maestro-knowledge

<required_reading> @~/.maestro/ref/knowledge-closeout.md </required_reading>

If the required file was not expanded by the host, or is no longer in context, Read @~/.maestro/ref/knowledge-closeout.md explicitly before knowledge closeout.

<purpose> Intent-driven knowledge-store management. No fixed grammar — state your intent; the command classifies it and runs the matching workflow or direct lifecycle command. Explicit keywords still work as deterministic shortcuts.

Operation	Keywords	Execution document or CLI
audit	audit / 审计 / 检查知识库	maestro knowledge audit --scope all（只读；细节见 ~/.maestro/workflows/knowledge-audit.md）
normalize	normalize / migrate / 规范化 / 迁移	maestro knowledge normalize --report <path>，审阅后才可原样追加 --apply
review	review / 审查 / 证据 / 下一步 / 匹配 / 去重 / 冲突检测 / 裁决 / 候选 / backlog	maestro knowledge review <session-id> [--refresh] --json → shared presentation/authorization
stage	stage / 暂存 / candidate / 沉淀候选 / cited / validated / contradicted / 记录命中关系	maestro knowledge stage ... [--signal <signal> --signal-ids <ids>]
promote	promote / 晋升 / 发布候选	Shared closeout protocol → maestro knowledge promote <session-id> --candidate <approved-id>
harvest	harvest / 提取 / 收割 / 从工件	~/.maestro/workflows/harvest.md
wiki	wiki / 知识图谱 / 连接 / 摘要 / 健康	~/.maestro/workflows/wiki-manage.md / ~/.maestro/workflows/wiki-connect.md / ~/.maestro/workflows/wiki-digest.md
extractors	extractors / 抽取器 / 生成抽取规则	~/.maestro/workflows/extractors.md
domain	domain / 领域术语 / 注册术语 / term	~/.maestro/workflows/domain-add.md

</purpose> <dispatch> Classify the intent in `$ARGUMENTS` into one operation. For an operation mapped to an execution document, read the path shown in the table directly and follow it; do not create a Session or Run merely to load instructions. For direct lifecycle operations, invoke the listed `maestro knowledge` CLI command.

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 use the corresponding maestro knowledge CLI. For candidate closeout, follow @~/.maestro/ref/knowledge-closeout.md (Review → Refresh → Present → Authorize → Execute → Verify), not an unreviewed direct mutation: review-only stays read-only; refresh only missing/stale receipts within authority; after confirming the disposition and any resulting publication, use promote --resolve; publish already eligible fixed explicitly approved IDs via promote --candidate. 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> --json as the closeout review surface. It shows fresh/missing/stale receipts and diversified evidence-backed matches. The shared protocol governs presentation, human authorization, inline promote --resolve, selected promote --candidate, and result/review readback; copyable commands alone are not owner closeout. review --resolve remains the deprecated compatibility fallback, including relationship-only adjudication without publication approval.
• 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 the relationship is explicitly confirmed and resolved via promote --resolve (or the deprecated review --resolve fallback).
• Agent closeout uses fixed approved candidate IDs, never -y or --all as publication approval. Zero candidates, rejection, or deferral leave backlog intact and permit otherwise valid completion; no question tool and no explicit authorization means defer.
• Promotion requires eligible candidates with fresh receipts. Run-source candidates require every source Run sealed. A canonical session/3.0 session-source candidate does not require Session completion: it requires immutable candidate version/content hash, bound Session identity/revision (later unrelated Session activity alone does not invalidate the snapshot), non-empty evidence roots/hash, and a fresh session reconciliation receipt for the candidate snapshot and current corpus fingerprint, revalidated at final commit. Normal Run/Session completion never implies approval or promotion. Older schema rules belong only to explicitly selected legacy compatibility.
• 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. </dispatch>