How skills work together
See who owns each step of a pull request, backlog run or delegated build. Find and install skills.
Deliver a pull request
Use the steps that fit your change.
- Settle open decisions
Use think for the brief and choose-design when a UI choice remains.
- Build the change
ship owns delivery. implement can do the coding in a separate worktree.
- Check and repair
review-loop uses independent findings to guide repairs and verify the result.
- Open and finish the PR
raise-pr prepares the evidence. babysitter handles checks, reviews and the merge.
- Review the run
ship verifies the requested endpoint, accounts for tracked bugs and records what to change next time.
Settle the choices that change the build
think asks only the questions whose answers would change the next step, then has a separate reviewer confirm nothing important is open. Plans have no model-family restriction. choose-design shows mockups from two model families, or three for a new screen, and you pick before UI work starts. An existing design that already answers the question needs no chooser.
Bring evidence to the review
review-loop repairs and re-reviews until the score target and required checks pass, or stops after two rounds that fail to beat the best score. review supplies the formal reviewer and the evidence and severity rules.
If two code reviews in a row find different material bugs that come from the same design assumption, the coordinator considers a repair that removes that assumption.
Use hillclimb to improve a workflow, artifact or product against a rubric. Confirm the judges before work begins. When the full panel joins, you can upgrade the lead judge before they all score the same version. UI judges use the running interface. One report shows the evidence and scores.
If either loop stops short, the implementer explains what stalled and proposes a change. You can approve continuing from the latest candidate. Judges keep access checks that still apply.
Verify delivery
ship hands the pull request to babysitter, which handles CI, review threads and merge verification. It works through actionable failures before registering a T3 pull request watch. Check deployment and live behaviour separately when the task requires them.
Decide how to handle findings
The coordinator fixes material findings, records reasons for declining invalid or out-of-scope requests, and files eligible deferred bugs as issues. Required approvals and changes-requested reviews still block the merge.
Drive open issues and pull requests to zero
backlog-burn handles open issues and pull requests together. It closes obsolete or completed work with evidence, sends worthwhile fixes through ship, and uses babysitter to repair and merge PRs without an active owner.
For an organisation, it starts one run per repository with open work, skips active runs and returns links to the threads.
- Sort open work
backlog-burn checks ownership and looks for obsolete, completed or superseded items it can close with evidence.
- Run two work queues
Send unowned PRs to babysitter and worthwhile issue fixes to ship. Keep work with a live owner on record.
- Keep ready work moving
Send decisions to the owner and continue independent work. Refresh the issue and PR lists on each wake.
- Verify what remains
Check for zero open issues and PRs. Record outstanding holds and live owners, then write one retrospective.
The coordinator owns PR delivery and monitoring. It uses the available delegate slots for ready work, up to 20 across coding and review, and checks for newly opened PRs when it resumes. Dependency bumps get small in-PR repairs. A migration that changes scope goes back to the owner.
A fresh inventory verifies zero open issues and PRs. The report records any items still held, awaiting a decision or actively owned elsewhere, and distinguishes delivered fixes from work closed with evidence.
Short role cards for delegates
The coordinator reads the workflow and gives each worker or reviewer the card for their assignment.
Assigned reviewer
The assigned-reviewer card tells the reviewer to stay read-only, check their model's eligibility and return evidence-backed findings by severity.
Implementation participant
The implementation-participant card tells a worker to repair what it was given, prove each repair and hand decisions back to its parent.
Check how much each role reads
python3 scripts/check_library.py --stats prints how many words each skill loads on its own, with its folder, and if an agent follows every link.
Implement in a separate worktree
implement starts one lead in its own T3 worktree. The lead codes, adds workers only for large independent parts, and returns reviewed commits. Your main conversation keeps acceptance, integration and shipping.
- Brief the lead
The main thread gives implement the goal, limits and expected evidence.
- Build in isolation
The lead codes in a separate worktree and adds workers when independent parts justify them.
- Review and repair
review-loop checks the actual diff and fixes supported findings.
- Hand back reviewed commits
The main thread checks the exact revision and evidence before integration and delivery.
Choose coding and review models
The lead weighs coding cost alongside review, retries and repairs. Code reviewers must come from a different model family from the code authors. The primary checks the full scope, including UI behaviour when it changes. The current routing and companion rules are in reviewer model selection and parallel review.
Decisions come back to you
If a product or design decision appears, the lead returns it to the main thread and continues only the work that does not depend on it.
Track deferred bugs
Every confirmed bug intentionally left unfixed needs a verified issue link before the workflow is complete. raise-issue files those issues in a background T3 task. Required fixes and checks still gate the merge.
- Confirm the bug can wait
Fix blockers before delivery. Track confirmed bugs you can defer.
- Track it and keep delivery moving
raise-issue files the bug in a background task while required checks and reviews continue.
- Verify both results
Read back the issue URL and confirm the authorised delivery endpoint before final handoff.
Progress reports and the retrospective
Task list and timeline
ship, implement and review-loop show phases in T3's Tasks row when it is available. The main thread records unfinished work, blockers and owner decisions in the handoff.
plan-report makes a timeline from a private ledger. ship, backlog-burn and implement use it at proposal, task completion, final handoff and on request. The final report compares actual build and review work with the original estimates.
Evidence at delivery
Visible changes get a captioned screenshot. The coordinator reports any owner-only steps still open and checks the allowed merge method before each merge. The optional one-round review policy lets the coordinator verify eligible repairs from that review. Later findings and new behaviour still follow the normal review rules. See project settings and the one-round review rules.
A retrospective from the logs
The coordinator compares the results with the plan and checks the logs for consequential events or missing evidence. Any proposed workflow changes go to the configured repository as issues when the operator has authorised posting.
AI-written text is labelled
github-writing adds a byline to AI-written GitHub posts and replies. It names the model that wrote the prose and the operator when verified.
T3 permissions and capacity
Before starting a T3 delegate, the coordinator reads the T3 dispatch preflight. Check that reference for current workarounds.
The coordinator checks LiteLLM's subscription snapshot before choosing a route. It skips routes reported as exhausted or near exhaustion, respects the allowed models and treats stale or missing observations as unknown.
OpenCode, Grok and Muse may fall back to asking
The affected T3 routes can drop from auto to ask. Use explicit full access only when the parent session and host allow it. T3 Code #11458.
Codex in full access rejects prompt rules
T3 maps full access to Codex's approvalPolicy: "never", so a command matching a Codex prompt rule fails before it runs. The agent reports the block. T3 Code #12248 and Codex #41068.