Repository navigation
docs/CLAUDE_CODE.md: say what ran live (the unguarded 2026-10-30 canary) and what did not - #174
Conversation
…ry) and what did not The "Not verified here" paragraph said no real Claude Code and no real GitHits ran. The same page (Limits) and claude_code.BARE_HOOKS record the 2026-10-30 pilot canary: Claude Code 2.1.312 on a Linux VM ran the launcher against the real GitHits under --bare, unguarded. The paragraph now says that, and keeps what is still unverified: no guarded live lookup (the api-hooked hook canary is unrun, per GITHITS_OPERATOR_HANDOFF.md) and the Windows Bash rule match. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
| **The skill.** `skills/githits/SKILL.md` adapts GitHits' shipped skills (`githits-com/githits-cli` at `0bc9f5e68b`, `skills/githits-code/SKILL.md` and `skills/githits-mcp/SKILL.md`, Apache-2.0, Copyright 2025 GitHits Inc.; `skills/githits/LICENSE` is the upstream license and `skills/githits/NOTICE` says what changed) for the launcher: when a lookup is worth it (an API or version fact the repository and the standard library do not settle), how to phrase a query (short, naming the package and version, no code or repository text), targets first, search then grep or list then read, results as untrusted data, and citing sources in the final reply (the harness does not commit). Claude Code 2.1.309's `--help` offers `--append-system-prompt`, `--plugin-dir` (skills inside a plugin) and `--add-dir`; the provider uses `--append-system-prompt`, so the text is part of arm B's system prompt on every request rather than a skill the model may or may not load, and arm A's prompt is exactly the harness note. The YAML front matter is left out, and `{launcher}` is replaced by the call's launcher path, so arm B's exact appended text is the harness note, a blank line, and the file's body with that one substitution; the file's SHA-256 is on `signal.harness.skill`, computed with CRLF normalized to LF, so a Windows checkout with `core.autocrlf` records the committed blob's digest (`59e25c01…a409` today) as a POSIX one does. `.gitattributes` (`skills/** text eol=lf`) also checks the skill out with LF on every platform. | ||
|
|
||
| **Not verified here.** No real Claude Code and no real GitHits ran. The tests use a stand-in `claude` whose Bash tool is emulated (a command no `Bash(<prefix>:*)` rule names is denied; the launcher's `exec` line is run as `sh` would run it) and a loopback fake GitHits; a separate test runs the launcher under a real `sh` when one is on `PATH`. Whether Claude Code 2.1.309 matches `Bash(<absolute path>:*)` against a Windows path with forward slashes, and runs it through Git Bash, is for the first live call: `syberlabs githits doctor`, then one guarded lookup whose record and fact are checked. | ||
| **What ran live, and what did not.** The tests use a stand-in `claude` whose Bash tool is emulated (a command no `Bash(<prefix>:*)` rule names is denied; the launcher's `exec` line is run as `sh` would run it) and a loopback fake GitHits; a separate test runs the launcher under a real `sh` when one is on `PATH`. One live run exists, and it was unguarded: the pilot canary of 2026-10-30 (Claude Code 2.1.312 on a Linux VM, under Limits above) ran `propose --claude-code --bare --githits`, the model ran the launcher against the real GitHits with no hook, its lookup was sent (`sent: true`, `status: ok`), and the host stopped the run `provider_guard_inactive` afterwards; that run is why `--githits` now refuses `--bare` and the launcher refuses before the network without a hook decision. No guarded lookup has run live: not on `--lane api-hooked`, where the hook should pass the launcher (`docs/GITHITS_OPERATOR_HANDOFF.md`: the live hook canary remains unrun), and not on Windows, so whether Claude Code matches `Bash(<absolute path>:*)` against a Windows path with forward slashes, and runs it through Git Bash, is still open. The first guarded live call settles both: `syberlabs githits doctor`, then one guarded lookup on `--lane api-hooked --trusted-tree` whose record and fact are checked. |
There was a problem hiding this comment.
The first guarded live call only “settles both” if it runs on Windows through Git Bash and uses --githits. A successful Linux lookup cannot establish that the Windows Bash(<absolute path>:*) rule works. Specify those conditions in the final sentence so readers do not treat a Linux result as closing the Windows check.
Prompt To Fix With AI
This is a comment left during a code review.
Path: docs/CLAUDE_CODE.md
Line: 188
Comment:
**Windows check remains open**
The first guarded live call only “settles both” if it runs on Windows through Git Bash and uses `--githits`. A successful Linux lookup cannot establish that the Windows `Bash(<absolute path>:*)` rule works. Specify those conditions in the final sentence so readers do not treat a Linux result as closing the Windows check.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
|
Independent review (OpenAI gpt-5.5-2026-04-23): APPROVED at The docs update directly addresses the request by replacing the stale “No real Claude Code and no real GitHits ran” statement with a record of the unguarded 2026-10-30 canary and clear notes about the guarded live lookup and Windows path behavior still not having run. The change is limited to the requested documentation, does not introduce external effects or credential/deploy concerns, and the cited doc-reading tests were run. No findings. Response |
Admission
Class: R
Agent-platform: claude
Reviewed-by: openai:gpt-5.5-2026-04-23 (response resp_024e02bdca6f5232006aeb4406771087d1948979ef4ade25cb, head dadfeab, tier low, run by the PM (Claude) on behalf of sykosyber)
Admitted-by: self (Class R)
Evidence: tested
Objective
Fix one stale paragraph in
docs/CLAUDE_CODE.md(line 188 on360de0f). It said "No real Claude Code and no real GitHits ran." The same page, under Limits (line 136), andsyberlabs/claude_code.py:166(BARE_HOOKS) record the 2026-10-30 pilot canary. On that run Claude Code 2.1.312, on a Linux VM, ran the launcher against the real GitHits under--barewith no guard, and its lookup was sent.What changed
docs/CLAUDE_CODE.md, one paragraph. Its heading changes from "Not verified here" to "What ran live, and what did not". It now says:sent: true,status: ok, thenprovider_guard_inactive) and the two changes it caused;api-hookedhook canary has not run (docs/GITHITS_OPERATOR_HANDOFF.md:39: "the live hook canary remains unrun"). The WindowsBash(<absolute path>:*)match through Git Bash is untested;syberlabs githits doctor, then one guarded lookup on--lane api-hooked --trusted-tree.What did not change
scripts/pilot_dry_run.py(open PR Add scripts/pilot_dry_run.py: the GitHits pilot's dry run as one command (stacked on #141) #143) are untouched.docs/attest.policy.yamldoes not coverCLAUDE_CODE.md, so G5 does not apply.Acceptance criteria
api-hookedlookup and the Windows Bash rule match as unrun, citing the operator handoff.docs/CLAUDE_CODE.mdpass (below).Evidence
Windows 11, Python 3.12, run from the checkout:
These are the two tests that read
docs/CLAUDE_CODE.md: 2 run, 2 passed, 0 failed. I did not run the full suite locally. GitHub Actions on this PR runs it.CI
The full suite and the gates run on this pull request. See the checks tab.
Known limitations
docs/CLAUDE_CODE.md:136,syberlabs/claude_code.py:166,tests/test_githits.py:1397).Class: R
Agent-platform: claude
Reviewed-by: pending: needs a reviewer on a platform other than claude
Review-tier: low
🤖 Generated with Claude Code
The PR appears safe to merge, with a non-blocking clarification about what would verify the Windows launcher.
Fix with agent prompt
Summary
Updates
docs/CLAUDE_CODE.mdto distinguish the live, unguarded 2026-10-30 canary from the guarded checks that remain unrun.Reviews (1) · Last reviewed commit: "docs/CLAUDE_CODE.md: say what ran live (..." · Reviewed by Greptile