Check Ollama connectivity plus fast/strong adaptive model availability and resource settings.

Classify as deterministic, local, local-supervised, or Claude. local-supervised means Claude already resolved a sensitive auth/credential/permission/security decision and only bounded implementation remains; full-diff Claude review is mandatory.

List a bounded safe workspace view and root package scripts without following symlinks.

Literal bounded text/code search inside a safe workspace.

Ask the fast local model for bounded code/analysis text without modifying repository files.

Compatibility full-result executor. Prefer execute_local_code_task_compact, which includes routing preflight, adaptive fast-to-strong retry, and supervised-sensitive review enforcement.

Compatibility full-result orchestrator. Supports local-supervised subtasks and adaptive fast-to-strong retry; prefer the compact orchestrator for enforced review behavior.

Build a compact evidence-first context capsule for a coding task using a persistent local repository index. Prefer this before broad Claude file exploration; verify cited file:line evidence for architectural or risky decisions.

Preferred bounded local executor. It preflights routing itself. Sensitive auth/credential/permission/security implementation may run as local-supervised only after Claude explicitly resolves the sensitive decision; supervised runs force full-diff Claude review.

Preferred large-feature orchestrator. Claude owns planning/decomposition. Already-resolved sensitive subtasks can run as local-supervised; any such task forces full aggregate-diff review by Claude.

Fetch a stored local execution result only when Claude needs more detail. Prefer summary first; request diff, validation, or full incrementally with offset/maxChars instead of loading a large result into context at once.

Aggregate privacy-preserving routing/execution/orchestration telemetry including exact per-model Ollama inference usage.