behavior-enhancer-generic
Health Uyari
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 5 GitHub stars
Code Basarisiz
- rm -rf — Recursive force deletion command in adapters/claude-code/hooks/register.test.ts
- rm -rf — Recursive force deletion command in adapters/claude-code/hooks/register.ts
- process.env — Environment variable access in adapters/codex/hooks/hooks.json
- fs.rmSync — Destructive file system operation in adapters/codex/hooks/post_tool_use.cjs
- fs module — File system access in adapters/codex/hooks/post_tool_use.cjs
- exec() — Shell command execution in adapters/codex/hooks/pre_tool_use.cjs
- fs module — File system access in adapters/codex/hooks/pre_tool_use.cjs
Permissions Gecti
- Permissions — No dangerous permissions requested
Bu listing icin henuz AI raporu yok.
Behavior discipline for coding agents: read-before-write enforcement, failure convergence, write verification with auto-rollback, and a high-risk command gate. Ships FULL plugin/hook adapters for DSH, Claude Code, Codex, and Pi, plus a host-agnostic SOFT skill with ADAPTIVE adapter guidance.
behavior-enhancer-generic
English · 中文
The behaviour design of dsh-behavior-enhancer, shipped as a cross-host dual-form product:
- Plugin form (FULL): where the host exposes runtime interception, you get real enforcement — vetoes, write verification, auto-rollback;
- Skill form (SOFT): where the host only gives an agent-facing channel, you get behavioural guidance (self-discipline rules).
Same semantics, different enforcement strength. No FULL claims we cannot back, and no pretending SOFT is FULL.
Which host gets which form
| Host | Form | What to install | Status |
|---|---|---|---|
| Claude Code | FULL (8 features) | adapters/claude-code/ — plugin (F3b is implemented as a concurrency-window veto) |
engine-level tests 14/14, dogfooding |
| Codex | FULL (8 features) | adapters/codex/ — config.toml hooks or plugin manifest |
protocol tests 13/13, ignition-tested on the desktop build |
| Pi | FULL (7 features + half of F9) | adapters/pi/ — pi extension |
protocol tests 11/11, live-tested on a real install |
| DSH | FULL | adapters/dsh/ — installable plugin (all modules of the original, plus F1 read-before-write and two fixes) |
all-module mock-ctx tests pass, real-install test pending |
| Any other host | SOFT + ADAPTIVE | skill/SKILL.md §8: the agent self-builds an adapter from core/adapter-template.md (conservative testing, conservative FULL, written report) |
published form |
Install on Claude Code (one command):
/plugin marketplace add Zoria-Lind/behavior-enhancer-generic
/plugin install behavior-enhancer@zoria-behavior
For local development you can instead use claude --plugin-dir "<repo>\adapters\claude-code" orCLAUDE_CODE_PLUGIN_DIRS=...\adapters\claude-code (seeadapters/claude-code/README.md) — pick one, never both.
The four-state activation model
Every feature activates independently, and its state is decided by the host's capability surface — never by user or agent preference:
| State | Meaning |
|---|---|
| NATIVE | The host already does this (only claimed when measured); the skill stays quiet |
| FULL | Plugin-level takeover (veto / write verification / rollback) |
| SOFT | Behavioural guidance (instructions, flow, decision trees) |
| DISABLED | We do not pretend a capability exists |
There are no if DSH / elif ClaudeCode branches — only capabilities. Full design ingeneric-skill_plan_v1.md; the feature × capability matrix incore/audit.md; cross-host shared semantics (tokens, pattern tables, thresholds, TTL, tier ladder) incore/semantics.md.
Note: those design documents (
core/,generic-skill_plan_v1.md) are Chinese-only for now. English readers can work from this README plus the per-adapter READMEs.
Relationship to the DSH plugin
The DSH adapter started from the original plugin dsh-behavior-enhancer (a separate repo) and then added the generic work. From there the two evolve independently — there is no sync mechanism: when one side produces a better rule, the other copies it by hand if it is worth it. Install one or the other, never both
(see adapters/dsh/README.md).
Repository layout
behavior-enhancer-generic/
├── .claude-plugin/ # Claude Code marketplace manifest (marketplace.json, name: zoria-behavior)
├── core/ # behaviour semantics + capability matrix (four states) + generic adapter template (ADAPTIVE mode)
├── skill/ # SOFT artifact: SKILL.md + per-host resident fragments (claude-code / codex / pi / cursor)
├── adapters/ # FULL artifacts: claude-code / codex / pi (DSH is only a pointer back to the original repo)
├── publish/ # Agensi publishing artifacts (the SOFT skill package)
└── README.md / README.zh-CN.md # English (this file) / Chinese
Tests
cd adapters/claude-code && claude plugin validate . && claude plugin test . # 14 engine-level tests
cd adapters/codex && node test/adapter.test.cjs # 13 protocol tests
cd adapters/pi && node test/adapter.test.cjs # 11 protocol tests
Links
- Free SOFT version (pure SKILL.md, works with any compatible host, no executable code): Agensi — Behavior Enhancer. The FULL version in this repo remains the only complete form; the two feed each other.
- Claude Market — Claude Code plugin directory (behavior-enhancer entry).
- Claude Plugin Hub — plugin directory (verified listing)
License
MIT — see LICENSE. The rules started from the DSH version and have evolved independently since.
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi