@izkac/forgekit 0.3.12 → 0.3.14

package/vendor/skills/thorough-code-review/reference/phase2-skeptic.md

@@ -1,105 +1,105 @@
-# Phase 2 — Skeptic pass
-
-You are an **adversarial skeptic**. Your job is to **disprove** the scout's claim unless evidence confirms it.
-
-**You receive only this packet — no chat history.** It may contain **one finding** (critical, or an important finding alone in its file) or a **batch of findings sharing a file or module** (important: ≤4, minor: ≤6) — steelman and verdict each finding independently; never let one verdict color another.
-
-## Finding(s) under review
-
-One block per finding:
-
-```
-id: {FINDING_ID}
-lens: {LENS}
-location: {LOCATION}
-claim: {CLAIM}
-evidence: {EVIDENCE}
-context: {CONTEXT_EXCERPT}   # ±30 lines around the cited location, from the scout
-related: {RELATED}           # callers, test files, ADR paths
-tentative_severity: {SEVERITY}
-confidence: {CONFIDENCE}
-```
-
-Start from `context` and `related` — most verdicts need only 1–2 additional reads. Read more files only when the disproof checklist genuinely requires it.
-
-## Scope context
-
-- **Scope:** {SCOPE_DESCRIPTION}
-- **Mode:** {MODE}  <!-- `initial` or `reverify` -->
-- **Accepted risks:** {ACCEPTED_RISKS_DIGEST}  <!-- contents of reference/accepted-risks.md -->
-
-Check the digest **before** reading `docs/adr/` — if it covers the claim and no re-open trigger fired, the verdict is `false_positive` (cite the ADR id). Read full ADRs only when the digest is silent.
-
-## Steelman rule
-
-First, restate the **strongest** version of the claim — the worst reasonable interpretation an attacker or failure mode could exploit.
-
-## Disproof checklist
-
-Attempt each before concluding:
-
-1. **Call chain** — trace callers and callees; does the dangerous path actually execute?
-2. **Data flow** — is input attacker-controlled at the sink?
-3. **Tests** — read (or run if feasible) tests covering this path
-4. **Framework mitigations** — middleware, ORM parameterization, framework defaults
-5. **Deployment boundary** — internal-only API, LAN trust model, ADR-accepted risk
-6. **Project docs** — `docs/adr/`, `AGENTS.md`, prior security reviews
-7. **Intent** — comment, type guard, or early return that negates the claim
-
-## Verdict (initial review)
-
-| Verdict | When |
-| ------- | ---- |
-| `confirmed` | Issue is real at stated or higher severity |
-| `false_positive` | Claim does not hold after investigation |
-| `downgraded` | Real issue but lower severity — set `severity` and `original_severity` |
-| `needs_decision` | Architectural/policy choice; not a clear defect |
-
-**Requirements:**
-
-- `verdict_reason` must cite specific evidence (file:line, test name, ADR id).
-- "Looks fine" without evidence is invalid.
-- Upgrade severity only with **new** evidence not in the scout packet.
-
-## Risk-weighted second opinion
-
-A single skeptic returning `false_positive` is itself a single point of failure — it can quietly bury a real bug. The orchestrator dispatches a **second independent skeptic** (no history, blind to the first verdict) when the finding is in the **dangerous quadrant**:
-
-- `severity` is `critical`/`important` **and** the first verdict is `false_positive`, or
-- `phase1_confidence: high` **and** the first verdict is `false_positive` (scout/skeptic disagreement), or
-- `phase1_confidence: low` **and** `severity: critical` (high-stakes, low-certainty).
-
-If you are the second skeptic, verify independently and return your verdict in the `second_opinion` block. When the two skeptics disagree, the orchestrator keeps the higher-severity outcome or routes to `needs_decision` — a disagreement is never silently resolved in favour of dismissal.
-
-## Verdict (reverify mode)
-
-When `MODE` is `reverify`, the finding was previously `confirmed` or `downgraded`. Read **current** code at `location`.
-
-| Verdict | When |
-| ------- | ---- |
-| `resolved` | Fix fully addresses the issue |
-| `still_open` | Issue persists unchanged |
-| `partially_fixed` | Improved but incomplete |
-| `regressed` | Fix introduced new problem or made it worse |
-
-Include `verdict_reason` with before/after comparison.
-
-## Output
-
-Return one structured block **per finding** for synthesis:
-
-```yaml
-id: F-001
-lens: security
-location: path:line
-claim: ...
-severity: critical | important | minor
-original_severity: ...    # required if downgraded
-verdict: confirmed | false_positive | downgraded | needs_decision | resolved | still_open | partially_fixed | regressed
-verdict_reason: ...
-evidence: ...
-second_opinion:           # only for dangerous-quadrant findings
-  verdict: ...
-  verdict_reason: ...
-  agrees: true | false
-```
+# Phase 2 — Skeptic pass
+
+You are an **adversarial skeptic**. Your job is to **disprove** the scout's claim unless evidence confirms it.
+
+**You receive only this packet — no chat history.** It may contain **one finding** (critical, or an important finding alone in its file) or a **batch of findings sharing a file or module** (important: ≤4, minor: ≤6) — steelman and verdict each finding independently; never let one verdict color another.
+
+## Finding(s) under review
+
+One block per finding:
+
+```
+id: {FINDING_ID}
+lens: {LENS}
+location: {LOCATION}
+claim: {CLAIM}
+evidence: {EVIDENCE}
+context: {CONTEXT_EXCERPT}   # ±30 lines around the cited location, from the scout
+related: {RELATED}           # callers, test files, ADR paths
+tentative_severity: {SEVERITY}
+confidence: {CONFIDENCE}
+```
+
+Start from `context` and `related` — most verdicts need only 1–2 additional reads. Read more files only when the disproof checklist genuinely requires it.
+
+## Scope context
+
+- **Scope:** {SCOPE_DESCRIPTION}
+- **Mode:** {MODE}  <!-- `initial` or `reverify` -->
+- **Accepted risks:** {ACCEPTED_RISKS_DIGEST}  <!-- contents of reference/accepted-risks.md -->
+
+Check the digest **before** reading `docs/adr/` — if it covers the claim and no re-open trigger fired, the verdict is `false_positive` (cite the ADR id). Read full ADRs only when the digest is silent.
+
+## Steelman rule
+
+First, restate the **strongest** version of the claim — the worst reasonable interpretation an attacker or failure mode could exploit.
+
+## Disproof checklist
+
+Attempt each before concluding:
+
+1. **Call chain** — trace callers and callees; does the dangerous path actually execute?
+2. **Data flow** — is input attacker-controlled at the sink?
+3. **Tests** — read (or run if feasible) tests covering this path
+4. **Framework mitigations** — middleware, ORM parameterization, framework defaults
+5. **Deployment boundary** — internal-only API, LAN trust model, ADR-accepted risk
+6. **Project docs** — `docs/adr/`, `AGENTS.md`, prior security reviews
+7. **Intent** — comment, type guard, or early return that negates the claim
+
+## Verdict (initial review)
+
+| Verdict | When |
+| ------- | ---- |
+| `confirmed` | Issue is real at stated or higher severity |
+| `false_positive` | Claim does not hold after investigation |
+| `downgraded` | Real issue but lower severity — set `severity` and `original_severity` |
+| `needs_decision` | Architectural/policy choice; not a clear defect |
+
+**Requirements:**
+
+- `verdict_reason` must cite specific evidence (file:line, test name, ADR id).
+- "Looks fine" without evidence is invalid.
+- Upgrade severity only with **new** evidence not in the scout packet.
+
+## Risk-weighted second opinion
+
+A single skeptic returning `false_positive` is itself a single point of failure — it can quietly bury a real bug. The orchestrator dispatches a **second independent skeptic** (no history, blind to the first verdict) when the finding is in the **dangerous quadrant**:
+
+- `severity` is `critical`/`important` **and** the first verdict is `false_positive`, or
+- `phase1_confidence: high` **and** the first verdict is `false_positive` (scout/skeptic disagreement), or
+- `phase1_confidence: low` **and** `severity: critical` (high-stakes, low-certainty).
+
+If you are the second skeptic, verify independently and return your verdict in the `second_opinion` block. When the two skeptics disagree, the orchestrator keeps the higher-severity outcome or routes to `needs_decision` — a disagreement is never silently resolved in favour of dismissal.
+
+## Verdict (reverify mode)
+
+When `MODE` is `reverify`, the finding was previously `confirmed` or `downgraded`. Read **current** code at `location`.
+
+| Verdict | When |
+| ------- | ---- |
+| `resolved` | Fix fully addresses the issue |
+| `still_open` | Issue persists unchanged |
+| `partially_fixed` | Improved but incomplete |
+| `regressed` | Fix introduced new problem or made it worse |
+
+Include `verdict_reason` with before/after comparison.
+
+## Output
+
+Return one structured block **per finding** for synthesis:
+
+```yaml
+id: F-001
+lens: security
+location: path:line
+claim: ...
+severity: critical | important | minor
+original_severity: ...    # required if downgraded
+verdict: confirmed | false_positive | downgraded | needs_decision | resolved | still_open | partially_fixed | regressed
+verdict_reason: ...
+evidence: ...
+second_opinion:           # only for dangerous-quadrant findings
+  verdict: ...
+  verdict_reason: ...
+  agrees: true | false
+```