mq（MQ 搜索）

mq Skill: Efficient Document Querying

mq doesn't compute answers - it externalizes document structure into your context so you can reason to answers yourself.

Documents → mq query → Structure enters your context → You reason → Results

The Pattern

1. See structure    →  mq <path> .tree            → Map enters your context
2. Find relevant    →  mq <path> ".search('x')"   → Locations enter your context
3. Extract content  →  mq <path> ".section('Y') | .text"  → Content enters your context
                       mq <path> ".search('x') | .text" → Flatten matched structured results
                       mq <path> ".search('x') | .nth(0)" → Show one raw matched result
                       mq <path> ".search('x') | .nth(0) | .raw" → Explicit raw record
4. Reason           →  You compute the answer from what's now in your context

Your context accumulates structure. You do the final reasoning.

Quick Reference

# Structure (your working index)
mq file.md .tree                    # Document structure (headings, sections, previews)
mq dir/ .tree                       # Directory overview (all files with sections + previews)

# Search
mq file.md ".search('term')"        # Find sections containing term
mq dir/ ".search('term')"           # Search across all files
mq log.jsonl ".search('error')"     # JSONL: line-level search with record context

# Extract
mq file.md ".section('Name') | .text"   # Get section content
mq file.md ".code('python')"            # Get code blocks by language
mq file.md .links                       # Get all links
mq file.md .metadata                    # Get YAML frontmatter
mq log.jsonl ".search('error') | .text"            # Flatten matching records
mq log.jsonl ".search('error') | .nth(0)"          # Narrow to one raw matching record
mq log.jsonl ".search('error') | .nth(0) | .raw"   # Explicit raw record

Efficient Workflow

Starting: Get the Map

# For a single file
mq README.md .tree

# For a directory (start here for multi-file exploration)
mq docs/ .tree

Output shows you the territory:

docs/ (7 files, 42 sections)
├── API.md (234 lines, 12 sections)
│   ├── # API Reference
│   │        "Complete reference for all REST endpoints..."
│   ├── ## Authentication
│   │        "All requests require Bearer token..."

Now you know: API.md has auth info, 234 lines, section called "Authentication".

Finding: Narrow Down

If you need something specific but don't know where:

mq docs/ ".search('OAuth')"

Output points you to exact locations:

Found 3 matches for "OAuth":

docs/auth.md:
  ## Authentication (lines 34-89)
     "...OAuth 2.0 authentication flow..."
  ## OAuth Flow (lines 45-67)

Now you know: auth.md, section "OAuth Flow", lines 45-67.

Extracting: Get Only What You Need

Don't read the whole file. Extract the section:

mq docs/auth.md ".section('OAuth Flow') | .text"

This returns just that section's content.

Anti-Patterns

Bad: Reading entire files

cat docs/auth.md  # Wastes tokens on irrelevant content

Good: Query then extract

mq docs/auth.md .tree                           # See structure
mq docs/auth.md ".section('OAuth Flow') | .text"  # Get only what's needed

Bad: Re-querying structure you already have

mq docs/ .tree    # First time - good
mq docs/ .tree    # Again - wasteful, you already have this in context

Good: Use what's in your context

mq docs/ .tree    # Once - now you know the structure
# Use the structure you learned to make targeted queries
mq docs/auth.md ".section('OAuth') | .text"

Context as Working Memory

Every mq output enters your context. Your context becomes a working index that grows as you explore:

Query 1: mq docs/ .tree
→ You now see: file list, line counts, section counts
→ You can reason: "auth.md looks relevant to my question"

Query 2: mq docs/auth.md .tree
→ You now see: auth.md's full section hierarchy
→ You can reason: "OAuth Flow section has what I need"

Query 3: mq docs/auth.md ".section('OAuth Flow') | .text"
→ You now have: the actual content
→ You can reason: compute the final answer

mq externalizes structure. You do the thinking. Don't re-query what you already see.

Examples by Task

"Find something in a JSONL session file"

mq session.jsonl ".search('deploy')"  # Line-level matches with record type
# → [line 5] user/user
#     content: Can you deploy the new version?
#     ts: 2026-02-01T20:25:29Z
# → [line 8] assistant/tool_use: Bash
#     ts: 2026-02-01T20:25:34Z

mq session.jsonl ".search('deploy') | .text"           # Flatten matching records
mq session.jsonl ".search('deploy') | .nth(1)"         # Narrow to one raw matching record
mq session.jsonl ".search('deploy') | .nth(1) | .raw"  # Explicit raw record

"Find how authentication works"

mq docs/ ".search('auth')"           # Find relevant files/sections
mq docs/auth.md ".section('Overview') | .text"  # Read the overview

"Get all Python examples"

mq docs/ .tree                       # Find files with examples
mq docs/examples.md ".code('python')"  # Extract all Python code

"Understand the API structure"

mq docs/api.md .tree                 # See all endpoints/sections
mq docs/api.md ".section('Endpoints') | .tree"  # Drill into endpoints
mq docs/api.md ".section('POST /users') | .text"  # Get specific endpoint

"Find configuration options"

mq . ".search('config')"             # Search entire project
mq config.md ".section('Options') | .text"  # Extract options