Perform Task

Run and publish

Execute references/pipeline.md exactly. A task that changes the repository and is approved produces:

1. one or more tested source implementation-attempt commits, each with an exact one-line subject using the pipeline's conditional [ai] prefix, blank line, and Task: <full-task-id>;
2. local tracked phase artifacts in the AI slot worktree, without phase commits;
3. one canonical Approve <full-task-id> commit containing all final AI artifacts and state.

New and unfinished tasks use the single adaptive implement path. Assessment must first confirm that the request is one cohesive implementation/review/test unit. If it contains independently useful and independently testable product boundaries, record Scope: split-required and a concrete split proposal before source edits. The same result may arise later from the bounded convergence assessment when the retained implementation proves that one review/evidence campaign is not coherent. Do not force the broad request through smaller implementation phases and call it one task. The independent assessment has veto authority over further implementation, not authority to create, retire, or rewrite tasks. The performer writes the split result, preserves any owned implementation and source refs, and publishes it with finish --status split-required. The checkout scheduler owns the later queue mutation and implementation transfer. A direct invocation returns the published proposal to the human.

For a cohesive task, use one mandatory general review, all five standard review lenses, and a falsifiable evidence plan. On the initial implementation the general reviewer and all lenses inspect the task and complete diff without seeing one another's findings. A lens may return a compact NOT_APPLICABLE immediately after that scan when it proves the diff affects no mechanism it owns; otherwise it reads the relevant changed files and adjacent code and returns CLEAN or FINDINGS. The evidence loop may use static readings, commands and artifacts, unit tests, a standalone probe or component binary, a Telegram Debug build with logged assertions, an in-app overlay, Computer Use, screenshots, or any necessary combination. Do not require a portable account, Telegram executable, or desktop unless a selected check uses it. Do not weaken a runtime or visual check merely because another instrument is cheaper.

The general reviewer examines every changed file in full and the evidence plan, may reject an unsupported NOT_APPLICABLE, require a named domain specialist or stronger instrument, and cannot defer its own concern. Its approval and every clean or proved-not-applicable lens result carry forward. A fix invalidates only the findings, changed invariants, specialists, validations, and evidence checks it actually affects. Review fixes receive a focused general delta review plus only those invalidated specialists; they do not restart the full review or evidence design.

Automatic replay is bounded. If two review verdicts need changes, findings are not converging, or a fix expands the architecture or owned paths, run the pipeline's independent convergence assessment instead of another broad round. It chooses a bounded focused repair, a coherent replan, or RESCOPE_REQUIRED; unresolved findings are never approved merely to meet the bound. A task whose desired outcome was already present may finish without a source commit only after the same general review and evidence loop prove Outcome: already-satisfied.

Only a genuine exhausted task blocker produces a canonical Block <full-task-id> commit. Agent interruption, tool loss, and global environment stops leave the task in-progress with its task-scoped local state intact for the next invocation.

A repeated evidence setup failure is not exhausted recovery by itself. Follow the shared directness ladder: forbid the failed command, fixture, probe, or capture technique and make the next run closer to the changed surface. The configured test-run cap closes one campaign: preserve prior passes, isolate the unmet checks, and start at most one focused recovery campaign unless a fresh assessment proves every direct strategy exhausted. A second campaign cap or a repeated non-converging focused signature stops automatic work for an explicit human/convergence decision; it does not start another campaign. A cap and a TEST_FLAW can never by themselves publish BLOCKED or approval.

A pre-Runner crash or DeadlockDetector event is not an evidence setup failure merely because the scenario did not start. Apply the shared crash diagnostics and debugger fallback before changing an account fixture. An empty or unusable dump requires live debugging after at most one confirmation run; it never supports a fixture verdict.

A locked macOS session is not an environment stop or evidence blocker for a selected Telegram runtime check. Skip interactive Computer Use and complete the same coverage through the in-binary overlay: drive the flow, log/assert, capture widgets or windows, quit, and assess the saved artifacts. Non-app instruments are unaffected.

A Windows build-output lock is not an immediate environment stop. Follow the shared bounded recovery contract, including exact-path cleanup before builds. Only its exhausted or unsafe outcome is a global hard stop; it never becomes a task Block.

Do not report success from a source commit alone. The final AI commit must be canonical. Retry ordinary concurrent-master publication races until success. On a semantic conflict, unsafe checkout, or unreachable remote, preserve resumable state and report a hard stop.

Return a compact result with the full task id, status or hard stop, attempts, touched files, canonical final-publication confirmation, and exact evidence or unverified behavior. Never persist or report commit hashes; the full task id is the only cross-repository link.