Execute these phases in order. The primary agent alone owns judgment, edits, commands that mutate the workspace, validation, Git, PR changes, and bot replies.

Use the reviewer count, model mix, and reasoning levels explicitly requested by the user. Otherwise use this default cohort for every initial and remediation pass:

sol-1

gpt-5.6-sol

high

terra-1

gpt-5.6-terra

xhigh

luna-1

gpt-5.6-luna

max

luna-2

gpt-5.6-luna

xhigh

luna-3

gpt-5.6-luna

xhigh

For a count from one through five, use that order. For another explicit mix, assign stable IDs from model tier plus one-based ordinal. Ask for a mix when a count above five is otherwise underspecified. Normalize task names from to for deterministic discovery. Keep the raw review prompt, target fingerprint, role boundary, and service tier identical across the cohort; use each reviewer's assigned reasoning level. Queue over concurrency limits without editing the target.

Do not send this skill, any reference, helper commands, run-log details, or the primary conversation to a reviewer. Give every reviewer the same self-contained packet containing only the review task: raw user prompt (or default criteria), target SHA and workspace fingerprint, relevant scope/diff or paths, conflict summary if applicable, read-only boundary, and required finding format (file, minimal line range, severity, scenario, rationale). Exclude other reviewers' findings, the primary's conclusions, remediation decisions, and telemetry instructions.

Launch every initial reviewer with , its self-contained packet, and the requested model and reasoning level; do not fork the primary conversation. Require the runtime to expose the exact applied controls and a stable resumable handle. Verify applied settings from runtime evidence, not requested arguments alone.

Launch concurrently where the runtime permits. Never use an ephemeral reviewer. A failed control verification blocks editing and finishes ; never treat requested arguments alone as verification. Record as soon as a handle is available and only after persisted verification. Record completed-task only when the runtime exposes it.

After each bounded wait (30 seconds by default), run the cohort watcher. It is a maximum polling interval, not a runtime floor: handle completions immediately.

node scripts/review-run-log.mjs inspect-reviewers \
  --log "$REVIEW_RUN_LOG" \
  --stale-after-ms 120000 \
  --soft-deadline-ms 600000 \
  --hard-deadline-ms 1200000 \
  --record

The soft deadline is a warning. The hard deadline begins at the current (or session start when absent). Before classifying a reviewer as failed, inspect its exact persisted session. , , and one result are not failures. The watcher does not interrupt workers.

For a hard-exceeded reviewer, immediately inspect the exact session once more. If it remains non-terminal, call with the inspection's , never its persisted ; confirm with that it stopped; then append with reviewer ID, persisted session ID, native handle, phase, reason, and deadline. Never use a broad kill, interrupt another reviewer, or probe an initial review with a follow-up. One fresh initial retry may use a distinct task name such as , but retains reviewer ID ; a second hard deadline finishes or . A hard-exceeded continuity session follows the full-cohort restart rule after clearing the exact handle.

For a missing or unreadable result, inspect the exact session before retrying:

node scripts/review-run-log.mjs inspect-native-session --log "$REVIEW_RUN_LOG" --reviewer-id "$REVIEWER_ID" --stale-after-ms 120000

The recovered result must match exactly one reviewer handle, repository, applied controls, and completed final-answer event. Use its recovered answer but never log the review body. Retry only after inspection says and follow the hard-deadline cleanup above. Keep raw results outside the run log; record only concise observations and outcomes.

Before fixes, create a gitignored ledger with stable reviewer ID, requested/applied controls, reviewer handle, initial fingerprint, and continuity state. Never store credentials, prompts, or review bodies. Before editing, resume every initial session with its original controls and read-only boundary; require only . Record a completed-task duration when available. If any handshake fails, discard every report and restart the full cohort once against the unchanged target; a second failure blocks editing. Remediation uses only these verified handles.

Resume every continuity-verified session with its original controls and read-only boundary. Never use an ephemeral or replacement session without user authorization.

Always attempt , even for a blocked/failed run, using event-derived reviewers/findings plus actual bot, validation, status, and SHA outcomes. Use , , or instead of when the cohort cannot finish. Use ; collection is per reviewer, so completed sessions still contribute real tokens, cost, and duration when another worker is unavailable. Do not report while any reviewer lacks both tokens and an exact duration. Run , resolve its per-reviewer session/ledger cause (including an allowed relaunch when needed), then run again. Generate the usage section with only after that gate passes; do not manually calculate or reformat it. Treat model comparisons as one-run observations.

Report the applied cohort/controls, persistent sessions and continuity/retries, log path and derived invocation/round/usage coverage, shared/unique findings and model comparison, base SHA/integration/conflicts, all finding dispositions, remediation rounds and disagreements, validation per push, commits/PR, bot-loop outcomes, and remaining blockers. refuses an incomplete-telemetry cohort, so append its table verbatim only after it succeeds. Do not add pricing or telemetry caveats. Keep immediately after and last; put nothing after it.