| name | llm-canvass |
| description | Apply an LLM judgment task to every in-scope file in the repo, in parallel. Bin-packs files into balanced token slices, fans out general-purpose subagents that read full files, synthesizes results. |
| argument-hint | --type|--glob|--all <task> |
| user-invocable | true |
| disable-model-invocation | false |
| effort | max |
| allowed-tools | ["Bash(.claude/skills/llm-canvass/plan-slices.py:*)","Agent"] |
LLM-Canvass: Parallel LLM Judgment over the Codebase
Run a structured LLM judgment task across every in-scope source file in
parallel. The user gives the task; the skill handles enumeration, slicing,
fan-out, and synthesis.
Phase 1: Plan slices
Run plan-slices.py via the Bash tool — single-quote the task description so
parens, backticks, quotes, and other shell metacharacters in the user's task
pass through unparsed. Flags (--type, --glob, --all, --agent,
--background, --max-agents, --per-agent-tokens) come before the task.
.claude/skills/llm-canvass/plan-slices.py --type py 'your task description'
The companion script:
- Enumerates files via
rg --files under the required scope (--type, --glob, or --all)
- Estimates tokens (bytes ÷ 4)
- Greedy bin-packs into balanced slices (~400K tokens each, default cap 8)
- Writes
brief.md and slice-N.txt to the session scratchpad
- Prints plan summary: file count, total tokens, slice distribution, scratchpad path
Phase 2: Fan out
In a single message, emit N parallel Agent calls. Use the
subagent_type Phase 1 printed (default general-purpose; user override
via --agent). Use the run_in_background value Phase 1 printed
(default false; user opts into background via --background).
For each slice N, the agent prompt is (substitute <scratchpad> with the
absolute path Phase 1 printed on its Scratchpad: line):
Read <scratchpad>/brief.md for your task. Then read
<scratchpad>/slice-N.txt for your file list. Read every file in the
list IN FULL. Apply the task.
Phase 3: Synthesize
When all agents return, synthesize the N tool_result blocks into the
shape the task asked for.
Operating defaults
- Scope: required —
--type, --glob, or --all.
- Agent type:
general-purpose (override --agent <type>).
- Execution mode: synchronous (override
--background).
- Per-agent budget: ~400K source tokens.
- Agent count:
ceil(total_tokens / 400_000), capped at 8.
- Output format: dictated by the task. The skill imposes none.
