plan exists, the checkout is clean, and the start record exists.

2. Split

Run the omt-ulw step 6 design check here: run a requested design pass with omt-design, get the owner's approval, and commit the spec with omt git commit before you record <base>.

Split the open units into at most maxParallel tracks, each with a kebab-case name, its units, and its scope paths. Two tracks share no file, and a unit goes in the track of the units that it depends on.

Keep <base> from git rev-parse HEAD. For each track, run omt plan new <slug>-<track> --light --profile <profile>, and copy in its units, their ## Tests lines, done conditions, final checks, the ## Summary, and design specs. Record the tier by omt-ulw step 3. Run omt plan limit <slug>-<track> <k> with <k> 1 or 2, and leave limits in the shared .omt3/ config as it is.

Log the split in the plan: each track, its units, scope, branch omt/<slug>/<track>, and <base>.

Done when: every open unit is in exactly one track, no two tracks share a file, and each track has a plan <slug>-<track>.

3. Launch

For each track, call t3_thread_launch with:

{
  "title": "<slug>: <track>",
  "clientRequestId": "<slug>-<track>",
  "workspaceStrategy": { "type": "worktree", "baseRef": "<base>", "branch": "omt/<slug>/<track>", "startFromOrigin": false },
  "message": "<launch message>"
}

The launch message is a plain task, not an omt brief:

1. You are track <track> of swarm plan <slug>, with plan <slug>-<track> and scope <scope paths>. Edit no file outside it.
2. Call t3_thread_configure with no arguments, and use your driver as <your family>. Read your plan and omt profile show --plan <slug>-<track> --json. Apply the F2 rule of omt-ulw step 5. Build from the approved design spec in <base>, else from the task, without omt-design.
3. Run omt git start --plan <slug>-<track>, then omt-ulw steps 7 to 12 for your units, with your own review and QA children. In step 12, run only omt git commit --plan <slug>-<track>: no push, no PR.
4. Report what changed, each check with its exit code and log path, and the commit hash.
5. End with one line: TRACK <track>: DONE <commit hash> or TRACK <track>: BLOCKED <reason>.

Keep each branch's path from git worktree list --porcelain as <track worktree>. Record each launch as dispatched, with the threadId as the note. After a failed or lost launch, launch again with the same clientRequestId only when t3_thread_list does not show the thread.

Done when: every track has exactly one thread id in the ledger and one worktree path.

4. Wait

For each track, call t3_thread_wait with threadId, untilDelegatedWorkSettled: true, and timeoutMs up to 3,600,000, again on waitTimedOut: true. Read its last message with t3_thread_read, and check the commit with git log --oneline <base>..omt/<slug>/<track>.

Record DONE as completed with the commit hash. Record BLOCKED or a question as failed with the reason, pass it to the owner, send the answer with t3_thread_send, and wait again.

Done when: every track reported DONE with a commit on its branch, or the owner has its blocker.

5. Review before the merge

Run omt git start --plan <slug> in each <track worktree>. Write a context file: the ## Summary, the done conditions, and each track's name, branch, worktree, and scope paths, without docs/design/.

The review task says:

• Review each git diff <base>...omt/<slug>/<track> against the Summary, and check that the tracks fit together: names, interfaces, and behavior.
• Fix each local finding, inside the track's changed files, in its <track worktree>, and leave it uncommitted. List each fix under FIXES with its track and absolute paths.
• Report each other finding.

Build the brief with omt brief review --task-file - --context-file <file> --scope <track worktree>/<path> ... --title "review: <slug> tracks" --plan <slug> --unit REVIEW --against <your family> --strict --save --json, and dispatch it with mode: "wait".

Done when: the review child returned a result block, and omt ledger done recorded it.

6. Read the fix diff and fix

For each track with fixes, read git -C <track worktree> diff and status --porcelain. Undo a fix that you disprove, and report why. Commit the rest in the track worktree with omt git commit -m "<message>" --plan <slug> and one --path per path under FIXES.

Apply the relevance check of omt-ulw step 9 to each reported finding, and log each dropped one with its reason. Fix each kept finding in its track worktree, run the track's verify commands, and commit there. Send a fix that needs deep track knowledge to that track with t3_thread_send, wait as in step 4, and check the new commit. When a finding changes the plan, mark the unit blocked and ask the owner.

Done when: every accepted fix and kept finding is committed on its track branch, or the owner has it.

7. Merge

Before each merge, run omt git check merge --plan <slug>. On auto, run git merge --no-ff omt/<slug>/<track>. On ask, ask the owner. On off, do not merge, and report it.

Resolve each conflict yourself. When it needs deep knowledge of one track, run git merge --abort, ask that track with t3_thread_send to merge your branch, resolve the conflict, verify, and commit without a push, then wait as in step 4 and merge again. After each merge, mark each unit of the track done.

Done when: every DONE track branch is merged, and its units are marked done.

8. QA

Dispatch one qa child on the merged code with the plan's verify commands, final checks, and device flows with their steps and expected results. Apply the design exclusion of omt-ulw step 10, and route device checks by Verification and devices. Build the brief with omt brief qa --task-file - --scope <path> ... --verify <command> ... --title "qa: <slug>" --plan <slug> --unit QA --round <n> --save --json, and dispatch it with mode: "wait".

Read only each CHECKS exit code and the tail of each failed log. Mark each D and F item done after its command passed in QA.

Done when: the QA result is recorded and every check passed, or step 9 has the failures.

9. Fix failed checks

For each failed check, find the cause in its log, fix it on your branch, and run the check yourself. Then dispatch QA for only the failed checks with the next --round, and mark the items that pass. After 2 re-runs, ask the owner, with each failed command, exit code, and log path.

Done when: every check passed in a QA run, or the owner has the failures.

10. Finish

Commit your QA fixes with omt git commit -m "<message>" --plan <slug>. Run omt git pr --title "<title>" --body-file <file> --plan <slug>, which opens a draft PR. On split:, run omt git stack by Git actions. Link and watch the PR by Pull requests.

The report contains what the omt-ulw step 12 report contains, plus each track with its units, branch, and commit, and each merge conflict with its resolution.

Done when: every done condition is in the report with evidence or as an open item, and the PR is linked or its policy reason reported.