gsd-pi 2.72.0 → 2.73.0-dev.27730dc

package/README.md

@@ -623,8 +623,10 @@ The best practice for working in teams is to ensure unique milestone names acros
 # ── GSD: Runtime / Ephemeral (per-developer, per-session) ──────────────────
 # Crash detection sentinel — PID lock, written per auto-mode session
 .gsd/auto.lock
-# Auto-mode dispatch tracker — prevents re-running completed units
-.gsd/completed-units.json
+# Auto-mode dispatch tracker — prevents re-running completed units (includes archived per-milestone files)
+.gsd/completed-units*.json
+# State manifest — workflow state for recovery
+.gsd/state-manifest.json
 # Derived state cache — regenerated from plan/roadmap files on disk
 .gsd/STATE.md
 # Per-developer token/cost accumulator
@@ -637,6 +639,14 @@ The best practice for working in teams is to ensure unique milestone names acros
 .gsd/worktrees/
 # Parallel orchestration IPC and worker status
 .gsd/parallel/
+# SQLite database and WAL sidecars — checkpoint state, forensics data
+.gsd/gsd.db*
+# Daily-rotated event journal — structured event log for forensics
+.gsd/journal/
+# Doctor run history — diagnostic check results
+.gsd/doctor-history.jsonl
+# Workflow event log — structured event stream
+.gsd/event-log.jsonl
 # Generated HTML reports (regenerable via /gsd export --html)
 .gsd/reports/
 # Session-specific interrupted-work markers

package/dist/cli.js

@@ -5,9 +5,7 @@ import { agentDir, sessionsDir, authFilePath } from './app-paths.js';
 import { initResources, buildResourceLoader, getNewerManagedResourceVersion } from './resource-loader.js';
 import { ensureManagedTools } from './tool-bootstrap.js';
 import { loadStoredEnvKeys } from './wizard.js';
-import { migratePiCredentials } from './pi-migration.js';
-import { validateConfiguredModel } from './startup-model-validation.js';
-import { shouldMigrateAnthropicToClaudeCode } from './provider-migrations.js';
+import { migratePiCredentials, getPiDefaultModelAndProvider } from './pi-migration.js';
 import { shouldRunOnboarding, runOnboarding } from './onboarding.js';
 import chalk from 'chalk';
 import { checkForUpdates } from './update-check.js';
@@ -102,6 +100,41 @@ function parseCliArgs(argv) {
     }
     return flags;
 }
+/**
+ * Validate the configured default model against the registry and reset it if
+ * it no longer exists.  Must run AFTER extensions have registered their
+ * providers so that extension models (e.g. pi-claude-cli) are visible.
+ */
+function validateConfiguredModel(modelRegistry, settingsManager) {
+    const configuredProvider = settingsManager.getDefaultProvider();
+    const configuredModel = settingsManager.getDefaultModel();
+    const allModels = modelRegistry.getAll();
+    const availableModels = modelRegistry.getAvailable();
+    const configuredExists = configuredProvider && configuredModel &&
+        allModels.some((m) => m.provider === configuredProvider && m.id === configuredModel);
+    const configuredAvailable = configuredProvider && configuredModel &&
+        availableModels.some((m) => m.provider === configuredProvider && m.id === configuredModel);
+    if (!configuredModel || !configuredExists) {
+        // Model not configured at all, or removed from registry — pick a fallback.
+        // Only fires when the model is genuinely unknown (not just temporarily unavailable).
+        const piDefault = getPiDefaultModelAndProvider();
+        const preferred = (piDefault
+            ? availableModels.find((m) => m.provider === piDefault.provider && m.id === piDefault.model)
+            : undefined) ||
+            availableModels.find((m) => m.provider === 'openai' && m.id === 'gpt-5.4') ||
+            availableModels.find((m) => m.provider === 'openai') ||
+            availableModels.find((m) => m.provider === 'anthropic' && m.id === 'claude-opus-4-6') ||
+            availableModels.find((m) => m.provider === 'anthropic' && m.id.includes('opus')) ||
+            availableModels.find((m) => m.provider === 'anthropic') ||
+            availableModels[0];
+        if (preferred) {
+            settingsManager.setDefaultModelAndProvider(preferred.provider, preferred.id);
+        }
+    }
+    if (settingsManager.getDefaultThinkingLevel() !== 'off' && !configuredExists) {
+        settingsManager.setDefaultThinkingLevel('off');
+    }
+}
 const cliFlags = parseCliArgs(process.argv);
 const isPrintMode = cliFlags.print || cliFlags.mode !== undefined;
 // Early resource-skew check — must run before TTY gate so version mismatch
@@ -320,8 +353,22 @@ if (!isPrintMode) {
 if (!isPrintMode && process.stdout.columns && process.stdout.columns < 40) {
     process.stderr.write(chalk.yellow(`[gsd] Terminal width is ${process.stdout.columns} columns (minimum recommended: 40). Output may be unreadable.\n`));
 }
-// --list-models: print available models and exit (no TTY needed)
+// --list-models: load extensions so that extension-registered providers (e.g.
+// pi-claude-cli) appear in the listing, then flush their pending registrations
+// into the model registry before printing.
 if (cliFlags.listModels !== undefined) {
+    exitIfManagedResourcesAreNewer(agentDir);
+    initResources(agentDir);
+    const listModelsLoader = new DefaultResourceLoader({
+        agentDir,
+        additionalExtensionPaths: cliFlags.extensions.length > 0 ? cliFlags.extensions : undefined,
+    });
+    await listModelsLoader.reload();
+    const listModelsExtensions = listModelsLoader.getExtensions();
+    for (const { name, config } of listModelsExtensions.runtime.pendingProviderRegistrations) {
+        modelRegistry.registerProvider(name, config);
+    }
+    listModelsExtensions.runtime.pendingProviderRegistrations = [];
     const models = modelRegistry.getAvailable();
     if (models.length === 0) {
         console.log('No models available. Set API keys in environment variables.');
@@ -407,29 +454,6 @@ if (isPrintMode) {
         isClaudeCodeReady: () => modelRegistry.isProviderRequestReady('claude-code'),
     });
     markStartup('createAgentSession');
-    // Migrate anthropic OAuth users to claude-code provider when CLI is available (#3772).
-    // Anthropic blocks third-party apps from using subscription quotas — routing through
-    // the local claude CLI binary is TOS-compliant.
-    if (shouldMigrateAnthropicToClaudeCode({
-        authStorage,
-        isClaudeCodeReady: modelRegistry.isProviderRequestReady('claude-code'),
-        defaultProvider: settingsManager.getDefaultProvider(),
-    })) {
-        const currentModelId = settingsManager.getDefaultModel();
-        if (currentModelId) {
-            const ccModel = modelRegistry.find('claude-code', currentModelId);
-            if (ccModel) {
-                try {
-                    await session.setModel(ccModel);
-                    // Only persist after successful session switch to avoid desync
-                    settingsManager.setDefaultModelAndProvider('claude-code', currentModelId);
-                }
-                catch {
-                    // claude-code provider not ready — leave both session and settings unchanged
-                }
-            }
-        }
-    }
     // Validate configured model AFTER extensions have registered their models (#2626).
     // Before this, extension-provided models (e.g. claude-code/*) were not yet in the
     // registry, causing the user's valid choice to be silently overwritten.
@@ -461,6 +485,10 @@ if (isPrintMode) {
             process.stderr.write(`[gsd] ${prefix}: ${err.error}\n`);
         }
     }
+    // Validate configured model now that extension providers are registered.
+    // Must run after createAgentSession() which flushes pendingProviderRegistrations
+    // so extension models (e.g. pi-claude-cli) are visible in the registry.
+    validateConfiguredModel(modelRegistry, settingsManager);
     // Apply --model override if specified
     if (cliFlags.model) {
         const available = modelRegistry.getAvailable();
@@ -595,29 +623,6 @@ const { session, extensionsResult, modelFallbackMessage: interactiveFallbackMsg
     isClaudeCodeReady: () => modelRegistry.isProviderRequestReady('claude-code'),
 });
 markStartup('createAgentSession');
-// Migrate anthropic OAuth users to claude-code provider when CLI is available (#3772).
-// Anthropic blocks third-party apps from using subscription quotas — routing through
-// the local claude CLI binary is TOS-compliant.
-if (shouldMigrateAnthropicToClaudeCode({
-    authStorage,
-    isClaudeCodeReady: modelRegistry.isProviderRequestReady('claude-code'),
-    defaultProvider: settingsManager.getDefaultProvider(),
-})) {
-    const currentModelId = settingsManager.getDefaultModel();
-    if (currentModelId) {
-        const ccModel = modelRegistry.find('claude-code', currentModelId);
-        if (ccModel) {
-            try {
-                await session.setModel(ccModel);
-                // Only persist after successful session switch to avoid desync
-                settingsManager.setDefaultModelAndProvider('claude-code', currentModelId);
-            }
-            catch {
-                // claude-code provider not ready — leave both session and settings unchanged
-            }
-        }
-    }
-}
 // Validate configured model AFTER extensions have registered their models (#2626).
 // Before this, extension-provided models (e.g. claude-code/*) were not yet in the
 // registry, causing the user's valid choice to be silently overwritten.
@@ -648,6 +653,10 @@ if (extensionsResult.errors.length > 0) {
         process.stderr.write(`[gsd] ${prefix}: ${err.error}\n`);
     }
 }
+// Validate configured model now that extension providers are registered.
+// Must run after createAgentSession() which flushes pendingProviderRegistrations
+// so extension models (e.g. pi-claude-cli) are visible in the registry.
+validateConfiguredModel(modelRegistry, settingsManager);
 // Restore scoped models from settings on startup.
 // The upstream InteractiveMode reads enabledModels from settings when /scoped-models is opened,
 // but doesn't apply them to the session at startup — so Ctrl+P cycles all models instead of

package/dist/help-text.js

@@ -148,7 +148,7 @@ export function printHelp(version) {
     process.stdout.write('  --print, -p              Single-shot print mode\n');
     process.stdout.write('  --continue, -c           Resume the most recent session\n');
     process.stdout.write('  --worktree, -w [name]    Start in an isolated worktree (auto-named if omitted)\n');
-    process.stdout.write('  --model <id>             Override model (e.g. claude-opus-4-6)\n');
+    process.stdout.write('  --model <id>             Override model (e.g. provider/model-id)\n');
     process.stdout.write('  --no-session             Disable session persistence\n');
     process.stdout.write('  --extension <path>       Load additional extension\n');
     process.stdout.write('  --tools <a,b,c>          Restrict available tools\n');

package/dist/onboarding.js

@@ -277,6 +277,16 @@ async function runLlmStep(p, pc, authStorage) {
         p.log.info('Your Claude subscription will be used for inference. No API key needed.');
         // Store sentinel so hasAuth('claude-code') returns true on future boots
         authStorage.set('claude-code', { type: 'api_key', key: 'cli' });
+        // Persist claude-code as the default provider so the startup migration in
+        // cli.ts does not need to fire and the user is not left on "anthropic".
+        const settingsPath = join(agentDir, 'settings.json');
+        try {
+            const raw = existsSync(settingsPath) ? JSON.parse(readFileSync(settingsPath, 'utf-8')) : {};
+            raw.defaultProvider = 'claude-code';
+            mkdirSync(dirname(settingsPath), { recursive: true });
+            writeFileSync(settingsPath, JSON.stringify(raw, null, 2), 'utf-8');
+        }
+        catch { /* non-fatal — startup migration will catch it */ }
         return true;
     }
     // ── Step 2: Which provider? ──────────────────────────────────────────────

package/dist/resources/extensions/async-jobs/await-tool.js

@@ -54,11 +54,14 @@ export function createAwaitTool(getManager) {
                     };
                 }
             }
-            // Mark all watched jobs as awaited upfront so the onJobComplete
-            // callback (which fires synchronously in the promise .then()) knows
-            // to suppress the follow-up message.
+            // Suppress follow-up notifications for all watched jobs upfront.
+            // suppressFollowUp() cancels the pending delivery timer (if any), which
+            // handles both the within-turn case (job completes while we await) and
+            // the cross-turn case (job already completed before await_job was called).
+            // Previously this only set j.awaited = true, which missed the cross-turn
+            // case because the queueMicrotask had already fired (#3787).
             for (const j of watched)
-                j.awaited = true;
+                manager.suppressFollowUp(j.id);
             // If all watched jobs are already done, return immediately
             const running = watched.filter((j) => j.status === "running");
             if (running.length === 0) {

package/dist/resources/extensions/async-jobs/job-manager.js

@@ -118,13 +118,38 @@ export class AsyncJobManager {
         }
     }
     // ── Private ────────────────────────────────────────────────────────────
+    /**
+     * Suppress follow-up notification for a job — cancels any pending delivery
+     * timer and marks the job as awaited. Safe to call at any time, including
+     * before or after the job completes (#3787).
+     */
+    suppressFollowUp(id) {
+        const job = this.jobs.get(id);
+        if (!job)
+            return;
+        job.awaited = true;
+        if (job.deliveryTimer !== undefined) {
+            clearTimeout(job.deliveryTimer);
+            job.deliveryTimer = undefined;
+        }
+    }
     deliverResult(job) {
         if (!this.onJobComplete)
             return;
-        // Defer delivery by one microtask so await_job's .then() chain runs first
-        // and can set job.awaited = true before onJobComplete checks it (#2762).
+        // Use setTimeout(0) instead of queueMicrotask so the handle is cancellable.
+        // suppressFollowUp() can clear this timer even when await_job is called in
+        // a later LLM turn (after the job already completed). queueMicrotask ran
+        // immediately and could not be cancelled (#2762, #3787).
         const cb = this.onJobComplete;
-        queueMicrotask(() => cb(job));
+        job.deliveryTimer = setTimeout(() => {
+            job.deliveryTimer = undefined;
+            if (!job.awaited)
+                cb(job);
+        }, 0);
+        // Allow process to exit even if timer is pending
+        if (typeof job.deliveryTimer === "object" && "unref" in job.deliveryTimer) {
+            job.deliveryTimer.unref();
+        }
     }
     scheduleEviction(id) {
         const existing = this.evictionTimers.get(id);

package/dist/resources/extensions/claude-code-cli/partial-builder.js

@@ -6,6 +6,44 @@
  */
 import { hasXmlParameterTags, repairToolJson } from "@gsd/pi-ai";
 // ---------------------------------------------------------------------------
+// MCP tool name parsing
+// ---------------------------------------------------------------------------
+/**
+ * Split a Claude Code MCP tool name (`mcp__<server>__<tool>`) into its parts.
+ * Returns null for non-prefixed names so callers can fall through unchanged.
+ *
+ * Server names may contain hyphens (`gsd-workflow`); the SDK uses the literal
+ * `__` delimiter between the server name and the tool name.
+ */
+export function parseMcpToolName(name) {
+    if (!name.startsWith("mcp__"))
+        return null;
+    const rest = name.slice("mcp__".length);
+    const delim = rest.indexOf("__");
+    if (delim <= 0 || delim === rest.length - 2)
+        return null;
+    return { server: rest.slice(0, delim), tool: rest.slice(delim + 2) };
+}
+/**
+ * Build a GSD ToolCall block from a Claude Code SDK tool_use block, stripping
+ * the `mcp__<server>__` prefix from the name so registered extension renderers
+ * (which use the unprefixed canonical names) can match. The original server
+ * name is preserved on the block for diagnostics and rendering.
+ */
+function toolCallFromBlock(id, rawName, input) {
+    const parsed = parseMcpToolName(rawName);
+    const toolCall = {
+        type: "toolCall",
+        id,
+        name: parsed ? parsed.tool : rawName,
+        arguments: input,
+    };
+    if (parsed) {
+        toolCall.mcpServer = parsed.server;
+    }
+    return toolCall;
+}
+// ---------------------------------------------------------------------------
 // Content-block mapping helpers
 // ---------------------------------------------------------------------------
 /**
@@ -22,12 +60,7 @@ export function mapContentBlock(block) {
                 ...(block.signature ? { thinkingSignature: block.signature } : {}),
             };
         case "tool_use":
-            return {
-                type: "toolCall",
-                id: block.id,
-                name: block.name,
-                arguments: block.input,
-            };
+            return toolCallFromBlock(block.id, block.name, block.input);
         case "server_tool_use":
             return {
                 type: "serverToolUse",
@@ -149,12 +182,7 @@ export class PartialMessageBuilder {
                 }
                 if (block.type === "tool_use") {
                     this.toolJsonAccum.set(streamIndex, "");
-                    this.partial.content.push({
-                        type: "toolCall",
-                        id: block.id,
-                        name: block.name,
-                        arguments: {},
-                    });
+                    this.partial.content.push(toolCallFromBlock(block.id, block.name, {}));
                     return { type: "toolcall_start", contentIndex, partial: this.partial };
                 }
                 if (block.type === "server_tool_use") {

package/dist/resources/extensions/claude-code-cli/stream-adapter.js

@@ -92,18 +92,34 @@ function extractMessageText(msg) {
  * call effectively stateless. This version serialises the complete
  * conversation history (system prompt + all user/assistant turns) so
  * Claude Code has full context for multi-turn continuity.
+ *
+ * History is wrapped in XML-tag structure rather than `[User]`/`[Assistant]`
+ * bracket headers. Bracket headers read to the model as an in-context
+ * demonstration of how turns are delimited, causing it to fabricate fake
+ * user turns in its own output. XML tags read as document structure and
+ * don't get mirrored in free text.
  */
 export function buildPromptFromContext(context) {
-    const parts = [];
+    const hasContent = Boolean(context.systemPrompt) || context.messages.some((m) => extractMessageText(m));
+    if (!hasContent)
+        return "";
+    const parts = [
+        "Respond only to the final user message below. " +
+            "Do not emit <user_message>, <assistant_message>, or <prior_system_context> tags in your response.",
+    ];
     if (context.systemPrompt) {
-        parts.push(`[System]\n${context.systemPrompt}`);
+        parts.push(`<prior_system_context>\n${context.systemPrompt}\n</prior_system_context>`);
     }
+    const turns = [];
     for (const msg of context.messages) {
         const text = extractMessageText(msg);
         if (!text)
             continue;
-        const label = msg.role === "user" ? "User" : msg.role === "assistant" ? "Assistant" : "System";
-        parts.push(`[${label}]\n${text}`);
+        const tag = msg.role === "user" ? "user_message" : msg.role === "assistant" ? "assistant_message" : "system_message";
+        turns.push(`<${tag}>\n${text}\n</${tag}>`);
+    }
+    if (turns.length > 0) {
+        parts.push(`<conversation_history>\n${turns.join("\n")}\n</conversation_history>`);
     }
     return parts.join("\n\n");
 }
@@ -389,32 +405,25 @@ export function makeAbortedMessage(model, lastTextContent) {
 /**
  * Resolve the Claude Code permission mode for the current run.
  *
- * - Auto-mode / headless runs bypass permissions so tool calls don't block
- *   on prompts the user isn't watching.
- * - Interactive runs default to `acceptEdits` so file/bash writes still
- *   land quickly but the SDK retains a permission gate.
- * - `GSD_CLAUDE_CODE_PERMISSION_MODE` forces a specific mode when set.
+ * GSD subagents run underneath a host Claude Code session the user has
+ * already consented to, and their work (edits, shell inspection, MCP calls)
+ * spans the full workflow toolset. Defaulting the inner SDK to
+ * `bypassPermissions` avoids per-tool approval prompts that offer no
+ * meaningful safety beyond what the host session and the subagent prompts
+ * already enforce. `GSD_CLAUDE_CODE_PERMISSION_MODE` lets security-conscious
+ * users opt into a stricter mode (`acceptEdits`, `default`, `plan`).
  *
- * Cross-extension coupling is kept minimal by dynamically importing
- * `isAutoActive` and falling back to the bypass default if the import
- * fails (e.g. in unit tests that load stream-adapter in isolation).
+ * Tradeoff: bypass means a prompt-injection payload read from an untrusted
+ * file could trigger tool calls without a second gate. Accepted for GSD
+ * because the workflow is explicit user intent and the alternative
+ * (#4099) is continuous approval fatigue that blocks real work.
  */
 export async function resolveClaudePermissionMode(env = process.env) {
     const override = env.GSD_CLAUDE_CODE_PERMISSION_MODE?.trim();
     if (override === "bypassPermissions" || override === "acceptEdits" || override === "default" || override === "plan") {
         return override;
     }
-    try {
-        const autoMod = (await import("../gsd/auto.js"));
-        if (typeof autoMod.isAutoActive === "function" && autoMod.isAutoActive()) {
-            return "bypassPermissions";
-        }
-        return "acceptEdits";
-    }
-    catch {
-        // auto.ts unavailable (tests, non-GSD contexts) — stay permissive.
-        return "bypassPermissions";
-    }
+    return "bypassPermissions";
 }
 /**
  * Build the options object passed to the Claude Agent SDK's `query()` call.
@@ -431,6 +440,21 @@ export function buildSdkOptions(modelId, prompt, overrides, extraOptions = {}) {
     const mcpServers = buildWorkflowMcpServers();
     const permissionMode = overrides?.permissionMode ?? "bypassPermissions";
     const disallowedTools = ["AskUserQuestion"];
+    // Pre-authorize the safe built-ins and every registered workflow MCP
+    // server's tools. `acceptEdits` mode (the interactive default) only
+    // auto-approves file edits — Read/Glob/Grep, basic shell inspection, and
+    // every `mcp__gsd-workflow__*` call still surface as "This command
+    // requires approval" and block GSD actions (#4099).
+    const allowedTools = [
+        "Read",
+        "Write",
+        "Edit",
+        "Glob",
+        "Grep",
+        "Bash(ls:*)",
+        "Bash(pwd)",
+        ...(mcpServers ? Object.keys(mcpServers).map((serverName) => `mcp__${serverName}__*`) : []),
+    ];
     return {
         pathToClaudeCodeExecutable: getClaudePath(),
         model: modelId,
@@ -442,6 +466,7 @@ export function buildSdkOptions(modelId, prompt, overrides, extraOptions = {}) {
         settingSources: ["project"],
         systemPrompt: { type: "preset", preset: "claude_code" },
         disallowedTools,
+        ...(allowedTools.length > 0 ? { allowedTools } : {}),
         ...(mcpServers ? { mcpServers } : {}),
         betas: modelId.includes("sonnet") ? ["context-1m-2025-08-07"] : [],
         ...extraOptions,

package/dist/resources/extensions/gsd/auto/loop.js

@@ -13,6 +13,69 @@ import { runPreDispatch, runDispatch, runGuards, runUnitPhase, runFinalize, } fr
 import { debugLog } from "../debug-logger.js";
 import { isInfrastructureError, isTransientCooldownError, getCooldownRetryAfterMs, COOLDOWN_FALLBACK_WAIT_MS, MAX_COOLDOWN_RETRIES } from "./infra-errors.js";
 import { resolveEngine } from "../engine-resolver.js";
+import { logWarning } from "../workflow-logger.js";
+import { gsdRoot } from "../paths.js";
+import { readFileSync, writeFileSync, mkdirSync } from "node:fs";
+import { join } from "node:path";
+// ── Stuck detection persistence (#3704) ──────────────────────────────────
+// Persist stuck detection state to disk so it survives session restarts.
+// Without this, restarting auto-mode resets all counters, allowing the
+// same blocked unit to burn a full retry budget each session.
+function stuckStatePath(basePath) {
+    return join(gsdRoot(basePath), "runtime", "stuck-state.json");
+}
+function loadStuckState(basePath) {
+    try {
+        const data = JSON.parse(readFileSync(stuckStatePath(basePath), "utf-8"));
+        return {
+            recentUnits: Array.isArray(data.recentUnits) ? data.recentUnits : [],
+            stuckRecoveryAttempts: typeof data.stuckRecoveryAttempts === "number" ? data.stuckRecoveryAttempts : 0,
+        };
+    }
+    catch (err) {
+        debugLog("autoLoop", { phase: "load-stuck-state-failed", error: err instanceof Error ? err.message : String(err) });
+        return { recentUnits: [], stuckRecoveryAttempts: 0 };
+    }
+}
+function saveStuckState(basePath, state) {
+    try {
+        const filePath = stuckStatePath(basePath);
+        mkdirSync(join(gsdRoot(basePath), "runtime"), { recursive: true });
+        writeFileSync(filePath, JSON.stringify({
+            recentUnits: state.recentUnits.slice(-20), // keep last 20 entries
+            stuckRecoveryAttempts: state.stuckRecoveryAttempts,
+            updatedAt: new Date().toISOString(),
+        }) + "\n");
+    }
+    catch (err) {
+        debugLog("autoLoop", { phase: "save-stuck-state-failed", error: err instanceof Error ? err.message : String(err) });
+    }
+}
+// ── Memory pressure monitoring (#3331) ──────────────────────────────────
+// Check heap usage every N iterations and trigger graceful shutdown before
+// the OS OOM killer sends SIGKILL. The threshold is 90% of the V8 heap
+// limit (--max-old-space-size or default ~1.5-4GB depending on platform).
+const MEMORY_CHECK_INTERVAL = 5; // check every 5 iterations
+const MEMORY_PRESSURE_THRESHOLD = 0.85; // 85% of heap limit
+function checkMemoryPressure() {
+    const mem = process.memoryUsage();
+    // v8.getHeapStatistics() gives heap_size_limit but requires import
+    // Use a conservative estimate: RSS > 3GB is danger zone on most systems
+    const heapMB = Math.round(mem.heapUsed / 1024 / 1024);
+    const rssMB = Math.round(mem.rss / 1024 / 1024);
+    // Try to get the actual V8 heap limit
+    let limitMB = 4096; // conservative default
+    try {
+        const v8 = require("node:v8");
+        const stats = v8.getHeapStatistics();
+        limitMB = Math.round(stats.heap_size_limit / 1024 / 1024);
+    }
+    catch {
+        limitMB = 4096; /* v8 stats unavailable — use conservative default */
+    }
+    const pct = heapMB / limitMB;
+    return { pressured: pct > MEMORY_PRESSURE_THRESHOLD, heapMB, limitMB, pct };
+}
 /**
  * Main auto-mode execution loop. Iterates: derive → dispatch → guards →
  * runUnit → finalize → repeat. Exits when s.active becomes false or a
@@ -24,7 +87,13 @@ import { resolveEngine } from "../engine-resolver.js";
 export async function autoLoop(ctx, pi, s, deps) {
     debugLog("autoLoop", { phase: "enter" });
     let iteration = 0;
-    const loopState = { recentUnits: [], stuckRecoveryAttempts: 0, consecutiveFinalizeTimeouts: 0 };
+    // Load persisted stuck state so counters survive session restarts (#3704)
+    const persisted = loadStuckState(s.basePath);
+    const loopState = {
+        recentUnits: persisted.recentUnits,
+        stuckRecoveryAttempts: persisted.stuckRecoveryAttempts,
+        consecutiveFinalizeTimeouts: 0,
+    };
     let consecutiveErrors = 0;
     let consecutiveCooldowns = 0;
     const recentErrorMessages = [];
@@ -44,6 +113,19 @@ export async function autoLoop(ctx, pi, s, deps) {
             await deps.stopAuto(ctx, pi, `Safety: loop exceeded ${MAX_LOOP_ITERATIONS} iterations — possible runaway`);
             break;
         }
+        // ── Memory pressure check (#3331) ──
+        // Graceful shutdown before OOM killer sends SIGKILL.
+        if (iteration % MEMORY_CHECK_INTERVAL === 0) {
+            const mem = checkMemoryPressure();
+            debugLog("autoLoop", { phase: "memory-check", ...mem });
+            if (mem.pressured) {
+                logWarning("dispatch", `Memory pressure: ${mem.heapMB}MB / ${mem.limitMB}MB (${Math.round(mem.pct * 100)}%) — stopping auto-mode to prevent OOM kill`);
+                await deps.stopAuto(ctx, pi, `Memory pressure: heap at ${mem.heapMB}MB / ${mem.limitMB}MB (${Math.round(mem.pct * 100)}%). ` +
+                    `Stopping gracefully to prevent OOM kill after ${iteration} iterations. ` +
+                    `Resume with /gsd auto to continue from where you left off.`);
+                break;
+            }
+        }
         if (!s.cmdCtx) {
             debugLog("autoLoop", { phase: "exit", reason: "no-cmdCtx" });
             break;
@@ -162,6 +244,7 @@ export async function autoLoop(ctx, pi, s, deps) {
                 consecutiveCooldowns = 0;
                 recentErrorMessages.length = 0;
                 deps.emitJournalEvent({ ts: new Date().toISOString(), flowId, seq: nextSeq(), eventType: "iteration-end", data: { iteration } });
+                saveStuckState(s.basePath, loopState); // persist across session restarts (#3704)
                 debugLog("autoLoop", { phase: "iteration-complete", iteration });
                 if (reconcileResult.outcome === "milestone-complete") {
                     await deps.stopAuto(ctx, pi, "Workflow complete");

package/dist/resources/extensions/gsd/auto-dispatch.js

@@ -18,6 +18,7 @@ import { logWarning, logError } from "./workflow-logger.js";
 import { join } from "node:path";
 import { hasImplementationArtifacts } from "./auto-recovery.js";
 import { buildDiscussMilestonePrompt, buildResearchMilestonePrompt, buildPlanMilestonePrompt, buildResearchSlicePrompt, buildPlanSlicePrompt, buildExecuteTaskPrompt, buildCompleteSlicePrompt, buildCompleteMilestonePrompt, buildValidateMilestonePrompt, buildReplanSlicePrompt, buildRunUatPrompt, buildReassessRoadmapPrompt, buildRewriteDocsPrompt, buildReactiveExecutePrompt, buildGateEvaluatePrompt, buildParallelResearchSlicesPrompt, checkNeedsReassessment, checkNeedsRunUat, } from "./auto-prompts.js";
+import { resolveModelWithFallbacksForUnit } from "./preferences-models.js";
 function missingSliceStop(mid, phase) {
     return {
         action: "stop",
@@ -330,7 +331,7 @@ export const DISPATCH_RULES = [
                 action: "dispatch",
                 unitType: "research-slice",
                 unitId: `${mid}/parallel-research`,
-                prompt: await buildParallelResearchSlicesPrompt(mid, midTitle, researchReadySlices, basePath),
+                prompt: await buildParallelResearchSlicesPrompt(mid, midTitle, researchReadySlices, basePath, resolveModelWithFallbacksForUnit("subagent")?.primary),
             };
         },
     },
@@ -401,7 +402,7 @@ export const DISPATCH_RULES = [
                 action: "dispatch",
                 unitType: "gate-evaluate",
                 unitId: `${mid}/${sid}/gates+${pending.map(g => g.gate_id).join(",")}`,
-                prompt: await buildGateEvaluatePrompt(mid, midTitle, sid, sTitle, basePath),
+                prompt: await buildGateEvaluatePrompt(mid, midTitle, sid, sTitle, basePath, resolveModelWithFallbacksForUnit("subagent")?.primary),
             };
         },
     },
@@ -436,6 +437,7 @@ export const DISPATCH_RULES = [
             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
             if (maxParallel <= 1)
@@ -478,7 +480,7 @@ export const DISPATCH_RULES = [
                     action: "dispatch",
                     unitType: "reactive-execute",
                     unitId: `${mid}/${sid}/reactive+${batchSuffix}`,
-                    prompt: await buildReactiveExecutePrompt(mid, midTitle, sid, sTitle, selected, basePath),
+                    prompt: await buildReactiveExecutePrompt(mid, midTitle, sid, sTitle, selected, basePath, subagentModel),
                 };
             }
             catch (err) {

package/dist/resources/extensions/gsd/auto-post-unit.js

@@ -16,6 +16,7 @@ import { loadFile, parseSummary, resolveAllOverrides } from "./files.js";
 import { loadPrompt } from "./prompt-loader.js";
 import { resolveSliceFile, resolveSlicePath, resolveTaskFile, resolveMilestoneFile, resolveTasksDir, buildTaskFileName, } from "./paths.js";
 import { invalidateAllCaches } from "./cache.js";
+import { rebuildState } from "./doctor.js";
 import { parseUnitId } from "./unit-id.js";
 import { closeoutUnit } from "./auto-unit-closeout.js";
 import { autoCommitCurrentBranch, } from "./worktree.js";
@@ -288,6 +289,11 @@ export async function postUnitPreVerification(pctx, opts) {
                 debugLog("postUnit", { phase: "browser-teardown", status: "closed" });
             }
         });
+        // Keep the on-disk STATE.md aligned with the live derived state after
+        // ordinary unit completion, before any worktree state is synced back.
+        await runSafely("postUnit", "state-rebuild", async () => {
+            await rebuildState(s.basePath);
+        });
         // Sync worktree state back to project root (skipped for lightweight sidecars)
         if (!opts?.skipWorktreeSync && s.originalBasePath && s.originalBasePath !== s.basePath) {
             await runSafely("postUnit", "worktree-sync", () => {