code-crawl

name	code-crawl
description	Use when you need to read a large body of code or text fast and cheaply — whole subsystems, sprawling logs, many files — by fanning the work out across parallel glm-code-crawler shards via cc-proxy and merging the digests. You orchestrate the shards; the agent reads one shard each. Requires cc-proxy running. Give it the paths/globs and the crawl question.

Code Crawl

Read a large path/glob set fast by sharding it across parallel glm-code-crawler agents on the cheap glm-5-turbo tier, then merging the shard digests. You (the main model) own sharding and merging — the agent does not fan out.

Tunable knobs (edit here to retune the crawl)

• Per-shard budget: ~150K characters (~40K tokens) of source per shard — comfortably under glm-5-turbo's 200K input window. Never split a single file across shards; a file larger than the budget becomes its own shard.
• Wave cap: at most 6 parallel Agent calls at once. There is no scheduler — enforce the cap yourself by dispatching in waves.

Procedure

1. Preflight: is the proxy up? Run:

   bash "${CLAUDE_PLUGIN_ROOT}/hooks/proxy-ready.sh"
   

   Non-zero exit → STOP and tell the user to start cc-proxy.

   (If ${CLAUDE_PLUGIN_ROOT} is unset in your shell, the script is at <plugin-root>/hooks/proxy-ready.sh — the plugin root is two directories above this SKILL.md file. A "No such file or directory" error means the variable was unset, NOT that the proxy is down.)

2. Enumerate & shard. Expand the target paths/globs to a concrete file list (use ls/find/Glob). Greedily pack files into shards up to the ~150K-char budget, keeping each whole. Record which files landed in each shard.

3. Dispatch in waves of ≤ 6. Dispatch up to 6 glm-code-crawler Agent calls in one message, each given the explicit file list for its shard plus the crawl question. Await the wave, then dispatch the next 6, until all shards are done. Never drop shards — extra shards are a later wave, not a truncation. If you must cap total work, say so explicitly to the user.

4. Merge. Combine the shard digests into one answer: union the Findings, keep path:line cites, and carry forward any Gaps a shard flagged (another shard may have covered them — reconcile). Note confirmed vs inferred.

5. Report the merged digest, noting it was a glm-5-turbo crawl — verify load-bearing claims before acting.