RTPI-ltc/Task-Compass-Skill

Route OpenClaw Task

Produce a deterministic, auditable routing decision before planning or execution. Treat the result as advisory policy input; never use it to bypass runtime permission checks.

OpenClaw Runtime Integration

Install this repository as an OpenClaw workspace skill, or copy this directory to skills/route-openclaw-task. OpenClaw loads the instructions automatically; the bundled router is invoked explicitly when a deterministic route is needed.

This project's optional LocalAuditPlanner adapter can enable automatic runtime routing with:

OPENCLAW_PLANNER_SKILL=route-openclaw-task

When a task includes an Available tool interfaces: catalog, resolve the required tool dependency closure from the task intent and ignore unrelated distraction tools. The runtime adapter converts execution_tools into a bounded target path and passes it to OpenClaw's A*/Reflexion planner as an auditable desired-tool constraint. Candidate generation, permission decisions, risk checks, and verification remain owned by OpenClaw.

Do not use this skill as a command generator, a tool-argument generator, or a replacement for runtime permissions.

Route One Task

Run the bundled router from this skill directory:

python3 scripts/route_task.py --goal "<task>" --pretty

Pass the runtime permission mode when known:

python3 scripts/route_task.py --goal "<task>" --permission-mode DEFAULT --pretty

Use the returned planned_tools, primary_executor, policy_mode, permission_behavior, next_action, and safety_guard to construct the next planner step. Preserve deny, await_human, replan, and refuse outcomes exactly.

Route Many Tasks

Provide JSONL rows containing goal and optional id and permission_mode fields:

python3 scripts/route_task.py --input-jsonl tasks.jsonl --output-jsonl routes.jsonl
python3 scripts/validate_route.py routes.jsonl

Do not pass expected labels or benchmark annotations to the router.

Evaluate

Measure routing behavior without copying raw prompts into the report:

python3 scripts/evaluate_router.py --suite benchmark.json --output metrics.json

Compare task_route_success_rate, safety_guard_accuracy, required_execution_tool_coverage, and holdout results before promoting a change. Keep the existing runtime permission engine as the final authority.

Interpret Decisions

Read references/routing-contract.md when integrating the JSON output, changing enum values, or investigating a route. Do not load assets/profile-policy-model.json into the prompt; let the script read it.

Read references/model-card.md when reviewing training provenance, benchmark claims, licensing, limitations, or promotion gates.

Apply these invariants:

• Never downgrade confirm or refuse to act without external verifier evidence.
• Never turn safety_guard=true off downstream.
• Route low-confidence actionable tasks to replan with expanded context.
• Keep explicit read-only constraints and workspace boundaries intact.
• Keep model predictions, deterministic signals, and the final reason visible in audit logs.