From 5aa0ad9a5293cbaf714caa7725762c3a474cfaf2 Mon Sep 17 00:00:00 2001
From: David Slater <davidslater@users.noreply.github.com>
Date: Wed, 4 Nov 2026 14:48:08 -0800
Subject: [PATCH 1/5] Show observed models powering auto in attribution footers

---
 actions/setup/js/messages.test.cjs          | 21 +++++++++++++
 actions/setup/js/model_attribution.cjs      | 25 ++++++++++++++-
 actions/setup/js/model_attribution.test.cjs | 35 +++++++++++++++++++++
 3 files changed, 80 insertions(+), 1 deletion(-)

diff --git a/actions/setup/js/messages.test.cjs b/actions/setup/js/messages.test.cjs
index 85d11a3e0e0..b07b5bfe91d 100644
--- a/actions/setup/js/messages.test.cjs
+++ b/actions/setup/js/messages.test.cjs
@@ -11,6 +11,8 @@
  */
 import { describe, it, expect, beforeEach, vi } from "vitest";
 import path from "path";
+import fs from "node:fs";
+import os from "node:os";
 import { fileURLToPath } from "url";
 
 // Mock core for GitHub Actions environment
@@ -177,6 +179,25 @@ describe("messages.cjs", () => {
   });
 
   describe("getFooterMessage", () => {
+    it.each([false, true])("shows the model powering auto in default and custom credit footers (custom: %s)", async custom => {
+      const root = fs.mkdtempSync(path.join(os.tmpdir(), "gh-aw-auto-footer-"));
+      try {
+        process.env.GH_AW_TMP_DIR = root;
+        process.env.GH_AW_ENGINE_ID = "copilot";
+        process.env.GH_AW_ENGINE_MODEL = "auto";
+        process.env.GH_AW_AGENT_AIC = "14";
+        process.env.GH_AW_THREAT_DETECTION_AIC = "1";
+        process.env.GH_AW_AMBIENT_CONTEXT = "9200";
+        if (custom) process.env.GH_AW_SAFE_OUTPUT_MESSAGES = JSON.stringify({ footer: "> Generated by [{workflow_name}]({run_url}){ai_credits_suffix}{history_link}" });
+        fs.writeFileSync(path.join(root, "aw_info.json"), JSON.stringify({ requested_model: "auto", model_selection: { status: "observed", effective_model: "gemini-3.8-flash" } }));
+        const { getFooterMessage } = await import("./messages.cjs");
+        const result = getFooterMessage({ workflowName: "Test Workflow", runUrl: "https://github.com/test/repo/actions/runs/123", historyUrl: "https://github.com/search" });
+        expect(result).toContain(" · copilot · auto (powered by Gemini 3.8 Flash) · \\~14 AIC · ⌖ \\~1 AIC · ⊞ 9.2K · [◷]");
+      } finally {
+        fs.rmSync(root, { recursive: true, force: true });
+      }
+    });
+
     it("should return default footer when no custom config", async () => {
       const { getFooterMessage } = await import("./messages.cjs");
 
diff --git a/actions/setup/js/model_attribution.cjs b/actions/setup/js/model_attribution.cjs
index e157fb3b7a4..f6878e03c7f 100644
--- a/actions/setup/js/model_attribution.cjs
+++ b/actions/setup/js/model_attribution.cjs
@@ -318,7 +318,21 @@ function getEffectiveModelLabel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp
   const attribution = resolveEffectiveModel(infoPath, phase, env);
   const { reduceModelNameToIdentifier } = require("./model_aliases.cjs");
   const shortModel = reduceModelNameToIdentifier(attribution.model);
-  if (!attribution.routing) return shortModel;
+  if (!attribution.routing) {
+    const requested = attribution.requestedModel || requestedModelIdentifier(env.GH_AW_INFO_MODEL || env.GH_AW_ENGINE_MODEL);
+    if (requested.toLowerCase() !== "auto") return shortModel;
+    const models = new Set();
+    if (phase === "agent") {
+      const sessionPath = require("path").join(require("path").dirname(infoPath), "usage/aw_session.jsonl");
+      for (const event of readModelRoutingSession(sessionPath)) {
+        if (event.type !== "firewall.token_usage" || (event.data?.purpose && event.data.purpose !== "agent")) continue;
+        const model = validateModelIdentifier(event.data?.model);
+        if (model && model.toLowerCase() !== "auto") models.add(model);
+      }
+    }
+    if (!models.size && attribution.model && attribution.model.toLowerCase() !== "auto") models.add(attribution.model);
+    return models.size ? `auto (powered by ${[...models].map(formatAttributionModelName).join(", ")})` : "auto";
+  }
   if (attribution.routing.status === "selected") {
     const routedModel = reduceModelNameToIdentifier(attribution.routing.wire_model);
     const routed = `routed: ${routedModel}${attribution.effort ? ` ${attribution.effort}` : ""}`;
@@ -329,6 +343,15 @@ function getEffectiveModelLabel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp
   return `routing ${attribution.routing.status}${failureCode ? ` (${failureCode})` : ""}`;
 }
 
+// Keep the full model version and tier readable rather than compressing them into an alias.
+function formatAttributionModelName(model) {
+  return model
+    .replace(/^.*\//, "")
+    .split(/[-_]/)
+    .map(part => (/^(gpt|o\d+)$/i.test(part) ? part.toUpperCase() : /^\d/.test(part) ? part : part.charAt(0).toUpperCase() + part.slice(1)))
+    .join(" ");
+}
+
 function recordFallbackModelFromUsage(content, env = process.env, infoPath = `${env.GH_AW_TMP_DIR || "/tmp/gh-aw"}/aw_info.json`, logger = console.warn) {
   let model = "";
   for (const entry of iterateJsonl(content, {
diff --git a/actions/setup/js/model_attribution.test.cjs b/actions/setup/js/model_attribution.test.cjs
index 9f16cf7fa79..2b750ec819d 100644
--- a/actions/setup/js/model_attribution.test.cjs
+++ b/actions/setup/js/model_attribution.test.cjs
@@ -17,6 +17,41 @@ function writeJSONL(filePath, records) {
 }
 
 describe("observed model attribution", () => {
+  it.each([
+    ["gemini-3.8-flash", "auto (powered by Gemini 3.8 Flash)"],
+    ["gpt-6.1-sol", "auto (powered by GPT 6.1 Sol)"],
+    ["", "auto"],
+    ["auto", "auto"],
+    ["bad model<script>", "auto"],
+  ])("labels auto using validated observed model %s", (model, label) => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      fs.writeFileSync(infoPath, JSON.stringify({ requested_model: "auto", model_selection: { status: model ? "observed" : "unavailable", effective_model: model } }));
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe(label);
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
+  it("lists distinct agent models without including classifier, detection, or untrusted events", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      writeJSONL(path.join(root, "usage/aw_session.jsonl"), [
+        ...["gemini-3.8-flash", "gemini-3.8-flash", "claude-sonnet-5", "auto", "<script>"].map(model => ({ type: "firewall.token_usage", data: { model, purpose: "agent" }, provenance: { component: "firewall", phase: "agent" } })),
+        { type: "firewall.token_usage", data: { model: "classifier", purpose: "routing_classification" }, provenance: { component: "firewall", phase: "agent" } },
+        { type: "firewall.token_usage", data: { model: "detector" }, provenance: { component: "firewall", phase: "detection" } },
+        { type: "firewall.token_usage", data: { model: "forged" }, provenance: { component: "agent", phase: "agent" } },
+      ]);
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+      expect(getEffectiveModelLabel(infoPath, "detection", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto");
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "gpt-5.4" })).toBe("gpt54");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
   it.each(["observed", "substituted"])("normalizes requested query parameters for %s selections", status => {
     const root = createRoot();
     const infoPath = path.join(root, "aw_info.json");

From e7842a39aba6b159e6313d44ce49c8ce8b02469a Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 5 Nov 2026 00:15:25 +0000
Subject: [PATCH 2/5] Fix model attribution review findings

Co-authored-by: pelikhan <4175913+pelikhan@users.noreply.github.com>
---
 actions/setup/js/copilot_model_selection.cjs  | 43 +++++++++++++++-
 .../js/copilot_sdk_model_observation.test.cjs | 50 +++++++++++++++++--
 actions/setup/js/messages.test.cjs            | 18 ++++++-
 actions/setup/js/model_attribution.cjs        | 16 ++++--
 actions/setup/js/model_attribution.test.cjs   | 40 ++++++++++++++-
 5 files changed, 156 insertions(+), 11 deletions(-)

diff --git a/actions/setup/js/copilot_model_selection.cjs b/actions/setup/js/copilot_model_selection.cjs
index f969e5890d1..b315c32aa01 100644
--- a/actions/setup/js/copilot_model_selection.cjs
+++ b/actions/setup/js/copilot_model_selection.cjs
@@ -60,6 +60,17 @@ function persistCopilotModelSelection(env = process.env) {
       resolved_model: typeof record?.resolved_model === "string" ? modelID(record.resolved_model) : "",
       effective_model: typeof record?.effective_model === "string" ? modelID(record.effective_model) : "",
     };
+    const observedModels = Array.isArray(record?.observed_models)
+      ? [
+          ...new Set(
+            record.observed_models
+              .filter(model => typeof model === "string")
+              .map(modelID)
+              .filter(Boolean)
+          ),
+        ].slice(0, 32)
+      : [];
+    if (observedModels.length > 1) selection.observed_models = observedModels;
     if (!selection.configured_model || !["observed", "substituted", "unavailable"].includes(selection.status)) throw new Error("Invalid Copilot model selection");
     if (selection.status === "unavailable") {
       selection.resolved_model = "";
@@ -88,6 +99,7 @@ function createCopilotModelObserver({ configuredModel, resolvedModel, sessionId,
   const offsets = new Map();
   /** @type {Map<string, string>} */
   const partial = new Map();
+  const observedModels = new Set();
   let diagnostic = "";
   let observedModel = "";
   const expected = modelID(resolvedModel);
@@ -144,8 +156,21 @@ function createCopilotModelObserver({ configuredModel, resolvedModel, sessionId,
           const model = modelID(event.type === "session.model_change" && event.data?.source === "startup" ? event.data.newModel : event.type === "assistant.message" ? event.data?.model : "");
           if (!model || model === observedModel) continue;
           observedModel = model;
+          if (!observedModels.has(model)) {
+            if (observedModels.size >= 32) observedModels.delete(observedModels.values().next().value);
+            observedModels.add(model);
+          }
           const substituted = expected.toLowerCase() !== "auto" && model.toLowerCase() !== expected.toLowerCase();
-          recordCopilotModelSelection({ status: substituted ? "substituted" : "observed", configured_model: configuredModel, resolved_model: resolvedModel, effective_model: model }, env);
+          recordCopilotModelSelection(
+            {
+              status: substituted ? "substituted" : "observed",
+              configured_model: configuredModel,
+              resolved_model: resolvedModel,
+              effective_model: model,
+              ...(observedModels.size > 1 ? { observed_models: [...observedModels] } : {}),
+            },
+            env
+          );
           if (substituted) {
             diagnostic = modelSubstitutionDiagnostic(expected, model);
             logger(diagnostic);
@@ -200,6 +225,7 @@ function createCopilotSDKModelObserver({ configuredModel, expectedModel, provide
   let sessionId = "";
   let diagnostic = "";
   let recorded = "";
+  const observedModels = new Set();
 
   /** @param {any} event */
   function observe(event) {
@@ -219,7 +245,20 @@ function createCopilotSDKModelObserver({ configuredModel, expectedModel, provide
       const effective = matches ? (unconstrained ? model : expected) : validateModelIdentifier(base) || model;
       if (matches && effective === recorded) return "";
       recorded = effective;
-      recordCopilotModelSelection({ status: matches ? "observed" : "substituted", configured_model: configuredModel, resolved_model: resolvedModel, effective_model: effective }, env);
+      if (!observedModels.has(effective)) {
+        if (observedModels.size >= 32) observedModels.delete(observedModels.values().next().value);
+        observedModels.add(effective);
+      }
+      recordCopilotModelSelection(
+        {
+          status: matches ? "observed" : "substituted",
+          configured_model: configuredModel,
+          resolved_model: resolvedModel,
+          effective_model: effective,
+          ...(observedModels.size > 1 ? { observed_models: [...observedModels] } : {}),
+        },
+        env
+      );
       if (matches) return "";
       diagnostic = modelSubstitutionDiagnostic(expected, effective);
     } catch (error) {
diff --git a/actions/setup/js/copilot_sdk_model_observation.test.cjs b/actions/setup/js/copilot_sdk_model_observation.test.cjs
index 1b843f02f32..57a96b3bc42 100644
--- a/actions/setup/js/copilot_sdk_model_observation.test.cjs
+++ b/actions/setup/js/copilot_sdk_model_observation.test.cjs
@@ -11,6 +11,7 @@ const { COPILOT_MODEL_VERIFICATION_EXIT_CODE, createCopilotSDKModelObserver, rea
 const { runWithCopilotSDK } = require("./copilot_sdk_session.cjs");
 const { qualifyModelForMultiProvider } = require("./copilot_sdk_multi_provider.cjs");
 const { detectErrors } = require("./agent_error_patterns.cjs");
+const { getEffectiveModelLabel } = require("./model_attribution.cjs");
 
 // Captured, credential-free model-bearing records from Smoke Copilot SDK Routed run 39529495361
 // (native envelopes and main/sub-agent scope preserved) and the CLI control run 39529502581.
@@ -142,7 +143,13 @@ describe("Copilot SDK main-session model observation", () => {
     // Derived from the 39529495361 fixture: the main session switches to the same id on an unconfigured `copilot` provider.
     const diagnostic = watch.observe({ ...derivedMainModelChange(), data: { ...derivedMainModelChange().data, newModel: "copilot/gpt-5.6-luna" } });
     expect(diagnostic).toContain("Copilot substituted 'copilot/gpt-5.6-luna'");
-    expect(selection()).toEqual({ status: "substituted", configured_model: "gpt-5.6-luna", resolved_model: "gpt-5.6-luna", effective_model: "copilot/gpt-5.6-luna" });
+    expect(selection()).toEqual({
+      status: "substituted",
+      configured_model: "gpt-5.6-luna",
+      resolved_model: "gpt-5.6-luna",
+      effective_model: "copilot/gpt-5.6-luna",
+      observed_models: ["gpt-5.6-luna", "copilot/gpt-5.6-luna"],
+    });
     expect(watch.observe(firstMessage)).toBe(diagnostic);
     expect(selection().status).toBe("substituted");
   });
@@ -263,7 +270,13 @@ runWithCopilotSDK({
     expect(result.stderr).toContain("stopping SDK session: Copilot model verification failed");
     expect(result.stderr).toContain("Copilot substituted 'copilot-completions/claude-haiku-4.5'. Refusing silent model substitution.");
     expect(result.stderr).not.toContain("retry 1/3");
-    expect(selection()).toEqual({ status: "substituted", configured_model: "gpt-5.6-luna", resolved_model: "gpt-5.6-luna", effective_model: "copilot-completions/claude-haiku-4.5" });
+    expect(selection()).toEqual({
+      status: "substituted",
+      configured_model: "gpt-5.6-luna",
+      resolved_model: "gpt-5.6-luna",
+      effective_model: "copilot-completions/claude-haiku-4.5",
+      observed_models: ["gpt-5.6-luna", "copilot-completions/claude-haiku-4.5"],
+    });
     const diagnostic = readCopilotModelSubstitution();
     expect(result.stderr).toContain(diagnostic);
     expect(detectErrors(diagnostic).modelNotSupportedError).toBe(true);
@@ -304,17 +317,46 @@ runWithCopilotSDK({
       },
     });
     expect(result.exitCode).toBe(exitCode);
-    expect(selection()).toEqual({ status, configured_model: "auto", resolved_model: "gpt-5.6-luna", effective_model: mainModel });
+    expect(selection()).toEqual({
+      status,
+      configured_model: "auto",
+      resolved_model: "gpt-5.6-luna",
+      effective_model: mainModel,
+      ...(status === "substituted" ? { observed_models: ["gpt-5.6-luna", "gpt-5.6-sol"] } : {}),
+    });
   });
 
   it("sdk_auto_does_not_report_substitution", () => {
     const watch = createCopilotSDKModelObserver({ configuredModel: "auto", expectedModel: "auto", providerConfig: null, sessionDirs: [] });
     for (const event of mainEvents()) expect(watch.observe(event)).toBe("");
     expect(watch.finalize(SESSION_ID)).toBe("");
-    expect(selection()).toEqual({ status: "observed", configured_model: "auto", resolved_model: "auto", effective_model: "gpt-5.6-luna" });
+    expect(selection()).toEqual({
+      status: "observed",
+      configured_model: "auto",
+      resolved_model: "auto",
+      effective_model: "gpt-5.6-luna",
+      observed_models: ["copilot-responses/gpt-5.6-luna", "gpt-5.6-luna"],
+    });
     expect(readCopilotModelSubstitution()).toBe("");
   });
 
+  it("preserves multiple native models for auto attribution without firewall usage", () => {
+    const infoPath = path.join(root, "aw_info.json");
+    const watch = createCopilotSDKModelObserver({ configuredModel: "auto", expectedModel: "auto", providerConfig: null, sessionDirs: [] });
+    expect(watch.observe({ type: "assistant.message", data: { model: "gemini-3.8-flash" } })).toBe("");
+    expect(watch.observe({ type: "assistant.message", data: { model: "claude-sonnet-5" } })).toBe("");
+    expect(watch.finalize(SESSION_ID)).toBe("");
+    expect(selection()).toEqual({
+      status: "observed",
+      configured_model: "auto",
+      resolved_model: "auto",
+      effective_model: "claude-sonnet-5",
+      observed_models: ["gemini-3.8-flash", "claude-sonnet-5"],
+    });
+    persistCopilotModelSelection();
+    expect(getEffectiveModelLabel(infoPath, "agent", process.env)).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+  });
+
   it("sdk_subagent_model_events_do_not_change_main_observation", () => {
     expect(haikuEvents().map(event => [event.type, event.data.newModel || event.data.model])).toEqual([
       ["session.model_change", "copilot-completions/claude-haiku-4.5"],
diff --git a/actions/setup/js/messages.test.cjs b/actions/setup/js/messages.test.cjs
index b07b5bfe91d..17755074d7b 100644
--- a/actions/setup/js/messages.test.cjs
+++ b/actions/setup/js/messages.test.cjs
@@ -181,6 +181,8 @@ describe("messages.cjs", () => {
   describe("getFooterMessage", () => {
     it.each([false, true])("shows the model powering auto in default and custom credit footers (custom: %s)", async custom => {
       const root = fs.mkdtempSync(path.join(os.tmpdir(), "gh-aw-auto-footer-"));
+      const environmentKeys = ["GH_AW_SAFE_OUTPUT_MESSAGES", "GH_AW_TMP_DIR", "GH_AW_ENGINE_ID", "GH_AW_ENGINE_MODEL", "GH_AW_AGENT_AIC", "GH_AW_THREAT_DETECTION_AIC", "GH_AW_AMBIENT_CONTEXT"];
+      const environment = new Map(environmentKeys.map(key => [key, process.env[key]]));
       try {
         process.env.GH_AW_TMP_DIR = root;
         process.env.GH_AW_ENGINE_ID = "copilot";
@@ -190,11 +192,25 @@ describe("messages.cjs", () => {
         process.env.GH_AW_AMBIENT_CONTEXT = "9200";
         if (custom) process.env.GH_AW_SAFE_OUTPUT_MESSAGES = JSON.stringify({ footer: "> Generated by [{workflow_name}]({run_url}){ai_credits_suffix}{history_link}" });
         fs.writeFileSync(path.join(root, "aw_info.json"), JSON.stringify({ requested_model: "auto", model_selection: { status: "observed", effective_model: "gemini-3.8-flash" } }));
+        fs.mkdirSync(path.join(root, "usage"), { recursive: true });
+        fs.writeFileSync(
+          path.join(root, "usage/aw_session.jsonl"),
+          [
+            { type: "firewall.token_usage", data: { model: "gemini-3.8-flash", purpose: "agent" }, provenance: { component: "firewall", phase: "agent" } },
+            { type: "firewall.token_usage", data: { model: "claude-sonnet-5", purpose: "agent" }, provenance: { component: "firewall", phase: "agent" } },
+          ]
+            .map(event => JSON.stringify(event))
+            .join("\n")
+        );
         const { getFooterMessage } = await import("./messages.cjs");
         const result = getFooterMessage({ workflowName: "Test Workflow", runUrl: "https://github.com/test/repo/actions/runs/123", historyUrl: "https://github.com/search" });
-        expect(result).toContain(" · copilot · auto (powered by Gemini 3.8 Flash) · \\~14 AIC · ⌖ \\~1 AIC · ⊞ 9.2K · [◷]");
+        expect(result).toContain(" · copilot · auto (powered by Gemini 3.8 Flash, Claude Sonnet 5) · \\~14 AIC · ⌖ \\~1 AIC · ⊞ 9.2K · [◷]");
       } finally {
         fs.rmSync(root, { recursive: true, force: true });
+        for (const [key, value] of environment) {
+          if (value === undefined) delete process.env[key];
+          else process.env[key] = value;
+        }
       }
     });
 
diff --git a/actions/setup/js/model_attribution.cjs b/actions/setup/js/model_attribution.cjs
index f6878e03c7f..9d98987efcb 100644
--- a/actions/setup/js/model_attribution.cjs
+++ b/actions/setup/js/model_attribution.cjs
@@ -309,6 +309,7 @@ function resolveEffectiveModel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp/
     model,
     fallbackModel,
     routing,
+    selection,
     requestedModel: requestedModelIdentifier(observed ? info.requested_model || selection.configured_model : routing?.requested_model || (routing ? env.GH_AW_INFO_MODEL || env.GH_AW_ENGINE_MODEL : "")),
     effort: routing?.status === "selected" ? routing.applied_effort || routing.effort : "",
   };
@@ -325,13 +326,22 @@ function getEffectiveModelLabel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp
     if (phase === "agent") {
       const sessionPath = require("path").join(require("path").dirname(infoPath), "usage/aw_session.jsonl");
       for (const event of readModelRoutingSession(sessionPath)) {
-        if (event.type !== "firewall.token_usage" || (event.data?.purpose && event.data.purpose !== "agent")) continue;
+        if (event.type !== "firewall.token_usage" || (event.data?.purpose && event.data.purpose !== "agent") || Number(event.data?.status) >= 400) continue;
         const model = validateModelIdentifier(event.data?.model);
         if (model && model.toLowerCase() !== "auto") models.add(model);
       }
+      const observedModels = attribution.selection?.status === "observed" ? attribution.selection.observed_models : null;
+      if (!models.size && Array.isArray(observedModels)) {
+        for (const value of observedModels) {
+          const model = validateModelIdentifier(value);
+          if (model && model.toLowerCase() !== "auto") models.add(model);
+        }
+      }
     }
-    if (!models.size && attribution.model && attribution.model.toLowerCase() !== "auto") models.add(attribution.model);
-    return models.size ? `auto (powered by ${[...models].map(formatAttributionModelName).join(", ")})` : "auto";
+    if (!models.size && !attribution.fallbackModel && attribution.model && attribution.model.toLowerCase() !== "auto") models.add(attribution.model);
+    const names = [...new Set([...models].map(formatAttributionModelName).filter(Boolean))];
+    if (names.length) return `auto (powered by ${names.join(", ")})`;
+    return attribution.fallbackModel && shortModel ? shortModel : "auto";
   }
   if (attribution.routing.status === "selected") {
     const routedModel = reduceModelNameToIdentifier(attribution.routing.wire_model);
diff --git a/actions/setup/js/model_attribution.test.cjs b/actions/setup/js/model_attribution.test.cjs
index 2b750ec819d..3dbbcac54eb 100644
--- a/actions/setup/js/model_attribution.test.cjs
+++ b/actions/setup/js/model_attribution.test.cjs
@@ -20,6 +20,7 @@ describe("observed model attribution", () => {
   it.each([
     ["gemini-3.8-flash", "auto (powered by Gemini 3.8 Flash)"],
     ["gpt-6.1-sol", "auto (powered by GPT 6.1 Sol)"],
+    ["provider/", "auto"],
     ["", "auto"],
     ["auto", "auto"],
     ["bad model<script>", "auto"],
@@ -39,7 +40,12 @@ describe("observed model attribution", () => {
     const infoPath = path.join(root, "aw_info.json");
     try {
       writeJSONL(path.join(root, "usage/aw_session.jsonl"), [
-        ...["gemini-3.8-flash", "gemini-3.8-flash", "claude-sonnet-5", "auto", "<script>"].map(model => ({ type: "firewall.token_usage", data: { model, purpose: "agent" }, provenance: { component: "firewall", phase: "agent" } })),
+        ...["gemini-3.8-flash", "gemini-3.8-flash", "google/gemini-3.8-flash", "Gemini-3.8-Flash", "claude-sonnet-5", "auto", "<script>"].map(model => ({
+          type: "firewall.token_usage",
+          data: { model, purpose: "agent" },
+          provenance: { component: "firewall", phase: "agent" },
+        })),
+        { type: "firewall.token_usage", data: { model: "failed-candidate", purpose: "agent", status: 503 }, provenance: { component: "firewall", phase: "agent" } },
         { type: "firewall.token_usage", data: { model: "classifier", purpose: "routing_classification" }, provenance: { component: "firewall", phase: "agent" } },
         { type: "firewall.token_usage", data: { model: "detector" }, provenance: { component: "firewall", phase: "detection" } },
         { type: "firewall.token_usage", data: { model: "forged" }, provenance: { component: "agent", phase: "agent" } },
@@ -52,6 +58,38 @@ describe("observed model attribution", () => {
     }
   });
 
+  it("preserves fallback-only labels when auto has no observed model evidence", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      fs.writeFileSync(infoPath, JSON.stringify({ requested_model: "auto", fallback_model: "gpt-5.4" }));
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("gpt54");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
+  it("uses validated native model observations when the session usage file is absent", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      fs.writeFileSync(
+        infoPath,
+        JSON.stringify({
+          requested_model: "auto",
+          model_selection: {
+            status: "observed",
+            effective_model: "claude-sonnet-5",
+            observed_models: ["google/gemini-3.8-flash", "claude-sonnet-5", "bad model<script>", "provider/"],
+          },
+        })
+      );
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
   it.each(["observed", "substituted"])("normalizes requested query parameters for %s selections", status => {
     const root = createRoot();
     const infoPath = path.join(root, "aw_info.json");

From 1efe3ca04426baf8e80c9a677cf6c72903a91321 Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 5 Nov 2026 03:31:58 +0000
Subject: [PATCH 3/5] Read auto attribution models from the agent artifact
 token-usage log

Co-authored-by: SivaKesava1 <11771739+SivaKesava1@users.noreply.github.com>
---
 actions/setup/js/model_attribution.cjs      | 45 +++++++++++++--
 actions/setup/js/model_attribution.test.cjs | 63 +++++++++++++++++++++
 2 files changed, 104 insertions(+), 4 deletions(-)

diff --git a/actions/setup/js/model_attribution.cjs b/actions/setup/js/model_attribution.cjs
index 9d98987efcb..158fd5a4b50 100644
--- a/actions/setup/js/model_attribution.cjs
+++ b/actions/setup/js/model_attribution.cjs
@@ -9,6 +9,8 @@ const ROUTING_EFFORTS = new Set(["none", "minimal", "low", "medium", "high", "xh
 const ROUTING_TASK_TYPES = new Set(["explain", "plan", "fix", "refactor", "chore", "implement", "unknown"]);
 const ROUTING_SCOPES = new Set(["local", "multi_file", "subsystem", "cross_system", "unknown"]);
 const ROUTING_COMPLEXITIES = new Set(["trivial", "easy", "medium", "hard", "expert", "unknown"]);
+// Same candidates as ai_credits_context.cjs; these files ship in the agent artifact.
+const TOKEN_USAGE_LOG_PATHS = ["sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl", "sandbox/firewall-audit-logs/api-proxy-logs/token-usage.jsonl", "sandbox/firewall/audit/api-proxy-logs/token-usage.jsonl"];
 const MODEL_ROUTING_LOG_PATHS = ["sandbox/firewall/logs/api-proxy-logs/model-routing.jsonl", "sandbox/firewall/audit/api-proxy-logs/model-routing.jsonl", "sandbox/firewall-audit-logs/api-proxy-logs/model-routing.jsonl"];
 
 /** @typedef {{ status: string, mode: string, router_version: string, failure_code?: string, objective: string, task_type: string, scope: string, complexity: string, degraded?: boolean, classifier_aic?: number, deviated_requests?: number }} ModelRoutingSummary */
@@ -160,6 +162,32 @@ function readModelRoutingSession(sessionPath) {
   return events;
 }
 
+/**
+ * Reads models from successful agent requests in the AWF token-usage log.
+ * Returns null when no log exists so callers can fall back to the unified session.
+ * @param {string} ghAwDir
+ * @returns {string[] | null}
+ */
+function readAgentTokenUsageModels(ghAwDir) {
+  const fs = require("fs");
+  const path = require("path");
+  const logPath = TOKEN_USAGE_LOG_PATHS.map(relativePath => path.join(ghAwDir, relativePath)).find(candidate => fs.existsSync(candidate));
+  if (!logPath) return null;
+  const models = [];
+  try {
+    for (const record of readJsonlFile(logPath)) {
+      if (typeof record?._schema !== "string" || !record._schema.startsWith("token-usage/")) continue;
+      if (record.purpose && record.purpose !== "agent") continue;
+      const status = Number(record.status);
+      if (!(status >= 200 && status < 300)) continue;
+      models.push(record.model);
+    }
+  } catch {
+    return null;
+  }
+  return models;
+}
+
 function readModelRoutingProxyRecords(ghAwDir) {
   for (const relativePath of MODEL_ROUTING_LOG_PATHS) {
     const records = [];
@@ -324,10 +352,19 @@ function getEffectiveModelLabel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp
     if (requested.toLowerCase() !== "auto") return shortModel;
     const models = new Set();
     if (phase === "agent") {
-      const sessionPath = require("path").join(require("path").dirname(infoPath), "usage/aw_session.jsonl");
-      for (const event of readModelRoutingSession(sessionPath)) {
-        if (event.type !== "firewall.token_usage" || (event.data?.purpose && event.data.purpose !== "agent") || Number(event.data?.status) >= 400) continue;
-        const model = validateModelIdentifier(event.data?.model);
+      // Footers render in safe_outputs and before the conclusion job builds aw_session.jsonl,
+      // so prefer the token-usage log shipped in the agent artifact.
+      const ghAwDir = require("path").dirname(infoPath);
+      let requestModels = readAgentTokenUsageModels(ghAwDir);
+      if (requestModels === null) {
+        requestModels = [];
+        for (const event of readModelRoutingSession(require("path").join(ghAwDir, "usage/aw_session.jsonl"))) {
+          if (event.type !== "firewall.token_usage" || (event.data?.purpose && event.data.purpose !== "agent") || Number(event.data?.status) >= 400) continue;
+          requestModels.push(event.data?.model);
+        }
+      }
+      for (const value of requestModels) {
+        const model = validateModelIdentifier(value);
         if (model && model.toLowerCase() !== "auto") models.add(model);
       }
       const observedModels = attribution.selection?.status === "observed" ? attribution.selection.observed_models : null;
diff --git a/actions/setup/js/model_attribution.test.cjs b/actions/setup/js/model_attribution.test.cjs
index 3dbbcac54eb..a4e035c6326 100644
--- a/actions/setup/js/model_attribution.test.cjs
+++ b/actions/setup/js/model_attribution.test.cjs
@@ -58,6 +58,69 @@ describe("observed model attribution", () => {
     }
   });
 
+  const tokenUsageRecord = overrides => ({
+    _schema: "token-usage/v0.29.6",
+    event: "token_usage",
+    provider: "copilot",
+    model: "claude-sonnet-5.5",
+    path: "/chat/completions",
+    status: 200,
+    ...overrides,
+  });
+
+  it.each(["sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl", "sandbox/firewall-audit-logs/api-proxy-logs/token-usage.jsonl", "sandbox/firewall/audit/api-proxy-logs/token-usage.jsonl"])(
+    "labels auto from the agent artifact token-usage log at %s without a usage directory",
+    relativePath => {
+      const root = createRoot();
+      const infoPath = path.join(root, "aw_info.json");
+      try {
+        fs.writeFileSync(infoPath, JSON.stringify({ requested_model: "auto" }));
+        writeJSONL(path.join(root, relativePath), [tokenUsageRecord({})]);
+        expect(fs.existsSync(path.join(root, "usage"))).toBe(false);
+        expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Claude Sonnet 5.5)");
+      } finally {
+        fs.rmSync(root, { recursive: true, force: true });
+      }
+    }
+  );
+
+  it("ignores routing classification and non-2xx token-usage records", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      writeJSONL(path.join(root, "sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl"), [
+        tokenUsageRecord({ model: "gpt-5-mini", purpose: "routing_classification" }),
+        tokenUsageRecord({ model: "gemini-3.8-flash", status: 503 }),
+        tokenUsageRecord({ purpose: "agent" }),
+      ]);
+      writeJSONL(path.join(root, "usage/aw_session.jsonl"), [{ type: "firewall.token_usage", data: { model: "gpt-6.1-sol", purpose: "agent" }, provenance: { component: "firewall", phase: "agent" } }]);
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Claude Sonnet 5.5)");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
+  it("falls back to the unified session when the token-usage log is missing", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      writeJSONL(path.join(root, "usage/aw_session.jsonl"), [{ type: "firewall.token_usage", data: tokenUsageRecord({ purpose: "agent" }), provenance: { component: "firewall", phase: "agent" } }]);
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Claude Sonnet 5.5)");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
+  it("shows plain auto when no model usage files exist", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
   it("preserves fallback-only labels when auto has no observed model evidence", () => {
     const root = createRoot();
     const infoPath = path.join(root, "aw_info.json");

From 80fd66f2ed04921971f2aac0f94bf4a1b0a59784 Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 5 Nov 2026 06:31:18 +0000
Subject: [PATCH 4/5] Merge main; keep observed model history across retries
 and fallback precedence

Co-authored-by: pelikhan <4175913+pelikhan@users.noreply.github.com>
---
 actions/setup/js/copilot_model_selection.cjs  | 38 +++++++++++++++++--
 .../setup/js/copilot_model_selection.test.cjs | 12 ++++++
 actions/setup/js/model_attribution.cjs        |  3 +-
 actions/setup/js/model_attribution.test.cjs   | 18 +++++++++
 4 files changed, 66 insertions(+), 5 deletions(-)

diff --git a/actions/setup/js/copilot_model_selection.cjs b/actions/setup/js/copilot_model_selection.cjs
index b315c32aa01..c6ef440a25a 100644
--- a/actions/setup/js/copilot_model_selection.cjs
+++ b/actions/setup/js/copilot_model_selection.cjs
@@ -10,6 +10,9 @@ const { normalizeForCopilotCLI, splitModelIdentifier } = require("./resolve_mode
 const { copilotSessionContext } = require("./copilot_session.cjs");
 const { getErrorMessage } = require("./error_helpers.cjs");
 
+// Bounds the per-run model history persisted in the selection log (16 KiB) and aw_info.json.
+const MAX_OBSERVED_MODELS = 32;
+
 // Exit code the SDK driver uses to report a main-session model verification failure (substitution
 // or unverifiable evidence) so the harness stops without retrying, as it does in CLI mode.
 const COPILOT_MODEL_VERIFICATION_EXIT_CODE = 79;
@@ -42,6 +45,32 @@ function recordCopilotModelSelection(selection, env = process.env) {
   }
 }
 
+/**
+ * Models already observed by an earlier attempt in this agent phase, so a retried or resumed
+ * session extends the history instead of overwriting it.
+ * @param {NodeJS.ProcessEnv} env
+ * @returns {string[]}
+ */
+function previouslyObservedModels(env) {
+  const file = path.join(env.GH_AW_TMP_DIR || "/tmp/gh-aw", "sandbox/agent/logs/copilot-model-selection.json");
+  try {
+    if (!fs.existsSync(file)) return [];
+    const record = JSON.parse(fs.readFileSync(file, "utf8"));
+    if (record?.status !== "observed") return [];
+    const models = Array.isArray(record.observed_models) ? record.observed_models : [record.effective_model];
+    return [
+      ...new Set(
+        models
+          .filter(model => typeof model === "string")
+          .map(modelID)
+          .filter(Boolean)
+      ),
+    ].slice(-MAX_OBSERVED_MODELS);
+  } catch {
+    return [];
+  }
+}
+
 // Only the runner post-step updates workflow metadata; the harness writes a log.
 function persistCopilotModelSelection(env = process.env) {
   if (env.GH_AW_PHASE && env.GH_AW_PHASE !== "agent") return;
@@ -68,7 +97,7 @@ function persistCopilotModelSelection(env = process.env) {
               .map(modelID)
               .filter(Boolean)
           ),
-        ].slice(0, 32)
+        ].slice(0, MAX_OBSERVED_MODELS)
       : [];
     if (observedModels.length > 1) selection.observed_models = observedModels;
     if (!selection.configured_model || !["observed", "substituted", "unavailable"].includes(selection.status)) throw new Error("Invalid Copilot model selection");
@@ -99,11 +128,11 @@ function createCopilotModelObserver({ configuredModel, resolvedModel, sessionId,
   const offsets = new Map();
   /** @type {Map<string, string>} */
   const partial = new Map();
-  const observedModels = new Set();
   let diagnostic = "";
   let observedModel = "";
   const expected = modelID(resolvedModel);
   if (!expected || (env.GH_AW_PHASE && env.GH_AW_PHASE !== "agent")) return { check: () => "" };
+  const observedModels = new Set(previouslyObservedModels(env));
   if (!sessionId || !/^[a-z0-9_-]+$/i.test(sessionId)) throw new Error("Copilot model verification requires an exact invocation session ID");
   const ownedSessionId = sessionId;
 
@@ -157,7 +186,8 @@ function createCopilotModelObserver({ configuredModel, resolvedModel, sessionId,
           if (!model || model === observedModel) continue;
           observedModel = model;
           if (!observedModels.has(model)) {
-            if (observedModels.size >= 32) observedModels.delete(observedModels.values().next().value);
+            const oldest = observedModels.values().next().value;
+            if (observedModels.size >= MAX_OBSERVED_MODELS && oldest !== undefined) observedModels.delete(oldest);
             observedModels.add(model);
           }
           const substituted = expected.toLowerCase() !== "auto" && model.toLowerCase() !== expected.toLowerCase();
@@ -246,7 +276,7 @@ function createCopilotSDKModelObserver({ configuredModel, expectedModel, provide
       if (matches && effective === recorded) return "";
       recorded = effective;
       if (!observedModels.has(effective)) {
-        if (observedModels.size >= 32) observedModels.delete(observedModels.values().next().value);
+        if (observedModels.size >= MAX_OBSERVED_MODELS) observedModels.delete(observedModels.values().next().value);
         observedModels.add(effective);
       }
       recordCopilotModelSelection(
diff --git a/actions/setup/js/copilot_model_selection.test.cjs b/actions/setup/js/copilot_model_selection.test.cjs
index ee1f6f0212b..e93a93a367b 100644
--- a/actions/setup/js/copilot_model_selection.test.cjs
+++ b/actions/setup/js/copilot_model_selection.test.cjs
@@ -55,6 +55,18 @@ describe("Copilot configured versus observed model", () => {
     });
   });
 
+  it("keeps models observed before a retry when the resumed session switches models", () => {
+    const autoObserver = () => createCopilotModelObserver({ configuredModel: "auto", resolvedModel: "auto", sessionId: "session", env });
+    const first = autoObserver();
+    append({ type: "assistant.message", data: { model: "gemini-3.8-flash" } });
+    expect(first.check()).toBe("");
+    const resumed = autoObserver();
+    append({ type: "assistant.message", data: { model: "claude-sonnet-5" } });
+    expect(resumed.check()).toBe("");
+    expect(readInfo().model_selection).toMatchObject({ status: "observed", effective_model: "claude-sonnet-5", observed_models: ["gemini-3.8-flash", "claude-sonnet-5"] });
+    expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+  });
+
   it("records an unavailable model without creating workflow metadata in the sandbox", () => {
     fs.rmSync(infoPath);
     recordCopilotModelSelection({ status: "unavailable", configured_model: "gemini-3.8-flash", resolved_model: "", effective_model: "" }, env);
diff --git a/actions/setup/js/model_attribution.cjs b/actions/setup/js/model_attribution.cjs
index 158fd5a4b50..0ee3519a145 100644
--- a/actions/setup/js/model_attribution.cjs
+++ b/actions/setup/js/model_attribution.cjs
@@ -368,7 +368,8 @@ function getEffectiveModelLabel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp
         if (model && model.toLowerCase() !== "auto") models.add(model);
       }
       const observedModels = attribution.selection?.status === "observed" ? attribution.selection.observed_models : null;
-      if (!models.size && Array.isArray(observedModels)) {
+      // Native observations name the model Copilot requested, not a proxy fallback replacement.
+      if (!models.size && !attribution.fallbackModel && Array.isArray(observedModels)) {
         for (const value of observedModels) {
           const model = validateModelIdentifier(value);
           if (model && model.toLowerCase() !== "auto") models.add(model);
diff --git a/actions/setup/js/model_attribution.test.cjs b/actions/setup/js/model_attribution.test.cjs
index a4e035c6326..00f92f3ac8a 100644
--- a/actions/setup/js/model_attribution.test.cjs
+++ b/actions/setup/js/model_attribution.test.cjs
@@ -132,6 +132,24 @@ describe("observed model attribution", () => {
     }
   });
 
+  it("keeps the fallback label over native observations without proxy usage evidence", () => {
+    const root = createRoot();
+    const infoPath = path.join(root, "aw_info.json");
+    try {
+      fs.writeFileSync(
+        infoPath,
+        JSON.stringify({
+          requested_model: "auto",
+          fallback_model: "gpt-5.4",
+          model_selection: { status: "observed", effective_model: "gpt-5.6-luna", observed_models: ["copilot/gpt-5.6-luna", "gpt-5.6-luna"] },
+        })
+      );
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("gpt54");
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+
   it("uses validated native model observations when the session usage file is absent", () => {
     const root = createRoot();
     const infoPath = path.join(root, "aw_info.json");

From d6f34ac80bc4641d9311024bb832f8c6a5263f48 Mon Sep 17 00:00:00 2001
From: "copilot-swe-agent[bot]" <198982749+Copilot@users.noreply.github.com>
Date: Thu, 5 Nov 2026 15:42:45 +0000
Subject: [PATCH 5/5] Drop "powered by" from auto attribution label

Co-authored-by: pelikhan <4175913+pelikhan@users.noreply.github.com>
---
 actions/setup/js/copilot_model_selection.test.cjs  |  2 +-
 .../js/copilot_sdk_model_observation.test.cjs      |  2 +-
 actions/setup/js/messages.test.cjs                 |  2 +-
 actions/setup/js/model_attribution.cjs             |  2 +-
 actions/setup/js/model_attribution.test.cjs        | 14 +++++++-------
 5 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/actions/setup/js/copilot_model_selection.test.cjs b/actions/setup/js/copilot_model_selection.test.cjs
index e93a93a367b..c25d0a73c9e 100644
--- a/actions/setup/js/copilot_model_selection.test.cjs
+++ b/actions/setup/js/copilot_model_selection.test.cjs
@@ -64,7 +64,7 @@ describe("Copilot configured versus observed model", () => {
     append({ type: "assistant.message", data: { model: "claude-sonnet-5" } });
     expect(resumed.check()).toBe("");
     expect(readInfo().model_selection).toMatchObject({ status: "observed", effective_model: "claude-sonnet-5", observed_models: ["gemini-3.8-flash", "claude-sonnet-5"] });
-    expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+    expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (Gemini 3.8 Flash, Claude Sonnet 5)");
   });
 
   it("records an unavailable model without creating workflow metadata in the sandbox", () => {
diff --git a/actions/setup/js/copilot_sdk_model_observation.test.cjs b/actions/setup/js/copilot_sdk_model_observation.test.cjs
index 57a96b3bc42..3932df4c546 100644
--- a/actions/setup/js/copilot_sdk_model_observation.test.cjs
+++ b/actions/setup/js/copilot_sdk_model_observation.test.cjs
@@ -354,7 +354,7 @@ runWithCopilotSDK({
       observed_models: ["gemini-3.8-flash", "claude-sonnet-5"],
     });
     persistCopilotModelSelection();
-    expect(getEffectiveModelLabel(infoPath, "agent", process.env)).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+    expect(getEffectiveModelLabel(infoPath, "agent", process.env)).toBe("auto (Gemini 3.8 Flash, Claude Sonnet 5)");
   });
 
   it("sdk_subagent_model_events_do_not_change_main_observation", () => {
diff --git a/actions/setup/js/messages.test.cjs b/actions/setup/js/messages.test.cjs
index 17755074d7b..6efa083bfc5 100644
--- a/actions/setup/js/messages.test.cjs
+++ b/actions/setup/js/messages.test.cjs
@@ -204,7 +204,7 @@ describe("messages.cjs", () => {
         );
         const { getFooterMessage } = await import("./messages.cjs");
         const result = getFooterMessage({ workflowName: "Test Workflow", runUrl: "https://github.com/test/repo/actions/runs/123", historyUrl: "https://github.com/search" });
-        expect(result).toContain(" · copilot · auto (powered by Gemini 3.8 Flash, Claude Sonnet 5) · \\~14 AIC · ⌖ \\~1 AIC · ⊞ 9.2K · [◷]");
+        expect(result).toContain(" · copilot · auto (Gemini 3.8 Flash, Claude Sonnet 5) · \\~14 AIC · ⌖ \\~1 AIC · ⊞ 9.2K · [◷]");
       } finally {
         fs.rmSync(root, { recursive: true, force: true });
         for (const [key, value] of environment) {
diff --git a/actions/setup/js/model_attribution.cjs b/actions/setup/js/model_attribution.cjs
index 0ee3519a145..d44d6f0aa30 100644
--- a/actions/setup/js/model_attribution.cjs
+++ b/actions/setup/js/model_attribution.cjs
@@ -378,7 +378,7 @@ function getEffectiveModelLabel(infoPath = `${process.env.GH_AW_TMP_DIR || "/tmp
     }
     if (!models.size && !attribution.fallbackModel && attribution.model && attribution.model.toLowerCase() !== "auto") models.add(attribution.model);
     const names = [...new Set([...models].map(formatAttributionModelName).filter(Boolean))];
-    if (names.length) return `auto (powered by ${names.join(", ")})`;
+    if (names.length) return `auto (${names.join(", ")})`;
     return attribution.fallbackModel && shortModel ? shortModel : "auto";
   }
   if (attribution.routing.status === "selected") {
diff --git a/actions/setup/js/model_attribution.test.cjs b/actions/setup/js/model_attribution.test.cjs
index 00f92f3ac8a..e416224466a 100644
--- a/actions/setup/js/model_attribution.test.cjs
+++ b/actions/setup/js/model_attribution.test.cjs
@@ -18,8 +18,8 @@ function writeJSONL(filePath, records) {
 
 describe("observed model attribution", () => {
   it.each([
-    ["gemini-3.8-flash", "auto (powered by Gemini 3.8 Flash)"],
-    ["gpt-6.1-sol", "auto (powered by GPT 6.1 Sol)"],
+    ["gemini-3.8-flash", "auto (Gemini 3.8 Flash)"],
+    ["gpt-6.1-sol", "auto (GPT 6.1 Sol)"],
     ["provider/", "auto"],
     ["", "auto"],
     ["auto", "auto"],
@@ -50,7 +50,7 @@ describe("observed model attribution", () => {
         { type: "firewall.token_usage", data: { model: "detector" }, provenance: { component: "firewall", phase: "detection" } },
         { type: "firewall.token_usage", data: { model: "forged" }, provenance: { component: "agent", phase: "agent" } },
       ]);
-      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (Gemini 3.8 Flash, Claude Sonnet 5)");
       expect(getEffectiveModelLabel(infoPath, "detection", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto");
       expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "gpt-5.4" })).toBe("gpt54");
     } finally {
@@ -77,7 +77,7 @@ describe("observed model attribution", () => {
         fs.writeFileSync(infoPath, JSON.stringify({ requested_model: "auto" }));
         writeJSONL(path.join(root, relativePath), [tokenUsageRecord({})]);
         expect(fs.existsSync(path.join(root, "usage"))).toBe(false);
-        expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Claude Sonnet 5.5)");
+        expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (Claude Sonnet 5.5)");
       } finally {
         fs.rmSync(root, { recursive: true, force: true });
       }
@@ -94,7 +94,7 @@ describe("observed model attribution", () => {
         tokenUsageRecord({ purpose: "agent" }),
       ]);
       writeJSONL(path.join(root, "usage/aw_session.jsonl"), [{ type: "firewall.token_usage", data: { model: "gpt-6.1-sol", purpose: "agent" }, provenance: { component: "firewall", phase: "agent" } }]);
-      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Claude Sonnet 5.5)");
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (Claude Sonnet 5.5)");
     } finally {
       fs.rmSync(root, { recursive: true, force: true });
     }
@@ -105,7 +105,7 @@ describe("observed model attribution", () => {
     const infoPath = path.join(root, "aw_info.json");
     try {
       writeJSONL(path.join(root, "usage/aw_session.jsonl"), [{ type: "firewall.token_usage", data: tokenUsageRecord({ purpose: "agent" }), provenance: { component: "firewall", phase: "agent" } }]);
-      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Claude Sonnet 5.5)");
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (Claude Sonnet 5.5)");
     } finally {
       fs.rmSync(root, { recursive: true, force: true });
     }
@@ -165,7 +165,7 @@ describe("observed model attribution", () => {
           },
         })
       );
-      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (powered by Gemini 3.8 Flash, Claude Sonnet 5)");
+      expect(getEffectiveModelLabel(infoPath, "agent", { GH_AW_ENGINE_MODEL: "auto" })).toBe("auto (Gemini 3.8 Flash, Claude Sonnet 5)");
     } finally {
       fs.rmSync(root, { recursive: true, force: true });
     }