route-gpt56-task

name	route-gpt56-task
description	Route GPT-5.6 engineering work across eight roles with bounded topologies.

Route GPT-5.6 engineering work

Use only when the user invokes $route-gpt56-task. Routing is advisory: the root owns the outcome, may use any model, and may work directly.

Choose topology first

Task shape	Topology
Small or tightly coupled	Root-direct or one engineer
Evidence-heavy question	Owner plus read-only explorers
Competing failure hypotheses	Debugger plus bounded probes
Stable disjoint components	Owner plus parallel workers
Cross-layer, unsettled contracts	One engineer; explorers only
Critical mutation	Implementer, then independent reviewer
Mechanical validation	Deterministic tools, not agents

Name one outcome/integration owner. Freeze shared contracts before parallel writes. Parallel workers need disjoint paths and contracts. Expansion requires an independent ownership boundary, hypothesis, or review charter; there is no fixed agent cap.

Choose role

Work	Role	Route
Repeatable low-risk mechanics	gpt56_router_luna_worker	Luna/high
Read-heavy repository mapping	gpt56_router_terra_explorer	Terra/medium
Bounded everyday implementation	gpt56_router_terra_worker	Terra/medium
Broad competing investigation	gpt56_router_terra_investigator	Terra/high
Bounded architecture advice	gpt56_router_sol_advisor	Sol/medium
Ambiguous, cross-layer, critical implementation	gpt56_router_sol_engineer	Sol/medium
Difficult debugging or failed lower route	gpt56_router_sol_debugger	Sol/high
Independent critical review	gpt56_router_sol_reviewer	Sol/high

For a managed route, pass the exact role identifier as agent_type to spawn_agent; a separate stable task_name uses only lowercase letters, digits, and underscores. Always pass fork_turns: "none". If agent_type is unavailable or persisted role/model/effort differs, report the route unavailable. Never treat a generic or task-name-only child as an equivalent success. Model override is diagnostic only; it never substitutes for role instructions, leaf behavior, or sandbox policy.

Security, auth, payment, privacy, destructive, concurrency, data-loss, and persistent-state changes use Sol/medium execution then one Sol/high review. After a lower-route failure escalate at most to Sol/high. A Sol/high failure makes the route unavailable until the root materially revises the plan, decomposition, or verifier evidence.

Execute in phases

Evidence -> contract decision -> implementation -> integration -> deterministic validation -> independent review -> targeted repair -> convergence.

Reuse an existing child through follow-up for repairs; replace it only when unavailable or genuine independence is required. Children remain leaves and lack commit, push, deploy, or external-mutation authority unless granted. Wait at phase barriers rather than polling repeatedly.

Default handoff:

Objective: <bounded outcome>
Owned paths: <canonical paths or read-only>
Owned contracts: <interfaces and semantics>
Read-only dependencies: <inputs>
Invariants: <must remain true>
Forbidden changes: <explicit exclusions; remain a leaf>
Verify: <commands or evidence>
Return: outcome, changed paths, validation, blockers.

Do not copy conversation history, source, diffs, logs, credentials, or secrets into a handoff.

Use one integrated reviewer by default. Add a specialist only for a genuinely independent critical domain. The original owner repairs findings, then one targeted rereview runs. Stop when required checks pass and review delta is zero; unresolved material findings stop progression.

Normal routing invokes no scripts. For reproducible one-workstream advice use scripts/route_task.py with a task-profile-v6 input. For persisted evidence use scripts/inspect_spawn.py or scripts/inspect_execution.py.