Verify Findings

Finding-Verifier Fleet Dispatch (stage_23_verify_findings)

This skill is the canonical finding-verifier dispatcher for /code-review at stage_23_verify_findings. It is split out of commands/start.md so the orchestration spine stays lean; the orchestrator invokes it when the walker reaches stage_23_verify_findings. The content below is authoritative for both MODE=local and MODE=github, with mode-specific Task scheduling: GitHub mode dispatches verifiers synchronously, while local mode preserves parallel background dispatch plus blocking TaskOutput collection.

Verifier Fleet (stage_23_verify_findings)

This stage runs when the walker reaches stage_23. It implements PLN-722's finding-verification pass: each eligible finding gets an independent second opinion from a verifier agent prompted to falsify (not confirm) the original claim. Findings that survive land in verified[]; findings rejected with positive evidence land in rejected[] and surface in the "Dismissed Findings" section so humans can falsify the dismissal.

Inputs

stage_22b_verify_prepare already wrote <CR_DIR>/verify_manifest.json and one input file per eligible finding at <CR_DIR>/verifier_inputs/<finding_id>.json. Read the manifest:

{
  "to_verify": [{"finding_id", "model", "input_path", "output_path", ...}, ...],
  "skipped_no_verification": [...],
  "deferred_budget": [...],
  "cache_hits": [...]
}

cache_hits[] entries have already been materialized at their output_path; do NOT respawn them. Only entries in to_verify[] need fleet dispatch.

If verify_manifest.json is absent, do NOT dispatch any verifier — stop and report. stage_22b_verify_prepare writes it last and exits 3 without it when it could not prove the review root, so a missing manifest means the verifiers would resolve source paths against a tree that is not the one under review.

Spawn contract

First branch on MODE. GitHub and local runs intentionally use different Task scheduling because GitHub headless mode cannot survive outstanding background verifiers after the assistant turn ends.

GitHub mode (MODE=github): dispatch synchronously. Spawn exactly one verifier at a time and wait for its Task response before spawning the next to_verify[] entry. Omit run_in_background or set run_in_background: false; never set it to true for GitHub verifiers. Do not use TaskOutput, watcher files, sleep loops, polling loops, or "wait for background task" turns in GitHub verifier dispatch. Each verifier must finish, write its verdict JSON to the manifest entry's output_path (<CR_DIR>/agent_verifier_<finding_id>.json), and return before the walker dispatches the next to_verify[] entry. When stage_23 completes in GitHub mode there must be no verifier task still running.

Local mode (MODE=local): spawn ALL verifiers at once. Use run_in_background: true on every verifier Task. You can spawn all agents in a single message or across a few messages.

Each spawned verifier — in either mode — uses subagent_type: "code-review:code-review-worker" (tool allowlist Read, Write, Grep, Glob — identical to the Reviewer Fleet's, so no permission changes are needed) and this prompt template:

You are the FINDING VERIFIER. Read your prompt at:
  {VERIFIER_PROMPT_PATH}

Your input file is at:
  {INPUT_PATH}

Read it for the finding to verify, the canonical output path, and the
per-output JSON shape. Write your verdict JSON to the output path the
input file specifies. Do not write anywhere else.

Substitute the resolved paths from the manifest entry (the verifier prompt is at <CR_DIR>/verifier_prompt.txt, copied by stage_02_prep_assets). Each input file also carries a review_root field — an absolute path stage_22b_verify_prepare proved holds this diff before writing the manifest, so it is never empty on a healthy run; the verifier prompt tells the agent to resolve every source path under it rather than against its own working directory (which is the invoking session's checkout, not the code under review) — no extra wiring is needed here. Set model to the entry's model field (currently uniform sonnet; future revisions may split by original-reviewer model for cross-model independence).