Tools
Plugin Skill Harness
By Pre-scans user intent before replies and injects routing hints via before_prompt_build hook.
Install
npm install --frozen-lockfileConfiguration Example
{ "query": "diagnose a deployment readiness failure", "limit": 5 }README
# Skill Harness
[](https://github.com/openclaw/openclaw)
[](https://opensource.org/licenses/MIT)
Skill Harness is an OpenClaw plugin that helps agents find relevant skills and reuse experience from previous work. Before an agent replies, it searches skills and experiences, selects the relevant candidates, and adds compact guidance to the conversation. An optional reviewer turns evidence from completed turns into reusable experience for future tasks.
The project combines **skill discovery**, **experience retrieval**, and **experience curation** around existing OpenClaw agents. Skills remain ordinary skill files; experience lives separately in runtime storage.
## Why this project exists
As a skill library grows, listing every skill in every prompt consumes context and makes it harder to find the right workflow. Short descriptions also leave out details that matter for selection: applicability, prerequisites, operating steps, and failure conditions often appear only in the full skill document.
Repeated work introduces another problem. An agent may solve the same issue again without retaining the useful procedure, while indiscriminate memory capture accumulates duplicate or outdated advice.
Skill Harness addresses these problems with:
- **On-demand discovery:** search skill metadata, bodies, and reference files before selecting a bounded set for the current task.
- **Reusable experience:** retrieve past procedures independently of skill discovery, including standalone guidance without skill associations.
- **Progressive loading:** inject descriptions and experience summaries; let the agent load detailed instructions when relevant.
- **Evidence-based curation:** optionally review completed turns, check existing coverage, and create, refine, merge, or delete experience when the evidence supports a change.
For example, a request to diagnose a deployment can surface a relevant operations skill alongside an experience describing a previously verified diagnostic sequence. The agent loads the skill through `skill_view`, queries further experience through `skill_experience`, and applies the guidance to the current environment.
## Research context
These papers provide context for the design choices and related directions. The table distinguishes the research ideas from the behavior implemented here; the papers' benchmark results are not measurements of this plugin.
| Paper | Research idea | Relationship to Skill Harness |
| ------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| [SkillRouter: Skill Routing for LLM Agents at Scale](https://arxiv.org/abs/2603.22455) | Skill bodies contain routing evidence that names and descriptions can miss; retrieve candidates before reranking. | Searches metadata, bodies, and references, then uses one constrained Jev selection call. Uses QMD and Jev rather than SkillRouter's trained retrieval/reranking models. |
| [SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale](https://arxiv.org/abs/2606.03056) | Typed directed relations add structure to skill selection, and execution can inform graph evolution. | Offers an imported author-relation graph and optional bounded candidate expansion. Relations remain unverified hints; runtime review curates experiences rather than evolving graph edges. |
| [Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose (SkillWeaver)](https://arxiv.org/abs/2606.18051) | Complex tasks can require decomposition, retrieval, and dependency-aware composition of multiple skills. | Selects multiple relevant skills; task decomposition and executable DAG planning remain related work, outside the current routing pipeline. |
| [SkillOpt: Executive Strategy for Self-Evolving Agent Skills](https://arxiv.org/abs/2605.23904) | Bounded document edits can improve an agent's external state when accepted through validation. | Reviewer changes pass schema, content, operation, and concurrency checks. This is experience curation; it does not optimize skill files using scored rollouts or held-out performance tests. |
## Architecture
```mermaid
flowchart TD
U[Conversation turn] --> H[OpenClaw before_prompt_build]
H --> W[Static working-set descriptions]
H --> S[Skill candidates: name matching and QMD search]
H --> E[Experience candidates: QMD search]
G[Optional author-relation graph] -.-> S
S --> J[One constrained Jev selection call]
E --> J
G -. Bounded relation relevance questions .-> J
W --> C[Agent context]
J --> C
C --> A[Agent loads guidance and executes tools]
A --> T[Session evidence and aggregate statistics]
T -. Review enabled .-> R[Background experience reviewer]
R --> V[Validate findings and file changes]
V --> X[Shared runtime experiences]
X --> E
```
| Component | Role |
| ------------------- | -------------------------------------------------------------------------------------------------------------------- |
| Skill catalog | Resolves the invoking agent's visible skills and provides discovery/loading tools. |
| QMD indexes | Hybrid retrieval over skill metadata, bodies and references, and experience keywords, summaries and bodies. |
| Jev selector | Evaluates the combined candidate pool against the conversation and configured relevance thresholds. |
| Prompt integration | Adds a static working set and compact, advisory dynamic context through OpenClaw hooks. |
| Experience reviewer | Optionally examines completed turns and proposes validated experience changes. |
| Runtime state | Stores experiences, session evidence, review outcomes, managed indexes, and aggregate telemetry outside the package. |
### Before a reply
1. Add skill-discovery guidance and the configured static working-set descriptions. Automatic workspace/workshop additions can be controlled separately.
2. For eligible agents and chats, search skill and experience candidates in parallel using the conversation context. Skill discovery combines typo-aware name matching with QMD retrieval.
3. Build one deduplicated skill pool from name/QMD hits and visible skills associated with all retrieved experience candidates. Optional relation expansion starts only from the original name/QMD hits. If candidates exist, make one constrained Jev call that independently scores every skill, experience and bounded relation; an empty pool skips selection.
4. Inject only individually selected skills and experiences, subject to their separate configured limits. Selecting an experience does not automatically select its associated skills.
5. The agent uses `skill_view` to load skill instructions and `skill_experience` to search detailed experience before applying relevant guidance.
Dynamic context contains optional `<matched_skills>` and standalone `<matched_experiences>` blocks. Each selected experience summary appears once: inside its highest-scored final associated skill, or standalone if no associated skill survives selection. Skills are ordered by Jev score with a locale-independent name tie-break (UTF-16 string order); experience score ties use the same ordering on IDs. Experiences with no associations, invisible associations or rejected associated skills remain eligible. Placement never boosts skill scores or requires loading a skill. The context is advisory, non-user input; full documents and the entire catalog are not inserted.
Runtime search or selection failures let the main agent continue with available static context. Failed selection does not inject an unchecked fallback list. Invalid configuration can still prevent the plugin from loading.
### After a completed turn
Review is disabled by default. When enabled, experience-health cadence, uncertain routing, or capability-fit evidence can schedule a background investigation. A trigger does not itself prove that an experience should be written.
The Review execution module loads only after Review is enabled and a trigger survives evaluation. After asynchronous loading and snapshot preparation, the plugin rechecks lifecycle cancellation, scheduler disposal, host hook validity, and the originating turn before queueing work. Pipeline events remain synchronous. QMD itself remains loaded on demand.
The reviewer checks existing coverage before creating an entry, prefers refining an existing procedure, and can merge overlapping experiences or remove guidance proven obsolete, useless, or superseded. A no-finding outcome is valid. Changes are checked against the declared operation, actual file changes, eligible observed skills, and concurrent updates before writeback.
Review operates on experience copies in a temporary workspace. Fi
... (truncated)
tools
Comments
Sign in to leave a comment