CLAUDE.md

Language

• Write commit messages, pull request titles and descriptions, and comments on GitHub (reviews, replies, issues) in English.

Security

kiroku reads private AI-agent history (prompts, file paths, commit messages). Treat security as a requirement of every change, not a separate task.

• Before opening a PR, review the diff for security: untrusted input from history files, git output and the network; escaping in the generated HTML; file paths; permissions of files kiroku writes; what kiroku serve exposes; downloads and self-replacement in kiroku update and install.sh.
• History never leaves the machine. Don't add telemetry or any network request that sends history or derived data. The only network access is checking and downloading releases from GitHub.
• Everything from history, git and the network is untrusted. Escape it in the HTML view (never insert it as HTML), never pass it to a shell, and validate paths before reading or serving files.
• Files kiroku writes that contain history (HTML, JSON, archive copies, caches) must not be readable by other users (0600 files, 0700 directories).
• kiroku serve listens on 127.0.0.1 by default and keeps the Host check; anything that widens access must be opt-in and documented in SECURITY.md. The one exception is "Analyze with AI", on by default within the limits SECURITY.md sets (a confirmation for each run, loopback only, --no-ai turns it off).
• Downloads must be verified (checksums.txt, and build provenance where available) before they replace anything.
• Keep govulncheck clean. When it reports a vulnerability in code kiroku calls, fix it (update the Go toolchain or the module) before other work, and release the fix.
• GitHub Actions: pin third-party actions to a full commit SHA, give each job the least permissions it needs, and don't run untrusted code with write tokens.
• Add a test for every security fix, and note it under ### Security in CHANGELOG.md.

Decision history

Use the decision-log skill (.claude/skills/decision-log/, installed from decision-continuity; update it with npx skills add MichinaoShimizu/decision-continuity --skill decision-log -a claude-code -y and copy agents/claude/decision-log-reader.md from the same release to .claude/agents/; npx skills update would replace the copy with a symlink into .agents/. Never edit it here) to recover why something is the way it is. Write Decision Logs in English (lang=en); they go to docs/decision-logs/. They are past observations, not rules: the rules are in this file, SECURITY.md and docs/compatibility.md.