goat-critique

分类: 数据与AI | 上传者: blundergoat | 下载: 0 | 版本: v1.0（最新）
当决策或分析需要多角度批评以在发货前暴露盲点时使用。
更新日志: Source: GitHub https://github.com/blundergoat/goat-flow
目录结构

当前层级: tree/main/.agents/skills/goat-critique/
📁 references/
- 📄 rubric-examples.md 4.1 KB
- 📄 sub-agent-directives.md 4.2 KB
📄 SKILL.md 17.0 KB
SKILL.md

---
name: goat-critique
description: "Use when a decision or analysis needs multi-lens critique to surface blind spots before shipping."
goat-flow-skill-version: "1.4.0"
---
# /goat-critique

## Shared Conventions

Read `.goat-flow/skill-reference/skill-preamble.md` and `.goat-flow/skill-reference/skill-conventions.md` for shared conventions before proceeding.

## When to Use

Use when a concrete artifact deserves multi-perspective critique before shipping. Takes ANY input artifact: a plan, security assessment, debug hypothesis set, review findings, test strategy, architecture proposal, or refactor approach.

**Use when:**
- The stakes justify structured critique before shipping
- You have a concrete artifact to critique (not vague ideas)
- You want competing perspectives, not just validation
- Called by another goat-* skill or directly by the user

**NOT this skill (pre-invocation routing):** Use when deciding which skill to invoke, not after explicit invocation.
- No artifact exists yet → create one first (goat-review, goat-debug, etc.)
- Simple factual question → answer directly
- Trivial artifact (hotfix, single-file change) → consider goat-review instead

| Excuse | Reality |
|--------|---------|
| "The artifact is trivial - a quick critique would cover it" | Quick mode was tried and removed. A single reviewer running lens passes in one context is self-talk under three labels, not multi-perspective critique. |
| "All three agents agree so it must be right" | Consensus without orchestrator verification is unverified self-declaration. The orchestrator's job is to verify claims, not count votes. |
| "Inline role-play is faster than spawning agents" | Agents that role-play SBAO inline produce indistinguishable perspectives. Isolated context is what makes findings independent. |
| "Closing checks happen after the main answer - skip them" | End-of-task rules have near-zero voluntary compliance. Phase 5.5 meta-audit and outcome capture exist because post-deliverable steps get skipped. |

**Direct invocation is binding.** `$goat-critique` or `/goat-critique` runs Phases 1-5 plus mandatory post-synthesis steps (5.5, 5.6). Dispatcher ambiguity rules do not override direct invocation; raise scope concerns after synthesis.

**Report-only by default.** `$goat-critique make X shorter` = critique only; `$goat-critique ... then apply it` = critique first, apply after gate. See Constraints for mutation and apply rules.

## Step 0 - Intake

goat-critique runs in one mode: full delegated, Phases 1-5 plus mandatory post-synthesis steps (5.5 meta-audit, 5.6 outcome capture), three critique sub-agents plus one lightweight meta-agent in 5.5. If an agent suggests adding a lighter mode, that suggestion is the failure this design prevents.

**Intake checklist:**
- Confirm the artifact exists and is concrete (a file, a plan document, a specific set of findings - not a vague idea).
- Select the critique rubric for the artifact type (see Critique Rubrics below). If unclear, ask the user.
- Use the preamble's grep-first learning-loop retrieval on relevant `.goat-flow/footguns/` and `.goat-flow/lessons/`; record explicit misses instead of broad-loading buckets.
- Delegation consent: proceed directly to Phase 1. Skill-chained entry: skip intake confirmation, use caller context; still run retrieval + rubric selection. All phases (1-5 + 5.5 + 5.6) always run.
- **Differential mode detection:** Check `.goat-flow/logs/critiques/` for prior critiques of the same artifact slug within 30 days. If found, offer differential mode: sub-agents A and B receive the prior critique log + artifact diff; Agent C stays cold. Phase 5 Verdict adds a delta block (Resolved/Regressed/New/Unchanged counts). Tag log as `[diff-of: <prior-uuid>]`.
- **Read context map:** Read the selected rubric's context map (see `references/rubric-examples.md`) and pass to each sub-agent's spawn directive.

## Phase 1 - Generate Competing Critiques

Spawn all three sub-agents in parallel via the Agent tool.

Context varies intentionally - informational diversity catches more than tonal diversity.

### The Core Trio Lens

Agents A and B both use the SKEPTIC/ANALYST/STRATEGIST combined lens. These three perspectives work as a unit - never split them into separate agents:

- **SKEPTIC** - "What could go wrong? What assumptions are unproven? What's the worst-case scenario?"
- **ANALYST** - "What does the evidence actually say? What's the cost/benefit? What do the numbers and code paths tell us?"
- **STRATEGIST** - "What's the fastest path to shipping? What can we defer? What's the highest-leverage change?"

All three perspectives must appear in every critique from Agents A and B. The tension between them is the point.

**Context split:**

| Agent | Reads | Does NOT read |
|---|---|---|
| A (Risk) | artifact + architecture.md + footguns + lessons + rubric | git history, config.yaml |
| B (Alternatives) | artifact + architecture.md + `git log --oneline -20` + config.yaml + rubric | footguns, lessons |
| C (Fresh Eyes) | artifact + rubric ONLY | everything else (isolation enforced) |

### Sub-Agent Definitions

Full directives: `references/sub-agent-directives.md`.

- **A (Risk):** SKEPTIC/ANALYST/STRATEGIST on risks, 2nd-order impacts, fastest safe path. Must cite downstream files by name.
- **B (Alternatives):** SKEPTIC/ANALYST/STRATEGIST on alternatives, ranked by implementation friction. Must surface at least one alternative.
- **C (Fresh Eyes):** No project context. Flags unstated assumptions. ISOLATION RULE enforced.

Each sub-agent MUST return 3-7 findings, each with: title, severity, evidence (file:line), confidence, Proof attempt, Evidence quality (OBSERVED/INFERRED/UNVERIFIED), SKEPTIC/ANALYST/STRATEGIST lines, and rubric dimensions covered. Plus: overall assessment (STRONG/ADEQUATE/WEAK/FLAWED) and one thing the artifact gets RIGHT.

**Lens-finding floor:** each lens must surface >= 1 finding per sub-agent or re-run once; convergence allowed after one re-run. See anti-fabrication constraint. Full floor spec in the sub-agent directives reference pack.

## Phase 2 - Rank and Compare

Execute in this order:

**1. Context leak scan.** Grep Agent C's output for `.goat-flow/`, `goat-*`, `architecture.md`, `config.yaml`, or project-specific namespace references. Only flag references that do NOT appear in the input artifact Agent C received — quoting terms from the artifact is expected, not a leak. Any match that cannot be traced to the artifact text = CONTEXT LEAK; discard and re-spawn with stricter isolation.

**1b. Completeness gate.** Verify each sub-agent returned required fields (see Constraints). Incomplete → re-spawn once.

**2. Classify each finding:** **Consensus** (≥2 agents, severity within ±1), **Split** (≥2 agents, severity differs ≥2 levels or explicit reject vs blocking), **Unique** (one agent only). Silence is not a dismiss; treat as Unique.

**3. Score each sub-agent's critique** on five axes: Grounding (file:line evidence?), Specificity (concrete?), Actionability (clear next step?), Coverage (rubric dimensions addressed?), Calibration (severity matches evidence?).

**4. Verify sub-agent dimension coverage.** Skim each agent's findings; confirm each claimed dimension has substantive content. Demote unsubstantiated claims. Use orchestrator-verified dimensions as input to step 5.

**5. Compute rubric coverage gates.** Unaddressed mandatory dimensions → auto-generate HIGH coverage-gap finding. Unaddressed optional → auto-generate MEDIUM.

**6. Spot-check OBSERVED claims.** For each finding marked OBSERVED, re-read the cited file:line or proof artifact. Findings that fail spot-check get tagged `[evidence-gap: spot-check failed]`; Phase 3 decides retract or upgrade.

**7. Label control group deltas.** For fresh-eyes-only findings, orchestrator assigns: **CONTEXT DRIFT** (wrong due to missing context), **READABILITY GAP** (valid for any reader), or **CONTEXT-LIMITED** (may be valid, cannot fully evaluate).

## Phase 3 - Cross-Examine

**Early exit:** If Phase 2 yields zero split findings and zero unique HIGH/CRITICAL findings, skip Phase 3. Note "no disputes - full consensus" in output and proceed to Phase 4.

If splits + unique HIGH/CRITICAL exceed the cross-examination budget, batch multiple disputes into a single agent prompt. Triage by severity - CRITICAL and HIGH first.

For each split finding, spawn a cross-exam agent: "Agent A says [X], Agent B says [Y]. Which is correct given the actual codebase?"

For unique HIGH/CRITICAL findings, spawn verification: "Only one critique raised [finding]. Genuine blind spot or false positive?"

Mark each: RESOLVED (with winner) / STILL DISPUTED / RETRACTED (false positive confirmed).

## Phase 4 - Clarify

**Persist before gate:** Before evaluating clarification questions, write Phase 1-3 results to `.goat-flow/logs/critiques/<YYYY-MM-DD>-<HHMM>-<artifact-slug>-<rand5>.md` - sub-agent summaries, comparison matrix, cross-examination outcomes. This runs regardless of whether Phase 3 took the early-exit branch.

Before synthesising, present the unresolved items to the human conversationally.

**Opener:** Lead with a one-line summary of how many decisions are needed and their titles. Example: "3 decisions before synthesis: (1) SEC-01 severity, (2) remediation path, (3) attacker model scope."

**Per-question format:** `Q[N]: [decision]? (A) [option] (B) [option] Default: [A/B]. Background: [1 sentence]`.

**Compact table (3+ questions):** `| # | Decision | Option A (default) | Option B | Why |`. Follow with: "Reply with numbers to override defaults; or approve to proceed."

Question types: (1) Disputes from Phase 3, (2) Trade-offs with two valid approaches, (3) Context drift findings - intentional vs oversight.

**Closer:** After all questions, end with: "Reply with your picks (e.g. 'A, B, go with defaults on the rest') or push back on any framing."

**If questions exist:** BLOCKING GATE - STOP and wait for human response.
**If no questions (full consensus, no trade-offs, no context drift):** CHECKPOINT - note "no disputes - proceeding to synthesis" and continue.

## Phase 5 - Synthesise

Produce the prime critique. Lead with a **Verdict** block:
- **Gate: BLOCK | CONCERNS | CLEAN** - derived from surviving findings: any CRITICAL → BLOCK, any HIGH (no CRITICAL) → CONCERNS, else CLEAN
- Assessment: STRONG / ADEQUATE / WEAK / FLAWED (synthesised from sub-agent assessments and cross-examination outcomes)
- Risk level: LOW / MEDIUM / HIGH / CRITICAL
- Top 1-3 blockers (if any) - one line each, linked to findings below
- If differential mode: append delta block (`Resolved: N | Regressed: M | New: K | Unchanged: J` vs prior critique)

Then the full critique:
- Consensus findings (preserved as-is)
- Resolved split findings (with resolution rationale)
- Human-directed findings (from Phase 4 clarification responses)
- Verified unique findings (survived cross-examination)
- Retracted findings (listed so user sees what was considered and dismissed)

**Open questions:** Items with INFERRED-only evidence, inconclusive single-agent findings, or unvalidated assumptions go here - not as recommendations. Each open question states: confidence, evidence needed to resolve, revisit trigger.

**Blind spot check:** List unaddressed artifact sections, unmapped rubric aspects, and unread referenced files as "What Wasn't Critiqued." Must never be empty.

**Proof Gate:** Apply the Proof Gate (see Constraints) to every synthesised finding before inclusion.

**Phase 5.5 - Meta-audit.** Spawn a lightweight meta-agent (budget: 2 tool calls, no context beyond the draft Phase 5 output). Audit the critique for internal consistency against the 10-point rubric in `references/rubric-examples.md`. If issues found, insert an `## Auto-Detected Issues` block before presenting. Verdict block updated with `Meta-score: N/100`.

**BLOCKING GATE:** Present the synthesised critique (including Meta-score if 5.5 produced one). "Options: (A) apply, (B) dig deeper, (C) re-run, (D) close. Default: D." After plan critique, suggest `/goat-plan`.

**Phase 5.6 - Outcome capture.** After the human picks A/B/C/D, tag each surviving finding: `accepted | rejected | deferred | partial`. Default: option (A) → all `accepted`; option (D) → all `deferred`. Persisted to the critique log under `## Outcomes`.

**Integration hooks.** Populate from surviving findings when applicable:
- `for-goat-plan` - milestone updates, reordering
- `for-goat-debug` - hypothesis seeds, evidence to capture
- `for-implementation` - immediate fixes, deferred items

Empty sections collapsed to `none`.

## Critique Rubrics

The rubric determines what sub-agents evaluate. Match to artifact type. Dimensions marked **[M]** are mandatory (unaddressed → auto-HIGH coverage-gap finding); dimensions marked **[O]** are optional (unaddressed → auto-MEDIUM). Each rubric has a context map (A/B/C file assignments) in `references/rubric-examples.md`; Step 0 reads the selected map.

**Plan:** correctness against codebase [M], integration safety [M], sequencing quality [M], validation coverage [O], task specificity [O]
**Security assessment:** threat model completeness [M], exploitability calibration [M], attack surface coverage [M], framework mitigation accuracy [O], data flow quality [O]
**Debug hypotheses:** hypothesis diversity [M], evidence quality (OBSERVED vs INFERRED) [M], elimination rigour [M], confidence calibration [O], reproduction completeness [O]
**Review findings:** severity calibration [M], diff coverage [M], pre-existing separation [M], false positive rate [O], cross-reference impact [O]
**Test strategy:** coverage gaps [M], risk-proportionate depth [M], doer-verifier separation [O], manual test specificity [O], mock awareness [O]
**Architecture/refactor:** blast radius accuracy [M], migration safety [M], backward compatibility [M], dependency impact [O], rollback feasibility [O]
**Generic (fallback):** internal consistency [M], evidence grounding [M], scope completeness [M], feasibility [M], risk identification [M]. All dimensions mandatory for the fallback rubric. If using the generic rubric, state why no specific rubric matched and which was closest.

## Constraints

- MUST run in one mode: full delegated, Phases 1-5 plus mandatory post-synthesis steps (5.5, 5.6), three critique sub-agents plus one meta-agent in 5.5. Phase 5.5 runs before the human gate; Phase 5.6 runs after the human responds. Quick/lite modes were tried and removed - a single reviewer running lens passes in one context is self-talk under three labels, not multi-perspective critique.
- Explicit `$goat-critique` or `/goat-critique` invocation IS consent to spawn sub-agents and the full protocol. Do NOT ask again.
- Report-only by default. Do not mutate the target artifact or committed files unless the user separately says to apply, edit, update, fix, or otherwise implement. If interrupted, freeze writes.
- MUST Spawn all three sub-agents in a single parallel batch. Sequential spawning loses the informational-diversity benefit.
- MUST enforce max 5 tool calls per sub-agent.
- MUST Scan Agent C output for context leaks before any other Phase 2 work. Only flag references absent from the input artifact. Any untraceable match = CONTEXT LEAK; discard and re-spawn.
- MUST Check sub-agent completeness: verify each sub-agent returned 3-7 findings plus required lens fields, severity, evidence, confidence, rubric dimensions, overall assessment, and preservation note. Incomplete → re-spawn once; if still incomplete, record `sub-agent completeness limited`.
- MUST enforce cross-examination budget: Max 3 cross-examination agents total, max 3 tool calls per agent.
- Recommendations are never auto-applied. After synthesis, stop. Do not enter implementation mode unless the user explicitly asks to apply changes.
- MUST apply the Proof Gate from `skill-preamble.md` to every synthesised finding. Sub-agent reports are inputs to verify, not evidence to launder. Re-read applies to findings surviving to Phase 5 (typically 3-7 after Phase 3/4 filtering), not to all findings raised in Phase 1.
- MUST NOT fabricate findings. Do not fabricate findings to meet the lens-finding floor; convergence allowed after one re-run.
- Universal constraints from skill-preamble.md apply.

## Output Format

**Terse-first directive:** Informational sections (Sub-Agent Comparison Matrix, Retracted Findings, What Wasn't Critiqued) default to terse: one sentence per bullet, no qualifiers, no closing offers. Gate prompts and evidence-tagged findings retain full detail.

```markdown
## Verdict  <!-- includes Gate: BLOCK|CONCERNS|CLEAN + Meta-score -->
## Critique Rubric
## Sub-Agent Comparison Matrix
## Sub-Agent Rankings
## Rubric Coverage Gaps
## Control Group Delta
## Validated Findings  <!-- source pool for Recommended Changes -->
## Cross-Examination Results
## Auto-Detected Issues  <!-- from Phase 5.5 meta-audit, if any -->
## Retracted Findings
## Human Decisions
## Strengths
## Recommended Changes  <!-- subset of Validated Findings; ordered by severity; each with concrete action -->
## Open Questions
## Integration Hooks  <!-- for-goat-plan, for-goat-debug, for-implementation -->
## What Wasn't Critiqued
## Outcomes  <!-- Phase 5.6: per-finding accepted|rejected|deferred|partial -->
```
登录后下载/点赞/收藏 ❤ 5 | ★ 0
goat-critique

目录结构

SKILL.md

举报内容

提示
