Quartermaster

Quartermaster

Quartermaster is a non-destructive lifecycle manager for coding-agent skills. It moves skills along a tiered ladder based on real usage and never removes anything from disk without explicit human approval.

Use the bundled qm CLI for all operations. It uses runtime adapters for Claude Code, Codex, Copilot CLI, VS Code, and generic command agents; logs every change to a reversible audit trail; and prints the token-saved number.

The state ladder

Demote, hide, and archive are safe and reversible. Delete is the only destructive action and requires explicit --yes confirmation.

Running the CLI

python3 "${CLAUDE_PLUGIN_ROOT}/bin/qm" <command>
# or, if bin/ is on PATH:
qm <command>

Commands

• qm runtimes — list supported runtime adapters.
• qm runtime-setup [runtime] [--all] — write workspace-local setup files for Claude, Codex, Copilot CLI, VS Code, or generic command agents.
• qm status [--layers] [--all] — table of every skill, state, last-used date, token cost, optional metadata layer/priority, and optional archived skills.
• qm compile "<project intent>" — propose an active loadout matched to the project. Shows layered keep/demote/block decisions, guardrails, conflicts, and a loadout manifest path when applied.
• qm review — show proposed demotions/promotions/hides/archives/restores and approve them in one batch.
• qm restore <skill> / qm activate <skill> — bring a skill back to active, including from archive.
• qm demote <skill> — take a skill out of auto-selection (manual-only).
• qm hide <skill> — remove a skill from context entirely.
• qm archive <skill> --yes — move a skill out of active roots into reversible archive storage.
• qm delete <skill> --yes — permanently remove a (hidden) skill. Destructive.
• qm revert [-n N] [--skill X] — undo the last N automatic changes (one-click revert). Skips deletions/admissions, which stay human-gated.
• qm history <skill> — inspect historical usage, selection, state, and metadata for a skill.
• qm conflicts — report explicit and inferred skill conflicts.
• qm sources — list curated external public skill sources to consider.
• qm intake <local_repo> --dry-run — scan a local external skill repo without executing code; --yes imports only accepted high-value candidates.
• qm log — print the audit trail of every transition.

Authoring arm (the capability-gap loop)

• qm gap "<what you needed>" — record a capability gap: a need with no matching skill. When the same gap recurs and nothing on the shelf covers it, Quartermaster recommends authoring a new skill.
• qm gaps — show clustered gaps and which ones warrant a new skill.
• qm author <name> [--desc ...] [--brief ...] — scaffold a new skill and admit it as active, probationary (on trial). This writes a stub only.
• qm graduate <skill> — end a skill's probation once it has proven useful.

Feedback (plain-language complaints)

• qm feedback "<what's off>" — route a natural-language complaint to the right lever. A style miss is appended to the always-on style file; a capability miss is recorded as a gap (feeding the authoring arm); "stop suggesting X" / "I keep using X" become demote/promote suggestions for a named skill. Only the local style note and gap log are written automatically — skill moves are suggested unless you pass --apply.

Probation

A newly authored skill is admitted probationary: it is active and usable immediately, but on trial. qm review proposes graduating it if it gets used, or demoting it if it sits unused past the probation window. qm status marks probationary skills with ◉ prob.

How to behave when using this skill

1. Default to non-destructive verbs. Reach for demote, hide, compile, archive, and review. These are always safe.
2. Never run qm delete on your own initiative. Only delete when the user explicitly asks to permanently remove a specific skill, and surface that it is irreversible before doing so. Pass --yes only after the user confirms.
3. Show the numbers. After a lifecycle change, run qm status so the user sees the updated loadout and token savings.
4. Apply in batches with approval. For compile and review, present the proposed plan first; apply only after the user agrees (or pass --yes when they've already said go).
5. Reassure on reversibility. Every non-delete change is undone with qm restore <skill>; mention this when you demote, hide, or archive something.
6. Author through skill-creator. When qm gaps recommends a new skill (or the user asks for one), run qm author <name> to create the probationary stub, then invoke the skill-creator skill to write the actual instructions into the generated SKILL.md. qm author only scaffolds — it never writes the skill's content itself. Leave the skill probationary until the user confirms it works, then qm graduate <name>.
7. Turn complaints into action. When the user gripes in plain language ("this isn't matching my style", "I keep needing X", "stop suggesting Y"), run qm feedback "<their words>". Apply the style note / gap automatically, but confirm before moving a skill (or pass --apply once they agree).
8. Undo is one command. Any automatic change can be reversed with qm revert (last change) or qm restore <skill>. Offer it freely.
9. Respect guardrails. If qm compile reports a blocked risky skill or a guardrail conflict, do not force the action. Explain the missing guardrail or conflict and ask the user how to resolve it.
10. Treat external repos as untrusted. Use qm sources to suggest candidate repos, ask the user to clone one, run qm intake <path> --dry-run, and only import with --yes after accepted candidates are reviewed.

Notes

• Telemetry is local-only (under ~/.quartermaster/, or $QM_HOME). Nothing is sent over the network.
• To manage a specific skills directory, pass --skills-dir <path> or set QM_SKILLS_DIR.
• To select an agent runtime, pass --runtime claude|codex|copilot|vscode|generic or set QM_RUNTIME. Claude remains the default adapter.
• To prepare workspace-local runtime files, run qm runtime-setup <runtime> or qm runtime-setup --all.