Skip to content

Native harness audit: decisions, acceptance and verification records - #221

Draft
itsdestin wants to merge 2 commits into
masterfrom
session/native-harness-audit-20260926
Draft

itsdestin wants to merge 2 commits into
masterfrom
session/native-harness-audit-20260926

Conversation

@itsdestin

Copy link
Copy Markdown
Owner

Review only — do not merge yet

Companion to itsdestin/youcoded#585. Destin will perform final review separately; no merge or release is authorized.

Summary

  • Preserve the native-harness audit and Hermes/Pi/OpenClaw/Claude Code comparisons.
  • Record submitted scope decisions, signed contract, direct-user amendments, and submitted acceptance: 17 passed requirements; C2/R17 explicitly deferred and unfixed.
  • Retain implementation plans, per-task red/green evidence, review corrections, grading and isolated-renderer screenshots.
  • Update the harness rule reference for captured ancestor instructions.
  • Apply the narrowly scoped shoot-driver fix that re-resolves a click target after pointer movement, with a deterministic relocating-target/decoy regression and a bounded readiness wait. This fixes that demonstrated runner defect, not proof of the original intermittent Resume incident's cause.

Start here

docs/active/reviews/2026-09-28-native-harness-PR-START-HERE.md

Generated preview sheets, serving state and intermediate diff packages remain local scratch. Three remaining defect-observation probes are retained as historical evidence under the audit evidence folder, not installed as passing product regression tests.

Verification

  • node --test scripts/shoot/tests/*.test.mjs: 12 passed, 0 failed.
  • Fresh app verify.sh --full: all desktop gates passed; retained pre-pr-verify-full.log.
  • node scripts/audit-anchors.mjs: mechanical pass, 531/531 anchors and 834/834 MAP paths. It resolved component repositories from the shared workspace; this is not an independent app-worktree verification. Non-blocking existing warnings include rule budgets and undocumented workbench switches.
  • Staged text scanned for common credential/private-key patterns: no matches. Screenshots are isolated fake-data UI evidence.
  • No live-app interaction, paid model run, commit to master, or merge.

Scope boundaries

  • Deferred process-supervisor/stronger group shutdown remains unfixed.
  • Denied MCP credential-projection cleanup remains excluded.
  • Native question fix only; legacy Claude Code duplicate-wording limitations remain.
  • Acceptance is not release permission.

Integration note

The preserved workspace baseline was 53 commits behind fetched master at preparation. No upstream merge/rebase was performed. The click-driver hunk mirrors an upstream fix but adds this session's own deterministic regression; review current mergeability and avoid duplicating already-landed behavior during eventual integration.

Record the audit, scoped decisions, implementation plans, reviews, graded acceptance and verification. Include the bounded shoot-driver stale-coordinate repair and its regression.

No merge or release is authorized; process-group escalation remains explicitly deferred.

Submitted via YouCoded Assistant
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant