skill

ForgeZero — When to Invoke

audit

Trigger when:

• User is about to commit changes that touch .agents/skills/, .agents/rules/, or .agents/workflows/
• User says "what changed in our agent config" / "is this PR safe to merge"
• Pre-commit context, before pushing

Run: forge0 audit Exit codes: 0 = clean, 1 = error, 2 = changes detected (CI gate use)

Output includes the honest scope caveat — surface it to the user, do not paraphrase it.

provenance

Trigger when:

• User asks "what skills/KIs were active in conversation <id>"
• User asks for decision lineage, audit trail, or "why did the agent do X"
• Post-incident: explaining what the agent saw

Run: forge0 provenance <conversation-id>

Always preserve the [DETECTED_WHEN_EXPLICIT] tag. Skills loaded via system-prompt injection are NOT detectable — never claim a complete picture; the tool output is a lower bound.

share

Trigger when:

• User wants to package .agents/ for a teammate / new hire / fork
• User says "bundle our config" / "share this setup"

Run: forge0 share --tag <semver> Default scrubs secrets. Refuses with exit 3 if any detected. Refuses with exit 4 if live SKILL.md drifted from repo. On exit 4, surface the drift to the user and ask whether to run forge0 sync-skill or pass --allow-skill-drift. Do not auto-resolve. NEVER suggest --allow-secrets or --allow-skill-drift unprompted.

Hard prohibitions

• Do not invent flags. The real subcommands are: status, doctor, kickoff, audit, verify, provenance, share, receipt, ledger, selftest, sync-skill, and trace.
• Do not promise detection completeness. The honesty bound is load-bearing.
• Do not run share without explicit user request — bundle creation is destructive-feeling for users who don't expect it.