Maestro Knowledge

1. Explicit keyword present → use its execution document or direct CLI lifecycle command (deterministic shortcut).
2. Otherwise infer from the intent (see the table above), e.g. "审计/检查知识库" → audit, "从工件/session 提取" → harvest, "知识图谱/wiki 健康" → wiki, "注册术语 X" → domain.
3. review / stage / promote map directly to the corresponding maestro knowledge CLI. review --refresh includes reconciliation; review --resolve includes disposition resolution; stage --signal --signal-ids includes signal recording. Preserve stable knowledge IDs, graph aliases, Run ID, Session ID, signal, candidate ID, disposition, target, and reason exactly; do not translate these operations into direct spec/knowhow writes.
4. For wiki, classify the sub-action: connect/连接 → ~/.maestro/workflows/wiki-connect.md; digest/摘要 → ~/.maestro/workflows/wiki-digest.md; health/search/cleanup/stats/健康/检查/(none) → ~/.maestro/workflows/wiki-manage.md.
5. Ambiguous → display the operation table and ask the user to pick.

Routing rules

• Remaining tokens after classification become the chosen step's own arguments.
• During an active Run, reusable knowhow is staged here with maestro knowledge stage knowhow ...; project knowhow is written only by explicit promotion. Outside a Run, direct /maestro-knowhow capture remains available.
• Minimal canonical creation surfaces are deliberately small: ordinary Knowhow is exactly maestro knowhow add --type <type> --title "<title>" --content-file <path>; ordinary Spec is exactly maestro spec add <category> "<title>" "<content>". Keywords, sourceRef, relatedPaths, applicability, language, decision state, ID, and tool flags are advanced optional metadata—not required creation parameters. The nine Knowhow types remain session|tip|template|recipe|reference|decision|asset|blueprint|document.
• Outside any Run entirely (no Run to bind), the minimal Knowhow command is the fast path: it writes .workflow/knowhow/ directly with no Session, --evidence, or review/promote cycle. Reserve stage → review → promote for candidates needing corpus adjudication.
• Repository authority is host-owned. Omit targetRepoId for an ordinary current-repository write. Pass it only for an explicitly selected linked physical write after the host supplied that target's exact stable UUID and a live write capability for the matching corpus. The value must equal that host-supplied UUID; never infer it from cwd, repository name, alias, or path, and never persist alias/path as identity. If explicit linked authority is absent, fail closed instead of guessing.
• maestro knowledge audit is always read-only. Its only scopes are spec|knowhow|all; pipeline/repository diagnostics always run, usage diagnostics run when MaestroGraph exists, and --prune only includes a deterministic prune_plan. Audit has no --apply, mutation, delete, or purge path. Compatibility normalization never happens during audit. First save maestro knowledge normalize --report <path>, review it, then run that same report path with --apply; changed sources or repository identity invalidate the report.
• Stage candidate content from a temp file or stdin, never inline: write the content to a file and pass maestro knowledge stage <target> "<title>" --content-file <path|->. Inline positional content containing spaces, quotes, unicode (e.g. …), newlines, or leading dashes is misparsed and shifts later arguments.
• --signal-ids takes comma-separated IDs (--signal-ids spec:project:a,knowhow:b); space-separated values leak into positional arguments and corrupt the stage call.
• Use maestro knowledge review <session-id> as the human review surface. It shows fresh/missing/stale receipts, diversified evidence-backed matches, and copyable promote commands. --refresh reconciles all candidate source Runs. --resolve <candidate-id> --as <choice> --reason "..." resolves a candidate inline before displaying the refreshed view.
• Reconciliation is mandatory before completion but is not a popularity vote: exact identity, diversified semantic matches, and recorded/KG associations are evaluated separately. Unresolved semantic duplicate/conflict/supersession candidates may be sealed, but promotion must fail closed until resolved via review --resolve.
• promote --all promotes all eligible pending candidates (observed-only emits a warning); --include-observed has been removed.
• Treat audit findings and prune suggestions as evidence only. Never translate them into automatic lifecycle changes, and never prune solely because knowledge has low usage.