cc-enslaver

cc-enslaver

A Claude Code plugin and LLM-agnostic rule pack that eliminates lazy AI behavior — reactive patches, guessed citations, surface-level "fixes", half-finished work — by enforcing systematic thinking, verification, and root-cause analysis at every layer of the agent loop.

中文用户请直接看 → 中文说明

---

What is this?

LLM coding agents (Claude Code, Cursor, Copilot, Cline, Aider, etc.) frequently fall into predictable lazy patterns:

Lazy pattern	What it looks like
Reactive patching	Sees a bug, slaps a try/except around it, declares done.
Guessed citations	Cites file paths, line numbers, or APIs that don't exist.
Keyword-search-only	Greps once, edits, never reads the surrounding architecture.
Memory dependence	Acts on stale recollection instead of re-reading the current file.
Root-cause bypass	Adds sleep for races, --no-verify for hooks, swallows exceptions.
Half-finished work	Stops at "should work", leaves TODOs, doesn't verify the whole flow.
Premature done-claim	Claims "fixed" without re-running the original failing case, no edge cases, no comparison evidence.

cc-enslaver ships a layered defense against all seven, currently 9 built-in rules + user-defined Imperial Edicts (圣旨) + 8 Stop-hook gates (v0.20.0):

New in v0.20 — 📋 Structured YAML reporting + plain-language TL;DR: every reply now ends with a fixed ```yaml cc-enslaver: block (改前 / 改中 / 收敛 / 忠实 / 收尾 / tldr; English mirror before / edits / convergence / fidelity / closing / tldr) — the audit trail is scannable at a glance instead of drifting free-form prose. A new Stop layer (h) hard-enforces a one-sentence tldr (大白话总结) on every done-claim reply, and every block reason now carries a 大白话: takeaway. The schema's field names ARE the existing Stop-hook detection markers, so no detector changed — old emoji-markdown and new YAML reply forms both pass.

From v0.18 — 🧹 Opt-in auto-GC on SessionStart: set CC_ENSLAVER_AUTO_GC_DAYS=30 and the SessionStart hook automatically prunes session-state files older than N days. Rate-limited to once per 24h via a marker file so rapid session restarts don't re-scan. Default off (backward-compatible); the manual /cc-enslaver:gc slash command still works and shares the same prune_old_sessions() deletion routine.

From v0.17 — 🌐 Imperial Edicts go bilingual: with CC_ENSLAVER_LANG=en, the soft-layer injection and the PreToolUse DENY reason both flip to English ("Imperial Edicts" / "Imperial Edict E01 violation"). Default Chinese ("圣旨") preserved. Plus Windows portability fixes: file-claim regex now matches drive-letter paths (C:\Users\...\x.py), and manage_edicts.py forces UTF-8 stdout.

From v0.16 — 🕵️ Stop Layer (g) file-claim verification: read_guard captures per-file mtime baselines on first encounter; stop_guard parses I edited X.py / 我修改了 Y.md claims and BLOCKs the Stop when the on-disk mtime contradicts. Conservative-by-design (no baseline / any ambiguity → pass). Escape hatch: CC_ENSLAVER_DISABLE_LAYER_G=1.

From v0.15 — 🌍 English prompts mirror: set CC_ENSLAVER_LANG=en and the hook injects prompts/en/{session-start,user-prompt}.md instead of the Chinese canonical.

From v0.14 — ⚡ Three more Bash bypass patterns (git rebase --skip, --break-system-packages, rm -rf on root/$HOME/~) get PreToolUse(Bash) DENY. 🏛️ Edicts --global flag: add --global writes to ~/.claude/cc-enslaver/edicts.toml for personal cross-project rules.

From v0.13 — 🔁 Rule-09 rolling-patch hard layer: PreToolUse(Edit|Write) physically DENYs the 4th small Edit (≤ 10 lines AND < 200 chars) to the same file in one session unless a systematic rewrite (≥ 50 lines OR ≥ 1500 chars) resets the counter. See rules/09-systematic-modification.md §"Edit/Write 频率层".

From v0.12 — 🏛️ Imperial Edicts (圣旨): user-defined per-project hard rules loaded from .claude/cc-enslaver/edicts.toml with PreToolUse(Edit|Write|Bash) DENY and /cc-enslaver:edict CRUD slash command. See docs/EDICTS.md. Stop-hook block reasons render as uniform status tables. Soft-layer prompts thinned 54%.