flow-consolidator-instructions

Files (1)

• SKILL.md

SKILL.md (first lines)

<!-- flow-instructions-sentinel: flow-consolidator-instructions -->

# Consolidator-Validator instructions

These instructions are read by the Consolidator-Validator subagent that
`/flow-pr-review`'s SKILL.md spawns via the Task tool at Step 3.5 — after
Step 3's six-agent multi-agent review and before Step 4's
finding-consumption pass. The subagent runs in an isolated context: its
per-agent JSON reads, second-opinion validation prose, and dedup
reasoning all stay inside its own session and are never returned to the
caller. The only outputs it produces are the structured artifact at
`<worktree>/.flow-tmp/consolidator-result.json` and a brief
one-paragraph summary it returns on completion.

## 1. Role and scope

The Consolidator-Validator's job:

- Read the six per-agent JSON outputs (one each from bug-detection,
  security, pattern-consistency, performance, supply-chain,
  test-coverage), plus the OPTIONAL seventh cross-model Gemini lens
  (`agent-output-gemini.json`, tolerated-absent) and the OPTIONAL eighth
  brief-gated product lens (`agent-output-product.json`,
  tolerated-absent), validate each against
  the per-agent schema, and merge them into a single findings array.
- Apply the confidence threshold (>=80 for non-praise; praise findings
  pass through unfiltered) and dedup by `(file, line ± 2 lines window,
issue-class)`.
- Apply praise specificity (drop content-free praise).
- Run a **second-opinion** validation pass on every >=80-confidence
  non-praise finding that survives dedup, moving false positives to
  `dropped_by_validation[]` with a `reason` string naming the rule
  that rejected them.
- Write the consolidated + validated artifact at
  `<worktree>/.flow-tmp/consolidator-result.json` using the atomic
  write-`.tmp` → validate-`.tmp` → `mv` idiom.

The Consolidator-Validator's job is NOT:

- It does **not** post PR comments — that's Step 8c of `/flow-pr-review`'s
  wrapper.
- It does **not** apply fixes — that's the Fix-Applier Subagent (Step 8).
- It does **not** read review-comment threads or static-analysis output
  — those are upstream agent inputs already digested at Step 3.

## 2. Inputs

The wrapper passes you these inputs in its spawn prompt:

- The absolute worktree path (your working directory).
- The absolute skill base directory (`SKILL_DIR`).
- Six absolute per-agent output paths, one each:
  - `$WORKTREE/.flow-tmp/agent-output-bug-detection.json`
  - `$WORKTREE/.flow-tmp/agent-output-security.json`
  - `$WORKTREE/.flow-tmp/agent-output-pattern-consistency.json`
  - `$WORKTREE/.flow-tmp/agent-output-performance.json`
  - `$WORKTREE/.flow-tmp/agent-output-supply-chain.json`
  - `$WORKTREE/.flow-tmp/agent-output-test-coverage.json`
- Two OPTIONAL paths (seventh and eighth):
  - `$WORKTREE/.flow-tmp/agent-output-gemini.json` — the cross-model Gemini
    lens, **tolerated-absent**. Present and valid only when the
    `review.gemini` opt-in is enabled AND the lens ran successfully
    (`/flow-pr-review` Step 3 "Cross-model (Gemini) lens"). When the file
    is missing (lens disabled or gracefully skipped) you proceed with the
    six Claude outputs and DO NOT escalate `consolidator-missing-artifact`
    — that escalation is scoped to the six MANDATORY Claude lenses only.
    When present, merge it identically to a Claude lens.
  - `$WORKTREE/.flow-tmp/agent-output-product.json` — the brief-gated
    product lens, **tolerated-absent**. Present and valid only when a
    product brief resolved (`/flow-pr-review` Step 3's Product Agent).
    When the file is missing (no brief resolved) you proceed with the six
    Claude outputs and DO NOT escalate `consolidator-missing-artifact` —
    same non-mandatory treatment as the Gemini lens. When present, merge
    it identically to a Claude lens.
- The static-analysis JSON path at
  `$WORKTREE/.flow-tmp/static-analysis.json` (for context only — you
  don't re-derive from it; the agents already absorbed it).
- The PR diff path (`DIFF_PATH`, typically
  `$WORKTREE/.flow-tmp/diff.txt` — `flow-review-scope`'s output, full or
  delta)
  and the PR metadata (`PR_METADATA`, the `gh pr view --json
number,title,headRefName,baseRefName` JSON the wrapper saved to
  scratch, or the same fields inlined). Both feed the second-opinion
  validation pass's in-scope-of-diff check below.
- `REVIEW_SCOPE_PATH` (`$WORKTREE/.flow-tmp/review-scope.json`) — the
  `flow-review-scope` verdict: `scope` (`full`|`delta`), `delta_files[]`,
  and the lens gate verdicts. Feeds (d2) below.
- The absolute artifact path to write
  (`$WORKTREE/.flow-tmp/consolidator-result.json`).
- Any cross-lens agreement or dedup claim in the spawn prompt is an
  unverified pointer — check it against each `agent-output-<lens>.json`
  yourself, never merge or drop a finding on the prompt's say-so, and record
  a claim that fails the check in `anti_patterns_found`.

No new input paths for the lens negative-findings channel. The tri-state
(`"populated" | "empty" | "absent"`) comes off the per-lens
`flow-agent-finding-schema --validate` stdout envelope you already invoke in
step (a) below, which reports a `negatives: {rejected_alternatives,
anti_patterns_found}` field alongside `{ok: true}` on the per-agent success
path — but the envelope carries only that tri-state plus a cumulative
`skipped` count, never the entries themselves. The entries you collect and
tag with `lens: "<name>"` in step (b) are read directly from each per-lens
artifact JSON file at the six/seven paths above, not from the envelope.

## 3. Procedure

### (a) Read and validate each per-agent output

For each of the six per-agent paths, run:

```bash
flow-agent-finding-schema --validate "$PER_AGENT_PATH"
```

The validator now coerces trivially-fixable label/decoration drift (strips a
single surrounding-paren pair from `decoration`; maps lens-name / `add-a-test`
/ `doc-fix` labels to `suggestion`) BEFORE validating, so
`consolidator-schema-failure` fires only on genuinely-unparseable findings;
`bin/lib/agent-finding-schema.ts` is the authoritative source for exactly what

From gavingolden/flow, MIT.

More from gavingolden/flow