Model routing: move Purelock, Front Page Copy Guard and Refactoring Cadence off GPT-5.6 Terra before it retires on 16 November #74642

GPT-5.6 Terra is retired across Copilot on 16 November 2026 (changelog). Three gh-aw workflows pin copilot/gpt-5.6-terra and will stop working then. Route them instead of re-pinning. Estimated saving: about 520-620 AIC/week.

Why

• Routing instead of a new pin. The router picks a model and effort for each run, so these workflows stop depending on one model name. GitHub's suggested replacements (GPT-6.1 Sol, GPT-6 Luna) are among its choices.
• economy. Purelock opens test PRs that a maintainer reviews, Front Page Copy Guard posts a review comment and Refactoring Cadence files an issue. For this work, economy picks small current models: GPT-5.6 Luna, or Claude Haiku 5.5 for pi. Daily Go Test Parallelizer, which also writes Go tests, picked gpt-5.6-sol:max only while it was on balanced (Route gh-aw#74188's workflow through cost/balanced model routing #74200: 500-616 AIC per run, against about 57 on Terra). For the labels its runs got (implement at subsystem or multi-file scope, hard), economy in AWF v0.29.6's tables picks GPT-5.6 Luna, and it moved to economy on 3 November (Route gh-aw#74188's workflow with economy instead of balanced #74586). Economy picks Sol only for hard cross-system work.
• Detection gets cheaper too. Threat detection inherits today's Terra pin (about 6-10 AIC per run); without a pin it uses its default model, about 1. For Front Page Copy Guard, detection costs more than the agent.

Change

In .github/workflows/purelock.md and .github/workflows/front-page-copy-guard.md, delete model: copilot/gpt-5.6-terra, add copilot-requests: write to permissions, and set:

engine:
  id: codex
  model-provider: github
  model-routing:
    goal: cost
    mode: economy

In .github/workflows/refactoring-cadence.md, replace engine: pi and model: copilot/gpt-5.6-terra with the block below, and add copilot-requests: write to permissions:

engine:
  id: pi
  model-provider: github
  model-routing:
    goal: cost
    mode: economy

Keep the prompts, steps, sub-agents and safe outputs. Don't add model restrictions.

Run gh aw compile purelock front-page-copy-guard refactoring-cadence. On main (0fffa78f25), these compile with no new warnings: routing is enabled, inference uses github.token, and detection uses its default model.

Check

• Audit the first routed runs with gh aw audit <run>: expect a Luna pick (Haiku for pi), and compare AIC per run with the column above. A gpt-5.6-sol pick means the run was labelled hard cross-system work; if that keeps happening, pin copilot/gpt-6.1-sol instead.
• Purelock currently fails often because its test-writer sub-agents fail before they start ([aw] PureLock: reported incomplete result #74513). That's separate from this change and routing doesn't touch the sub-agents, so compare successful runs.
• Fallback: if routing misbehaves before 16 November, pin copilot/gpt-6.1-sol instead.

Also affected by the retirement (not in this issue)

• windows.md and cloclo.md pin copilot/gpt-5.6-terra: swap to copilot/gpt-6.1-sol.
• smoke-copilot-sub-agents.md: the terra-whoami sub-agent expects an exact gpt-5.6-terra match and will fail after 16 November.

Estimates use gh-aw usage from 19 October to 1 November and AWF v0.29.6's routing tables; the routed cost per run is estimated.