zob-tool-router

name	zob-tool-router
description	Use before non-trivial work, uncertain tool selection, orchestration, delegation, context lookup, factory work, autonomous-runtime claims, or evidence-gated completion.

ZOB Tool Router Skill

Purpose

This skill is the compact playbook for deciding which ZOB tool families, modes, and domain skills to use. It is not a runtime dispatcher and not a reason to use every tool. Use it to choose the smallest sufficient tool set, then follow each selected tool family's own safety gates.

When to use

Use this skill when any of these are true:

• the user ask is non-trivial, multi-step, or evidence-gated;
• the right mode/tool family is ambiguous;
• the work may need delegation, orchestration, ZTeam/ZAgent creation, tmux-backed agent teams, ZPeer room mapping, factory design, context retrieval, ProjectDNA, coms, workspace/merge coordination, autonomous-runtime evidence, or oracle review;
• you are about to claim completion for broad or risky work.

Do not use this skill for a small factual answer when no tools are needed.

Core principles

1. Do not use every tool. Use the smallest sufficient set.
2. The registry is the source of truth for available tool families, mode allowlists, skill refs, docs, and no-ship notes: .pi/capabilities/zob-public-runtime-capabilities.json.
3. This router selects families and domain skills; it does not bypass mode, approval, sandbox, budget, evidence, or oracle gates.
4. If a family is applicable, either use it, delegate it, or explicitly skip it with a reason.
5. Prefer evidence-backed progress over autonomy claims.

Routing loop

1. Classify the request:
   • trivial answer;
   • bounded code/doc edit;
   • multi-step goal;
   • multi-agent orchestration;
   • ZTeam/ZAgent team, tmux-backed team, ZPeer rooms, aliases, or room membership;
   • factory/repeated workflow;
   • context/reference lookup;
   • validation/oracle;
   • autonomous-runtime claim.
2. Identify applicable tool families from the table below.
3. Load specialized skills for selected families when their behavior matters.
4. Pick the smallest safe next action: answer directly, inspect, plan, implement, delegate, orchestrate, validate, block, or ask the user.
5. Attach evidence to TODOs or final reports when work is broad, delegated, or completion-gated.
6. If a human-decision blocker is already recorded (score >=90, no nextAgent, paused goal, visible blocker), report it once, do not repeat/redispatch, and wait for explicit /goal resume or resume_goal.
7. Escalate to oracle when no-ship, final completion, security, autonomous, sandbox, or promotion risk is involved.

Tool family routing table

Family	Use when	Tools / skills to consider
goal/TODO	Long-running, multi-step, delegated, factory, or evidence-gated work	create_goal, add_goal_todos, get_goal_todos, resolve_goal_todo; skill zob-goal-todo-tree
delegation	Specialist review, parallel discovery, implementation handoff, independent QA, or uncertain broad work	zob_delegation_catalog, delegate_task, delegate_agent; skill zob-delegation-routing
orchestration	Lead/Worker lanes, Chief Vision coordination, parent-owned dispatch, multi-agent workgraph	/zmode orchestrator, orchestrate_run, chain_run, goal/TODO tools
zteam/zagent	Owner asks for full Pi-session agents, ZTeams, ZAgents, tmux-backed teams, ZPeer rooms/aliases/membership, /zteam launch-plan, hot-add/remove, or team launchers	/zteam, /zagent, zob_zteam_hot_add, zob_zteam_remove; skills zob-zagent-creator, zob-coms-v2-live, zob-coms-safety
compute	Complexity, budget, model class, max profile, or multi-agent sizing matters	zob_compute_preview, zob_compute_resolve_profile, zob_compute_plan_workflow, zob_compute_validate_profile; skill zob-compute-profile
context / ProjectDNA	Need bounded repo/reference context, scan artifacts, cited context packs, or writeback proposals	zob_context_search, zob_context_validate_scope, zob_context_readiness, zob_project_dna_query, zob_project_dna_federated_query, zob_project_dna_readiness; skills zob-context-discovery, zob-project-dna
factory	Repeated workflow, smoke/pilot/batch gates, manifests, checkpoints, sentinels	factory_run, factory quarantine tools, autonomous factory read-only smokes; skill zob-factory
coms / goal-room	Parent-visible coordination, live required-local handoff, blockers, TODO claims, status refs	zob_goal_room_*, zob_coms_*; skills zob-coms-v2-live, zob-coms-safety
workspace / merge queue	Parallel write intent, sandbox diff review, parent-owned manual apply decisions	zob_workspace_claim, zob_workspace_release, zob_merge_candidate_submit, zob_merge_queue_decide; skill zob-sandbox
worker-pool	Parent-owned same-type/read-only or write-by-owner pool planning, path ownership/read-across coordination, owner requests/decisions	zob_worker_pool_plan, zob_worker_pool_status, zob_worker_pool_owner_request, zob_worker_pool_owner_decision; skills zob-harness, zob-delegation-routing, zob-coms-safety
worklist / liveness	Append-only metadata-only blackboard: append events, read derived directives, lease/satisfy by content hash, deliver idempotent DIRECTIVE_READY, run the liveness watchdog (observe/escalate), build/query the dependency DAG	zob_worklist (append/directives/claim/satisfy/deliver/validate/observe/escalate/dag); skill zob-worklist
autonomous runtime	Dry-run, readonly smoke, validation evidence, or any autonomy readiness claim	zob_autonomous_*; skill zob-autonomous-runtime
Mission Control	Dashboard snapshots, readiness, pause/resume/replan/request-oracle proposals	zob_mission_control_*, zob_coms_readiness; skill zob-mission-control-coms
oracle	Final readiness, no-ship, safety/security, release, sandbox apply, or autonomous enablement	oracle agent/review, propose_goal_completion, record_goal_oracle; skill zob-oracle

Anti-over-orchestration rules

• Do not create a runtime goal for a simple question or tiny one-shot answer.
• Do not delegate if direct answer or local inspection is enough.
• Do not run compute profiling for a small obvious edit.
• Do not use coms/goal-room when no other agent/run is involved.
• Do not create custom tmux/file-room communication when the request needs full-session ZAgents, ZTeams, ZPeer rooms, aliases, or membership; route to the ZTeam/ZAgent workflow and validate with /zteam launch-plan <team-id>.
• Do not invoke factory tooling unless the workflow is repeatable or explicitly factory-shaped.
• Do not treat dry-run, readonly smoke, plan-only orchestration, or mocked dispatch as final autonomous production readiness.
• Do not use workspace/merge queue for single-agent non-overlapping edits unless sandbox governance is required.
• Do not bypass oracle when completion, no-ship, sandbox promotion, autonomous enablement, or security risk is in scope.

Routing verdict format

For non-trivial or ambiguous work, produce a compact internal or visible verdict before the first substantive action when useful:

TOOL ROUTING VERDICT
- complexity: trivial | small | non-trivial | high-risk
- selected mode: explore | plan | implement | oracle | factory | orchestrator
- applicable families:
  - goal/TODO: use | delegate | skip — reason
  - delegation: use | delegate | skip — reason
  - orchestration: use | delegate | skip — reason
  - zteam/zagent: use | delegate | skip — reason
  - compute: use | delegate | skip — reason
  - context/ProjectDNA: use | delegate | skip — reason
  - factory: use | delegate | skip — reason
  - coms/goal-room: use | delegate | skip — reason
  - workspace/merge: use | delegate | skip — reason
  - worker-pool: use | delegate | skip — reason
  - autonomous: use | delegate | skip — reason
  - oracle: use | delegate | skip — reason
- selected skills:
- smallest sufficient next action:

Keep the verdict short. For routine small work, summarize only the selected next action instead of printing the full matrix.

Completion reminder

Before claiming broad work complete, verify evidence, required TODO closure, validation commands, and oracle/no-ship requirements. If evidence is missing, continue, block, or ask the user instead of claiming success.