engine: expose apiProxy.modelFallback in compiler frontmatter (currently requires lock-file editing to disable middle_power rewrites that break BYOK Azure deployments)

[zarenner]

Summary

The compiler emits an inline awf-config.json for the api-proxy sidecar but provides no frontmatter knob to override apiProxy.modelFallback. The default ({enabled: true, strategy: middle_power}) silently rewrites Azure OpenAI deployment names to base-catalog names, causing HTTP 404 DeploymentNotFound (see the proxy-side bug in gh-aw-firewall filed separately and the umbrella symptom in #35483).

Until the proxy-side default is fixed, BYOK Azure / custom-provider users have no supported way to opt out — the only workaround is a post-compile sed against the lock file, which is non-idempotent and gets clobbered by the next gh aw compile.

Why an env var doesn't work (subtle)

A natural attempt is to set AWF_MODEL_FALLBACK in engine.env and rely on awf --env-all. This does not work: awf --env-all propagates host env vars only into the agent container, not into the api-proxy sidecar. The sidecar's env is explicitly constructed by buildApiProxyService() in gh-aw-firewall/src/services/api-proxy-service.ts (~L86–170), which reads AWF_MODEL_FALLBACK only if config.modelFallback is defined in awf-config.json. Therefore the only ways to override are:

1. Edit the inline awf-config.json in the generated lock file (current sed workaround), or
2. Have gh-aw expose a frontmatter field that the compiler plumbs into the emitted awf-config.json.

Schema reference

gh-aw-firewall/src/awf-config-schema.json (~L163–182) defines:

{
  "apiProxy": {
    "modelFallback": {
      "enabled": true,
      "strategy": "middle_power"
    }
  }
}

Proposed change

Add an engine.firewall.apiProxy.modelFallback (or engine.apiProxy.modelFallback) field to the gh-aw frontmatter schema and emit the value into the inline awf-config.json written into the generated lock file. Example:

engine:
  id: copilot
  env: { ... }
  firewall:
    apiProxy:
      modelFallback:
        enabled: false
        # strategy: middle_power   # optional

For minimal surface area, a single boolean engine.firewall.modelFallback (default true) is sufficient.

Context

• Parent / umbrella: BYOK Azure OpenAI routing fails with 404 or 401 across base URL variants #35483
• Related (proxy-side root cause): gh-aw-firewall middle_power Azure rewrite (filed concurrently)
• Related existing compiler PRs that already plumb apiProxy.* fields: Sync AWF v0.25.43 spec updates and wire apiProxy.modelMultipliers usage #31398, Bump AWF to v0.25.44 and add firewall.effective-token-steering compiler support #31796, Bump gh-aw-firewall to v0.25.52 and sync embedded AWF schema #34114, feat: emit models section in AWF config JSON (under apiProxy) #30280, Stop emitting unsupported apiProxy.maxEffectiveTokens in generated AWF config #31117

[zarenner]

Cross-references:

• api-proxy: middle_power modelFallback rewrites Azure OpenAI deployment names to base-catalog names, causing HTTP 404 DeploymentNotFound on BYOK targets gh-aw-firewall#3975 — the proxy-side default-on middle_power rewrite that this compiler knob exists to disable.
• gh-aw BYOK: pathSetup export COPILOT_API_KEY="$COPILOT_DUMMY_BYOK" shadows COPILOT_PROVIDER_API_KEY, breaking AWF Copilot key resolution (HTTP 503/401) #35575, gh-aw BYOK: COPILOT_GITHUB_TOKEN is auto-injected even when COPILOT_PROVIDER_BASE_URL is set, leaking as fallback Bearer to external providers (HTTP 401) #35576 — the auth-side companion bugs in the same BYOK Azure OpenAI routing fails with 404 or 401 across base URL variants #35483 chain.