Return evidence that the feature meets its acceptance contract: the finite requirements, observable behavior, failure cases, and runtime gates agreed for this run. Name the part of the contract that decided a result: "The contract requires the result to survive a reload, so a successful HTTP response is insufficient."

Keep progress in the ledger, not the transcript: the run's folder beside the feature's tasks () holds , , and . Resume from it. is in a source checkout and a sibling skill in a flat install.

Confirm the repository, worktree, intended base, current diff including uncommitted changes, and environment. Inventory only changed surfaces, their direct callers, and named external interfaces. Record each requirement, operation, relevant variants, expected result, and the check and browser observation IDs that prove it. Keep the denominator finite: say whether "100%" means requirements, a closed operation inventory, or measured line or branch coverage with named thresholds. Timebox discovery; a missing source or policy becomes a named blocker, not a wider search.

Keep behavior, browser, security, migration, performance, and runtime gates separate. A source finding is not execution evidence, and a business pass never overrides a failed runtime gate.

Choose the mode. reports defects without touching product code. fixes them within named paths and needs a request to fix defects. Invocation already permits local test preparation, execution, and disposable fixtures; production writes, external messages, deployments, commits, and publication need their own authority.

Write after the inventory:

Every requirement maps to a required unit. The ceilings bound what you start; they do not stop a running call, so they are not a spend cap.

Build, lint, type checks, and existing suites run through repository commands with no model. A model is for a decision: missing tests, unsettled expected behavior, an ambiguous browser state, or an unexplained failure. When the work is command-only, the plan names no roles.

For a worker, resolve the route that fits with :

analysis

routine-code

standard-code

lookup

hard-code

routine-code

review

A strong role settles expected behavior before a cheaper role writes the tests. A browser worker needs real browser tools in its own process. The reviewer's model differs from any model that wrote the assertions.

Show the scope, each role with its model, effort, and host route, and the attempt, call, and time ceilings. Ask: "Approve this plan, or specify changes." A request to use defaults and run approves the unchanged resolved plan. Then route each role by host: T3 Code (continue with ; a Muse route uses ); Claude Code or Codex a native subagent; otherwise , , or . Give each role one persistent context for the run. An unavailable route blocks its own work; never swap in another model.

Read only the applicable rows of references/test-methods.md. Reuse evidence only when requirements, inputs, dependency versions, environment, and check definitions all match; a commit ID change alone does not invalidate it. A declared fresh check always runs.

Write each attempt into before running it; a failed or interrupted attempt still counts. Run cheap relevant checks first, then the real integration and user flows the contract needs. Capture every command with into . A browser observation records its ID, result, and screenshot and trace paths; check browser access first and record it with ; a blocked preflight consumes no attempt. A unit passes only when each of its checks has a current capture with the expected result. A prose pass does not execute a command or pass a unit.

Parallelize only independent work with isolated fixtures, accounts, ports, and browser contexts; one coordinator writes state. Keep full logs in files and read counts, failures, and paths.

Classify each failure as product behavior, test harness, environment, runtime gate, or unresolved policy, with actual and expected values. Cascading failures from one setup error are one cause until evidence shows more. Use for an unknown cause, and never rerun an unchanged failure without a new hypothesis.

Allow two validation attempts per unit by default. Test, evidence, and product repair allowances are set in the plan and share its total attempt, call, and time limits; each retry records its reason. At a ceiling, stop and report the remaining work. Never start a new run, rename a unit, or weaken an assertion, exclusion, timeout, or diagnostic policy to reset counts or obtain a pass. A stronger model can resolve uncertainty; it cannot choose an unsettled product policy.

On resume, find out what pending calls and attempts did before any retry. requires a closed inventory, every required unit passed with current captures of one tree, and no required failure, skip, pending call, or blocker. Otherwise report , , , or . Captures prove facts, not their meaning: get an independent review of the same tree before claiming readiness. Never describe a scoped pass as universal coverage.

Return: