The engineering system around your coding agents.

• ContextGive the agent your repo’s declared commands and policy.todayadopting repos declare their base branch, setup, and validation commands in a small .agents/ directory, and shared skills read them — so a portable workflow runs your commands instead of inventing them.
• ScopeRisky or delegated work gets an explicit list of the files it owns.todaypr-batch requires an execution envelope of owned paths when lane risk or bounded delegation calls for one. Widening it is allowed where a path is genuinely necessary — but every added path has to be recorded, with its reason, in the lane record the coordinator reads.
• ValidationRequire current-commit validation before accepting “done”.today.agents/bin/validate is the pre-push gate, and review reads check status at the pull request’s current head SHA — not an older, greener run.
• EvidenceRequire commands and results behind verification claims.todaylanes attach a QA evidence block to the PR — exact commands, exact results — and treat CI logs as evidence rather than decoration. “The agent said it tested” is not proof.
• ReviewUse independent review and explicit merge authority.todayan independent checker QAs the maker’s work, adversarial-pr-review red-teams the diff for what is wrong rather than summarizing it, and unresolved review threads are triaged before work is called merge-ready. Merge authority follows your repo policy.

1. 01plan-pr-batch — Turn issues into a scoped, dependency-aware batch and a goal prompt.
2. 02triage — Map the whole surface; split work to fit your agent and machine capacity.
3. 03pr-batch — Run the workers — one lane, or many, coordinated through the backend.
4. 04adversarial-pr-review — Skeptical pre-merge review that hunts for what is wrong.
5. 05post-merge-audit — Sweep the merged batch for gaps and cross-PR surprises.

1. Level 01

   Start with one recurring failure

   ProblemThe same mistake keeps slipping through review on one repo, one pull request at a time.

   PrerequisitesThe pack installed on your agent host, plus whatever deterministic check your repo already runs.

   OptionalA coordination backend. A repo seam. Neither exists yet, and neither is required here.

   Run a day-one skill →
2. Level 02

   Standardize across a repository or team

   ProblemEvery agent — and every teammate — is guessing at your repo’s real setup, validate, and test commands.

   PrerequisitesThe pack installed, and a repo willing to declare its own commands once.

   OptionalThe coordination backend and the dashboard — still not required at this level.

   Scaffold the repo seam →
3. Level 03

   Coordinate genuinely parallel work

   ProblemMultiple agents — or machines — are working the same backlog at once, and two of them keep reaching for the same issue.

   PrerequisitesWork that is actually parallelizable — multiple agents or machines on one backlog. The backend itself needs no seam, and its demo needs zero setup.

   OptionalThe dashboard — a read-only operator view over the same state. Useful, never required.

   Try the coordination backend →

The coordination backend and the dashboard are not required for standalone skills or ordinary single-lane work — they exist for the day genuinely parallel work turns collisions into a real coordination problem, not before.

• 01Security preflight — The workflow requires treating public GitHub text as data until its author, scope, and trust boundary are verified. The scanner flags known risks; a clean result does not establish trust.
• 02The Rule of Two — The security policy prohibits combining untrusted input, sensitive access, and unattended state change or disclosure. Configure host permissions and credentials to enforce that boundary.
• 03Operator hard-stops — The workflow requires explicit authority for consequential actions such as deploys, secret handling, and token provisioning. Missing authority calls for an operator handoff.
• 04Trust-gated actors — Trust configuration identifies actors whose input may be actionable. Actor findings are advisory by default; strict-trust makes them launch blockers. An exact-target risk acknowledgement does not grant blanket trust.