cqs-batch - SKILL.md Agent Skill

name: cqs-batch description: Execute multiple cqs queries in a single persistent session via stdin batch mode. disable-model-invocation: false argument-hint: ""

Batch Mode

cqs batch reads commands from stdin, outputs compact JSONL. Persistent Store + lazy Embedder — amortizes startup cost across N commands. Supports pipeline syntax for chaining commands.

Usage

# Single command
echo 'callers search_filtered' | cqs batch

# Multiple commands
printf 'callers gather\nexplain gather\nstats\n' | cqs batch

# Pipeline — chain commands via |
echo 'callees dispatch | callers | test-map' | cqs batch

# From file
cqs batch < commands.txt

Pipeline Syntax

Chain commands with | — upstream names feed downstream commands via fan-out:

# Search → get callers of each result
echo 'search "error handling" --limit 5 | callers' | cqs batch

# 3-stage: callees → callers of each → test coverage
echo 'callees main | callers | test-map' | cqs batch

# Dead code → explain each
echo 'dead --min-confidence high | explain' | cqs batch

Pipeable downstream commands: blame, callers, callees, deps, explain, similar, impact, test-map, related, scout.

Pipeline output is a JSON envelope (wrapped like every batch line, see Output below):

{"data": {"pipeline": "...", "stages": N, "results": [{"_input": "name", "data": ...}], "errors": [...], "total_inputs": N, "truncated": false}}

Limits: Max 50 names per stage (prevents fan-out explosion; override with CQS_PIPELINE_FAN_OUT, clamped 10-1000). Quoted pipes (search "foo | bar") are not treated as separators.

Supported Commands

Command	Example
`search <query>`	`search "error handling" --limit 3`
`callers <name>`	`callers search_filtered`
`callees <name>`	`callees gather`
`deps <name>`	`deps Store` or `deps --reverse search_filtered`
`explain <name>`	`explain search_filtered`
`similar <name>`	`similar gather --limit 3`
`gather <query>`	`gather "retry logic" --expand 2`
`impact <name>`	`impact search_filtered --depth 2`
`test-map <name>`	`test-map search_filtered`
`trace <src> <tgt>`	`trace main search_filtered`
`dead`	`dead --min-confidence high`
`related <name>`	`related gather --limit 3`
`context <path>`	`context src/lib.rs --compact`
`stats`	`stats`
`scout <query>`	`scout "error handling" --limit 10`
`where <desc>`	`where "new CLI command"`
`read <path>`	`read src/lib.rs --focus enumerate_files`
`stale`	`stale`
`health`	`health`
`notes`	`notes --warnings` or `notes --patterns`
`blame <name>`	`blame search_filtered`
`onboard <concept>`	`onboard "indexing pipeline"`
`task <description>`	`task "add a --rerank flag"`
`review` / `ci`	`review --base main` / `ci --gate high`
`impact-diff`	`impact-diff --base main`
`diff` / `drift`	`drift gemma --min-drift 0.1`
`plan <description>`	`plan "fix panic in gather"`
`suggest`	`suggest`
`gc`	`gc` (MUTATING — removes stale chunks)

Input Format

One command per line
# comments and empty lines are skipped
quit or exit stops processing
Quoted strings supported: search "multi word query"
| chains commands (pipeline syntax)

Output

Compact JSONL — one JSON object per line, flushed after each command. Every line is an envelope: success is {"data": <command output>}, errors are {"error":{"code":"...","message":"..."}}. Pipelines produce a single envelope per pipeline.