← ClaudeAtlas

workbench-sessionlisted

Run the AI-workbench session loop in the current repo. Use at the start of a work session to orient, brief and triage, and at the end to capture decisions, lessons and entity updates before committing. Applies to any repo whose AGENTS.md carries the ai-workbench pattern marker.
maurice-jobst/ai-workbench · ★ 0 · AI & Automation · score 73
Install: claude install-skill maurice-jobst/ai-workbench
# Workbench Session Execute the session protocol the current repo declares. This skill carries the procedure, never the doctrine: the steps come from the repo's own `AGENTS.md`, so a repo that deviates from the template gets its deviation honoured. ## 1. Resolve the pattern root First hit wins: 1. `$WORKBENCH_HOME`, if it contains `templates/AGENTS.md`. 2. This skill's base directory, walked up until a directory holds both `templates/` and `scripts/workbench_lint.py`. 3. Neither. Carry on without it and say the lint comparison was skipped. The pattern root is only needed for that comparison, and a repo-local copy of this skill normally lands here: a workbench holds `scripts/` but not `templates/`, so the walk up finds nothing. `workbench-scaffold` audit mode catches a drifted lint or a stale copy of this skill. ## 2. Preflight Get the repo root with `git rev-parse --show-toplevel`, then read `AGENTS.md`. Stop if either holds: - `AGENTS.md` is missing, or contains no `<!-- ai-workbench pattern` marker. - It has no numbered session protocol under the working agreement. - The folder is not a git repository at all, so there is no state to read. Output only which of the two failed and that `workbench-scaffold` is the fix. No briefing, no lint run, no partial loop. Never operate a folder that is not a workbench, and never substitute a default protocol for a missing one. ## 3. Pick the phase An argument overrides detection: `brief` forces open, `close` force