Channels
GSD OC
By GSD lifecycle orchestration as a native OpenClaw plugin — enforced research→plan→execute→verify→ship, hybrid retrieval, 0 Discord slots
Install
npm install
npmConfiguration Example
{ "plugins": { "entries": { "gsd-oc": { "hooks": { "allowPromptInjection": true } } } } }README
# GSD-OC — GSD for OpenClaw
**Get-Shit-Done lifecycle orchestration as a native OpenClaw plugin.**
Turns a coding/big-work prompt into a driven, *enforced* GSD lifecycle — research → map → plan →
execute → code-review → verify → ship — for **any** OpenClaw agent, **without typing a single `/command`**.
GSD-OC is its own thing: **no Claude Code runtime dependency, no `@opengsd/*` runtime dependency,
no ACP-into-Claude.** The GSD state engine is reimplemented natively in TypeScript, the GSD doc
corpus is snapshotted and bundled at build time, and the whole surface goes through tool-search +
routers so it consumes **zero** of Discord's 100 global slash-command slots.
---
## Why it's different
Most "methodology" plugins inject advisory text an agent can ignore. GSD-OC **enforces** the lifecycle
deterministically through OpenClaw's `before_tool_call` hook — a host-honored block, not a suggestion.
| Pillar | What it does |
|---|---|
| **Retrieval** | Hybrid search (spark embeddings + LanceDB + BM25 + trigram, fused via RRF) maps a free-text intent to the right GSD skills/subagents — long-tail aware (`"the build is flaky"` → `gsd-debugger`). Degrades gracefully to lexical when embeddings are offline. |
| **Enforcement** | `before_tool_call` **blocks file edits** until the current phase is planned, and **injects the matching GSD persona** into every subagent spawn — so no agent edits code out of order or spawns a bare, instruction-less subagent. |
| **State engine** | A native TS reimplementation of `gsd-tools`: `route()` reads `STATE.md` + phase artifacts to decide the next action; `gsd_state` advances it atomically under a lockfile. The read half and the write half are both live. |
## How it works
```
coding intent
│ auto-engage (before_prompt_build) injects GSD policy
▼
gsd_orchestrate ── retrieves relevant skills (gsd_retrieve) ──► selects a finite path
│ discuss → map-codebase → research →
│ plan → execute → code-review → verify → ship
│ (+ conditional: debug / secure / ui / ai / spike …)
▼
per step: dispatch the matching gsd-* subagent (persona injected, model-tier routed)
│
before_tool_call gate: edits blocked until planned · personas injected on spawns · FAILED verification halts
```
A prompt like *"add OAuth login"* auto-engages, retrieves the relevant skills, emits a path that
**includes the `secure` threat-model stage**, and gates code edits until the phase is planned.
## Quick start
```bash
npm install
npm run build # tsc → dist/ (ESM) + bundles corpus/vectors for self-contained runtime
npm test # 342 tests (node:test)
npx openclaw plugins build # generates openclaw.plugin.json
npx openclaw plugins validate # → "Plugin gsd-oc is valid."
```
Semantic retrieval needs a spark NIM embeddings endpoint via env (`SPARK_HOST` +
`SPARK_BEARER_TOKEN`/`SPARK_API_KEY`, never inlined); without it, retrieval degrades to BM25 + trigram.
See [docs/USAGE.md](docs/USAGE.md) for install, configuration, and a worked end-to-end example.
## The 10 tools (0 Discord slash slots)
| Tool | Purpose |
|---|---|
| `gsd_orchestrate` | Route a coding intent through the GSD lifecycle; `drive:true` dispatches the path's subagents. |
| `gsd_retrieve` | Hybrid retrieval of the relevant GSD skills/subagents for a free-text intent. |
| `gsd_state` | Advance `.planning/STATE.md` (status / progress / decisions / blockers) atomically. |
| `gsd_settings` | Inspect / bootstrap the project's GSD config. |
| `gsd_workflow` · `gsd_project` · `gsd_quality` · `gsd_context` · `gsd_manage` · `gsd_ideate` | 6 namespace routers — state-aware next-verb routing. |
`registerCommand` is **0** (asserted by a slot-audit test) — the entire surface is tool-search-reachable.
## Operator configuration
Auto-engage uses the `before_prompt_build` hook, which OpenClaw gates for non-bundled plugins. The
**operator** enables it in `~/.openclaw/openclaw.json` (the plugin never writes host config — it only
reads it via the SDK):
```jsonc
{ "plugins": { "entries": { "gsd-oc": { "hooks": { "allowPromptInjection": true } } } } }
```
Per-project opt-out: a `.gsd-off` file, `pluginConfig`, or `workflow.enforce_tool_gate: false` in
`.planning/config.json`. See [docs/USAGE.md](docs/USAGE.md).
## Documentation
- [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) — modules, the three pillars, lifecycle flow
- [docs/RETRIEVAL.md](docs/RETRIEVAL.md) — the hybrid retrieval engine (corpus, modalities, RRF, incremental re-index)
- [docs/ENFORCEMENT.md](docs/ENFORCEMENT.md) — the `before_tool_call` gate + spawn-persona model
- [docs/SUBAGENTS.md](docs/SUBAGENTS.md) — the 33 ported GSD subagents + model routing
- [docs/USAGE.md](docs/USAGE.md) — install, configure, use
- [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) · [CONTRIBUTING.md](CONTRIBUTING.md) — build pipeline, testing, contributing
## Project status
Milestone **M1 (Retrieval Core)** complete, with the finite-path orchestrator + enforcement layer built.
Hardened through **9 rounds of adversarial multi-agent code review** to convergence — the full
round-by-round log (CRITICAL → HIGH → converged, ~37 bugs fixed) is in
[docs/REVIEW-LOG.md](docs/REVIEW-LOG.md). The plugin was itself built with the GSD methodology it ports.
## Constraints (by design)
- OpenClaw plugin SDK — `definePluginEntry` (hooks + service + dynamic tools + runtime). Node ≥ 22,
TypeScript ESM, `typebox` runtime dep, `openclaw ≥ 2026.5.17`.
- No Claude Code in the loop; native OpenClaw orchestration.
- No `@opengsd/*` at runtime — the state engine is reimplemented in native TS.
- Discord's 100-slash-command cap respected — the GSD surface goes through tool-search + routers.
- Never mutates the host OpenClaw config or storage; no secrets inlined.
## License
[MIT](LICENSE) © Klein Panic
channels
Comments
Sign in to leave a comment