handoff-routing-policy

name	handoff-routing-policy
description	Route the next allowed role after one explicit subAgent result, or stop, without encoding full workflows, registries, or runtime semantics.
complexity	medium

Purpose

Choose the next route after one explicit subAgent result.

This skill works only after a real dispatch has already occurred and a result payload has been returned.

Trigger / When to use

Use this skill when:

• one explicit subAgent result has been returned
• the Observer is in ROUTING
• the result includes one frozen verdict value

Do not use this skill when:

• the task is still deciding whether to dispatch
• the task is to build the context package
• the task requires full workflow reconstruction or runtime orchestration logic

Inputs

• result_role: one of Planner, Implementer, Reviewer, Correction Planner
• verdict: one of
  • PASS
  • PATCH_REQUIRED
  • REPLAN_REQUIRED
  • MISSING_EVIDENCE
  • BLOCKED
• bounded evidence summary
• explicit blocker list, if any
• optional evidence owner for MISSING_EVIDENCE

Process

1. Confirm the result came from real dispatch rather than Observer simulation.
2. Confirm the verdict is one of the frozen allowed values.
3. If the result reveals runtime semantics, registry behavior, workflow binding, or another out-of-scope expansion, stop.
4. Route by verdict without inventing a broader workflow model:
   • PASS: choose the next role only if one more bounded role handoff is needed; otherwise stop
   • PATCH_REQUIRED: route to Implementer
   • REPLAN_REQUIRED: route to Correction Planner
   • MISSING_EVIDENCE: route only to the bounded role that can supply the missing evidence; if that owner is unknown, stop
   • BLOCKED: stop unless a bounded route to Planner is explicitly justified
5. Emit exactly one next role or stop, with a short factual reason.

Examples

• Positive: An Implementer returns PATCH_REQUIRED with concrete bounded evidence, and the skill routes to Implementer.
• Negative: A result says "probably approved" with no explicit verdict, and the skill refuses to invent one.

Outputs

• next_role: Planner | Implementer | Reviewer | Correction Planner | stop
• reason: short factual reason
• stop_condition: none or exact blocker

Validation

Required Checks

• PASS: the result came from real dispatch, the Observer is in ROUTING, the verdict is exactly one frozen allowed value, and the skill returns exactly one allowed next_role or stop consistent with the stated verdict.
• BLOCKED: stop when the verdict is unknown, unstructured, or unsupported; when the evidence owner for MISSING_EVIDENCE is unknown; or when proceeding would require invented workflow state, registry behavior, runtime semantics, or a broader routing model than this skill allows.

Quality Checks (best effort)

• SOFT FAIL: mark status as INCOMPLETE when the allowed verdict is clear enough to route or stop, but the bounded evidence summary or blocker detail is incomplete.
• Under SOFT FAIL, keep the routing decision within the frozen verdict set, state the missing evidence explicitly, and avoid inventing additional workflow state.

Failure Handling

Missing Context

• If the verdict is valid but supporting evidence is partial, return the bounded best-effort route or stop with the limitation stated explicitly.
• If missing context could reasonably change the next allowed role, mark the result BLOCKED and stop.

Ambiguous Requirement

• If the returned result does not contain one frozen verdict value, do not infer or normalize it into the allowed set.
• If MISSING_EVIDENCE lacks a known bounded owner, stop instead of inventing one.

Execution Limitation

• If the request depends on hidden workflow state, file-path ownership, registry identifiers, launcher-specific targets, or runtime orchestration details, return stop or BLOCKED with the limitation stated explicitly.
• Do not fabricate broader workflow progression beyond the single returned subAgent result.

Boundaries

• Do not choose the initial dispatch role.
• Do not build the context package.
• Do not reconstruct the full existing workflow from a step artifact.
• Do not invent verdict values outside the frozen set.
• Do not emit registry identifiers, file paths, or launcher-specific targets.

Local references

• examples.md: verdict-driven routing examples, including stop conditions for missing evidence and out-of-scope expansion