bulk

name	bulk
description	Bulk-parallel CLIs — turn N sequential gh/bd/git/file tool calls into a single fan-out JSON call. Use when the session is about to fire ≥3 similar sequential calls (gh pr view, bd show, Read of small files, up-to-date diagnose across repos).
allowed-tools	Bash, Read

Bulk Parallel Tools

One tool call firing N parallel sub-calls beats N sequential calls on wall-clock almost every time, and it keeps main-thread context cleaner. This skill ships five fan-out CLIs that shell out to gh / bd / diagnose.py / filesystem across a ThreadPoolExecutor and emit one JSON array.

When to use

Trigger when you're about to make ≥3 similar sequential tool calls:

• "Check the state of these N PRs" → bulk-gh-pr-details
• "List open PRs across these N repos" → bulk-gh-prs-open
• "Show me these N beads" → bulk-bd-show
• "Diagnose git state across these N repos" → bulk-up-to-date
• "Give me quick text of these N files" (each likely < 1 MB) → bulk-file-read

If the count is 1 or 2, use the direct tool call — bulk wins on wall-clock but loses on boilerplate for small Ns.

Install

Packaged as chop-bulk. Install via:

cd ~/gits/chop-conventions
uv tool install --force --reinstall ./skills/bulk/

Once PR #169 (packaged skill CLIs) lands, install-tools.py will pick up chop-bulk from its REGISTRY and install it alongside the others.

After install, five binaries land on $PATH:

• bulk-gh-pr-details
• bulk-gh-prs-open
• bulk-bd-show
• bulk-up-to-date
• bulk-file-read

Usage

Every tool takes inputs three ways — pick whatever's convenient:

1. Positional args (most common when the list is in plain text):

   bulk-gh-pr-details idvorkin/chop-conventions#169 idvorkin/chop-conventions#168
   

2. --input-file path.json (when the list is already on disk as JSON):

   bulk-bd-show --input-file /tmp/beads.json
   

3. stdin JSON array (for piping from another tool):

   echo '["igor2-bgt","igor2-88g","igor2-5i1"]' | bulk-bd-show
   

All tools support --max-workers N (default 8) and --pretty. Output is JSON to stdout; progress/errors go to stderr. Per-item failures are captured inline as {..., "error": "..."}. The batch never partial-fails.

Tools

bulk-gh-pr-details

Input: owner/repo#N specs. Output per PR:

{"repo": "idvorkin/chop-conventions",
 "number": 169,
 "title": "feat(packaging): migrate ...",
 "state": "OPEN",
 "mergeable": "MERGEABLE",
 "mergeStateStatus": "CLEAN",
 "url": "https://github.com/idvorkin/chop-conventions/pull/169"}

bulk-gh-prs-open

Input: owner/repo slugs. Output per repo:

{"repo": "idvorkin/chop-conventions",
 "open_prs": [
   {"number": 169, "title": "feat(packaging)...", "headRefName": "feat/uv-tool-packaging"}
 ]}

bulk-bd-show

Input: bead IDs. Output per bead:

{"id": "igor2-bgt",
 "title": "Telegram Infrastructure",
 "status": "open",
 "priority": 1,
 "type": "epic",
 "parent": null,
 "blocks": [],
 "blocked_by": []}

Handles the bd show --json array-vs-object gotcha internally — callers don't need to re-unwrap.

bulk-up-to-date

Input: absolute repo paths. Output per repo:

{"repo": "/home/developer/gits/chop-conventions",
 "diagnose_json": { ... full diagnose.py output ... }}

Invokes up-to-date-diag if packaged (PR #169) else the script at ~/.claude/skills/up-to-date/diagnose.py. cwd is set to the target repo so diagnose introspects the right directory.

bulk-file-read

Input: absolute file paths. Output {path: {size_bytes, content_utf8_or_null, error_or_null}}. Files > --max-bytes (default 1 MB) are skipped with an error, not loaded — keeps output sane on accidental big-file globs.

Design

• One pyproject.toml, five entry points. Matches the per-skill package pattern from PR #169 (gen-tts, harden-telegram, up-to-date each bundle their own tools).
• _build_app() lazy-import of Typer. Tests and pre-commit hooks import the pure-function layer (fetch_pr, fetch_bead, diagnose_repo, _read_one) in system Python without ModuleNotFoundError. Reference: skills/harden-telegram and skills/up-to-date in chop-conventions.
• Shared common.py owns input parsing, the ThreadPoolExecutor wiring, and the "never partial-fail" contract. The abstraction is justified at N=5 (the "wait for N=2" rule from CLAUDE.md).
• Subprocess injection for tests. Each worker fn takes a run=subprocess.run kwarg so tests mock it without patching the module global. Matches the test_diagnose.py and test_telegram_debug.py style in this repo.

Tests

python3 -m unittest discover -s skills/bulk/tests -p 'test_*.py'

Covered: bulk-gh-pr-details spec parsing + fetch + fan-out, and bulk-bd-show normalization of the array-vs-object JSON + dependency slicing. Other tools use the same common.py plumbing; the two tested paths validate the shared contract.

Related

• up-to-date — single-repo version of bulk-up-to-date.
• learn-from-session — reflection prompts now include a "≥3 sequential similar calls → propose bulk" heuristic.