Run go reviews in fresh Claude and Codex processes

[haacked]

• go runs /review-code --fix in a fresh process of the active harness, using Claude Code or Codex as appropriate. It saves review state and a validated result under the worktree's .notes/ directory so the parent can resume after a context clear.
• Each attempt keeps its sessions, reports, temporary worktrees, and hook files under .notes/go-reviews/. After validation, the parent archives the report in shared review history.
• Codex installs go and uses its own bounded CI route. Its review child keeps native sandbox permissions and has no write grant to shared review caches.

Depends on haacked/review-code#171. The installed review-code skill must support REVIEW_CODE_REVIEW_DIR before the fresh review runner can start.

Test plan

• Passed 38 runner tests, 30 skill spec checks, 5 canonical skill checks, 75 installer checks, portable skill checks, Ruff, and git diff --check.
• A find-only smoke test passed with Claude Code and the app-bundled Codex CLI. A full review and fix cycle has not been run.

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved critical and moderate issues affect Claude compatibility, fix execution, path safety, and test coverage.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 4 · 2

Open (6)

• Use the explicit skill path for the review runner · New
• Replace unsupported Claude permission-prompts option · New
• Grant workspace-write sandbox access to Codex subprocess · New
• Reject symlinked notes directories before creating state · New
• Select the correct review-code install path by harness · New
• Select the archive root for the active harness · New

What changed in this PR

Adds fresh Claude/Codex review execution for /go, including resumable state, isolated artifacts, Codex support, and expanded testing.

Changes:

• Adds the fresh-process review runner and tests.
• Updates /go documentation and Codex CI guidance.
• Enables Codex installation and CI coverage.

File	Review summary
ai/​tests/​test-ai-installers.sh	Installer coverage reviewed; no final comments.
ai/​skills/​go/​SKILL.md	Critical (1 vote): Commands at lines 284 and 291 resolve the runner from the implementation worktree instead of the installed skill path.
ai/​skills/​go/​scripts/​tests/​test_run_review.py	Moderate (1 vote): The symlink test does not target the randomized temporary file actually opened by save_in_directory.
ai/​skills/​go/​scripts/​run-review.py	Critical: unsupported Claude option (1 vote); missing Codex workspace-write sandbox (1 vote); unsafe pre-existing .notes symlink handling (1 vote). Moderate: hard-coded Codex install lookup (2 votes); hard-coded Codex archive destination (2 votes).
ai/​skills/​go/​references/​codex-ci.md	Codex CI guidance reviewed; no final comments.
ai/​README.md	Documentation reviewed; no final comments.
ai/​codex/​excluded-skills.txt	Codex skill configuration reviewed; no final comments.
.github/​workflows/​test.yml	CI workflow reviewed; no final comments.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved critical and moderate findings affect artifact safety, permissions, and review correctness.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 1

Open (2)

• Reject symlinked cached review artifacts before hashing · New
• Select child and archive paths from the active harness root · New

Resolved since last review (6)

• Reject symlinked notes directories before creating state
• Grant workspace-write sandbox access to Codex subprocess
• Replace unsupported Claude permission-prompts option
• Use the explicit skill path for the review runner
• Select the archive root for the active harness
• Select the correct review-code install path by harness

Previously missed (1)

In code that hasn't changed since last review

Configure Claude permissions for noninteractive review execution

ai/​skills/​go/​scripts/​run-review.py:171

--permission-prompts none makes any Claude tool call that is not already allowed by the user's permission configuration fail immediately instead of prompting. Since this process must run the full review and apply --fix, a normal/default Claude configuration can leave the child with no ability to run the review scripts or edit the checkout, while the prompt below says to keep normal permission checks and the existing noninteractive review path does not force this flag (bin/run-pr-reviews.sh:709-710). Please use an explicit noninteractive permission configuration that grants only the review's required operations, or do not start the child when that configuration is unavailable.