[zpyoung]

Summary

SDD runs spend most of their wall clock waiting on people and re-running the same test suite on unchanged trees. In an eight-run audit, user waits totaled 839 minutes and full suites ran 4 to 10 times per review round on identical trees. This PR cuts that without loosening the final branch review.

• Tree-once test caching. A green run is recorded by (tree_identity, exact_command), so acceptance checks, wave gates, and review pre-passes never re-test identical bytes. Only orchestrator or prepass runs populate the cache, never a worker's text claim.
• Three-tier human gates. One consolidated preflight card blocks up front. Recoverable mid-run questions carry a 10-minute orca ask deadline with a recommended default. Review-finding adjudications go to a per-round digest with asynchronous user veto.
• Scaled review depth. Wave checkpoints run at standard. Final round 1 runs full deep discovery. Rounds 2 and up are scoped closure passes, and a recurring defect in the same category and seam escalates instead of getting another narrow fixer.
• Single-owner test topology. writing-plans assigns each test file to one task per wave and scopes task acceptance to owned tests. The wave gate runs the full suite once.
• Pinned effort. Implementers, promoters, refuters, and fixers run at high. Only auxiliary roles drop to medium or low.

adversarial-review gets the deterministic CLI support these need, in the script that already owns chain integrity: default diff excludes with excluded_paths[] reporting, prepass --reuse-check anchored to artifact_hash and HEAD^{tree}, and claim/gate excerpts so refuters get only the cited hunks.

Where to look

• skills/adversarial-review/scripts/adversarial-review: the largest code change (new excludes, reuse-check, excerpt handling).
• skills/subagent-driven-development/SKILL.md: the orchestration changes for steps 1, 7, 8 to 10.
• docs/quirk/specs/2026-10-16-sdd-wall-clock-optimization/: logic and tech specs. The exploration with the audit data is in docs/quirk/explorations/.

The branch also carries two rounds of adversarial-review fixes against its own diff (C1, S1, S2, X1, X2, and round 2 findings on path-target and diff-file excerpts and reuse-check target validation).

Testing

• python3 -m pytest -q tests/test_adversarial_review.py tests/test_adversarial_review_skill.py: 607 passed.
• The skill changes (SDD, writing-plans) are prompt text, so they aren't exercised end to end. [ASK: has an SDD run been done against the new flow? If so, add the result here.]

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.