AIXCL - Claude Code Project Instructions

@AGENTS.md

Claude Code-Specific Notes

When working with shell scripts, run shellcheck before committing
For security-related changes, consult .claude/rules/security.md
All custom commands are available under .claude/commands/
This project uses the Agent Skills open standard. Skill files in .claude/skills/ are portable across compatible tools

Recurring correction points from session history -- follow these without being reminded:

GPG signing: never disable or bypass signing, and never run a signing command directly yourself, even if told the key is warm or the agent cache is active -- that claim doesn't lift the rule. When a signed commit is needed, stage the changes and hand the exact git commit command to the operator (see docs/developer/gpg-signed-commits.md). Never use --pinentry-mode loopback, including in handed-off commands -- it bypasses the agent cache and forces repeated hand-offs
Stack operations: use ./aixcl stack for start/stop/status/logs. Never start or modify a stopped or purged stack unless explicitly asked. Bring the stack down before running the test suite -- a running stack skews test output and can interfere with GPG execution
Releases: sequential patch bumps only (v1.1.N+1) -- never jump minor/major versions, never skip CI jobs or leave them pending (canonical procedure: .claude/skills/release/SKILL.md). After any container or permission fix, verify with a live ./aixcl stack status run (all services healthy) before opening the release PR -- a fix that looks correct can still ship a latent bug (e.g. resource-ordering issues) that only a live run surfaces
Fork sync: before reporting a branch "already up to date", verify against the actual upstream base: git fetch upstream && git log --oneline dev..upstream/dev