---
name: token-mizer-handoff
description: For selected or explicitly activated Token Mizer, including explicit shared invocation through token-mizer-integrate. Delegate substantial bounded work with compact context and objective checks.
---

# Lean handoff

Apply only while Token Mizer is selected or explicitly activated, including task-scoped invocation through `token-mizer-integrate` by an explicitly invoked consuming agent. Discover actual host-listed dependency names. Use `token-mizer-route` first when the route is not established. Explicitly select a runtime ID accepted by the host. Foundry workers require a verified provider mapping. Gemini builders require the policy-authorized `gemini-3.8-flash` route or a provider-qualified equivalent supplied by the host; never invent a provider prefix.

## Before spawning

- Keep work requiring five or fewer direct calls in the current session only when coordinator context capacity and route eligibility are established.
- Delegate a self-contained objective, not an open-ended exploration.
- Reuse a suitable worker and its checkpoints for follow-ups.
- Estimate the complete context before delegation. Pool candidates perform the same task role. Allocate and then freshly admit immediately before each host call; admission must revalidate current availability, authorization, context and immutable role/provider/family/runtime identity. Pass the fresh `handoff.selected_runtime_id` to the host, never `selected_model` or a cached handoff. No allocator spawn command exists. Admission is not a spending reservation or exactly-once execution guarantee. Prefer one bounded Gemini builder for substantial implementation only when host capacity and the separate budget policy permit it. Otherwise use Foundry Luna when the task fits; never use Sol unless explicitly selected. Large-context work excludes every Foundry route, including coordinator/direct/Astra paths.
- For a large-context task, denied or unknown GitHub capacity blocks rather than silently choosing a small-context Foundry worker. Offer only a lossless bounded decomposition when correctness is preserved.
- An authorized Astra escalation uses the same select → fresh admit flow: supply `family: "astra"`, `provider: "Foundry"`, and `explicit_model` equal to the verified qualified `gpt-6-astra` runtime. Explicit selection does not grant permission or bypass context/budget gates. Astra never enters automatic Flash/Luna rotation; preserve the task's role and acceptance target.
- Allow only one active worker per constrained Foundry connection when throttling or shared-capacity contention is suspected. Do not launch a Sol, Astra, or Luna replacement on the same connection as a workaround unless the host proves capacity is independent.
- Preserve required specialist agents, skills, permissions, and approvals.
- Never create an unrelated repository session merely to test routing.

## Assignment contract

Provide:

1. **Outcome** - deliverable and acceptance criteria.
2. **Scope** - owned files or bounded question, plus exclusions.
3. **Evidence** - relevant paths, sources, reproduction, errors, and decisions.
4. **Execution** - permitted actions, accepted runtime model ID, paid authorization when applicable, validation, and approvals.
5. **Stop** - finish when checks pass. In an opted-in bounded-RUG task, allow one initial implementation and at most one repair after coordinator diagnosis; if the repaired result fails verification, return the checkpoint and blocker rather than looping. Authentication or environment errors stop immediately with recovery guidance. Explicit throttling follows the bounded retry policy in `token-mizer-route`.
6. **Return** - result, artifacts, checks and outcomes, and unresolved risks.

Aim for 500 words or fewer in the assignment and 200 words or fewer in the response unless failure evidence requires more.

## Accept results

Require verifiable artifacts and the same applicable repository-wide CI-equivalent checks, not a success claim or edited-file tests alone. Include affected-package tests for shared code or configuration, plus actual typecheck, lint, generated-file, architecture, coverage, and deploy-safety gates when the repository has them. Do not invent unrelated full-suite gates or add automatic reviewer agents. Independent review is risk-based for substantial, risky, or previously failed work after deterministic gates, and is skipped when Forge or another validator already supplies equivalent evidence. Honor running-app verification and merge policies. Do not repeat a successful investigation with another model. While a constrained Foundry worker is waiting, the coordinator yields instead of polling or consuming the same pool. Astra verifies the returned evidence and escalates only the unresolved decision or discrepancy.