Orchestration

Build, test, and lint commands are NOT prescribed here. Every step that needs them defers to the target repo’s own conventions (CONTRIBUTING.md, AGENTS.md, README, package.json / Makefile / language tooling, etc.). Discover and follow the repo's existing practice — never invent build commands.

Time-boxing & extension protocol

When the cap is reached and the work is still partial, the parent narrows the input (batch smaller in step 4 / split fix scope in step 5) or takes the step over itself.

Sub-agent delegation map

The loop: one round = steps 1 → 2 → 3 → 4 → 5 → 6 → 7 → 8 → 9. After step 9, if Converged: false, go back to step 1 for another round. Repeat until step 9 returns Converged: true; then run step 10 once and exit. At every 10th round, the parent runs the round-cap recap gate before looping back — the circuit breaker that recaps all prior rounds and stops the loop if it has drifted out of the PR’s original scope. See 09-convergence.md for the convergence definition, the loop-exit / loop-back decision, and the recap gate.

Canonical order per round: request → wait → list → triage → fix → build → commit + push → reply + resolve (citing pushed SHA) → convergence check. Reply/resolve runs AFTER push so replies can cite the pushed commit SHA.

Single-iteration fallback

When 01-request-review.ps1 throws because Copilot Code Review isn’t enabled on the repo / account (the GraphQL mutation reports the bot isn’t a valid reviewer), the agent falls back to single-iteration mode:

• Skip step 2 (no Copilot review to wait for).
• Run steps 3 – 8 once against whatever review threads already exist (human, advanced-security, other bots).
• At step 9, pass -SingleIteration to 02-check-review-status.ps1 so the convergence check ignores the stale-review checks that can never advance without a new Copilot review. Converged: true collapses to OpenThreadsAwaitingReply == 0.
• Re-iteration happens only when a human posts new comments later — re-run the skill at that point.

Single-iteration mode is the agent’s decision after the trigger fails, not an auto-detected state — the script can’t reliably tell “Copilot disabled” from “Copilot enabled but not yet triggered” from API state alone.

Convergence proof

Print the proof of convergence in the task_complete message — proof, not assertion:

• HeadOid
• LatestCopilotReview.commitOid
• submittedAt
• OpenThreadsAwaitingReply: 0
• The list of any open escalate-to-user threads if OpenThreadCount > 0.

Notes

• Re-request is first-class. 01-request-review.ps1 does not silently skip when Copilot has already reviewed; it issues the same mutation and verifies via a new copilot_work_started event (the script enforces this — see api-quirks.md for the GraphQL surface and the silent-drop trap).
• Outdated threads still need reply + resolve. They show up in the PR UI as unresolved until you explicitly close them; step 10 is a safety net, not the primary mechanism.
• Reopened / revisit requests reset the thread to step 4. If a declined finding is reopened by the user (or by a follow-up Copilot review), pull it back into triage with the prior rationale as input rather than re-running the whole loop.
• Resumability after interruption. On restart, snapshot HEAD, the latest Copilot review’s commit.oid + submittedAt, the open-threads list, and any uncommitted local changes. Discard cached triage / drafts if HEAD or the open-threads set changed.
• Local-build patches. For projects with uncommitted local-build patches held out of the PR: git stash push -m "local-build" -- <paths> before committing, git stash pop after. Note -m must come BEFORE -- (see api-quirks.md).