fix(workflows): give plan validation adversarial teeth

test/gh-aw-plan-lifecycle.test.ts

@@ -3,7 +3,8 @@
  *
  * Guards the long-path planning lifecycle in `workflows/squad.md` against the
  * three defects tracked by #1758, the Owner/Agent cast-Name binding of #1759,
- * and the structural research contract of #1756.
+ * the structural research contract of #1756, and adversarial validation from
+ * #1757.
  *
  * These assertions target the workflow's structural contract (labeled sections,
  * ordering hints, binding rules) rather than incidental prose — a single
@@ -25,11 +26,26 @@ function readText(filePath: string): string {
 
 /** Slice a `## skill: \`name\`` block out of the workflow markdown. */
 function skillBlock(markdown: string, name: string): string {
-  const start = markdown.indexOf(`## skill: \`${name}\``);
+  const marker = `## skill: \`${name}\``;
+  const start = markdown.indexOf(marker);
   if (start === -1) throw new Error(`skill block "${name}" not found`);
-  const rest = markdown.slice(start + 1);
-  const nextIdx = rest.indexOf('\n## skill: `');
-  return nextIdx === -1 ? markdown.slice(start) : rest.slice(0, nextIdx);
+  const bodyStart = start + marker.length;
+  const nextSkillIdx = markdown.indexOf('\n## skill: `', bodyStart);
+  const endMarkerIdx = markdown.indexOf(`\n## end skill: \`${name}\``, bodyStart);
+  const candidates = [nextSkillIdx, endMarkerIdx].filter(index => index >= 0);
+  const endIdx = candidates.length === 0 ? -1 : Math.min(...candidates);
+  return endIdx === -1 ? markdown.slice(start) : markdown.slice(start, endIdx);
+}
+
+/** Slice a `## agent: \`name\`` block out of the workflow markdown. */
+function agentBlock(markdown: string, name: string): string {
+  const marker = `## agent: \`${name}\``;
+  const start = markdown.indexOf(marker);
+  if (start === -1) throw new Error(`agent block "${name}" not found`);
+  const bodyStart = start + marker.length;
+  const rest = markdown.slice(bodyStart);
+  const nextH2 = rest.search(/\n## /);
+  return nextH2 === -1 ? rest : rest.slice(0, nextH2);
 }
 
 const squad = readText(SQUAD_WORKFLOW);
@@ -344,6 +360,147 @@ describe('#1758.3: validate precedes both accept steps', () => {
   });
 });
 
+// ---------------------------------------------------------------------------
+// #1757 — validation owns an adversarial gate fed by Fact Checker DA evidence
+// ---------------------------------------------------------------------------
+
+describe('#1757: squad-plan-validate has adversarial teeth', () => {
+  const validation = skillBlock(squad, 'squad-plan-validate');
+  const factChecker = agentBlock(squad, 'fact-checker');
+  const adversarialChecks = [
+    'Opposition steelman',
+    'Load-bearing assumptions',
+    '30-day pre-mortem',
+    'Alternative approach',
+    'Remaining risk acceptance',
+  ];
+
+  function numberedChecks(block: string): Map<number, string> {
+    return new Map(
+      [...block.matchAll(/^\|\s*(\d+)\s*\|\s*([^|]+?)\s*\|/gm)].map(match => [
+        Number(match[1]),
+        match[2].trim(),
+      ]),
+    );
+  }
+
+  function hasSubstantiveAdversarialValidation(artifact: string): boolean {
+    const requiredSections = [
+      'Steelman of the opposition',
+      'Load-bearing assumptions',
+      '30-day pre-mortem',
+      'Alternative approach',
+      'Remaining risk acceptance',
+      'Validator Synthesis',
+    ];
+
+    const sectionsAreSubstantive = requiredSections.every((heading, index) => {
+      const escaped = heading.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+      const next = requiredSections
+        .slice(index + 1)
+        .map(item => item.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'))
+        .join('|');
+      const boundary = next ? `(?=\\n### (?:${next})|$)` : '$';
+      const body = artifact.match(new RegExp(`### ${escaped}\\n([\\s\\S]*?)${boundary}`))?.[1] ?? '';
+      return /\bWHAT:/i.test(body) && /\bWHY:/i.test(body) && /\bHOW:/i.test(body);
+    });
+    const checkNumbers = [...artifact.matchAll(/^\|\s*(\d+)\s*\|/gm)].map(match =>
+      Number(match[1]),
+    );
+
+    return (
+      sectionsAreSubstantive &&
+      checkNumbers.join(',') === '1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16' &&
+      /### 30-day pre-mortem[\s\S]*\b30 days?\b/i.test(artifact) &&
+      /### Remaining risk acceptance[\s\S]*\bACCEPTED\b[\s\S]*\bowner:/i.test(artifact) &&
+      /### Validator Synthesis[\s\S]*\bvalidation decision:/i.test(artifact)
+    );
+  }
+
+  it('extracts the validation skill with its complete marker and no following skill', () => {
+    expect(validation).toMatch(/^## skill: `squad-plan-validate`/);
+    expect(validation).not.toContain('## skill: `squad-plan-accept-scope`');
+  });
+
+  it('preserves structural checks 1-10 and augments them with five adversarial checks', () => {
+    const checks = numberedChecks(validation);
+    expect([...checks.keys()].slice(0, 10)).toEqual([1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
+    expect([...checks.values()].slice(10, 15)).toEqual(adversarialChecks);
+    expect(checks.get(16)).toBe('Validator synthesis');
+  });
+
+  it('uses Fact Checker DA output as advisory evidence, never as the verdict owner', () => {
+    expect(validation).toMatch(/Use the `fact-checker` sub-agent exactly once/);
+    expect(validation).toMatch(/brief is input evidence, not a verdict/i);
+    expect(validation).toMatch(/Validation — not Fact Checker —[\s\S]*final verdict/);
+    expect(factChecker).not.toMatch(/^RESULT: (?:PASS|FAIL)$/m);
+    expect(factChecker).toMatch(/Never emit `RESULT: PASS`, `RESULT: FAIL`/);
+  });
+
+  it('requires all five DA elements with concrete semantic thresholds', () => {
+    for (const section of [
+      '##### Steelman of the opposition',
+      '##### Load-bearing assumptions',
+      '##### 30-day pre-mortem',
+      '##### Alternative approach',
+      '##### Remaining risk acceptance',
+    ]) {
+      expect(factChecker, `Fact Checker must emit "${section}"`).toContain(section);
+    }
+    expect(validation).toMatch(/strongest credible opposition steelman/i);
+    expect(validation).toMatch(/falsifiable\s+conditions/);
+    expect(validation).toMatch(/exactly 30 days after\s+execution begins/);
+    expect(validation).toMatch(/materially different alternative approach\s+sketch/);
+    expect(validation).toMatch(/explicitly\s+`ACCEPTED` with rationale and an accountable owner/);
+  });
+
+  it('fails closed when adversarial evidence or verdict ownership is missing', () => {
+    expect(validation).toMatch(
+      /sub-agent is unavailable, errors, returns an empty brief,[\s\S]*force `RESULT: FAIL`/,
+    );
+    expect(validation).toMatch(/copied verdict,[\s\S]*cannot become `RESULT: PASS`/);
+    expect(validation).toMatch(/Structural PASS alone cannot produce overall PASS/);
+  });
+
+  it('distinguishes a neatly formatted bad plan from a genuinely validated plan', () => {
+    const validationHeader = [
+      '## ✅ Squad Plan Validation — PASSED',
+      'RESULT: PASS',
+      '| Check | Status | Details |',
+      '|---|---|---|',
+    ];
+    const structurallyValidBadPlan = [
+      ...validationHeader,
+      ...Array.from({ length: 10 }, (_, index) => `| ${index + 1} | ✅ | — |`),
+    ].join('\n');
+
+    const genuinelyValidatedPlan = [
+      ...validationHeader,
+      ...Array.from({ length: 16 }, (_, index) => `| ${index + 1} | ✅ | evidence |`),
+      '### Steelman of the opposition',
+      'WHAT: Replace the batch design with streaming. WHY: the latency target makes batching unsafe. HOW: prove the target with a prototype.',
+      '### Load-bearing assumptions',
+      'WHAT: queue ordering is stable. WHY: reordering breaks reconciliation. HOW: add an ordering probe before implementation.',
+      '### 30-day pre-mortem',
+      'WHAT: The rollout is reverted in 30 days. WHY: retries amplify duplicate writes. HOW: ship idempotency keys and alerts.',
+      '### Alternative approach',
+      'WHAT: Use a durable outbox. WHY: it narrows the consistency boundary. HOW: compare operational cost before choosing.',
+      '### Remaining risk acceptance',
+      'WHAT: delayed delivery. WHY: the queue can lag. HOW: ACCEPTED with bounded SLO; owner: EECOM.',
+      '### Validator Synthesis',
+      'WHAT: Structural and adversarial evidence align. WHY: the identified failure modes are bounded. HOW: validation decision: proceed.',
+    ].join('\n');
+
+    expect(hasSubstantiveAdversarialValidation(structurallyValidBadPlan)).toBe(false);
+    expect(hasSubstantiveAdversarialValidation(genuinelyValidatedPlan)).toBe(true);
+  });
+
+  it('keeps the sub-agent outside every skill block and uses safe nested headings', () => {
+    expect(squad).toContain('## end skill: `squad-plan-activate`\n\n## agent: `fact-checker`');
+    expect(factChecker).not.toMatch(/^## (?!agent:)/m);
+  });
+});
+
 // ---------------------------------------------------------------------------
 // #1756 — structural research contract replaces the >=200-char floor
 // ---------------------------------------------------------------------------

workflows/squad.md

@@ -1244,7 +1244,10 @@ Set Implementation Plan = `✅ Done`, state = Implementation planned, next = `/s
 description: Validate the implementation plan and emit a PASS/FAIL validation artifact.
 ---
 
-Readiness gate checking plan artifacts for structural issues.
+Readiness gate checking plan artifacts for structural integrity and adversarial viability.
+Validation owns the final `RESULT: PASS` or `RESULT: FAIL`. Fact Checker's
+Devil's Advocate mode supplies advisory evidence to this gate; it never owns or
+emits the validation verdict.
 
 **Acknowledge:** `🤖 Squad is validating the plan…`
 
@@ -1270,8 +1273,14 @@ Find the latest `program`, `implementation`, and `triage` artifacts. At minimum
 | 8 | Orphaned items | Both | Triage items not in program |
 | 9 | Milestone gaps | Program | Epics not in any milestone |
 | 10 | Non-roster owner/agent | Both | An `Owner`/`Agent` value is absent from the roster (see below) |
+| 11 | Opposition steelman | Adversarial | The strongest credible counter-argument is absent, weakened, or unanswered |
+| 12 | Load-bearing assumptions | Adversarial | Assumptions whose failure invalidates the plan are missing, unfalsifiable, or lack impact |
+| 13 | 30-day pre-mortem | Adversarial | No concrete causal failure scenario explains how this plan fails within 30 days |
+| 14 | Alternative approach | Adversarial | No materially different approach is sketched with trade-offs against the chosen plan |
+| 15 | Remaining risk acceptance | Adversarial | A remaining risk lacks explicit acceptance with rationale and an accountable owner |
+| 16 | Validator synthesis | All evidence | The verdict is copied from advisory output, lacks independent synthesis, or advisory evidence is unavailable |
 
-Severity: ❌ Critical (blocks acceptance): 1–6, 8, 10. ⚠️ Warning: 7, 9, borderline 5.
+Severity: ❌ Critical (blocks acceptance): 1–6, 8, 10–16. ⚠️ Warning: 7, 9, borderline 5.
 
 ###### Check 10 — roster binding (the `Name` column is the sole source of truth)
 
@@ -1289,13 +1298,60 @@ Run mechanically; never accept a value because it "looks like" a teammate.
    Never report a value as a valid roster name unless TG-2 emitted it as a
    `ROSTER_MEMBER:`.
 
+###### Checks 11–15 — Fact Checker Devil's Advocate evidence
+
+Use the `fact-checker` sub-agent exactly once. Give it the complete latest
+program, implementation, and triage artifacts and request its Devil's Advocate
+brief. The brief is input evidence, not a verdict: Fact Checker is advisory and
+must not emit `RESULT: PASS`, `RESULT: FAIL`, or decide whether acceptance may
+proceed.
+
+Fail closed when the sub-agent is unavailable, errors, returns an empty brief,
+or omits any required evidence. Do not invent or silently backfill missing
+evidence. Report the affected check ❌ Critical and force `RESULT: FAIL`.
+
+Evaluate the returned evidence against the actual plan:
+
+1. **Check 11:** Require the strongest credible opposition steelman, not a
+   strawman. It must identify what the opposition would choose instead and why.
+2. **Check 12:** Require load-bearing assumptions stated as falsifiable
+   conditions, with the plan impact if each is untrue.
+3. **Check 13:** Require a concrete failure scenario exactly 30 days after
+   execution begins, including a causal chain and observable warning signs.
+4. **Check 14:** Require at least one materially different alternative approach
+   sketch with its principal trade-offs against the chosen plan.
+5. **Check 15:** Name every remaining risk and assign the validation gate's
+   disposition. PASS requires each remaining risk to be explicitly
+   `ACCEPTED` with rationale and an accountable owner, or eliminated by a plan
+   revision. `MITIGATION REQUIRED`, `REJECTED`, or an omitted disposition is
+   ❌ Critical.
+
+Each evidence item must state WHAT the challenge is, WHY it could invalidate or
+outperform the plan, and HOW the plan addresses it or consciously accepts it.
+A structurally clean plan with weak, generic, or missing adversarial evidence
+fails checks 11–15.
+
+###### Check 16 — validator synthesis and verdict ownership
+
+After scoring checks 1–15, independently synthesize the structural findings,
+Fact Checker evidence, plan responses, and remaining-risk dispositions. Explain
+why the whole plan should or should not proceed. Validation — not Fact Checker —
+then determines and emits the final verdict.
+
+Never forward an advisory conclusion as the gate decision. A Fact Checker
+verdict, a copied verdict, unavailable advisory evidence, or a verdict without
+this synthesis is Check 16 ❌ Critical and cannot become `RESULT: PASS`.
+
 ##### Step 3: Post Result
 
 `add-comment` with `data: {"squad_artifact":"validation","schema_version":"1","origin_issue":{issue_number},"phases":[]}`. Keep `RESULT: PASS` or `RESULT: FAIL` in the human-readable body.
 
-Structure: `## ✅/❌ Squad Plan Validation — PASSED/FAILED` → Validated artifacts + timestamp → Results table (Check|Status|Details) → Issues Found (Critical ❌ then Warnings ⚠️ with fix instructions) → Summary (counts + verdict) → Next action.
+Structure: `## ✅/❌ Squad Plan Validation — PASSED/FAILED` → Validated artifacts + timestamp → Results table (Check|Status|Details; all checks 1–16) → Adversarial Evidence (Steelman of the opposition; Load-bearing assumptions; 30-day pre-mortem; Alternative approach; Remaining risk acceptance, each with WHAT/WHY/HOW) → Validator Synthesis (independent reasoning across structural and adversarial evidence) → Issues Found (Critical ❌ then Warnings ⚠️ with fix instructions) → Summary (counts + validator-owned verdict) → Next action.
 
 Rules: any ❌ = FAILED heading+verdict. Warnings alone ≠ failure.
+Structural PASS alone cannot produce overall PASS. Overall PASS requires checks
+1–16 to have run, complete adversarial evidence, explicit remaining-risk
+acceptance, and validator-owned synthesis.
 
 ##### Step 4: Update Lifecycle
 
@@ -1515,3 +1571,49 @@ Terminal (last phase): emit `data: {"squad_artifact":"activated","schema_version
 
 Phase: `🔄 Phase {N} of {total} activated`. Next: accept/activate next phase.
 Full/last: `✅ Done`, state = Activated. Terminal — no next action needed.
+
+## end skill: `squad-plan-activate`
+
+## agent: `fact-checker`
+---
+description: "Produces advisory Devil's Advocate evidence for plan validation"
+model: inherited
+---
+
+Operate only in Fact Checker's Devil's Advocate mode. Review the complete
+program, implementation, and triage artifacts supplied by the parent validator.
+Challenge the plan rigorously but constructively. Every finding states WHAT the
+challenge is, WHY it could invalidate or outperform the plan, and HOW the plan
+could address it.
+
+Return exactly these five evidence sections:
+
+##### Steelman of the opposition
+
+Give the strongest credible counter-argument and say what its advocates would
+choose instead. Never substitute an easy-to-dismiss strawman.
+
+##### Load-bearing assumptions
+
+List falsifiable assumptions whose failure would invalidate the plan, and state
+the plan impact of each failure.
+
+##### 30-day pre-mortem
+
+Describe a concrete failure exactly 30 days after execution begins, including
+the causal chain and observable warning signs.
+
+##### Alternative approach
+
+Sketch at least one materially different approach and compare its principal
+trade-offs with the chosen plan.
+
+##### Remaining risk acceptance
+
+Name the remaining risks and the conditions, rationale, and accountable owner
+needed to accept or mitigate each one.
+
+This brief is advisory evidence only. Never emit `RESULT: PASS`, `RESULT: FAIL`,
+or a final recommendation to accept/reject the plan. If an artifact or required
+analysis is unavailable, write `EVIDENCE UNAVAILABLE: {reason}` in the affected
+section instead of fabricating content.