gsd-pi 2.76.0 → 2.77.0

package/README.md

@@ -27,47 +27,67 @@ One command. Walk away. Come back to a built project with clean git history.
 
 ---
 
-## What's New in v2.75
+## What's New in v2.76
 
-### Knowledge Graph & Learning Extraction
+### Memory System
 
-- **Knowledge graph system** — GSD now builds a structured knowledge graph from project artifacts. Learnings, decisions, and patterns are parsed into queryable graph nodes.
-- **`/gsd extract-learnings`** — new command extracts decisions, lessons, patterns, and surprises from completed phase artifacts into `LEARNINGS.md`, which feeds the knowledge graph automatically.
+- **Persistent agent memory** — new `capture_thought`, `memory_query`, and `gsd_graph` tools let GSD store, retrieve, and relate thoughts across sessions (Phase 1).
+- **Ingestion + scopes/tags** — memories can be scoped (project, milestone, slice) and tagged for filtering (Phase 2).
+- **Hybrid retrieval** — keyword + semantic search over stored memory for higher-recall lookup (Phase 3).
+- **Knowledge graph relationships** — memories link into a queryable graph, feeding planning and dispatch context (Phase 4).
+- **Maintenance & observability** — cap cascade, decay metrics, and export/import for memory stores (Phase 5).
 
-### Unified Orchestration Kernel (UOK)
+### Progressive Planning (ADR-011)
 
-- **UOK is now the default** — the unified orchestration kernel replaces the legacy execution path. Plan-v2 compile gates, unified audit envelopes, turn-level git transaction modes, reactive/parallel scheduling via execution graph, and model policy filtering are all enforced by default. Legacy fallback remains as an emergency escape.
+- **Sketch-then-refine planning** — Phase 1 introduces progressive plan elaboration so GSD drafts a lightweight sketch before locking a full plan (#3789).
+- **Mid-execution escalation** — Phase 2 lets tasks escalate complexity mid-run with compensating rollback on write failure (#3789).
 
-### Extension API
+### Workflow & Onboarding
 
-- **GSD Extension API** — third-party extensions can now be loaded from `.gsd/extensions/`, with a formal API surface for hooking into the GSD lifecycle (#3338).
+- **Unified workflow plugin system** — new plugin architecture with modes and remote install. Manage via `/gsd workflow list|run|install|info|validate`.
+- **`/gsd onboarding` re-entry** — revisit setup without restarting the TUI; setup hub disambiguation prevents duplicate banners.
+- **`/gsd scan`** — new lightweight command for rapid codebase assessment, writing to `.gsd/codebase/`.
 
-### v1 Command Parity
+### Remote Control & TUI
 
-- **12 missing commands added** — GSD v2 now covers all v1 commands, closing the migration gap.
+- **Telegram command interface** — control auto-mode remotely via Telegram alongside existing Slack/Discord routes.
+- **Chat, footer, and welcome refresh** — TUI visuals updated; onboarding no longer hangs on re-entry.
+- **New themes** — `tui-classic` and web `classic`/`vivid` palettes added (#4301).
 
-### TUI Improvements
+### Models & Providers
 
-- **Chat frame redesign** — compaction notices, tool execution cards, and the chat frame now share a unified styling with timestamps and model headers.
-- **Inline tool calls** — assistant tool calls render inline with text instead of grouped at the end.
-- **Compaction and success fixes** — tool cards no longer stick after compaction; success notifications are properly promoted.
+- **Claude Opus 4.7** — added across Anthropic, Bedrock, Antigravity, and OpenRouter (#4348).
+- **GPT-5.4-mini** — added to the OpenAI Codex model list (#1215).
+- **Anthropic proxy support** — `ANTHROPIC_BASE_URL` env var honored for custom proxy endpoints (#4140).
+- **Flat-rate provider opt-in** — `allow_flat_rate_providers` routing toggle (#4386).
+- **Ask User Questions** — optional markdown preview panel with side-by-side layout.
 
 ### Auto-Mode & Reliability
 
-- **Session timeout recovery** — auto-resume timer handles session creation timeouts; timeout counter resets on resume.
-- **Compaction checkpoint fix** — all session phases are checkpointed during compaction, not just the executing phase.
-- **MCP worktree routing** — tool writes are routed to the active worktree when a milestone has one; worktree paths are accepted in the project root guard.
-- **Single-writer DB invariant** — the engine database now enforces a single-writer invariant, preventing corruption from concurrent access.
+- **Stuck-loop hardening** — repeated units detected across the stuck window; research-dispatch stuck loop fixed (#4414).
+- **Branch isolation** — milestone branch is created on entry in `isolation: branch` mode with `main_branch` validation (#4389).
+- **Model fallback on limits** — auto-mode limit errors now trigger model fallback (#4373).
+- **Python/Windows verification** — python invocations normalized on Windows in the verification gate (#4416).
+- **Self-healing `.gsd/`** — symlinked `.gsd` staging self-heals to prevent silent data loss (#4423); `.gsd.migrating` healed on resume.
+- **Stale OAuth recovery** — stale Anthropic OAuth credentials self-heal (#4399).
 
-### Providers & CI
+See the full [Changelog](./CHANGELOG.md) for details on every release.
 
-- **Alibaba DashScope** — added as a standalone provider (#3891).
-- **Persistent language preference** — `/gsd language` sets a persistent language preference.
-- **Flat-rate provider detection** — extended to custom and externalCli providers.
-- **Thinking level as effort** — Claude Code now passes thinking level as an effort parameter.
-- **Hardened release pipeline** — workspace versions synced in release commits, package-lock.json regenerated during bumps, incremental build cache issues resolved.
+<details>
+<summary>v2.75 highlights</summary>
+
+- **Knowledge graph system** — structured knowledge graph built from project artifacts
+- **`/gsd extract-learnings`** — extracts decisions, lessons, patterns, and surprises into `LEARNINGS.md`
+- **Unified Orchestration Kernel (UOK)** — now the default execution path with plan-v2 compile gates and reactive/parallel scheduling
+- **GSD Extension API** — third-party extensions loadable from `.gsd/extensions/` (#3338)
+- **v1 command parity** — 12 missing commands added, closing the migration gap
+- **Chat frame redesign** — unified styling for compaction notices, tool cards, and chat frame with timestamps and model headers
+- **Single-writer DB invariant** — engine database enforces single-writer to prevent corruption
+- **MCP worktree routing** — tool writes routed to active worktree; worktree paths accepted in project root guard
+- **Alibaba DashScope** — added as a standalone provider (#3891)
+- **Persistent language preference** — `/gsd language`
 
-See the full [Changelog](./CHANGELOG.md) for details on every release.
+</details>
 
 <details>
 <summary>v2.74 highlights</summary>

package/dist/claude-cli-check.js

@@ -12,12 +12,41 @@ import { execFileSync } from 'node:child_process';
  * `src/resources/extensions/gsd/pre-execution-checks.ts`.
  */
 export const CLAUDE_COMMAND = process.platform === 'win32' ? 'claude.cmd' : 'claude';
+/**
+ * Ordered list of binary names to probe for the Claude Code CLI.
+ *
+ * Windows installs vary: npm-global installs produce a `claude.cmd` shim,
+ * direct binary installs produce `claude.exe`, and Git Bash wrappers may
+ * expose a bare `claude` shim. Try all three so no valid install is missed.
+ */
+const CLAUDE_COMMAND_CANDIDATES = process.platform === 'win32' ? [CLAUDE_COMMAND, 'claude.exe', 'claude'] : [CLAUDE_COMMAND];
+/**
+ * Try to run `args` against each candidate binary.
+ * Returns the output buffer on first success, throws the last error if all fail.
+ */
+function execClaudeCheck(args) {
+    let lastError;
+    for (const command of CLAUDE_COMMAND_CANDIDATES) {
+        try {
+            return execFileSync(command, args, { timeout: 5_000, stdio: 'pipe' });
+        }
+        catch (error) {
+            lastError = error;
+            const code = error?.code;
+            // EINVAL can surface on Windows Git Bash for .cmd spawn failures.
+            if (code === 'ENOENT' || code === 'EINVAL')
+                continue;
+            throw error;
+        }
+    }
+    throw lastError ?? new Error(`Claude CLI not found (tried: ${CLAUDE_COMMAND_CANDIDATES.join(', ')})`);
+}
 /**
  * Check if the `claude` binary is installed (regardless of auth state).
  */
 export function isClaudeBinaryInstalled() {
     try {
-        execFileSync(CLAUDE_COMMAND, ['--version'], { timeout: 5_000, stdio: 'pipe' });
+        execClaudeCheck(['--version']);
         return true;
     }
     catch {
@@ -29,13 +58,13 @@ export function isClaudeBinaryInstalled() {
  */
 export function isClaudeCliReady() {
     try {
-        execFileSync(CLAUDE_COMMAND, ['--version'], { timeout: 5_000, stdio: 'pipe' });
+        execClaudeCheck(['--version']);
     }
     catch {
         return false;
     }
     try {
-        const output = execFileSync(CLAUDE_COMMAND, ['auth', 'status'], { timeout: 5_000, stdio: 'pipe' })
+        const output = execClaudeCheck(['auth', 'status'])
             .toString()
             .toLowerCase();
         return !(/not logged in|no credentials|unauthenticated|not authenticated/i.test(output));

package/dist/mcp-server.d.ts

@@ -1,6 +1,11 @@
 /**
  * Minimal tool interface matching GSD's AgentTool shape.
  * Avoids a direct dependency on @gsd/pi-agent-core from this compiled module.
+ *
+ * `details` and `isError` are optional fields that runtime tool implementations
+ * may populate. The MCP transport drops non-standard fields, so the wrapper at
+ * the call site mirrors `details` into `structuredContent` and forwards
+ * `isError` directly. See #4472.
  */
 export interface McpToolDef {
     name: string;
@@ -13,6 +18,8 @@ export interface McpToolDef {
             data?: string;
             mimeType?: string;
         }>;
+        details?: Record<string, unknown>;
+        isError?: boolean;
     }>;
 }
 /**

package/dist/mcp-server.js

@@ -1,3 +1,21 @@
+/**
+ * Strict plain-object guard. True only for object literals and
+ * `Object.create(null)` — not for `Date`, `URL`, `Map`, `Set`, class instances,
+ * or arrays. Used to gate `structuredContent` forwarding so the MCP transport
+ * receives only true JSON objects (the protocol contract). See #4477 review.
+ *
+ * Mirrored in `packages/mcp-server/src/workflow-tools.ts` for the
+ * `adaptExecutorResult` adapter on the workflow path. Keep both copies in
+ * sync if the contract definition needs to evolve.
+ */
+function isPlainObject(value) {
+    if (value === null || typeof value !== 'object')
+        return false;
+    if (Array.isArray(value))
+        return false;
+    const proto = Object.getPrototypeOf(value);
+    return proto === null || proto === Object.prototype;
+}
 // MCP SDK subpath imports use wildcard exports (./*) in @modelcontextprotocol/sdk's
 // package.json export map. The wildcard maps "./foo" → "./dist/cjs/foo" (no .js
 // suffix), so bare subpath specifiers like `${MCP_PKG}/server/stdio` resolve to
@@ -83,7 +101,23 @@ export async function startMcpServer(options) {
                 // by stringifying into a text block so clients see the payload.
                 return { type: 'text', text: JSON.stringify(block) };
             });
-            return { content };
+            // Forward a tool's runtime `details` field to MCP's `structuredContent`
+            // channel. The protocol drops non-standard fields on the wire, so tools
+            // that populate `details` for client-side renderers (e.g. save_gate_result)
+            // would otherwise arrive empty on the other side. See #4472.
+            //
+            // Use a strict plain-object guard (prototype-chain check) rather than just
+            // `typeof === 'object' && !Array.isArray()` — Date, URL, Map, Set, and
+            // class instances would otherwise pass through and end up as
+            // `structuredContent`, violating the protocol's JSON-object contract.
+            // The mirror discipline applies in `workflow-tools.ts adaptExecutorResult`.
+            const base = { content };
+            if (isPlainObject(result.details)) {
+                base.structuredContent = result.details;
+            }
+            if (result.isError === true)
+                base.isError = true;
+            return base;
         }
         catch (err) {
             // AbortError from a cancelled tool surfaces as a normal error — MCP

package/dist/onboarding.js

@@ -60,6 +60,8 @@ const OTHER_PROVIDERS = [
     { value: 'xai', label: 'xAI (Grok)', hint: 'console.x.ai' },
     { value: 'openrouter', label: 'OpenRouter', hint: '200+ models — openrouter.ai/keys' },
     { value: 'mistral', label: 'Mistral', hint: 'console.mistral.ai/api-keys' },
+    { value: 'minimax', label: 'MiniMax', hint: 'platform.minimax.io (Anthropic-compatible recommended)' },
+    { value: 'minimax-cn', label: 'MiniMax CN', hint: 'api.minimaxi.com (Anthropic-compatible)' },
     { value: 'ollama-cloud', label: 'Ollama Cloud' },
     { value: 'custom-openai', label: 'Custom (OpenAI-compatible)', hint: 'Ollama, LM Studio, vLLM, proxies — see docs/providers.md' },
 ];
@@ -131,6 +133,36 @@ function persistDefaultProvider(providerId) {
         // Non-fatal: startup fallback logic will still run.
     }
 }
+/**
+ * Persist the selected default model to settings.json.
+ */
+function persistDefaultModel(modelId) {
+    const settingsPath = join(agentDir, 'settings.json');
+    try {
+        const raw = existsSync(settingsPath) ? JSON.parse(readFileSync(settingsPath, 'utf-8')) : {};
+        raw.defaultModel = modelId;
+        mkdirSync(dirname(settingsPath), { recursive: true });
+        writeFileSync(settingsPath, JSON.stringify(raw, null, 2), 'utf-8');
+    }
+    catch {
+        // Non-fatal: startup fallback logic will still run.
+    }
+}
+function detectNativeProviderFromBaseUrl(baseUrl) {
+    try {
+        const hostname = new URL(baseUrl).hostname.toLowerCase();
+        if (hostname === 'api.minimax.io' || hostname.endsWith('.minimax.io')) {
+            return 'minimax';
+        }
+        if (hostname === 'api.minimaxi.com' || hostname.endsWith('.minimaxi.com')) {
+            return 'minimax-cn';
+        }
+    }
+    catch {
+        // ignore parse failures; handled by prior validation
+    }
+    return null;
+}
 /** Sentinel returned by runStep when the user cancels — tells the caller
  *  to abort the entire wizard. */
 const STEP_CANCELLED = Symbol('step-cancelled');
@@ -558,9 +590,22 @@ async function runCustomOpenAIFlow(p, pc, authStorage) {
     if (p.isCancel(modelId) || !modelId)
         return false;
     const trimmedModelId = modelId.trim();
+    const nativeProvider = detectNativeProviderFromBaseUrl(trimmedUrl);
+    if (nativeProvider) {
+        const envVar = nativeProvider === 'minimax' ? 'MINIMAX_API_KEY' : 'MINIMAX_CN_API_KEY';
+        authStorage.set(nativeProvider, { type: 'api_key', key: trimmedKey });
+        persistDefaultProvider(nativeProvider);
+        persistDefaultModel(trimmedModelId);
+        process.env[envVar] = trimmedKey;
+        p.log.success(`${pc.green('MiniMax')} detected — configured as native provider (${pc.cyan(nativeProvider)})`);
+        p.log.info(`Model: ${pc.cyan(trimmedModelId)}`);
+        p.log.info(pc.dim('Using Anthropic-compatible MiniMax integration for full model metadata and clean thinking output.'));
+        return true;
+    }
     // Save API key to auth storage
     authStorage.set('custom-openai', { type: 'api_key', key: trimmedKey });
     persistDefaultProvider('custom-openai');
+    persistDefaultModel(trimmedModelId);
     // Write or merge into models.json
     const modelsJsonPath = join(agentDir, 'models.json');
     let config = { providers: {} };

package/dist/resource-loader.d.ts

@@ -21,6 +21,6 @@ export declare function getNewerManagedResourceVersion(agentDir: string, current
  *
  * Inspectable: `ls ~/.gsd/agent/extensions/`
  */
-export declare function initResources(agentDir: string): void;
+export declare function initResources(agentDir: string, skillsDir?: string): void;
 export declare function hasStaleCompiledExtensionSiblings(extensionsDir: string, sourceDir?: string): boolean;
 export declare function buildResourceLoader(agentDir: string): DefaultResourceLoader;

package/dist/resource-loader.js

@@ -506,7 +506,7 @@ function pruneRemovedBundledExtensions(manifest, agentDir) {
  *
  * Inspectable: `ls ~/.gsd/agent/extensions/`
  */
-export function initResources(agentDir) {
+export function initResources(agentDir, skillsDir = join(homedir(), '.agents', 'skills')) {
     mkdirSync(agentDir, { recursive: true });
     const currentVersion = getBundledGsdVersion();
     const manifest = readManagedResourceManifest(agentDir);
@@ -538,13 +538,7 @@ export function initResources(agentDir) {
     // Sync bundled resources — overwrite so updates land on next launch.
     syncResourceDir(bundledExtensionsDir, join(agentDir, 'extensions'));
     syncResourceDir(join(resourcesDir, 'agents'), join(agentDir, 'agents'));
-    // Skills are no longer force-synced here. Users install skills via the
-    // skills.sh CLI (`npx skills add <repo>`) into ~/.agents/skills/ which
-    // is the industry-standard Agent Skills ecosystem directory.
-    //
-    // Migration from the legacy ~/.gsd/agent/skills/ directory is handled
-    // above the manifest check so it runs on every launch (including retries
-    // after partial copy failures).
+    syncResourceDir(join(resourcesDir, 'skills'), skillsDir);
     // Sync GSD-WORKFLOW.md to agentDir as a fallback for when GSD_WORKFLOW_PATH
     // env var is not set (e.g. fork/dev builds, alternative entry points).
     const workflowSrc = join(resourcesDir, 'GSD-WORKFLOW.md');

package/dist/resources/agents/researcher.md

@@ -1,7 +1,7 @@
 ---
 name: researcher
 description: Web researcher that finds and synthesizes current information using Brave Search
-tools: search-the-web, bash
+tools: search-the-web, bash, capture_thought, memory_query, gsd_graph
 ---
 
 You are a web researcher. You find current, accurate information using web search and synthesize it into a clear, well-structured report.

package/dist/resources/extensions/claude-code-cli/readiness.js

@@ -10,15 +10,38 @@
  */
 import { execFileSync } from "node:child_process";
 /**
- * Platform-correct binary name for the Claude Code CLI.
+ * Candidate executable names for the Claude Code CLI.
  *
- * On Windows, npm-global binaries are installed as `.cmd` shims and
- * `execFileSync` does not auto-resolve the extension — calling bare
- * `claude` would fail with ENOENT even when the CLI is installed and
- * authenticated. Mirrors the `NPM_COMMAND` pattern in
- * `src/resources/extensions/gsd/pre-execution-checks.ts`.
+ * Keep the explicit win32 ternary selector for regression coverage (Issue #4424):
+ * Node's execFileSync must target `claude.cmd` directly on Windows.
  */
 const CLAUDE_COMMAND = process.platform === "win32" ? "claude.cmd" : "claude";
+/**
+ * Windows installs vary: some environments expose `claude.cmd` (npm shim),
+ * `claude.exe` (direct binary install), or a bare `claude` shim on PATH
+ * (for example Git Bash wrappers). Try all three to avoid false "not
+ * installed" results in readiness checks.
+ */
+const CLAUDE_COMMAND_CANDIDATES = process.platform === "win32" ? [CLAUDE_COMMAND, "claude.exe", "claude"] : [CLAUDE_COMMAND];
+function execClaude(args) {
+    let lastError;
+    for (const command of CLAUDE_COMMAND_CANDIDATES) {
+        try {
+            return execFileSync(command, args, { timeout: 5_000, stdio: "pipe" });
+        }
+        catch (error) {
+            lastError = error;
+            const code = error?.code;
+            // Windows Git Bash can surface `.cmd` spawn failures as EINVAL instead
+            // of ENOENT. Treat both as "try next candidate".
+            if (code === "ENOENT" || code === "EINVAL") {
+                continue;
+            }
+            throw error;
+        }
+    }
+    throw lastError ?? new Error(`Claude CLI executable not found (tried: ${CLAUDE_COMMAND_CANDIDATES.join(", ")})`);
+}
 let cachedBinaryPresent = null;
 let cachedAuthed = null;
 let lastCheckMs = 0;
@@ -32,7 +55,7 @@ function refreshCache() {
     lastCheckMs = now;
     // Check binary presence
     try {
-        execFileSync(CLAUDE_COMMAND, ["--version"], { timeout: 5_000, stdio: "pipe" });
+        execClaude(["--version"]);
         cachedBinaryPresent = true;
     }
     catch {
@@ -42,7 +65,7 @@ function refreshCache() {
     }
     // Check auth status — exit code 0 with non-error output means authenticated
     try {
-        const output = execFileSync(CLAUDE_COMMAND, ["auth", "status"], { timeout: 5_000, stdio: "pipe" })
+        const output = execClaude(["auth", "status"])
             .toString()
             .toLowerCase();
         // The CLI outputs "not logged in", "no credentials", or similar when unauthenticated

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

@@ -484,25 +484,23 @@ export function makeAbortedMessage(model, lastTextContent) {
 /**
  * Resolve the Claude Code permission mode for the current run.
  *
- * 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`).
+ * Defaults to `acceptEdits`, which auto-approves file reads/edits but
+ * surfaces a permission dialog for dangerous operations (e.g. general Bash,
+ * Agent, WebFetch). This prevents tools outside the allowlist from being
+ * silently denied — the SDK emits an `extension_ui_request` event so the
+ * user sees a prompt instead of a silent refusal that Claude Code mistakes
+ * for user rejection (#4383).
  *
- * 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.
+ * Set `GSD_CLAUDE_CODE_PERMISSION_MODE` to `bypassPermissions` to restore
+ * the old always-approve behaviour, or to `default` / `plan` for stricter
+ * modes.
  */
 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;
     }
-    return "bypassPermissions";
+    return "acceptEdits";
 }
 // NOTE: These helpers intentionally mirror @gsd/pi-ai anthropic-shared
 // behavior so this extension remains typecheck-stable even when the published
@@ -554,7 +552,7 @@ function mapThinkingLevelToAnthropicEffort(level, modelId) {
 export function buildSdkOptions(modelId, prompt, overrides, extraOptions = {}) {
     const { reasoning, ...sdkExtraOptions } = extraOptions;
     const mcpServers = buildWorkflowMcpServers();
-    const permissionMode = overrides?.permissionMode ?? "bypassPermissions";
+    const permissionMode = overrides?.permissionMode ?? "acceptEdits";
     const disallowedTools = ["AskUserQuestion"];
     // Pre-authorize the safe built-ins and every registered workflow MCP
     // server's tools. `acceptEdits` mode (the interactive default) only
@@ -637,6 +635,68 @@ function normalizeToolResultContent(content) {
     }
     return blocks.length > 0 ? blocks : [{ type: "text", text: "" }];
 }
+/**
+ * Extract a `details` payload from an MCP tool-result block.
+ *
+ * MCP's `CallToolResult` carries structured data in `structuredContent` — the
+ * protocol's supported channel for non-text payloads. Claude Code's synthetic
+ * user message may surface that field in one of two shapes depending on SDK
+ * version: as a sibling on the `mcp_tool_result` block itself, or as a
+ * dedicated content sub-block with `type: "structuredContent"`. Snake-case
+ * (`structured_content`) is accepted defensively in case a transport hop
+ * rewrites casing. All other shapes fall back to an empty object so callers
+ * can rely on `details` being present.
+ */
+function extractStructuredDetailsFromBlock(block) {
+    const sibling = block.structuredContent ?? block.structured_content;
+    if (sibling && typeof sibling === "object" && !Array.isArray(sibling)) {
+        return sibling;
+    }
+    if (Array.isArray(block.content)) {
+        for (const item of block.content) {
+            if (!item || typeof item !== "object")
+                continue;
+            const sub = item;
+            if (sub.type !== "structuredContent" && sub.type !== "structured_content")
+                continue;
+            const payload = sub.structuredContent ?? sub.structured_content ?? sub.data ?? sub.value;
+            if (payload && typeof payload === "object" && !Array.isArray(payload)) {
+                return payload;
+            }
+        }
+    }
+    // Return undefined (not {}) when no structured payload is present, matching
+    // the pre-#4477 contract where `details` was nullable. An empty-object
+    // sentinel is truthy and breaks downstream consumers that gate on
+    // `if (details)`. `undefined` matches the type of the field these results
+    // flow into (`Record<string, unknown> | undefined`).
+    return undefined;
+}
+/**
+ * True for items that are MCP `structuredContent` pseudo-blocks living inside
+ * a tool-result `content[]` array. These blocks carry the structured payload
+ * (extracted separately by `extractStructuredDetailsFromBlock`) and must NOT
+ * leak into the visible content rendered to the user — otherwise the renderer
+ * stringifies the JSON pseudo-block and shows it next to the actual tool
+ * output. See PR #4477 review (CodeRabbit, post-fix-round).
+ */
+function isStructuredContentPseudoBlock(item) {
+    if (!item || typeof item !== "object")
+        return false;
+    const type = item.type;
+    return type === "structuredContent" || type === "structured_content";
+}
+/**
+ * Strip `structuredContent` pseudo-blocks from a tool-result content array
+ * before normalization. The structured payload is extracted via the sibling
+ * `structuredContent` field (or a dedicated extractor pass on the raw block);
+ * the visible content path must not include the pseudo-block itself.
+ */
+function stripStructuredContentPseudoBlocks(content) {
+    if (!Array.isArray(content))
+        return content;
+    return content.filter((item) => !isStructuredContentPseudoBlock(item));
+}
 /** Extract tool result payloads from an SDK synthetic user message, keyed by tool-use ID. */
 export function extractToolResultsFromSdkUserMessage(message) {
     const extracted = [];
@@ -657,8 +717,8 @@ export function extractToolResultsFromSdkUserMessage(message) {
         extracted.push({
             toolUseId,
             result: {
-                content: normalizeToolResultContent(block.content),
-                details: {},
+                content: normalizeToolResultContent(stripStructuredContentPseudoBlocks(block.content)),
+                details: extractStructuredDetailsFromBlock(block),
                 isError: block.is_error === true,
             },
         });
@@ -672,8 +732,8 @@ export function extractToolResultsFromSdkUserMessage(message) {
                 extracted.push({
                     toolUseId,
                     result: {
-                        content: normalizeToolResultContent(toolResult.content),
-                        details: {},
+                        content: normalizeToolResultContent(stripStructuredContentPseudoBlocks(toolResult.content)),
+                        details: extractStructuredDetailsFromBlock(toolResult),
                         isError: toolResult.is_error === true,
                     },
                 });

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

@@ -30,6 +30,13 @@ function stuckStatePath(basePath) {
 function loadStuckState(basePath) {
     try {
         const data = JSON.parse(readFileSync(stuckStatePath(basePath), "utf-8"));
+        // Only load state written by a DIFFERENT process (real session restart).
+        // If the PID matches the current process, this state was written by an earlier
+        // autoLoop call in the same process (e.g., a test that completed before this
+        // one), not by a crashed session — skip it to prevent test state pollution.
+        if (data.pid === process.pid) {
+            return { recentUnits: [], stuckRecoveryAttempts: 0 };
+        }
         return {
             recentUnits: Array.isArray(data.recentUnits) ? data.recentUnits : [],
             stuckRecoveryAttempts: typeof data.stuckRecoveryAttempts === "number" ? data.stuckRecoveryAttempts : 0,
@@ -45,6 +52,7 @@ function saveStuckState(basePath, state) {
         const filePath = stuckStatePath(basePath);
         mkdirSync(join(gsdRoot(basePath), "runtime"), { recursive: true });
         writeFileSync(filePath, JSON.stringify({
+            pid: process.pid,
             recentUnits: state.recentUnits.slice(-20), // keep last 20 entries
             stuckRecoveryAttempts: state.stuckRecoveryAttempts,
             updatedAt: new Date().toISOString(),
@@ -497,6 +505,7 @@ export async function autoLoop(ctx, pi, s, deps, options) {
             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 (#4382)
             debugLog("autoLoop", { phase: "iteration-complete", iteration });
             finishTurn("completed");
         }