菜单
Use when a long session has compacted or handed off, before any Edit, to confirm Claude's belief about file contents still matches disk. Diffs current sha256 signatures against the last-recorded snapshot in `.claude/state/belief.jsonl` for the N most-recently-edited files (default 5).
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。
基于 SOC 职业分类
| name | belief-audit |
| description | Use when a long session has compacted or handed off, before any Edit, to confirm Claude's belief about file contents still matches disk. Diffs current sha256 signatures against the last-recorded snapshot in `.claude/state/belief.jsonl` for the N most-recently-edited files (default 5). |
| when_to_use | Reach for this after session compaction, after a handoff from another agent, or manually when returning to a file after a long absence. Also fires automatically on PostCompact. Do NOT use for runtime behavior verification — use `/verify` instead. |
| disable-model-invocation | true |
| allowed-tools | ["Read","Bash"] |
| scheduling | invoke after compaction, after agent handoff, or before re-editing files touched more than one session ago |
| structural | ["Read .claude/state/belief.jsonl for the N most-recent unique paths","Re-compute sha256 for each path on disk now","Compare recorded signature vs current signature per path","Emit drift report (drifted paths + both hashes) and exit 1 if any drift found"] |
| logical | drift report emitted; exit 0 when all signatures match, exit 1 when any path has drifted |
After compaction, Claude's internal model of what's in a file can differ from disk reality (arXiv:2605.18747 §4.3 formalizes this as |Bk − Sk| — the gap between the agent's belief Bk and the actual system state Sk). Every Edit/Write is snapshotted automatically; this skill compares the stored signatures against current disk to surface divergence before it causes bugs.
bash scripts/audit.sh [N] (default N=5) — reads .claude/state/belief.jsonl, takes the latest entry per unique path, limits to N most-recent unique pathsExample 1 — No drift
Input:
.claude/state/belief.jsonl (last 5 unique paths):
{"ts":"2026-05-20T10:00:00Z","path":"plugins/context-engine/hooks/track-edits.sh","sha256":"abc123...","agent":"main","op":"post","session_id":"s-1"}
{"ts":"2026-05-20T10:01:00Z","path":"README.md","sha256":"def456...","agent":"main","op":"post","session_id":"s-1"}
Output:
## Belief-State Audit (2 files checked)
All 2 files match recorded signatures.
Exit: 0 (no drift)
Example 2 — Drift detected
Input:
.claude/state/belief.jsonl (last 5 unique paths):
{"ts":"2026-05-20T09:00:00Z","path":"plugins/context-engine/hooks/track-edits.sh","sha256":"abc123...","agent":"main","op":"post","session_id":"s-1"}
(file was edited externally between sessions)
Output:
## Belief-State Audit (1 file checked)
DRIFT DETECTED — 1 file(s) have changed since last snapshot:
| Path | Recorded sha256 | Current sha256 |
|------|-----------------|----------------|
| plugins/context-engine/hooks/track-edits.sh | abc123... | xyz789... |
Re-read these files before editing. Run /belief-audit again to confirm.
Exit: 1 (drift found)
Use when the user describes a multi-step feature or refactor and you need to decide whether to handle it solo, dispatch a parallel `/fan-out`, or run a `/worktree-team` planner→generator→reviewer pipeline. Outputs a routing recommendation with the reasoning behind it.
Use whenever the user asks to apply the same operation to many independent files or components — bulk migration, batch refactor across controllers, parallel exploration of unrelated subsystems. Dispatches one subagent per file with a shared prompt template, then collects the results.
Use to toggle the safe-mode flag at .claude/safe-mode. The flag is written automatically by consecutive-failure-guard once tool failures hit FORGE_SAFE_MODE_THRESHOLD (default 5), and block-destructive.sh denies every Bash/Write/Edit while it exists. /safe-mode off clears the flag, logs the exit, and prompts /postmortem; /safe-mode on enters manually; /safe-mode status reports current state.
Run the 5-stage Token Transformation Pipeline (TRAE §5.2.2) over the current session state — Collection, Ranking, Compression, Budgeting, Assembly. Emits a concrete next-action recommendation (/compact, /lean-md, or /progress-log + fresh session) instead of a generic warning.
Audit or scaffold a CLAUDE.md against the Karpathy 4-section structure (Think Before Coding · Simplicity First · Surgical Changes · Goal-Driven Execution). Companion to /lean-md — structure audit, not size trim.
Print every policy enforcement point declared in plugins/diagnostics/registry/policies.json — id, verdict, plugin, hook event, severity, bypass — grouped by verdict. Single discoverable index of what the harness blocks, anchors, nudges, or logs at runtime.