一键在 Manus 中运行任何 Skill

orchestrate-vault-work

星标29

分支0

更新时间2026年7月19日 15:39

Coordinate advanced Loreloom work across project-scoped Codex agents only after the owner explicitly requests parallel/custom agents or accepts a stated delegation benefit. Keeps every child read-only and the root as sole writer.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

yi-john-huang

yi-john-huang/loreloom

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Orchestrate Vault Work

Keep the root agent responsible for scope, decisions, writes, validation, and the final answer. Delegate bounded evidence gathering and specialist review.

Workflow

Read AGENTS.md, .agents/policies/vault-policy.md, and docs/MULTI_AGENT.md.
Confirm the owner's explicit request or acceptance of advanced delegation.
State the exact input paths, output paths, user-authorized mutations, and concrete delegation benefit.
Split work only along independent boundaries and select the narrowest available read-only roles from .codex/agents/.
Verify no live parent permission override broadens child access. Spawn at most three subagents concurrently and prohibit recursive delegation.
Give each role minimum context: exact paths, question, prohibited actions, and return format. Do not include the expected answer.
If a role is unavailable, do not substitute silently. Low-risk extraction or proposal work may return to the root with reduced coverage disclosed; material synthesis or review requires direct owner review and an incomplete-specialist-coverage report.
Reconcile disagreements against Sources and policy, never majority vote.
Before writing, record the immutable 40-character base commit and run uv run python -I scripts/validate_change.py --check-clean --base <40-character-commit> --snapshot-file <absolute-temp-path> with one exact --allow per authorized output path. Let only the root write.
Run uv run --locked python scripts/validate_vault.py, then run every final uv run python -I scripts/validate_change.py --target-root <candidate-vault-root> ... from a separate clean detached worktree at the immutable base with the same snapshot and every exact changed path. Request independent read-only review for material changes when available.
If a gated path needs approval, stop for an owner- or trusted-UI-created receipt under protected Git metadata. Never create or alter that receipt.
Report assignments, accepted and rejected findings, changes, uncertainty, human review, specialist coverage, and validation.

Do not delegate

user decisions, approval requests, or final truth promotion;
deletion, archival, bulk rename, commit, push, or publication authority;
overlapping edits to the same note or generated page;
a task whose coordination cost exceeds doing it directly.

Never substitute a model or role silently. Disclose reduced coverage and follow the owner-review rules above.

同仓库更多 Skills

同仓库

audit-vault-health

yi-john-huang/loreloom

Perform a read-only Loreloom health audit covering schema errors, broken or ambiguous links, provenance gaps, stale generated pages, privacy risks, orphan notes, and workflow drift. Use for weekly reviews, pre-publication checks, unexplained vault degradation, or broad cleanup requests before any repair work.

2026-07-1929

capture-vault-source

yi-john-huang/loreloom

Create editable Loreloom Source notes from exact local Assets or URLs. Use for provenance capture and faithful machine-assisted summaries before Source review; never overwrites assets or reviewed Sources, fetches URLs silently, or creates Knowledge.

2026-07-1929

connect-vault-notes

yi-john-huang/loreloom

Find and add meaningful relationships among a bounded set of Loreloom Knowledge notes and MOCs. Use for orphan review, near-duplicate detection, cross-topic navigation, reciprocal context, or approved backlink improvements. Starts read-only and does not mass-link keywords, create placeholders, merge, rename, or delete notes.

2026-07-1929

distill-vault-sources

yi-john-huang/loreloom

Transform owner-reviewed Loreloom Source or Daily notes into traceable atomic Knowledge drafts. Use when extracting durable concepts, comparing evidence, detecting duplicate concepts, or preparing human-reviewable notes from exact Source paths. Never rewrites Sources, fills gaps from model memory, or promotes drafts to evergreen.

2026-07-1929

regenerate-vault-wiki

yi-john-huang/loreloom

Build or regenerate a Loreloom generated Wiki page from declared Knowledge inputs. Use for overview pages, comparisons, guides, conflict synthesis, changed-input refreshes, or stale Wiki review. Preserves human-owned blocks exactly and never adds facts absent from authorized inputs.

2026-07-1929

triage-vault-inbox

yi-john-huang/loreloom

Classify unprocessed Loreloom Inbox captures and propose safe dispositions. Use when reviewing one or more files under Inbox, detecting duplicates, identifying provenance, or deciding whether a capture belongs in Sources, Knowledge, Projects, Areas, deferred work, or Archive. Starts read-only and never deletes or moves notes without exact approval.

2026-07-1929

name	orchestrate-vault-work
description	Coordinate advanced Loreloom work across project-scoped Codex agents only after the owner explicitly requests parallel/custom agents or accepts a stated delegation benefit. Keeps every child read-only and the root as sole writer.

Orchestrate Vault Work

Keep the root agent responsible for scope, decisions, writes, validation, and the final answer. Delegate bounded evidence gathering and specialist review.

Workflow

Read AGENTS.md, .agents/policies/vault-policy.md, and docs/MULTI_AGENT.md.
Confirm the owner's explicit request or acceptance of advanced delegation.
State the exact input paths, output paths, user-authorized mutations, and concrete delegation benefit.
Split work only along independent boundaries and select the narrowest available read-only roles from .codex/agents/.
Verify no live parent permission override broadens child access. Spawn at most three subagents concurrently and prohibit recursive delegation.
Give each role minimum context: exact paths, question, prohibited actions, and return format. Do not include the expected answer.
If a role is unavailable, do not substitute silently. Low-risk extraction or proposal work may return to the root with reduced coverage disclosed; material synthesis or review requires direct owner review and an incomplete-specialist-coverage report.
Reconcile disagreements against Sources and policy, never majority vote.
Before writing, record the immutable 40-character base commit and run uv run python -I scripts/validate_change.py --check-clean --base <40-character-commit> --snapshot-file <absolute-temp-path> with one exact --allow per authorized output path. Let only the root write.
Run uv run --locked python scripts/validate_vault.py, then run every final uv run python -I scripts/validate_change.py --target-root <candidate-vault-root> ... from a separate clean detached worktree at the immutable base with the same snapshot and every exact changed path. Request independent read-only review for material changes when available.
If a gated path needs approval, stop for an owner- or trusted-UI-created receipt under protected Git metadata. Never create or alter that receipt.
Report assignments, accepted and rejected findings, changes, uncertainty, human review, specialist coverage, and validation.

Do not delegate

user decisions, approval requests, or final truth promotion;
deletion, archival, bulk rename, commit, push, or publication authority;
overlapping edits to the same note or generated page;
a task whose coordination cost exceeds doing it directly.

Never substitute a model or role silently. Disclose reduced coverage and follow the owner-review rules above.