---
name: token-mizer-bounded-rug
description: Only for the selected Token Mizer agent after an explicit bounded-RUG pilot opt-in. Run one build, deterministic verification, and at most one diagnosed repair with durable acceptance evidence.
---

# Bounded build-verify-repair

Apply only while Token Mizer is selected and the user has opted this task into the bounded-RUG pilot. The normal Token Mizer workflow remains the default.

## Contract

1. Define one coherent task, its class, coordinator/worker/validator scopes, exact acceptance boundary, starting revision, and environment.
2. Keep tasks needing five or fewer direct calls in the coordinator only when its context capacity and route eligibility are established. For substantial work, estimate total context first and use one suitable context-fitting builder. Reuse that worker and compact checkpoints. Large-context work requires confirmed GitHub capacity; it must block rather than silently use small-context Luna when Flash is denied. Use a fresh admission's exact `handoff.selected_runtime_id` for each host invocation; never reuse cached permission.
3. Run one initial implementation, then the deterministic gates for the selected boundary. Bind verification to the exact built revision and environment; the starting revision is not a substitute.
4. If verification fails because of code, the Astra coordinator diagnoses the failure as a narrowly bounded rubber duck. Permit at most one repair implementation, then verify again.
5. If the repaired result fails, record `blocked` with both failures and stop automatic implementation attempts. Return the checkpoint and next decision. Never reset counters by resuming or recreating the same task record.
6. Stop earlier for budget, authorization, provider, authentication, or environment constraints. These are blockers, not code-repair attempts. Only explicit throttling follows the routing retry policy.
7. Accept only with evidence for the selected boundary and the same exact target as the preceding verification. Unit tests, CI, merge, deployment, and verified running behavior are distinct milestones and never imply one another.
8. Treat ownership completeness independently from task state. It defaults to `unknown`; attest `complete` only after every participant scope is present, non-overlapping, and inside the task window. Use `partial` when known incomplete. Attestation cannot be reset or downgraded.

Use `scripts\bounded_rug.py` for runtime-enforced state transitions and atomic private records. The private registry defaults to `$COPILOT_HOME\token-mizer\task-registry.json`, falling back to `~\.copilot\token-mizer\task-registry.json` when `COPILOT_HOME` is unset. Prompt instructions govern agent behavior; the helper enforces record structure, transition order, exact-target consistency, stale/replayed events, irreversible scope attestation, and the one-build/one-repair bounds. It validates evidence annotations but does not independently prove that an annotation is true.

## Validation and review

Run deterministic repository gates first. Add independent review only for substantial, risky, or previously failed work, and include its session scope in the record. Do not create a reviewer for every file, duplicate Forge or an equivalent validator, or rerun evidence that already proves the same boundary.

## Pilot measurement

For optional typed routing evidence, first export the allocator at an explicit exclusive UTC cutoff, then use `add-assignment --assignment-export <file> --assignment-id <id>` on the matching task record. Task identity/class/boundary and event-derived snapshots must match; `attach-rug` stores reference strings only. A snapshot newer than the ledger cutoff is excluded, not treated as zero attempts or success. Use an already reconstructed allocator as-of export to preserve valid earlier observations. The report's `--rug-ledger` returns matching observations under `rug_ingestion`. This optional path does not gate ordinary routing.

Export the record to the existing v1.0 task-ledger format and run `token-mizer-report`. Capture baseline tasks observationally through the existing ledger; do not run them through the bounded-RUG treatment helper. Record mode, task class, identical acceptance boundary, all coordinator/builder/reviewer/validator scopes, failed attempts, and provider/model observations including unknowns. The reporter supplies tokens, known credits, elapsed time, model-active time, and coverage. Missing cost remains unknown, never zero. Compare separate mode × task-class × acceptance-boundary strata, count every applicable attempt, and suppress ratios without the relevant complete coverage. Report subsequent repairs or reopens only when their observation timestamp is no later than the analysis cutoff. Synthetic runs prove mechanics only; real-world pilot outcomes remain pending until measured. Do not launch paid benchmarks or claim causality or savings.