gsd-pi 2.76.0-dev.4c866b677 → 2.76.0-dev.7218806ab

package/src/resources/extensions/gsd/guided-flow.ts

@@ -130,8 +130,20 @@ interface PendingAutoStartEntry {
   milestoneId: string; // the milestone being discussed
   step?: boolean; // preserve step mode through discuss → auto transition
   createdAt: number; // timestamp for staleness detection (#3274)
+  // #4573: counter for how many times the LLM emitted the ready phrase
+  // without writing the required artifacts. Cleared on entry delete/recreate.
+  readyRejectCount?: number;
 }
 
+// #4573: cap for how many times we nudge the LLM after a premature ready
+// phrase before giving up and asking the user to re-run /gsd.
+const MAX_READY_REJECTS = 2;
+
+// #4573: matches the canonical ready phrase the discuss prompt asks the LLM
+// to emit. Accepts any M-prefixed milestone ID (three digits + optional
+// suffix) with optional trailing punctuation.
+const READY_PHRASE_RE = /\bMilestone\s+M\d{3}[A-Z0-9-]*\s+ready\.?/i;
+
 const pendingAutoStartMap = new Map<string, PendingAutoStartEntry>();
 
 /**
@@ -279,6 +291,215 @@ export function checkAutoStartAfterDiscuss(): boolean {
   return true;
 }
 
+/**
+ * Extract the concatenated text content from an assistant message, whether it
+ * stores content as a string or as an array of text blocks.
+ */
+function extractAssistantText(msg: any): string {
+  if (!msg) return "";
+  const content = msg.content;
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+  const parts: string[] = [];
+  for (const block of content) {
+    if (!block || typeof block !== "object") continue;
+    if (block.type === "text" && typeof block.text === "string") parts.push(block.text);
+  }
+  return parts.join("\n");
+}
+
+/**
+ * Return true if the assistant message contains any tool-use block.
+ */
+function hasToolUse(msg: any): boolean {
+  if (!msg) return false;
+  const content = msg.content;
+  if (!Array.isArray(content)) return false;
+  return content.some((b: any) => b && typeof b === "object" && (b.type === "tool_use" || b.type === "tool-use"));
+}
+
+/**
+ * #4573 — Detect and recover from the "ready phrase without files" failure mode.
+ *
+ * When the LLM emits "Milestone {{id}} ready." but has not written CONTEXT.md
+ * or ROADMAP.md, `checkAutoStartAfterDiscuss()` silently returns false and the
+ * next /gsd invocation loops into the "All milestones complete" warning.
+ *
+ * This function, called from `handleAgentEnd` after `checkAutoStartAfterDiscuss`
+ * returns false, pattern-matches the ready phrase on the last assistant message.
+ * If it fired AND neither CONTEXT.md nor ROADMAP.md exists, it:
+ *   1. Notifies the user that the signal was rejected.
+ *   2. Injects a system message via `pi.sendMessage(..., {triggerTurn:true})`
+ *      telling the LLM the signal was premature and to emit the writes now.
+ *   3. Caps at `MAX_READY_REJECTS` per-entry; beyond that, gives up and asks
+ *      the user to re-run /gsd.
+ *
+ * Returns true when a nudge (or give-up) was emitted, signaling the caller to
+ * skip `resolveAgentEnd`.
+ */
+export function maybeHandleReadyPhraseWithoutFiles(event: { messages: any[] }): boolean {
+  const entry = _getPendingAutoStart();
+  if (!entry) return false;
+  const { ctx, pi, basePath, milestoneId } = entry;
+
+  // Gate: last assistant message must contain the ready phrase
+  const lastMsg = event.messages[event.messages.length - 1];
+  const text = extractAssistantText(lastMsg);
+  if (!READY_PHRASE_RE.test(text)) return false;
+
+  // Gate: artifacts must still be missing — if they exist, the happy path
+  // already fired and we have nothing to do.
+  const contextFile = resolveMilestoneFile(basePath, milestoneId, "CONTEXT");
+  const roadmapFile = resolveMilestoneFile(basePath, milestoneId, "ROADMAP");
+  if (contextFile || roadmapFile) return false;
+
+  entry.readyRejectCount = (entry.readyRejectCount ?? 0) + 1;
+
+  if (entry.readyRejectCount > MAX_READY_REJECTS) {
+    // Give up: clear state and tell the user to re-run /gsd. Avoids an
+    // infinite nudge loop when the LLM never produces the writes.
+    pendingAutoStartMap.delete(basePath);
+    ctx.ui.notify(
+      `Milestone ${milestoneId}: LLM signaled "ready" ${entry.readyRejectCount} times without writing files. ` +
+      `Stopping auto-nudge. Run /gsd to try again.`,
+      "error",
+    );
+    return true;
+  }
+
+  ctx.ui.notify(
+    `Milestone ${milestoneId}: "ready" signal rejected — CONTEXT.md and ROADMAP.md are missing. Asking the LLM to complete the writes.`,
+    "warning",
+  );
+
+  const nudge =
+    `You emitted "Milestone ${milestoneId} ready." but neither ` +
+    `.gsd/milestones/${milestoneId}/${milestoneId}-CONTEXT.md nor ` +
+    `.gsd/milestones/${milestoneId}/${milestoneId}-ROADMAP.md exists on disk. ` +
+    `The ready phrase is a POST-WRITE signal and has been rejected. ` +
+    `In this turn: (1) write PROJECT.md, REQUIREMENTS.md, and the milestone ` +
+    `CONTEXT.md, (2) call gsd_plan_milestone, then (3) emit the ready phrase. ` +
+    `Do not describe these steps — execute them as tool calls. ` +
+    `This is retry ${entry.readyRejectCount}/${MAX_READY_REJECTS}; further ` +
+    `premature signals will clear the session.`;
+
+  try {
+    pi.sendMessage(
+      { customType: "gsd-ready-no-files", content: nudge, display: false },
+      { triggerTurn: true },
+    );
+  } catch (e) {
+    logWarning("guided", `ready-phrase nudge sendMessage failed: ${(e as Error).message}`);
+    return false;
+  }
+  return true;
+}
+
+/**
+ * #4573 — Detect and recover from the "announces tool, never calls it" stall.
+ *
+ * The LLM emits text like "I'll now write the CONTEXT.md file" but the turn
+ * ends with zero tool-use blocks. The harness has no post-turn tool-call
+ * validation, so the unit promise resolves and the user sees a stalled state.
+ *
+ * This function, called from `handleAgentEnd`, inspects the last assistant
+ * message. If ALL of the following are true, it injects a recovery message:
+ *   - Text-only (no tool-use blocks)
+ *   - Contains a commit-intent phrase ("I'll write", "I'll call", etc.)
+ *   - Auto-mode is active OR a discussion autostart is pending
+ *   - `emptyTurnRetryCount` is under the cap
+ *
+ * Per-handler state is held on the `PendingAutoStartEntry` when present, and
+ * on a module-level map otherwise. The counter resets on any successful
+ * tool-use turn via `resetEmptyTurnCounter`.
+ */
+const emptyTurnCounterByBase = new Map<string, number>();
+const MAX_EMPTY_TURN_RETRIES = 2;
+
+// Phrases that indicate the LLM is about to do something but has not yet.
+// Kept tight to avoid flagging legitimate narration like "I'll wait for your answer."
+const COMMIT_INTENT_RE =
+  /\b(?:I['’]ll|I will|Next,? I['’]ll|Now I['’]ll|Let me|I['’]m going to|I am going to)\s+(?:now\s+)?(?:write|create|call|invoke|update|add|make|run|execute|generate|produce|emit|compose|implement|save|apply|commit)\b/i;
+
+/**
+ * Reset the empty-turn counter for a basePath after a successful tool-use turn.
+ * Called from handleAgentEnd when the last message contains tool_use blocks.
+ */
+export function resetEmptyTurnCounter(basePath?: string): void {
+  if (basePath) emptyTurnCounterByBase.delete(basePath);
+  else emptyTurnCounterByBase.clear();
+}
+
+export function maybeHandleEmptyIntentTurn(
+  event: { messages: any[] },
+  isAuto: boolean,
+): boolean {
+  // Gate: only fire when there is system-driven work in flight. Interactive
+  // /gsd discuss (user-driven) produces legitimate text-only turns.
+  if (!isAuto && pendingAutoStartMap.size === 0) return false;
+
+  const lastMsg = event.messages[event.messages.length - 1];
+  if (!lastMsg) return false;
+  if (hasToolUse(lastMsg)) return false;
+
+  const text = extractAssistantText(lastMsg).trim();
+  if (!text) return false;
+
+  // Skip if the LLM is emitting the ready phrase — that is the ready-no-files
+  // path, handled by maybeHandleReadyPhraseWithoutFiles.
+  if (READY_PHRASE_RE.test(text)) return false;
+
+  // Skip if the LLM is clearly handing back to the user. Keep the heuristic
+  // tight: a trailing question mark on the last non-empty line is the common
+  // signal for "I asked the user a question and stopped."
+  const lines = text.split(/\r?\n/).map((l) => l.trim()).filter(Boolean);
+  const lastLine = lines[lines.length - 1] ?? "";
+  if (lastLine.endsWith("?")) return false;
+
+  // Must contain a commit-intent phrase — this is the stall we care about.
+  if (!COMMIT_INTENT_RE.test(text)) return false;
+
+  // Resolve the target basePath + pi for injection. Prefer the pending
+  // autostart entry (discuss flow); otherwise we cannot inject.
+  const entry = _getPendingAutoStart();
+  if (!entry) return false;
+  const { ctx, pi, basePath } = entry;
+
+  const count = (emptyTurnCounterByBase.get(basePath) ?? 0) + 1;
+  emptyTurnCounterByBase.set(basePath, count);
+
+  if (count > MAX_EMPTY_TURN_RETRIES) {
+    ctx.ui.notify(
+      `Empty-turn recovery: LLM announced intent ${count} times without calling any tool. ` +
+      `Stopping auto-nudge.`,
+      "error",
+    );
+    return false; // let the normal flow resolve/pause the unit
+  }
+
+  ctx.ui.notify(
+    `Empty-turn detected: LLM announced intent but called no tool. Prompting it to execute.`,
+    "info",
+  );
+
+  const nudge =
+    `Your last turn announced an action (e.g. "I'll write…" or "Let me call…") ` +
+    `but contained no tool call. The system records zero tool-use blocks for ` +
+    `that turn. Execute the announced action NOW as a tool call in this turn. ` +
+    `Do not describe it again. Retry ${count}/${MAX_EMPTY_TURN_RETRIES}.`;
+
+  try {
+    pi.sendMessage(
+      { customType: "gsd-empty-turn-recovery", content: nudge, display: false },
+      { triggerTurn: true },
+    );
+  } catch (e) {
+    logWarning("guided", `empty-turn nudge sendMessage failed: ${(e as Error).message}`);
+    return false;
+  }
+  return true;
+}
+
 /**
  * Extract milestone IDs from PROJECT.md milestone sequence table.
  * Looks for rows like "| M001 | Name | Status |" and extracts the ID column.

package/src/resources/extensions/gsd/health-widget.ts

@@ -108,6 +108,7 @@ export function initHealthWidget(ctx: ExtensionContext): void {
     let data = initialData;
     let cachedLines: string[] | undefined;
     let refreshInFlight = false;
+    let isDisposed = false;
 
     const refresh = async () => {
       if (refreshInFlight) return;
@@ -115,7 +116,7 @@ export function initHealthWidget(ctx: ExtensionContext): void {
       try {
         data = loadHealthWidgetData(basePath);
         cachedLines = undefined;
-        _tui.requestRender();
+        if (!isDisposed) _tui.requestRender();
       } catch { /* non-fatal */ } finally {
         refreshInFlight = false;
       }
@@ -140,6 +141,7 @@ export function initHealthWidget(ctx: ExtensionContext): void {
       },
       invalidate(): void { cachedLines = undefined; cachedWidth = undefined; },
       dispose(): void {
+        isDisposed = true;
         clearInterval(refreshTimer);
       },
     };

package/src/resources/extensions/gsd/key-manager.ts

@@ -35,6 +35,12 @@ export interface ProviderInfo {
 export const PROVIDER_REGISTRY: ProviderInfo[] = [
   // LLM Providers
   { id: "anthropic",        label: "Anthropic (Claude)",      category: "llm", envVar: "ANTHROPIC_API_KEY",      prefixes: ["sk-ant-"], hasOAuth: true, dashboardUrl: "console.anthropic.com" },
+  // Claude Code CLI: routes through the local `claude` binary — no API key,
+  // authentication is handled by the CLI's own OAuth flow.
+  // Referenced by doctor-providers.ts, auto-model-selection.ts, and others;
+  // must be in the canonical registry so all consumers see the same catalog.
+  // See: https://github.com/gsd-build/gsd-2/issues/4541
+  { id: "claude-code",      label: "Claude Code CLI",         category: "llm",                                   hasOAuth: true },
   { id: "openai",           label: "OpenAI",                  category: "llm", envVar: "OPENAI_API_KEY",         prefixes: ["sk-"],     dashboardUrl: "platform.openai.com/api-keys" },
   { id: "github-copilot",   label: "GitHub Copilot",          category: "llm", envVar: "GITHUB_TOKEN",           hasOAuth: true },
   { id: "openai-codex",     label: "ChatGPT Plus/Pro (Codex)",category: "llm",                                   hasOAuth: true },

package/src/resources/extensions/gsd/model-router.ts

@@ -561,6 +561,23 @@ function bareModelId(modelId: string): string {
   return modelId.includes("/") ? modelId.split("/").pop()! : modelId;
 }
 
+// ─── Provider-specific Tool Limits ─────────────────────────────────────────
+
+/**
+ * Groq enforces a hard limit of 128 tools per request.
+ * Requests exceeding this limit receive a 400 error:
+ * "maximum number of items is 128"
+ * @see https://console.groq.com/docs/tool-use
+ */
+export const GROQ_MAX_TOOLS = 128;
+
+/**
+ * Provider IDs that map to the Groq API backend.
+ * Used to detect Groq at the GSD routing layer where only the provider string
+ * is available (the pi-ai openai-completions adapter is shared across providers).
+ */
+const GROQ_PROVIDER_IDS = new Set(["groq"]);
+
 // ─── Tool Compatibility Filter (ADR-005 Phase 3) ───────────────────────────
 
 /**
@@ -588,10 +605,17 @@ export function isToolCompatibleWithProvider(
 /**
  * Filter a list of tool names to only those compatible with a provider.
  * Used by the routing pipeline to adjust tool sets when switching providers.
+ *
+ * @param toolNames - The full list of active tool names to filter.
+ * @param providerApi - The pi-ai API string (e.g. "openai-completions").
+ * @param provider - Optional provider ID (e.g. "groq"). Used to apply
+ *   provider-specific limits that can't be expressed as API-level capabilities
+ *   (e.g. Groq's 128-tool hard limit on the shared openai-completions adapter).
  */
 export function filterToolsForProvider(
   toolNames: string[],
   providerApi: string,
+  provider?: string,
 ): { compatible: string[]; filtered: string[] } {
   const providerCaps = getProviderCapabilities(providerApi);
 
@@ -611,6 +635,17 @@ export function filterToolsForProvider(
     }
   }
 
+  // Groq enforces a hard limit of 128 tools per request (#4376).
+  // Trim the compatible list to GROQ_MAX_TOOLS and move the excess to filtered.
+  if (provider && GROQ_PROVIDER_IDS.has(provider) && compatible.length > GROQ_MAX_TOOLS) {
+    const trimmed = compatible.splice(GROQ_MAX_TOOLS);
+    filtered.push(...trimmed);
+    console.warn(
+      `[gsd] Groq tool limit: ${compatible.length + trimmed.length} tools active but Groq allows at most ${GROQ_MAX_TOOLS}. ` +
+        `Trimming to the first ${GROQ_MAX_TOOLS} tools. Removed: ${trimmed.join(", ")}`,
+    );
+  }
+
   return { compatible, filtered };
 }
 
@@ -620,11 +655,17 @@ export function filterToolsForProvider(
  *
  * This is a hard filter only — it removes tools that would fail at the
  * provider level. It does NOT remove tools based on soft heuristics.
+ *
+ * @param activeToolNames - The full list of currently active tool names.
+ * @param selectedModelApi - The pi-ai API string for the selected model.
+ * @param provider - Optional provider ID (e.g. "groq") for provider-specific
+ *   limits beyond what the API-level capability profile expresses.
  */
 export function adjustToolSet(
   activeToolNames: string[],
   selectedModelApi: string,
+  provider?: string,
 ): { toolNames: string[]; removedTools: string[] } {
-  const { compatible, filtered } = filterToolsForProvider(activeToolNames, selectedModelApi);
+  const { compatible, filtered } = filterToolsForProvider(activeToolNames, selectedModelApi, provider);
   return { toolNames: compatible, removedTools: filtered };
 }

package/src/resources/extensions/gsd/pre-execution-checks.ts

@@ -91,8 +91,13 @@ export function extractPackageReferences(description: string): string[] {
     }
   }
 
-  // require('pkg') or import from 'pkg' in code blocks
-  const importPattern = /(?:require\s*\(\s*['"]|from\s+['"])([a-zA-Z0-9@/_-]+)['"\)]/g;
+  // require('pkg') or `import ... from 'pkg'` in code blocks.
+  // The `from\s+['"]` branch MUST be preceded by an `import` keyword so that
+  // natural-language prose like `from "What's Next"` or `from 'master'` does
+  // not produce false package-existence failures.  Requiring the leading import
+  // keyword anchors the match to JavaScript/TypeScript syntax.
+  // See: https://github.com/gsd-build/gsd-2/issues/4388
+  const importPattern = /(?:require\s*\(\s*['"]|import\b[\s\S]*?\bfrom\s+['"])([a-zA-Z0-9@/_-]+)['"\)]/g;
   let importMatch: RegExpExecArray | null;
   while ((importMatch = importPattern.exec(description)) !== null) {
     // Skip relative imports and node builtins
@@ -325,7 +330,12 @@ function extractPathFromAnnotation(raw: string): string {
 
   const annotatedMatch = trimmed.match(/^(.+?)\s+[—–-]\s+.+$/);
   if (annotatedMatch) {
-    return annotatedMatch[1].trim();
+    const prefix = annotatedMatch[1].trim();
+    const prefixBacktickMatch = prefix.match(/`([^`]+)`/);
+    if (prefixBacktickMatch && looksLikePathOrUrl(prefixBacktickMatch[1].trim())) {
+      return prefixBacktickMatch[1].trim();
+    }
+    return prefix.replace(/`/g, "").trim();
   }
 
   // Fallback: scan all backticked tokens and return the first one that looks
@@ -388,13 +398,19 @@ function containsGlobPattern(candidate: string): boolean {
 
 /**
  * Build a set of files that will be created by tasks up to (but not including) taskIndex.
+ * Also includes outputs of completed tasks at any position — a completed task has already
+ * run and its outputs are available regardless of sequence position or disk state (#4071).
  * All paths are normalized for consistent comparison.
  */
 function getExpectedOutputsUpTo(tasks: TaskRow[], taskIndex: number): Set<string> {
   const outputs = new Set<string>();
-  for (let i = 0; i < taskIndex; i++) {
-    for (const file of tasks[i].expected_output) {
-      outputs.add(normalizeFilePath(file));
+  for (let i = 0; i < tasks.length; i++) {
+    const task = tasks[i];
+    // Include prior tasks (i < taskIndex) OR completed tasks at any position
+    if (i < taskIndex || task.status === "completed") {
+      for (const file of task.expected_output) {
+        outputs.add(normalizeFilePath(file));
+      }
     }
   }
   return outputs;
@@ -481,13 +497,19 @@ export function checkTaskOrdering(
   const results: PreExecutionCheckJSON[] = [];
 
   // Build map: normalized file → task index that creates it
-  const fileCreators = new Map<string, { taskId: string; index: number; originalPath: string }>();
+  const fileCreators = new Map<string, { taskId: string; index: number; originalPath: string; completed: boolean }>();
   for (let i = 0; i < tasks.length; i++) {
     const task = tasks[i];
     for (const file of task.expected_output) {
       const normalizedFile = normalizeFilePath(file);
-      if (!fileCreators.has(normalizedFile)) {
-        fileCreators.set(normalizedFile, { taskId: task.id, index: i, originalPath: file });
+      const existing = fileCreators.get(normalizedFile);
+      if (!existing || (!existing.completed && task.status === "completed")) {
+        fileCreators.set(normalizedFile, {
+          taskId: task.id,
+          index: i,
+          originalPath: file,
+          completed: task.status === "completed",
+        });
       }
     }
   }
@@ -511,7 +533,11 @@ export function checkTaskOrdering(
       const creator = fileCreators.get(normalizedFile);
       const absolutePath = resolve(basePath, normalizedFile);
       const existsOnDisk = existsSync(absolutePath);
-      if (creator && creator.index > i && !existsOnDisk) {
+      // Skip if the creating task has already completed — its output is available
+      // regardless of disk state (e.g. file was a temp artifact cleaned up after
+      // the task ran, or a replan introduced a new earlier-sequence task that
+      // reads this pre-execution output). (#4071)
+      if (creator && creator.index > i && !existsOnDisk && !creator.completed) {
         // Task reads file that is created later — impossible ordering
         results.push({
           category: "file",

package/src/resources/extensions/gsd/preferences-types.ts

@@ -28,6 +28,37 @@ export interface ContextManagementConfig {
   compaction_threshold_percent?: number;  // default: 0.70, range: 0.5-0.95
   tool_result_max_chars?: number;         // default: 800, range: 200-10000
 }
+
+/**
+ * Opt-in tool-output sandboxing for sub-sessions. When enabled, the gsd_exec
+ * MCP tool runs scripts in an isolated subprocess and returns only a short
+ * digest to the calling agent's context window; full stdout/stderr persist
+ * in the project memory store and can be retrieved by id later.
+ *
+ * Inspired by mksglu/context-mode (Elastic License 2.0). This is an
+ * independent implementation — no upstream code is incorporated.
+ */
+export interface ContextModeConfig {
+  /** Master switch. Default: true (opt-out via `enabled: false`). */
+  enabled?: boolean;
+  /** Per-invocation timeout in milliseconds. Default: 30_000. Range: 1_000–600_000. */
+  exec_timeout_ms?: number;
+  /** Cap on persisted stdout bytes per invocation. Default: 1_048_576 (1 MiB). Range: 4_096–16_777_216. */
+  exec_stdout_cap_bytes?: number;
+  /** Number of trailing stdout characters returned in the digest. Default: 300. Range: 0–4_000. */
+  exec_digest_chars?: number;
+  /** Environment variables forwarded to sandboxed processes (case-sensitive names). PATH and HOME are always forwarded. */
+  exec_env_allowlist?: string[];
+}
+
+/**
+ * Resolve whether context-mode features (gsd_exec sandbox + compaction
+ * snapshot) should be active. Default is ON: missing config or missing
+ * `enabled` is treated as true. Only `enabled: false` disables.
+ */
+export function isContextModeEnabled(prefs: { context_mode?: ContextModeConfig } | null | undefined): boolean {
+  return prefs?.context_mode?.enabled !== false;
+}
 import type { GitHubSyncConfig } from "../github-sync/types.js";
 
 // ─── Workflow Modes ──────────────────────────────────────────────────────────
@@ -117,6 +148,7 @@ export const KNOWN_PREFERENCE_KEYS = new Set<string>([
   "flat_rate_providers",
   "language",
   "context_window_override",
+  "context_mode",
 ]);
 
 /** Canonical list of all dispatch unit types. */
@@ -300,6 +332,12 @@ export interface GSDPreferences {
    */
   context_window_override?: number;
   context_management?: ContextManagementConfig;
+  /**
+   * Tool-output sandboxing via gsd_exec. Keeps sub-session context windows
+   * clean by running scripts in a subprocess and only surfacing a short
+   * digest. See `ContextModeConfig`. Default: disabled.
+   */
+  context_mode?: ContextModeConfig;
   token_profile?: TokenProfile;
   phases?: PhaseSkipPreferences;
   auto_visualize?: boolean;

package/src/resources/extensions/gsd/preferences-validation.ts

@@ -644,6 +644,50 @@ export function validatePreferences(preferences: GSDPreferences): {
     }
   }
 
+  // ─── Context Mode (gsd_exec sandbox) ────────────────────────────────────
+  if (preferences.context_mode !== undefined) {
+    if (typeof preferences.context_mode === "object" && preferences.context_mode !== null) {
+      const cmode = preferences.context_mode as unknown as Record<string, unknown>;
+      const validCmode: Record<string, unknown> = {};
+
+      if (cmode.enabled !== undefined) {
+        if (typeof cmode.enabled === "boolean") validCmode.enabled = cmode.enabled;
+        else errors.push("context_mode.enabled must be a boolean");
+      }
+      if (cmode.exec_timeout_ms !== undefined) {
+        const t = cmode.exec_timeout_ms;
+        if (typeof t === "number" && t >= 1000 && t <= 600_000) validCmode.exec_timeout_ms = Math.floor(t);
+        else errors.push("context_mode.exec_timeout_ms must be a number between 1000 and 600000");
+      }
+      if (cmode.exec_stdout_cap_bytes !== undefined) {
+        const b = cmode.exec_stdout_cap_bytes;
+        if (typeof b === "number" && b >= 4096 && b <= 16_777_216) validCmode.exec_stdout_cap_bytes = Math.floor(b);
+        else errors.push("context_mode.exec_stdout_cap_bytes must be a number between 4096 and 16777216");
+      }
+      if (cmode.exec_digest_chars !== undefined) {
+        const c = cmode.exec_digest_chars;
+        if (typeof c === "number" && c >= 0 && c <= 4000) validCmode.exec_digest_chars = Math.floor(c);
+        else errors.push("context_mode.exec_digest_chars must be a number between 0 and 4000");
+      }
+      if (cmode.exec_env_allowlist !== undefined) {
+        if (
+          Array.isArray(cmode.exec_env_allowlist) &&
+          cmode.exec_env_allowlist.every((v) => typeof v === "string" && /^[A-Z_][A-Z0-9_]*$/i.test(v))
+        ) {
+          validCmode.exec_env_allowlist = cmode.exec_env_allowlist;
+        } else {
+          errors.push("context_mode.exec_env_allowlist must be an array of valid env var names");
+        }
+      }
+
+      if (Object.keys(validCmode).length > 0) {
+        validated.context_mode = validCmode as any;
+      }
+    } else {
+      errors.push("context_mode must be an object");
+    }
+  }
+
   // ─── Parallel Config ────────────────────────────────────────────────────
   if (preferences.parallel && typeof preferences.parallel === "object") {
     const p = preferences.parallel as unknown as Record<string, unknown>;
@@ -697,6 +741,41 @@ export function validatePreferences(preferences: GSDPreferences): {
     }
   }
 
+  // ─── Slice Parallel Config ───────────────────────────────────────────────
+  if (preferences.slice_parallel !== undefined) {
+    if (typeof preferences.slice_parallel === "object" && preferences.slice_parallel !== null) {
+      const sp = preferences.slice_parallel as Record<string, unknown>;
+      const validSp: NonNullable<GSDPreferences["slice_parallel"]> = {};
+
+      if (sp.enabled !== undefined) {
+        if (typeof sp.enabled === "boolean") validSp.enabled = sp.enabled;
+        else errors.push("slice_parallel.enabled must be a boolean");
+      }
+
+      if (sp.max_workers !== undefined) {
+        const maxWorkers = typeof sp.max_workers === "number" ? sp.max_workers : Number(sp.max_workers);
+        if (Number.isFinite(maxWorkers) && maxWorkers >= 1 && maxWorkers <= 8) {
+          validSp.max_workers = Math.floor(maxWorkers);
+        } else {
+          errors.push("slice_parallel.max_workers must be a number between 1 and 8");
+        }
+      }
+
+      const knownSliceParallelKeys = new Set(["enabled", "max_workers"]);
+      for (const key of Object.keys(sp)) {
+        if (!knownSliceParallelKeys.has(key)) {
+          warnings.push(`unknown slice_parallel key "${key}" — ignored`);
+        }
+      }
+
+      if (Object.keys(validSp).length > 0) {
+        validated.slice_parallel = validSp;
+      }
+    } else {
+      errors.push("slice_parallel must be an object");
+    }
+  }
+
   // ─── Reactive Execution ─────────────────────────────────────────────────
   if (preferences.reactive_execution !== undefined) {
     if (typeof preferences.reactive_execution === "object" && preferences.reactive_execution !== null) {