pi-skill-search 0.1.0 → 0.2.0

package/CHANGELOG.md

@@ -1,5 +1,20 @@
 # Changelog
 
+## 0.2.0 (2026-05-23)
+
+### Fixed
+- **Packaging conflict**: Renamed `skills/` to `data/` to prevent Pi from discovering and inject-all-ing 284 skills
+- **Install script**: Created `install.mjs` to copy `skill-search` skill to `~/.pi/agent/skills/` (where Pi discovers skills)
+- **Scoring weights**: Changed from +5/+2 to +3/+1 per SPEC §5.1 for better relevance scoring
+- **Unused nameIndex**: Removed from SkillIndex interface and buildIndex() function
+- **EOF newline**: Added missing newline in categories.ts
+- **SPEC §7.4**: Updated documentation to always return `{ systemPrompt }`, never undefined
+
+### Changed
+- Scanner now scans `data/` directory instead of `skills/`
+- Extension registers `skill-search` tool via `before_agent_start` event
+- 284 skills corpus in `data/` (Pi does not discover this directory)
+
 ## 0.1.0 (2026-05-17)
 
 ### Added
@@ -11,10 +26,7 @@
 - Strip regex to remove Pi's `<available_skills>` block from system prompt
 - Category summary injection via `before_agent_start` event
 - Proactive suggestion hook for Python package detection (opt-in)
-- **258 bundled skills** from 10 source repos:
-  - scientific-agent-skills (137), oh-my-claudecode (29), vetc-dev-kit (33)
-  - pi-crew (27), superpowers (14), oh-my-pi (2), claude-mem (12)
-  - gstack (1), context-mode (1), spec-kit (1)
+- **284 bundled skills** from 10 source repos
 - 14/14 categories, 0 Other
 - 123 tests across 14 test files including search quality benchmarks
-- Biome lint + TypeScript strict mode clean
+- Biome lint + TypeScript strict mode clean

package/README.md

@@ -4,7 +4,7 @@
 
 ## Problem
 
-Pi injects **all** skill `<available_skills>` descriptions into every agent's system prompt. With 258 skills across 14 categories, that's ~40,000+ tokens wasted per turn — even when the agent never uses any skill.
+Pi injects **all** skill `<available_skills>` descriptions into every agent's system prompt. With 258+ skills across 14 categories, that's ~40,000+ tokens wasted per turn — even when the agent never uses any skill.
 
 ## Solution
 
@@ -14,12 +14,21 @@ Pi injects **all** skill `<available_skills>` descriptions into every agent's sy
 
 Result: **~97% token reduction** in skill-related system prompt content.
 
-## Quick Start
+## Installation
 
 ```bash
-# Install
-npm install
+# Install from local source
+npx pi install .
+
+# Or from npm (when published)
+npm install pi-skill-search
+```
+
+The `install.mjs` script automatically sets up the `skill-search` skill in `~/.pi/agent/skills/` so Pi can discover it.
+
+## Quick Start
 
+```bash
 # Run tests
 npm test
 
@@ -31,7 +40,9 @@ npx tsx scripts/build-skills-fixture.ts
 
 ```
 index.ts              ← Extension entry (before_agent_start + skill-search tool)
+install.mjs           ← Post-install: copy skill-search to ~/.pi/agent/skills/
 src/
+  scanner.ts          ← findCorpusPath() + scanSkillDirectory()
   types.ts            ← PiSkill, SkillEntry, SearchResult, SkillIndex
   text.ts             ← Tokenizer (SPEC §5.3)
   synonyms.ts         ← Synonym dictionary + expandQuery
@@ -41,19 +52,22 @@ src/
   strip.ts            ← Regex strip <available_skills>
   format.ts           ← Category summary + result formatting
   proactive.ts        ← Python package detection (Phase 10, opt-in)
+data/                 ← 284 skills corpus for search (not discovered by Pi)
 test/
   text.test.ts        ← 12 tests
   synonyms.test.ts    ← 8 tests
   classify.test.ts    ← 8 tests
   indexer.test.ts     ← 12 tests
   search.test.ts      ← 11 tests
-  format.test.ts      ← 12 tests
+  format.test.ts     ← 12 tests
   strip.test.ts       ← 6 tests
   lifecycle.test.ts   ← 15 tests
-  integration.test.ts ← 8 tests (258-skill corpus)
+  integration.test.ts ← 8 tests (284-skill corpus)
   failure-isolation.test.ts ← 3 tests
   coverage.test.ts    ← 3 tests (classifier coverage)
   proactive.test.ts   ← 9 tests
+  scanner.test.ts     ← 7 tests
+  search-quality.test.ts ← 10 tests
 ```
 
 ## Tool: `skill-search`
@@ -61,37 +75,28 @@ test/
 | Parameter | Type | Required | Description |
 |-----------|------|----------|-------------|
 | `query` | string | ✅ | Search query (1–500 chars) |
-| `limit` | number | ❌ | Max results (1–20, default 5) |
+| `limit` | number | ❌ | Max results (1-20, default 5) |
 
 Returns formatted skill descriptions with name, description, category, and file path.
 
-## Configuration
+## How It Works
 
-No configuration needed — the extension auto-indexes skills from Pi's `systemPromptOptions.skills` on the first `before_agent_start` event.
+1. **At session start**: Extension scans `data/` (284 skills) and builds a search index
+2. **On `before_agent_start`**: Strips `<available_skills>` block, injects category summary, registers `skill-search` tool
+3. **On `skill-search` call**: Returns top matching skills with path for `read` tool
 
-## Bundled Skills (`skills/`)
+## Configuration
 
-The extension ships with **25 bundled skills** covering 14 scientific domains. These live in `skills/` and are scanned by the extension directly — Pi does **not** discover them at startup.
+No configuration needed — the extension auto-indexes skills from `data/` on startup.
 
-| Category | Skills |
-|---|---|
-| Cheminformatics & Drug Discovery | rdkit |
-| Bioinformatics & Genomics | scanpy, biopython, huggingface, openephys, rdkit |
-| Machine Learning & AI | pytorch, huggingface, xgboost, xgboost-ts |
-| Clinical & Medical | clinical-trial, pubmed-search |
-| Physics & Quantum | astropy |
-| Databases & Data Sources | sqlite-pandas, fastapi, pyspark, pubmed-search |
-| Data Analysis & Visualization | matplotlib, statsmodels, xgboost-ts, nilearn, pyspark |
-| Scientific Writing & Communication | matplotlib, pubmed-search, research-reproducibility |
-| Geospatial & Remote Sensing | landsat, astropy, matplotlib |
-| Lab Automation & Integration | docker-sandbox, celery-pipeline |
-| Time Series & Forecasting | statsmodels, xgboost-ts, openephys |
-| Materials Science & Engineering | pymatgen |
-| Research Methodology | research-reproducibility, statsmodels |
-| Integration Platforms | rdflib, markitdown-docs, sqlite-pandas |
-| Meta | skill-search (this tool usage guide) |
+## Design Decisions
 
-To add new skills: create `skills/<name>/SKILL.md` with YAML frontmatter (`name`, `description`). The extension picks them up automatically.
+| Decision | Rationale |
+|----------|-----------|
+| Skills in `data/` not `skills/` | Pi would discover them and inject-all again |
+| skill-search in `~/.pi/agent/skills/` | Pi discovers skills here, not in extension dirs |
+| install.mjs for skill setup | pi install doesn't auto-discover extension skills |
 
 ## License
-MIT
+
+MIT