find

Files (7)

• SKILL.md
• advanced-features.md · references/
• architecture.md · references/
• auto-detection.md · references/
• consolidated-features.md · references/
• output-format.md · references/
• subagent-delegation.md · references/

SKILL.md (first lines)

# Unified Find

## Purpose

Single entry point to find across all CSF NIP data stores in parallel - **local sources only**. For web research, use `/web`.

**Scope**: CKS, CHS, CDS, Code/Grep, DOCS, SKILLS. No external dependencies.

## Project Context

- **Solo-dev constraints apply** (CLAUDE.md)
- **No external dependencies**: All search is local, privacy-preserving
- **Efficiency over complexity**: Parallel execution, fast responses
- **Evidence-based**: Search results provide citation context
- **Canonical CLI**: `core/cli.py` (runs from the cache-resolved plugin root; see execution directive below)
- **Backends**: CKS, CHS (claude-history/Rust+FTS5), VAULT (claude-vault/FTS5), CDS, Code/Grep, DOCS, SKILLS
- **Search methods**: FTS5 (~10ms), Hybrid (~50ms), Semantic (~200ms)
- **Output formats**: Human-readable (default), JSON for scripting
- **CHS bootstrap**: If `chat_history.db` exists but FTS5 tables are missing, build it with `python -m core.chs.scripts.reindex_from_jsonl --db-path "P:/.data/chat_history.db" --history-path "~/.claude/history.jsonl"`

See `references/architecture.md` for backend architecture details and NotebookLM integration.

## Your Workflow

1. **Auto-detect intent** - Analyze query for chat/web/combined source needs
2. **Select backends** - Automatically choose based on query patterns
3. **Choose chat method** - Auto-detect vector (semantic) vs grep (recent)
4. **Expand boundary search** - Prefer producers, consumers, schemas, validators, replay, and invalidation logic when query implies contracts or handoffs
5. **Execute search** - Parallel query across selected backends
6. **Check CHS results** - If empty/unavailable, trigger JSONL fallback
7. **Collect results** - Gather with relevance scores
8. **Format output** - Group by source, show relevance, citations, match explanations
9. **Present findings** - Human-readable or JSON format

## Auto-Detection Rules

The system **automatically** detects search intent. No flags needed.

| Query Pattern | Source | Example |
|---------------|--------|---------|
| "we discussed", "you mentioned", pronouns | `chat` | "what did we discuss about auth" |
| "recent", "yesterday", "earlier today" | `chat` | "recent changes to the find skill" |
| "how does X work", "tutorial", code terms | `docs` | "how does FAISS work" |
| "compare X vs Y", "evolved over time" | `chat` + `docs` | "how our auth approach evolved" |
| URLs (http/https) | Suggest `/web` | direct fetch |

Chat method is also auto-selected: `grep` for recent/exact matches, `vector` for semantic/conceptual queries.

See `references/auto-detection.md` for full detection rules, query indicators, and backend weighting.

## Validation Rules

- **Before claiming search works**: Verify actual query execution, not assumption
- **Before citing results**: Ensure result actually contains claimed content
- **Empty results**: Report clearly, don't fabricate findings
- **Relevance scores**: Report actual scores from search engine
- **Do not invent module paths**: Never call `knowledge.systems.chs.cli` (does not exist)
- **CHS readiness**: A non-empty `chat_history.db` file is not enough; verify schema/FTS tables or reindex first

When the query is about architecture, handoffs, resume, stale data, fields, payloads, schemas, or consumers, `/find` should preferentially surface:

- existing producers
- existing consumers
- schemas/validators
- replay/invalidation logic
- transcript/resume artifacts

## Routing Behavior

`/find` is a discovery skill. It may suggest downstream skills once relevant evidence is surfaced:

- `/arch` when search shows unresolved contract/state boundaries
- `/planning` when execution shape is unclear after discovery
- `/verify` when the main question becomes "does this actually work?"

`/find` should not absorb architecture, planning, or verification responsibilities.

## Execution Directive

Run local-first find for `/find` requests (`mode='auto'` — checks local stores, then escalates to web when local results are unsatisfactory):

```bash
cd "$(ls -d "$HOME/.claude/plugins/cache/local/search-research/"*/ | sort -V | tail -1)" && python -c "
from core.unified_router import UnifiedAsyncRouter
import asyncio

async def search():
    router = UnifiedAsyncRouter(mode='auto')
    results = await router.search_async('$query', limit=10)
    for r in results:
        print(f'[{r.score:.2f}] {r.title}')
        print(f'  {r.content[:200]}...')
        print()

asyncio.run(search())
"
```

**Session-chain queries** (last 7 days):

```bash
cd "$(ls -d "$HOME/.claude/plugins/cache/local/search-research/"*/ | sort -V | tail -1)" && python -c "
from pathlib import Path
from datetime import datetime, timedelta

project = Path.home() / '.claude/projects/P--'
handoff_dir = Path('P://.claude/state/handoff')
cutoff = datetime.now() - timedelta(days=7)

transcripts = [p for p in project.glob('*.jsonl') if datetime.fromtimestamp(p.stat().st_mtime) > cutoff]
transcripts.sort(key=lambda p: p.stat().st_mtime, reverse=True)
print(f'=== Session Transcripts ({len(transcripts)} last 7 days) ===')
for t in transcripts:
    mtime = datetime.fromtimestamp(t.stat().st_mtime).strftime('%Y-%m-%d %H:%M')
    print(f'  [{mtime}] {t.name}')

handoffs = [p for p in handoff_dir.glob('console_*_handoff.json') if datetime.fromtimestamp(p.stat().st_mtime) > cutoff]
handoffs.sort(key=lambda p: p.stat().st_mtime, reverse=True)
print(f'\n=== Handoff Files ({len(handoffs)} last 7 days) ===')
for h in handoffs:

From EndUser123/search-research, MIT.

More from EndUser123/search-research