feat(#1797): extract challenger pass into dedicated sub-agent

[fullsend-ai-coder]

The challenger pass (step 6e in the pr-review skill) previously ran inline within the orchestrator's LLM context. The same context that synthesized findings also challenged them, reducing false-positive detection effectiveness since the LLM had already formed opinions during synthesis.

This change extracts the challenger into a dedicated sub-agent with fresh context, as specified in the parent issue #1085. The challenger sub-agent receives only the raw merged findings and the PR diff — not the orchestrator's synthesis — preserving context isolation.

Changes:

• Add sub-agents/challenger.md with adversarial review instructions,
  structured output format, and evidence-based removal requirements
• Update SKILL.md step 6e to dispatch the challenger as a sub-agent
  after dimension sub-agents complete (sequential, not parallel)
• Add challenger to the sub-agent roster table
• Define graceful degradation: if the challenger fails, fall back to
  the pre-challenger finding set with an info-level disclosure

Note: make lint could not run (Go toolchain permission denied in sandbox). Changes are markdown-only; no Go code modified.

---

Closes #1797

Post-script verification

• Branch is not main/master (agent/1797-challenger-sub-agent)
• Secret scan passed (gitleaks — 972ce7f6654f0b7edaf699298901b10db69e4e4c..HEAD)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[github-actions]

Site preview

Preview: https://fc9756fd-site.fullsend-ai.workers.dev

Commit: 38bddcd6418016360fcdfbc1e526adf37cf74574

[fullsend-ai-review]

Review

Findings

Medium

• [logic-error] internal/scaffold/fullsend-repo/skills/pr-review/SKILL.md:541 — The challenger sub-agent receives the full merged finding set from steps 6a–6d, including orchestrator-only PR-specific findings from step 6d (protected-path, scope-authorization, PR body injection). Step 6e.4 instructs the orchestrator to "Replace the merged finding set with the challenger's adjudicated_findings," but the challenger has no instruction exempting these orchestrator-only findings from removal or downgrade. If the challenger removes a protected-path finding, the orchestrator's invariant ("the presence of a protected-path finding at high or medium severity means the outcome MUST NOT be approve") would be violated. The post-review.sh script provides a backup safety net, but the review agent's own constraint would be broken.
  Remediation: Either (a) exclude step 6d findings from the context package sent to the challenger and re-merge them after consuming the challenger's output, or (b) add an explicit instruction in challenger.md that protected-path, scope-authorization, and PR-body-injection findings from orchestrator-only checks must never be removed or downgraded.

Low

• [edge-case] internal/scaffold/fullsend-repo/skills/pr-review/sub-agents/challenger.md:48 — The challenger output schema defines adjudicated_findings with extra fields (challenger_action, challenger_reason) not present in the standard finding format validated by fullsend-check-output. Step 6e.4 says to replace the merged finding set with the challenger's output, but does not instruct the orchestrator to strip challenger-specific metadata fields before writing agent-result.json. If these fields propagate, schema validation (additionalProperties: false) would reject them.

• [edge-case] internal/scaffold/fullsend-repo/skills/pr-review/SKILL.md:549 — Step 6e.5 specifies fallback when the challenger fails (timeout, error, empty response), but does not define how to handle a valid JSON response with an empty adjudicated_findings array. An empty array could mean "all findings legitimately removed" or indicate a malfunction. Consider adding a sanity check: if the challenger removes all findings when the input set contained critical/high findings, treat it as suspicious and fall back to the pre-challenger set.

Info

• [sub-agent-failure] N/A — The style-conventions sub-agent did not return findings: model claude-sonnet-4-5@20250929 not available on vertex deployment.

• [sub-agent-failure] N/A — The docs-currency sub-agent did not return findings: model claude-sonnet-4-5@20250929 not available on vertex deployment.

• [sub-agent-failure] N/A — The intent-coherence sub-agent did not return findings: model claude-sonnet-4-5@20250929 not available on vertex deployment.

[rh-hemartin]

Suggested change

	**Do not own:** Generating new findings. You only challenge, downgrade,
	or remove existing ones. If you discover a genuine issue not covered by
	any finding, note it — but your primary job is quality control of the
	existing set.
	**Do not own:** Generating new findings.
	You only challenge, downgrade,
	or remove existing ones. If you discover a genuine issue not covered by
	any finding, note it — but your primary job is quality control of the
	existing set.