Treat expression-backed engine.model as omitted when empty

actions/setup/js/claude_harness.cjs

@@ -53,6 +53,7 @@ const { emitMissingToolPermissionIssue, hasExpectedSafeOutputs, hasNoopInSafeOut
 const { countPermissionDeniedIssues, hasNumerousPermissionDeniedIssues, extractDeniedCommands, buildMissingToolPermissionIssuePayload } = require("./permission_denied_helpers.cjs");
 const { detectNonRetryableHarnessGuard, buildSoftTimeoutGuard, emitSoftTimeoutSignal, isAuthenticationFailedError } = require("./harness_retry_guard.cjs");
 const { MODEL_NOT_SUPPORTED_PATTERN: INVALID_MODEL_ERROR_PATTERN } = require("./detect_agent_errors.cjs");
+const { applyModelFallback } = require("./model_fallback.cjs");
 
 // Pattern to detect Anthropic API overload errors (HTTP 529).
 // Matches "overloaded_error" from the Anthropic error type field, and the
@@ -318,6 +319,7 @@ function stripContinueArgs(args) {
  */
 async function buildClaudeChildEnv() {
   const childEnv = { ...process.env };
+  applyModelFallback(childEnv, "ANTHROPIC_MODEL", log);
   const provider = normalizeReflectProviderName(process.env.GH_AW_LLM_PROVIDER, "anthropic");
   try {
     const raw = fs.readFileSync(AWF_REFLECT_OUTPUT_PATH, "utf8");
@@ -611,6 +613,7 @@ if (typeof module !== "undefined" && module.exports) {
     hasExpectedSafeOutputs,
     resolveRetryConfig,
     resolveStartupRetryLimit,
+    applyModelFallback,
   };
 }
 

actions/setup/js/codex_harness.cjs

@@ -54,6 +54,7 @@ const { countPermissionDeniedIssues, hasNumerousPermissionDeniedIssues, extractD
 const { detectNonRetryableHarnessGuard, buildSoftTimeoutGuard, emitSoftTimeoutSignal, isAuthenticationFailedError } = require("./harness_retry_guard.cjs");
 const { MODEL_NOT_SUPPORTED_PATTERN: INVALID_MODEL_ERROR_PATTERN } = require("./detect_agent_errors.cjs");
 const { resolveRetryConfig } = require("./harness_retry_config.cjs");
+const { applyModelFallback, injectModelFlagAfterExec } = require("./model_fallback.cjs");
 
 // Pattern to detect OpenAI rate-limit errors.
 // Matches the JSON error type field ("rate_limit_exceeded"), the HTTP status code
@@ -221,6 +222,16 @@ function injectJsonFlag(args) {
   return ["exec", "--json", ...args.slice(1)];
 }
 
+function getCodexModelEnvVar(env = process.env) {
+  if ("GH_AW_MODEL_DETECTION_CODEX" in env) {
+    return "GH_AW_MODEL_DETECTION_CODEX";
+  }
+  if ("GH_AW_MODEL_AGENT_CODEX" in env) {
+    return "GH_AW_MODEL_AGENT_CODEX";
+  }
+  return "";
+}
+
 /**
  * Build child process environment for Codex execution.
  * Preserve API keys captured at harness startup, even if the parent environment
@@ -425,6 +436,10 @@ async function main() {
     process.exit(1);
   }
 
+  const codexModelEnvVar = getCodexModelEnvVar(process.env);
+  const resolvedModel = codexModelEnvVar ? applyModelFallback(process.env, codexModelEnvVar, log) : "";
+  resolvedArgs = injectModelFlagAfterExec(resolvedArgs, resolvedModel);
+
   // Safe arg list for logging: when --prompt-file was present, the last element of
   // resolvedArgs is the resolved prompt content. Replace it with a placeholder so that
   // task instructions are never written to stderr or captured in agent logs.
@@ -641,6 +656,9 @@ if (typeof module !== "undefined" && module.exports) {
     hasNoopInSafeOutputs,
     hasExpectedSafeOutputs,
     resolveRetryConfig,
+    applyModelFallback,
+    injectModelFlagAfterExec,
+    getCodexModelEnvVar,
   };
 }
 

actions/setup/js/copilot_harness.cjs

@@ -65,6 +65,7 @@ const { runSafeOutputsCLI, buildMissingToolAlternatives, emitMissingToolPermissi
 const { countPermissionDeniedIssues, hasNumerousPermissionDeniedIssues, extractDeniedCommands, buildMissingToolPermissionIssuePayload } = require("./permission_denied_helpers.cjs");
 const { detectNonRetryableHarnessGuard, buildSoftTimeoutGuard, emitSoftTimeoutSignal, isAuthenticationFailedError: isCommonAuthenticationFailedError } = require("./harness_retry_guard.cjs");
 const { isCAPIQuotaExceededError } = require("./detect_agent_errors.cjs");
+const { applyModelFallback } = require("./model_fallback.cjs");
 const { loadModelsJson } = require("./model_costs.cjs");
 const { resolveConfiguredCopilotModel } = require("./resolve_model_alias.cjs");
 
@@ -950,6 +951,7 @@ async function main() {
     }
   }
 
+  applyModelFallback(process.env, "COPILOT_MODEL", log);
   applyCopilotModelAliasResolution({ awfReflectData, logger: log });
   applyCopilotWireAPI({ modelsJson: loadModelsJson(), logger: log });
 
@@ -1456,6 +1458,7 @@ if (typeof module !== "undefined" && module.exports) {
     parseCopilotSDKServerArgsFromEnv,
     isCAPIQuotaExceededError,
     hasTerminalSafeOutput,
+    applyModelFallback,
     applyCopilotModelAliasResolution,
     applyCopilotWireAPI,
     loadAwfConfigData,

actions/setup/js/copilot_sdk_driver.cjs

@@ -30,6 +30,7 @@ const fs = require("fs");
 const { runWithCopilotSDK, extractPromptFromArgs } = require("./copilot_sdk_session.cjs");
 const { parsePermissionConfigFromServerArgs } = require("./copilot_sdk_permissions.cjs");
 const { parseMultiProviderJson } = require("./copilot_sdk_multi_provider.cjs");
+const { applyModelFallback } = require("./model_fallback.cjs");
 
 // Re-export the session and permission helpers so that existing callers that
 // require("./copilot_sdk_driver.cjs") (e.g. copilot_harness.cjs) continue to work.
@@ -103,7 +104,7 @@ async function main() {
   const providers = multiProviderConfig.providers;
   /** @type {import("@github/copilot-sdk").ProviderModelConfig[]} */
   const sdkModels = multiProviderConfig.models;
-  let model = process.env.COPILOT_MODEL || multiProviderConfig.model || undefined;
+  let model = applyModelFallback(process.env, "COPILOT_MODEL", log) || multiProviderConfig.model || undefined;
   log(`multi-provider mode: ${providers.length} providers, ${sdkModels.length} models, model=${model ?? "(env)"}`);
   for (const p of providers) {
     log(`  provider: name=${p.name} type=${p.type} baseUrl=${p.baseUrl}${p.wireApi ? ` wireApi=${p.wireApi}` : ""}`);

actions/setup/js/copilot_sdk_session.cjs

@@ -36,6 +36,7 @@ const fs = require("fs");
 const path = require("path");
 const os = require("os");
 const { buildCopilotSDKPermissionHandler, getEnvPositiveIntOrDefault, parseMaxToolDenialsLimit, MAX_TOOL_DENIALS_DEFAULT } = require("./copilot_sdk_permissions.cjs");
+const { resolveModelWithFallback } = require("./model_fallback.cjs");
 const { extractShellCommandFromToolData } = require("./tool_call_details.cjs");
 
 // Default timeout for a single sendAndWait call: 10 minutes.
@@ -254,7 +255,7 @@ async function runWithCopilotSDK({ sdkUri, prompt, logger, attempt = 0, model, c
     // Build session config using the multi-provider surface.
     /** @type {import("@github/copilot-sdk").SessionConfig} */
     const sessionConfig = {
-      model: model || process.env.COPILOT_MODEL || undefined,
+      model: model || resolveModelWithFallback(process.env, "COPILOT_MODEL") || undefined,
       providers,
       models: providerModels,
       onPermissionRequest,

actions/setup/js/model_fallback.cjs

@@ -0,0 +1,48 @@
+"use strict";
+
+const MODEL_FALLBACK_ENV_VAR = "GH_AW_MODEL_FALLBACK";
+
+function readTrimmedEnv(env, name) {
+  return typeof env?.[name] === "string" ? env[name].trim() : "";
+}
+
+function resolveModelWithFallback(env, primaryEnvVar) {
+  return readTrimmedEnv(env, primaryEnvVar) || readTrimmedEnv(env, MODEL_FALLBACK_ENV_VAR);
+}
+
+/**
+ * @param {NodeJS.ProcessEnv} env
+ * @param {string} primaryEnvVar
+ * @param {(message: string) => void} [logger]
+ * @returns {string}
+ */
+function applyModelFallback(env, primaryEnvVar, logger = () => {}) {
+  const primary = readTrimmedEnv(env, primaryEnvVar);
+  if (primary) {
+    return primary;
+  }
+  const fallback = readTrimmedEnv(env, MODEL_FALLBACK_ENV_VAR);
+  if (fallback) {
+    env[primaryEnvVar] = fallback;
+    logger(`applied ${MODEL_FALLBACK_ENV_VAR} to ${primaryEnvVar}`);
+  }
+  return fallback;
+}
+
+function injectModelFlagAfterExec(args, model) {
+  if (!model || args.includes("--model")) {
+    return args;
+  }
+  const execIndex = args.indexOf("exec");
+  if (execIndex === -1) {
+    return [...args, "--model", model];
+  }
+  return [...args.slice(0, execIndex + 1), "--model", model, ...args.slice(execIndex + 1)];
+}
+
+module.exports = {
+  MODEL_FALLBACK_ENV_VAR,
+  resolveModelWithFallback,
+  applyModelFallback,
+  injectModelFlagAfterExec,
+};

actions/setup/js/model_fallback.test.js

@@ -0,0 +1,25 @@
+import { describe, expect, it } from "vitest";
+import { createRequire } from "module";
+
+const require = createRequire(import.meta.url);
+const { MODEL_FALLBACK_ENV_VAR, resolveModelWithFallback, applyModelFallback, injectModelFlagAfterExec } = require("./model_fallback.cjs");
+
+describe("model_fallback.cjs", () => {
+  it("prefers the primary model env var when it is non-empty", () => {
+    expect(resolveModelWithFallback({ COPILOT_MODEL: "gpt-5", [MODEL_FALLBACK_ENV_VAR]: "claude-sonnet-4.6" }, "COPILOT_MODEL")).toBe("gpt-5");
+  });
+
+  it("uses the fallback model when the primary resolves empty", () => {
+    const env = { COPILOT_MODEL: "", [MODEL_FALLBACK_ENV_VAR]: "claude-sonnet-4.6" };
+    expect(applyModelFallback(env, "COPILOT_MODEL")).toBe("claude-sonnet-4.6");
+    expect(env.COPILOT_MODEL).toBe("claude-sonnet-4.6");
+  });
+
+  it("injects --model immediately after exec when needed", () => {
+    expect(injectModelFlagAfterExec(["exec", "--json", "prompt"], "gpt-5")).toEqual(["exec", "--model", "gpt-5", "--json", "prompt"]);
+  });
+
+  it("does not inject --model when already present", () => {
+    expect(injectModelFlagAfterExec(["exec", "--model", "gpt-5", "--json"], "claude-sonnet-4.6")).toEqual(["exec", "--model", "gpt-5", "--json"]);
+  });
+});

actions/setup/js/run_evals.cjs

@@ -35,6 +35,7 @@ const path = require("path");
 
 const { ERR_VALIDATION } = require("./error_codes.cjs");
 const { EVALS_OUTPUT_PATH } = require("./evals_constants.cjs");
+const { resolveModelWithFallback } = require("./model_fallback.cjs");
 
 const EVALS_DIR = "/tmp/gh-aw/evals";
 const EVALS_LOG_PATH = "/tmp/gh-aw/evals/evals.log";
@@ -118,7 +119,7 @@ async function setupMain() {
  */
 async function parseMain() {
   const questionsRaw = process.env.GH_AW_EVALS_QUESTIONS;
-  const model = process.env.GH_AW_EVALS_MODEL || "";
+  const model = resolveModelWithFallback(process.env, "GH_AW_EVALS_MODEL") || "";
   const runID = process.env.GITHUB_RUN_ID || "unknown";
 
   /** @type {Array<{id: string, question: string}>} */

actions/setup/js/run_evals.test.cjs

@@ -6,6 +6,7 @@ const EVALS_DIR = "/tmp/gh-aw/evals";
 const EVALS_LOG_PATH = `${EVALS_DIR}/evals.log`;
 const EVALS_OUTPUT_PATH = "/tmp/gh-aw/evals.jsonl";
 const require = createRequire(import.meta.url);
+const { MODEL_FALLBACK_ENV_VAR } = require("./model_fallback.cjs");
 const { parseMain, extractAssistantTextFromJsonlLog } = require("./run_evals.cjs");
 
 const mockCore = {
@@ -75,6 +76,19 @@ describe("run_evals.cjs", () => {
     expect(JSON.parse(line).runid).toBe("unknown");
   });
 
+  it("uses GH_AW_MODEL_FALLBACK when GH_AW_EVALS_MODEL resolves empty", async () => {
+    vi.stubEnv("GH_AW_EVALS_QUESTIONS", JSON.stringify([{ id: "labels-applied", question: "Did labels get applied?" }]));
+    vi.stubEnv("GH_AW_EVALS_MODEL", "");
+    vi.stubEnv(MODEL_FALLBACK_ENV_VAR, "claude-sonnet-4.6");
+    vi.stubEnv("GITHUB_RUN_ID", "123456789");
+    fs.writeFileSync(EVALS_LOG_PATH, "labels-applied: YES\n", "utf8");
+
+    await parseMain();
+
+    const [line] = fs.readFileSync(EVALS_OUTPUT_PATH, "utf8").trim().split("\n");
+    expect(JSON.parse(line).model).toBe("claude-sonnet-4.6");
+  });
+
   it("parses answers from Pi v3 JSONL turn_end events (positional format)", async () => {
     vi.stubEnv(
       "GH_AW_EVALS_QUESTIONS",

pkg/constants/engine_constants.go

@@ -243,6 +243,11 @@ const (
 	// EnvVarModelAgentPi configures the default Pi model for agent execution
 	EnvVarModelAgentPi = "GH_AW_MODEL_AGENT_PI"
 
+	// EnvVarModelFallback carries the standard org/enterprise/built-in model fallback
+	// expression for the current engine step. Runtime JavaScript harnesses can use this
+	// when a configured model expression resolves to an empty string.
+	EnvVarModelFallback = "GH_AW_MODEL_FALLBACK"
+
 	// CopilotCLIModelEnvVar is the native environment variable name supported by the Copilot CLI
 	// for selecting the model. Setting this env var is equivalent to passing --model to the CLI.
 	CopilotCLIModelEnvVar = "COPILOT_MODEL"