models.providers custom pricing never reaches the api-proxy; custom/BYOK models still 400 with unknown_model_ai_credits

[strawgate]

The models.providers frontmatter field (added in #38276, documented in the FAQ from #44450) is meant to let you price a model missing from the models.dev catalog so it passes the AI-credits guard. In practice the override reaches the agent-side catalog but is stripped from the api-proxy config, so the proxy still rejects the model with unknown_model_ai_credits before the first turn.

Environment

gh-aw v0.82.14, AWF firewall/api-proxy 0.27.37. claude engine pointed at an Anthropic-compatible endpoint via ANTHROPIC_BASE_URL (BYOK), with ANTHROPIC_MODEL set to a model not in the models.dev catalog.

Repro (pricing supplied exactly per the FAQ)

engine:
  id: claude
  api-target: api.fireworks.ai
  env:
    ANTHROPIC_BASE_URL: https://api.fireworks.ai/inference
    ANTHROPIC_MODEL: accounts/fireworks/models/minimax-m3
models:
  providers:
    anthropic:
      models:
        accounts/fireworks/models/minimax-m3:
          cost: { input: '3e-07', output: '1.5e-06', cache_read: '3e-08', cache_write: '3.75e-07' }

Expected: the api-proxy prices the model from the override and the run proceeds.
Actual: API Error: 400 … unknown_model_ai_credits on turn 1; the agent spends 0 tokens; the harness retries 4× with the same 400.

Root cause (from run artifacts)

The compiler emits the override to the agent-side catalog and to the lock's apiProxy.providers template, but not to what the proxy actually enforces against:

Where	minimax priced?	used for
activation/models.json (GH_AW_MODELS_JSON_PATH)	yes	agent-side AIC reporting only
container awf-config.json → apiProxy	no — providers stripped, maxAiCredits: 1000	credit-enforcement gate
proxy api-proxy-logs/models.json	no (anthropic.models: null)	proxy pricing lookup

So apiProxy.providers is dropped during config resolution, and the agent-side override never propagates into the api-proxy's pricing catalog — leaving the model unknown to the guard.

Also affects threat-detection

The detection sub-agent runs the same model behind its own guardrail (GH_AW_DEFAULT_DETECTION_MAX_AI_CREDITS, default 400) with no override reaching it, so it 400s and every safe-output gets the agentic threat detected / results could not be parsed banner. The main-agent workaround (max-ai-credits: -1, which makes the compiler omit maxAiCredits) isn't available there: the detection config hard-codes maxAiCredits, and -1 is then rejected by the firewall as maxAiCredits must be > 0.

Related

#38276, #44450, #46986 (the BYOK guardrail-skip only checks COPILOT_PROVIDER_*, not custom ANTHROPIC_BASE_URL/OPENAI_BASE_URL), #45923 / #46077.