gsd-pi 2.73.0 → 2.73.1-dev.06e4302

package/src/resources/extensions/claude-code-cli/tests/stream-adapter.test.ts

@@ -10,8 +10,10 @@ import {
 	mergePendingToolCalls,
 	resolveClaudePermissionMode,
 	buildPromptFromContext,
+	buildSdkQueryPrompt,
 	buildSdkOptions,
 	createClaudeCodeElicitationHandler,
+	extractImageBlocksFromContext,
 	extractToolResultsFromSdkUserMessage,
 	getClaudeLookupCommand,
 	parseAskUserQuestionsElicitation,
@@ -167,6 +169,92 @@ describe("stream-adapter — full context prompt (#2859)", () => {
 	});
 });
 
+describe("stream-adapter — image prompt forwarding (#4183)", () => {
+	test("extractImageBlocksFromContext maps user image parts to Anthropic base64 image blocks", () => {
+		const context: Context = {
+			messages: [
+				{
+					role: "user",
+					content: [
+						{ type: "text", text: "look" },
+						{
+							type: "image",
+							data: "data:image/png;base64,abc123",
+							mimeType: "image/png",
+						},
+					],
+				} as Message,
+			],
+		};
+
+		const imageBlocks = extractImageBlocksFromContext(context);
+		assert.deepEqual(imageBlocks, [
+			{
+				type: "image",
+				source: {
+					type: "base64",
+					media_type: "image/png",
+					data: "abc123",
+				},
+			},
+		]);
+	});
+
+	test("buildSdkQueryPrompt returns plain string when no images exist in context", () => {
+		const context: Context = {
+			messages: [{ role: "user", content: "hello" } as Message],
+		};
+		const textPrompt = buildPromptFromContext(context);
+
+		const prompt = buildSdkQueryPrompt(context, textPrompt);
+		assert.equal(typeof prompt, "string");
+		assert.equal(prompt, textPrompt);
+	});
+
+	test("buildSdkQueryPrompt wraps images and prompt text in an SDK user message iterable", async () => {
+		const context: Context = {
+			messages: [
+				{
+					role: "user",
+					content: [
+						{ type: "image", data: "ZmFrZQ==", mimeType: "image/jpeg" },
+						{ type: "text", text: "What is in this image?" },
+					],
+				} as Message,
+			],
+		};
+		const textPrompt = buildPromptFromContext(context);
+
+		const prompt = buildSdkQueryPrompt(context, textPrompt);
+		assert.notEqual(typeof prompt, "string");
+		assert.ok(prompt && typeof (prompt as any)[Symbol.asyncIterator] === "function");
+
+		const messages: any[] = [];
+		for await (const item of prompt as AsyncIterable<any>) {
+			messages.push(item);
+		}
+		assert.equal(messages.length, 1);
+		assert.deepEqual(messages[0], {
+			type: "user",
+			message: {
+				role: "user",
+				content: [
+					{
+						type: "image",
+						source: {
+							type: "base64",
+							media_type: "image/jpeg",
+							data: "ZmFrZQ==",
+						},
+					},
+					{ type: "text", text: textPrompt },
+				],
+			},
+			parent_tool_use_id: null,
+		});
+	});
+});
+
 // ---------------------------------------------------------------------------
 // Bug #4102 — transcript fabrication regression tests
 // ---------------------------------------------------------------------------
@@ -343,6 +431,26 @@ describe("stream-adapter — session persistence (#2859)", () => {
 		);
 	});
 
+	test("buildSdkOptions maps reasoning to effort for adaptive Claude Code models (#3917)", () => {
+		const options = buildSdkOptions("claude-sonnet-4-6", "test", undefined, { reasoning: "high" });
+		assert.equal(options.effort, "high");
+	});
+
+	test("buildSdkOptions upgrades xhigh reasoning to max for opus 4.6 (#3917)", () => {
+		const options = buildSdkOptions("claude-opus-4-6", "test", undefined, { reasoning: "xhigh" });
+		assert.equal(options.effort, "max");
+	});
+
+	test("buildSdkOptions omits effort when reasoning is undefined (#3917)", () => {
+		const options = buildSdkOptions("claude-sonnet-4-6", "test");
+		assert.equal("effort" in options, false);
+	});
+
+	test("buildSdkOptions omits effort for non-adaptive Claude models (#3917)", () => {
+		const options = buildSdkOptions("claude-sonnet-4-20250514", "test", undefined, { reasoning: "high" });
+		assert.equal("effort" in options, false);
+	});
+
 	test("buildSdkOptions includes workflow MCP server config when env is set", () => {
 		const prev = {
 			GSD_WORKFLOW_MCP_COMMAND: process.env.GSD_WORKFLOW_MCP_COMMAND,
@@ -774,11 +882,12 @@ describe("stream-adapter — MCP elicitation bridge", () => {
 			},
 		};
 
+		const secureValue = "ui-collected-value";
 		const inputCalls: Array<{ opts?: { secure?: boolean } }> = [];
 		const handler = createClaudeCodeElicitationHandler({
 			input: async (_title: string, _placeholder?: string, opts?: { secure?: boolean }) => {
 				inputCalls.push({ opts });
-				return "example-secure-input";
+				return secureValue;
 			},
 		} as any);
 		assert.ok(handler);
@@ -787,7 +896,7 @@ describe("stream-adapter — MCP elicitation bridge", () => {
 		assert.deepEqual(result, {
 			action: "accept",
 			content: {
-				TEST_SECURE_FIELD: "example-secure-input",
+				TEST_SECURE_FIELD: secureValue,
 			},
 		});
 		assert.equal(inputCalls.length, 1);

package/src/resources/extensions/gsd/auto/phases.ts

@@ -481,10 +481,13 @@ export async function runPreDispatch(
       );
     } else if (state.phase === "blocked") {
       const blockerMsg = `Blocked: ${state.blockers.join(", ")}`;
-      await deps.stopAuto(ctx, pi, blockerMsg);
-      ctx.ui.notify(`${blockerMsg}. Fix and run /gsd auto.`, "warning");
-      deps.sendDesktopNotification("GSD", blockerMsg, "error", "attention", basename(s.originalBasePath || s.basePath));
-      deps.logCmuxEvent(prefs, blockerMsg, "error");
+      // Pause instead of hard-stop so the session is resumable with `/gsd auto`.
+      // Hard-stop here was causing premature termination when slice dependencies
+      // were temporarily unresolvable (e.g. after reassessment added new slices).
+      await deps.pauseAuto(ctx, pi);
+      ctx.ui.notify(`${blockerMsg}. Fix and run /gsd auto to resume.`, "warning");
+      deps.sendDesktopNotification("GSD", blockerMsg, "warning", "attention", basename(s.originalBasePath || s.basePath));
+      deps.logCmuxEvent(prefs, blockerMsg, "warning");
     } else {
       const ids = incomplete.map((m: { id: string }) => m.id).join(", ");
       const diag = `basePath=${s.basePath}, milestones=[${state.registry.map((m: { id: string; status: string }) => `${m.id}:${m.status}`).join(", ")}], phase=${state.phase}`;
@@ -583,13 +586,23 @@ export async function runPreDispatch(
     return { action: "break", reason: "milestone-complete" };
   }
 
-  // Terminal: blocked
+  // Terminal: blocked — pause instead of hard-stop so the session is resumable.
   if (state.phase === "blocked") {
     const blockerMsg = `Blocked: ${state.blockers.join(", ")}`;
-    await closeoutAndStop(ctx, pi, s, deps, blockerMsg);
-    ctx.ui.notify(`${blockerMsg}. Fix and run /gsd auto.`, "warning");
-    deps.sendDesktopNotification("GSD", blockerMsg, "error", "attention", basename(s.originalBasePath || s.basePath));
-    deps.logCmuxEvent(prefs, blockerMsg, "error");
+    if (s.currentUnit) {
+      await deps.closeoutUnit(
+        ctx,
+        s.basePath,
+        s.currentUnit.type,
+        s.currentUnit.id,
+        s.currentUnit.startedAt,
+        deps.buildSnapshotOpts(s.currentUnit.type, s.currentUnit.id),
+      );
+    }
+    await deps.pauseAuto(ctx, pi);
+    ctx.ui.notify(`${blockerMsg}. Fix and run /gsd auto to resume.`, "warning");
+    deps.sendDesktopNotification("GSD", blockerMsg, "warning", "attention", basename(s.originalBasePath || s.basePath));
+    deps.logCmuxEvent(prefs, blockerMsg, "warning");
     debugLog("autoLoop", { phase: "exit", reason: "blocked" });
     deps.emitJournalEvent({ ts: new Date().toISOString(), flowId: ic.flowId, seq: ic.nextSeq(), eventType: "terminal", data: { reason: "blocked", blockers: state.blockers } });
     return { action: "break", reason: "blocked" };

package/src/resources/extensions/gsd/auto-dispatch.ts

@@ -52,6 +52,7 @@ import {
   checkNeedsReassessment,
   checkNeedsRunUat,
 } from "./auto-prompts.js";
+import { resolveModelWithFallbacksForUnit } from "./preferences-models.js";
 
 // ─── Types ────────────────────────────────────────────────────────────────
 
@@ -306,8 +307,11 @@ export const DISPATCH_RULES: DispatchRule[] = [
   {
     name: "reassess-roadmap (post-completion)",
     match: async ({ state, mid, midTitle, basePath, prefs }) => {
-      if (prefs?.phases?.skip_reassess || !prefs?.phases?.reassess_after_slice)
-        return null;
+      if (prefs?.phases?.skip_reassess) return null;
+      // Default reassess_after_slice to true — reassessment after slice completion
+      // is essential for roadmap integrity. Opt-out via explicit `false`.
+      const reassessEnabled = prefs?.phases?.reassess_after_slice ?? true;
+      if (!reassessEnabled) return null;
       const needsReassess = await checkNeedsReassessment(basePath, mid, state);
       if (!needsReassess) return null;
       return {
@@ -423,6 +427,7 @@ export const DISPATCH_RULES: DispatchRule[] = [
           midTitle,
           researchReadySlices,
           basePath,
+          resolveModelWithFallbacksForUnit("subagent")?.primary,
         ),
       };
     },
@@ -510,6 +515,7 @@ export const DISPATCH_RULES: DispatchRule[] = [
           sid,
           sTitle,
           basePath,
+          resolveModelWithFallbacksForUnit("subagent")?.primary,
         ),
       };
     },
@@ -548,6 +554,7 @@ export const DISPATCH_RULES: DispatchRule[] = [
       const sid = state.activeSlice.id;
       const sTitle = state.activeSlice.title;
       const maxParallel = reactiveConfig.max_parallel ?? 2;
+      const subagentModel = reactiveConfig.subagent_model ?? resolveModelWithFallbacksForUnit("subagent")?.primary;
 
       // Dry-run mode: max_parallel=1 means graph is derived and logged but
       // execution remains sequential
@@ -618,6 +625,7 @@ export const DISPATCH_RULES: DispatchRule[] = [
             sTitle,
             selected,
             basePath,
+            subagentModel,
           ),
         };
       } catch (err) {
@@ -872,11 +880,14 @@ export async function resolveDispatch(
     }
   }
 
-  // No rule matched — unhandled phase
+  // No rule matched — unhandled phase.
+  // Use level "warning" so the loop pauses (resumable) instead of hard-stopping.
+  // Hard-stop here was causing premature termination for transient phase gaps
+  // (e.g. after reassessment modifies the roadmap and state needs re-derivation).
   return {
     action: "stop",
     reason: `Unhandled phase "${ctx.state.phase}" — run /gsd doctor to diagnose.`,
-    level: "info",
+    level: "warning",
     matchedRule: "<no-match>",
   };
 }

package/src/resources/extensions/gsd/auto-model-selection.ts

@@ -15,6 +15,7 @@ import { resolveModelForComplexity, escalateTier, getEligibleModels, loadCapabil
 import { getLedger, getProjectTotals } from "./metrics.js";
 import { unitPhaseLabel } from "./auto-dashboard.js";
 import { getSessionModelOverride } from "./session-model-override.js";
+import { logWarning } from "./workflow-logger.js";
 
 export interface ModelSelectionResult {
   /** Routing metadata for metrics recording */
@@ -25,9 +26,7 @@ export interface ModelSelectionResult {
 
 export function resolvePreferredModelConfig(
   unitType: string,
-  autoModeStartModel: { provider: string; id: string } | null,
-  /** When false, only return explicit per-phase model configs — do not
-   *  synthesize a routing ceiling from dynamic_routing.tier_models (#3962). */
+  autoModeStartModel: { provider: string; id: string; flatRateCtx?: FlatRateContext } | null,
   isAutoMode = true,
 ) {
   const explicitConfig = resolveModelWithFallbacksForUnit(unitType);
@@ -41,7 +40,7 @@ export function resolvePreferredModelConfig(
   if (!routingConfig.enabled || !routingConfig.tier_models) return undefined;
 
   // Don't synthesize a routing config for flat-rate providers (#3453).
-  if (autoModeStartModel && isFlatRateProvider(autoModeStartModel.provider)) return undefined;
+  if (autoModeStartModel && isFlatRateProvider(autoModeStartModel.provider, autoModeStartModel.flatRateCtx)) return undefined;
 
   const ceilingModel = routingConfig.tier_models.heavy
     ?? (autoModeStartModel ? `${autoModeStartModel.provider}/${autoModeStartModel.id}` : undefined);
@@ -68,7 +67,7 @@ export async function selectAndApplyModel(
   basePath: string,
   prefs: GSDPreferences | undefined,
   verbose: boolean,
-  autoModeStartModel: { provider: string; id: string } | null,
+  autoModeStartModel: { provider: string; id: string; flatRateCtx?: FlatRateContext } | null,
   retryContext?: { isRetry: boolean; previousTier?: string },
   /** When false (interactive/guided-flow), skip dynamic routing and use the session model.
    *  Dynamic routing only applies in auto-mode where cost optimization is expected. (#3962) */
@@ -79,6 +78,17 @@ export async function selectAndApplyModel(
   const effectiveSessionModelOverride = sessionModelOverride === undefined
     ? getSessionModelOverride(ctx.sessionManager.getSessionId())
     : (sessionModelOverride ?? undefined);
+  // Enrich the start model with a flat-rate context up front so routing
+  // synthesis and the dispatch-time guard see the same signals (built-in
+  // list + user `flat_rate_providers` preference + externalCli auto-
+  // detection).  The dispatch-time primary-model check below builds its
+  // own per-provider context when it has a resolved primary model.
+  if (autoModeStartModel) {
+    autoModeStartModel = {
+      ...autoModeStartModel,
+      flatRateCtx: buildFlatRateContext(autoModeStartModel.provider, ctx, prefs),
+    };
+  }
   const modelConfig = effectiveSessionModelOverride
     ? undefined
     : resolvePreferredModelConfig(unitType, autoModeStartModel, isAutoMode);
@@ -107,12 +117,16 @@ export async function selectAndApplyModel(
     if (routingConfig.enabled) {
       const primaryModel = resolveModelId(modelConfig.primary, availableModels, ctx.model?.provider);
       if (primaryModel) {
-        if (isFlatRateProvider(primaryModel.provider)) {
+        const primaryFlatRateCtx = buildFlatRateContext(primaryModel.provider, ctx, prefs);
+        if (isFlatRateProvider(primaryModel.provider, primaryFlatRateCtx)) {
           routingConfig.enabled = false;
         }
       } else if (
-        (autoModeStartModel && isFlatRateProvider(autoModeStartModel.provider))
-        || (ctx.model?.provider && isFlatRateProvider(ctx.model.provider))
+        (autoModeStartModel && isFlatRateProvider(autoModeStartModel.provider, autoModeStartModel.flatRateCtx))
+        || (ctx.model?.provider && isFlatRateProvider(
+          ctx.model.provider,
+          buildFlatRateContext(ctx.model.provider, ctx, prefs),
+        ))
       ) {
         // Primary model unresolvable but provider signals indicate flat-rate —
         // disable routing to prevent quality degradation.
@@ -416,8 +430,68 @@ export function resolveModelId<T extends { id: string; provider: string }>(
  * Uses case-insensitive matching with alias support to prevent fail-open on
  * provider naming variations (e.g. "copilot" vs "github-copilot").
  */
-const FLAT_RATE_PROVIDERS = new Set(["github-copilot", "copilot", "claude-code"]);
+const BUILTIN_FLAT_RATE = new Set(["github-copilot", "copilot", "claude-code"]);
+
+/**
+ * Optional context that lets callers extend flat-rate detection beyond the
+ * hard-coded built-in list.  Either signal on its own is enough to classify
+ * a provider as flat-rate.
+ */
+export interface FlatRateContext {
+  /**
+   * Auth mode for the specific provider being checked, as returned by
+   * `ctx.modelRegistry.getProviderAuthMode(provider)`.  Any provider that
+   * wraps a local CLI (externalCli) is, by definition, a flat-rate
+   * subscription wrapper — every request costs the same regardless of
+   * model, so dynamic routing only degrades quality.
+   */
+  authMode?: "apiKey" | "oauth" | "externalCli" | "none";
+  /**
+   * Case-insensitive list of extra provider IDs the user has declared as
+   * flat-rate via `preferences.flat_rate_providers`.  Used for private
+   * subscription-backed proxies and enterprise-gated deployments that the
+   * built-in list doesn't know about.
+   */
+  userFlatRate?: readonly string[];
+}
+
+export function isFlatRateProvider(provider: string, opts?: FlatRateContext): boolean {
+  const p = provider.toLowerCase();
+  if (BUILTIN_FLAT_RATE.has(p)) return true;
+  if (opts?.userFlatRate?.some(id => id.toLowerCase() === p)) return true;
+  if (opts?.authMode === "externalCli") return true;
+  return false;
+}
 
-export function isFlatRateProvider(provider: string): boolean {
-  return FLAT_RATE_PROVIDERS.has(provider.toLowerCase());
+/**
+ * Build a FlatRateContext for a given provider from live runtime state.
+ * Safe to call when ctx or prefs are undefined — missing pieces are
+ * treated as "no signal".
+ */
+export function buildFlatRateContext(
+  provider: string,
+  ctx?: { modelRegistry?: { getProviderAuthMode?: (p: string) => string } },
+  prefs?: { flat_rate_providers?: readonly string[] },
+): FlatRateContext {
+  let authMode: FlatRateContext["authMode"];
+  const getAuthMode = ctx?.modelRegistry?.getProviderAuthMode;
+  if (typeof getAuthMode === "function") {
+    try {
+      const mode = getAuthMode(provider);
+      if (mode === "apiKey" || mode === "oauth" || mode === "externalCli" || mode === "none") {
+        authMode = mode;
+      }
+    } catch (err) {
+      // Registry lookup failure must never break flat-rate detection —
+      // fall through with authMode undefined and surface the cause.
+      logWarning(
+        "dispatch",
+        `flat-rate auth-mode lookup failed for ${provider}: ${err instanceof Error ? err.message : String(err)}`,
+      );
+    }
+  }
+  return {
+    authMode,
+    userFlatRate: prefs?.flat_rate_providers,
+  };
 }

package/src/resources/extensions/gsd/auto-post-unit.ts

@@ -104,6 +104,7 @@ import {
   updateSliceProgressCache,
   unitVerb,
   hideFooter,
+  describeNextUnit,
 } from "./auto-dashboard.js";
 import { existsSync, unlinkSync } from "node:fs";
 import { join } from "node:path";
@@ -233,6 +234,18 @@ export function detectRogueFileWrites(
   return rogues;
 }
 
+export const STEP_COMPLETE_FALLBACK_MESSAGE =
+  "Step complete. Run /clear, then /gsd to continue (or /gsd auto to run continuously).";
+
+export function buildStepCompleteMessage(nextState: import("./types.js").GSDState): string {
+  if (nextState.phase === "complete") {
+    return "Step complete — milestone finished. Run /gsd status to review, or start the next milestone.";
+  }
+  const next = describeNextUnit(nextState);
+  return `Step complete. Next: ${next.label}\n`
+    + `Run /clear, then /gsd to continue (or /gsd auto to run continuously).`;
+}
+
 export interface PreVerificationOpts {
   skipSettleDelay?: boolean;
   skipWorktreeSync?: boolean;
@@ -619,6 +632,30 @@ export async function postUnitPreVerification(pctx: PostUnitContext, opts?: PreV
           s.verificationRetryCount.set(retryKey, attempt);
 
           if (attempt > MAX_VERIFICATION_RETRIES) {
+            // #4175: For complete-milestone, a blocker placeholder is harmful —
+            // the stub SUMMARY has no recovery value (milestone is terminal),
+            // it does not update DB status (so deriveState never advances),
+            // and it fools stopAuto's presence check into merging a milestone
+            // that was never legitimately completed. Pause auto-mode with a
+            // clear single failure signal and preserve the worktree branch.
+            if (s.currentUnit.type === "complete-milestone") {
+              debugLog("postUnit", {
+                phase: "artifact-verify-pause-complete-milestone",
+                unitType: s.currentUnit.type,
+                unitId: s.currentUnit.id,
+                attempt,
+                maxRetries: MAX_VERIFICATION_RETRIES,
+              });
+              s.verificationRetryCount.delete(retryKey);
+              s.pendingVerificationRetry = null;
+              ctx.ui.notify(
+                `Milestone ${s.currentUnit.id} verification failed after ${MAX_VERIFICATION_RETRIES} retries — worktree branch preserved. Re-run /gsd auto once blockers are resolved.`,
+                "error",
+              );
+              await pauseAuto(ctx, pi);
+              return "dispatched";
+            }
+
             // Retries exhausted — write a blocker placeholder so the pipeline
             // can advance past this stuck unit (#2653).
             debugLog("postUnit", {
@@ -1025,8 +1062,17 @@ export async function postUnitPostVerification(pctx: PostUnitContext): Promise<"
     }
   }
 
-  // Step mode → show wizard instead of dispatch
+  // Step mode → show wizard instead of dispatch.
+  // Without this notify(), /gsd in step mode finishes a unit and silently
+  // exits the loop, leaving the user with no hint to /clear and /gsd again.
   if (s.stepMode) {
+    try {
+      const nextState = await deriveState(s.basePath);
+      ctx.ui.notify(buildStepCompleteMessage(nextState), "info");
+    } catch (e) {
+      debugLog("postUnit", { phase: "step-wizard-notify", error: String(e) });
+      ctx.ui.notify(STEP_COMPLETE_FALLBACK_MESSAGE, "info");
+    }
     return "step-wizard";
   }
 

package/src/resources/extensions/gsd/auto-prompts.ts

@@ -1926,6 +1926,7 @@ export async function buildReassessRoadmapPrompt(
 export async function buildReactiveExecutePrompt(
   mid: string, midTitle: string, sid: string, sTitle: string,
   readyTaskIds: string[], base: string,
+  subagentModel?: string,
 ): Promise<string> {
   const { loadSliceTaskIO, deriveTaskGraph, graphMetrics } = await import("./reactive-graph.js");
 
@@ -1970,10 +1971,11 @@ export async function buildReactiveExecutePrompt(
       { carryForwardPaths: depPaths },
     );
 
+    const modelSuffix = subagentModel ? ` with model: "${subagentModel}"` : "";
     subagentSections.push([
       `### ${tid}: ${tTitle}`,
       "",
-      "Use this as the prompt for a `subagent` call:",
+      `Use this as the prompt for a \`subagent\` call${modelSuffix}:`,
       "",
       "```",
       taskPrompt,
@@ -2049,15 +2051,17 @@ export async function buildParallelResearchSlicesPrompt(
   midTitle: string,
   slices: Array<{ id: string; title: string }>,
   basePath: string,
+  subagentModel?: string,
 ): Promise<string> {
   // Build individual research-slice prompts for each slice
   const subagentSections: string[] = [];
+  const modelSuffix = subagentModel ? ` with model: "${subagentModel}"` : "";
   for (const slice of slices) {
     const slicePrompt = await buildResearchSlicePrompt(mid, midTitle, slice.id, slice.title, basePath);
     subagentSections.push([
       `### ${slice.id}: ${slice.title}`,
       "",
-      "Use this as the prompt for a `subagent` call (agent: `gsd-executor` or the default agent):",
+      `Use this as the prompt for a \`subagent\` call${modelSuffix} (agent: \`gsd-executor\` or the default agent):`,
       "",
       "```",
       slicePrompt,
@@ -2077,6 +2081,7 @@ export async function buildParallelResearchSlicesPrompt(
 export async function buildGateEvaluatePrompt(
   mid: string, midTitle: string, sid: string, sTitle: string,
   base: string,
+  subagentModel?: string,
 ): Promise<string> {
   // Pull only the gates this turn actually owns (Q3/Q4). Filter via the
   // registry so that scope:"slice" gates owned by other turns (Q8) can't
@@ -2128,10 +2133,11 @@ export async function buildGateEvaluatePrompt(
       "- `findings`: detailed markdown findings (or empty if omitted)",
     ].join("\n");
 
+    const modelSuffix = subagentModel ? ` with model: "${subagentModel}"` : "";
     subagentSections.push([
       `### ${def.id}: ${def.question}`,
       "",
-      "Use this as the prompt for a `subagent` call:",
+      `Use this as the prompt for a \`subagent\` call${modelSuffix}:`,
       "",
       "```",
       subPrompt,

package/src/resources/extensions/gsd/auto-start.ts

@@ -83,7 +83,11 @@ import { join } from "node:path";
 import { sep as pathSep } from "node:path";
 
 import { resolveProjectRootDbPath } from "./bootstrap/dynamic-tools.js";
-import { resolveDefaultSessionModel, resolveDynamicRoutingConfig } from "./preferences-models.js";
+import {
+  isCustomProvider,
+  resolveDefaultSessionModel,
+  resolveDynamicRoutingConfig,
+} from "./preferences-models.js";
 import type { WorktreeResolver } from "./worktree-resolver.js";
 import { getSessionModelOverride } from "./session-model-override.js";
 
@@ -274,8 +278,18 @@ export async function bootstrapAutoSession(
   //
   // This preserves #3517 defaults while honoring explicit runtime model
   // selection for subsequent /gsd runs in the same session.
+  //
+  // Exception (#4122): when the session provider is a custom provider declared
+  // in ~/.gsd/agent/models.json (Ollama, vLLM, OpenAI-compatible proxy, etc.),
+  // PREFERENCES.md is skipped entirely. PREFERENCES.md cannot reference custom
+  // providers, so honoring it would silently reroute auto-mode to a built-in
+  // provider the user is not logged into and surface as "Not logged in · Please
+  // run /login" before pausing and resetting to claude-code/claude-sonnet-4-6.
   const manualSessionOverride = getSessionModelOverride(ctx.sessionManager.getSessionId());
-  const preferredModel = resolveDefaultSessionModel(ctx.model?.provider);
+  const sessionProviderIsCustom = isCustomProvider(ctx.model?.provider);
+  const preferredModel = sessionProviderIsCustom
+    ? null
+    : resolveDefaultSessionModel(ctx.model?.provider);
   // Validate the preferred model against the live registry + provider auth so
   // an unconfigured PREFERENCES.md entry (no API key / OAuth) can't become the
   // start-model snapshot. Without this, every subsequent unit would try to
@@ -792,6 +806,9 @@ export async function bootstrapAutoSession(
 
     ctx.ui.setStatus("gsd-auto", s.stepMode ? "next" : "auto");
     ctx.ui.setFooter(hideFooter);
+    // Hide gsd-health during AUTO — gsd-progress is the single source of truth
+    // for last-commit / cost / health signal while auto is running.
+    ctx.ui.setWidget("gsd-health", undefined);
     const modeLabel = s.stepMode ? "Step-mode" : "Auto-mode";
     const pendingCount = (state.registry ?? []).filter(
       (m) => m.status !== "complete" && m.status !== "parked",
@@ -811,12 +828,19 @@ export async function bootstrapAutoSession(
       ? `${s.autoModeStartModel.provider}/${s.autoModeStartModel.id}`
       : ctx.model ? `${ctx.model.provider}/${ctx.model.id}` : "default";
 
-    // Flat-rate providers (e.g. GitHub Copilot, claude-code) suppress routing
-    // at dispatch time (#3453) — reflect that in the banner.
-    const { isFlatRateProvider } = await import("./auto-model-selection.js");
+    // Flat-rate providers (e.g. GitHub Copilot, claude-code, user-declared
+    // subscription proxies, externalCli CLIs) suppress routing at dispatch
+    // time (#3453) — reflect that in the banner.  Thread the same
+    // FlatRateContext used by selectAndApplyModel so user-declared
+    // flat-rate providers and externalCli auto-detection are respected.
+    const { isFlatRateProvider, buildFlatRateContext } = await import("./auto-model-selection.js");
+    const bannerPrefs = loadEffectiveGSDPreferences()?.preferences;
     const effectiveProvider = s.autoModeStartModel?.provider ?? ctx.model?.provider;
     const effectivelyEnabled = routingConfig.enabled
-      && !(effectiveProvider && isFlatRateProvider(effectiveProvider));
+      && !(effectiveProvider && isFlatRateProvider(
+        effectiveProvider,
+        buildFlatRateContext(effectiveProvider, ctx, bannerPrefs),
+      ));
 
     // The actual ceiling may come from tier_models.heavy, not the start model.
     const effectiveCeiling = (routingConfig.enabled && routingConfig.tier_models?.heavy)

package/src/resources/extensions/gsd/auto-timeout-recovery.ts

@@ -230,6 +230,23 @@ export async function recoverTimedOutUnit(
     return "recovered";
   }
 
+  // #4175: For complete-milestone, never write a blocker placeholder — a stub
+  // SUMMARY has no recovery value (milestone is terminal), it does not update
+  // DB status, and downstream merge paths can treat the stub as a legitimate
+  // completion signal. Pause instead so the worktree branch is preserved.
+  if (unitType === "complete-milestone") {
+    writeUnitRuntimeRecord(basePath, unitType, unitId, currentUnitStartedAt, {
+      phase: "paused",
+      recoveryAttempts: recoveryAttempts + 1,
+      lastRecoveryReason: reason,
+    });
+    ctx.ui.notify(
+      `Milestone ${unitId} ${reason}-recovery exhausted ${maxRecoveryAttempts} attempt(s) — worktree branch preserved. Re-run /gsd auto once blockers are resolved.`,
+      "error",
+    );
+    return "paused";
+  }
+
   // Retries exhausted — write a blocker placeholder and advance the pipeline
   // instead of silently stalling.
   const placeholder = writeBlockerPlaceholder(