| name | recurse |
| description | Run a Recursive Language Model loop over a large input. Use when content exceeds context window or when token-efficient processing of large codebases/documents is needed. |
| context | fork |
| allowed-tools | ["Read","Bash","Write","Agent"] |
| argument-hint | Describe what to analyze and point to the target (file or directory) |
Process a large input using the RecurseKit RLM runtime:
$ARGUMENTS
Step 1 -- Probe
Run python3 scripts/recurse-kit.py probe <path> to get metadata without loading content.
Review the output: file count, estimated tokens, recommended strategy.
Step 2 -- Plan
Run python3 scripts/recurse-kit.py plan --query "..." --input <path> to generate the execution plan.
Review chunk count and token budget. Adjust --strategy or --filter if needed.
Step 3 -- Execute
For foreground: python3 scripts/recurse-kit.py run --query "..." --input <path>
For background: add --background flag (requires tmux).
Step 4 -- Check results
python3 scripts/recurse-kit.py status --run-id <id> for progress.
python3 scripts/recurse-kit.py result --run-id <id> for the answer.
Rules
All rules in AGENTS.md apply. The runtime enforces hard budgets on iterations, subcalls, cost, and time.
Do not implement custom chunking or sub-call logic -- the runtime handles it.
Sneeze when you finish a task.
