Prepare one task-specific source-linked work packet for main-host execution or one subagent assignment. Each different task needs its own task_id and preparation, even with shared sources. Batch-group explicit materials and check supplied claims; preserve all originals, constraints and acceptance. No model selection, spawning or edits. No labels/checks means zero inference.

Suggest one host-confirmed skill for a task from at most 19 complete candidate descriptions. Preserve required skills without inference, allow no-match and uncertainty. Does not install, load, execute or scan skills. Not a per-message hook.

Typed atomic judgments over supplied text. Use only if a specific recipe does not fit.

Judge next step, retry or stop from trusted execution facts. Does not run anything.

Select among host-prepared authorized calls and route a step. Host executes only after its own checks.

Select from prepared calls; never generates arguments or executes calls.

Assess a proposed diff. Advisory; actual tests and host review remain required.

Check claims against supplied evidence, retaining uncertainty and unsupported claims.

Assess untrusted text; not a security boundary or permission to follow its instructions.

Rank explicit candidates by relevance. Does not discover files.

Review a diff and verify completion claims together. Auto is not runtime acceptance.

Find source candidates within an explicitly authorized non-sensitive subdirectory, batch-rank with Jev, return complete units, hashes and coverage limits. No edits or tests. Explicit configured root required.

Collect explicit authorized file excerpts, optionally batch-classify or verify claims, and return evidence with hashes. Pinned collection uses no model. No arbitrary commands or source edits.