gsd-pi 2.76.0-dev.fe143342a → 2.77.0-dev.1d17f366c

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

@@ -13,7 +13,7 @@ import { loadFile, saveFile } from "./files.js";
 import { isDbAvailable, getMilestoneSlices } from "./gsd-db.js";
 import { parseRoadmapSlices } from "./roadmap-slices.js";
 import { loadPrompt, inlineTemplate } from "./prompt-loader.js";
-import { buildSkillActivationBlock } from "./auto-prompts.js";
+import { buildDiscussMilestonePrompt, buildSkillActivationBlock } from "./auto-prompts.js";
 import { deriveState } from "./state.js";
 import { invalidateAllCaches } from "./cache.js";
 import { startAutoDetached } from "./auto.js";
@@ -38,7 +38,7 @@ import { isInheritedRepo } from "./repo-identity.js";
 import { ensureGitignore, ensurePreferences, untrackRuntimeFiles } from "./gitignore.js";
 import { loadEffectiveGSDPreferences } from "./preferences.js";
 import { resolveUokFlags } from "./uok/flags.js";
-import { ensurePlanV2Graph } from "./uok/plan-v2.js";
+import { ensurePlanV2Graph, isMissingFinalizedContextResult } from "./uok/plan-v2.js";
 import { detectProjectState } from "./detection.js";
 import { showProjectInit, offerMigration } from "./init-wizard.js";
 import { validateDirectory } from "./validate-directory.js";
@@ -93,24 +93,29 @@ function needsPlanV2Gate(state: GSDState): boolean {
     || state.phase === "completing-milestone";
 }
 
+type PlanV2GateDecision = "pass" | "recover-missing-context" | "block";
+
 function runPlanV2Gate(
   ctx: ExtensionContext,
   basePath: string,
   state: GSDState,
-): boolean {
+): PlanV2GateDecision {
   const prefs = loadEffectiveGSDPreferences()?.preferences;
   const uokFlags = resolveUokFlags(prefs);
-  if (!uokFlags.planV2 || !needsPlanV2Gate(state)) return true;
+  if (!uokFlags.planV2 || !needsPlanV2Gate(state)) return "pass";
   const compiled = ensurePlanV2Graph(basePath, state);
   if (!compiled.ok) {
+    if (isMissingFinalizedContextResult(compiled)) {
+      return "recover-missing-context";
+    }
     const reason = compiled.reason ?? "plan-v2 compilation failed";
     ctx.ui.notify(
-      `Plan gate failed-closed: ${reason}. Complete plan/discuss artifacts before execution.`,
+      `Plan gate failed-closed: ${reason}. Complete plan/discuss artifacts before execution.\n\nIf this keeps happening, try: /gsd doctor heal`,
       "error",
     );
-    return false;
+    return "block";
   }
-  return true;
+  return "pass";
 }
 
 // ─── Commit Instruction Helpers ──────────────────────────────────────────────
@@ -130,8 +135,20 @@ 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 +296,215 @@ 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.
@@ -604,8 +830,13 @@ export async function showHeadlessMilestoneCreation(
   // Set pending auto start (auto-mode triggers on "Milestone X ready." via checkAutoStartAfterDiscuss)
   pendingAutoStartMap.set(basePath, { ctx, pi, basePath, milestoneId: nextId, createdAt: Date.now() });
 
-  // Dispatch — headless milestone creation is a planning activity
-  await dispatchWorkflow(pi, prompt, "gsd-run", ctx, "plan-milestone");
+  // Dispatch as discuss-milestone. The LLM writes PROJECT.md, REQUIREMENTS.md,
+  // and CONTEXT.md, then calls gsd_plan_milestone — this is semantically the
+  // discuss path, just non-interactive. Using "plan-milestone" here caused
+  // model/tool routing to skip discuss-flow tool scoping and
+  // `checkAutoStartAfterDiscuss` guardrails that rely on the
+  // "discuss-"-prefixed unitType.
+  await dispatchWorkflow(pi, prompt, "gsd-run", ctx, "discuss-milestone");
 }
 
 
@@ -1352,7 +1583,8 @@ export async function showSmartEntry(
     logWarning("guided", `STATE.md rebuild failed: ${(err as Error).message}`);
   }
 
-  if (!runPlanV2Gate(ctx, basePath, state)) return;
+  const planV2GateDecision = runPlanV2Gate(ctx, basePath, state);
+  if (planV2GateDecision === "block") return;
 
   if (!state.activeMilestone?.id) {
     // Guard: if a discuss session is already in flight, don't re-inject the prompt.
@@ -1440,6 +1672,23 @@ export async function showSmartEntry(
   const milestoneId = state.activeMilestone.id;
   const milestoneTitle = state.activeMilestone.title;
 
+  if (planV2GateDecision === "recover-missing-context") {
+    pendingAutoStartMap.set(basePath, { ctx, pi, basePath, milestoneId, step: stepMode, createdAt: Date.now() });
+    await dispatchWorkflow(
+      pi,
+      await buildDiscussMilestonePrompt(
+        milestoneId,
+        milestoneTitle,
+        basePath,
+        getStructuredQuestionsAvailability(pi, ctx),
+      ),
+      "gsd-discuss",
+      ctx,
+      "discuss-milestone",
+    );
+    return;
+  }
+
   // ── All milestones complete → New milestone ──────────────────────────
   if (state.phase === "complete") {
     const choice = await showNextAction(ctx, {

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/journal.ts

@@ -12,8 +12,18 @@
  * - Silent failure: journal writes never throw — absence of events is the failure signal
  */
 
-import { appendFileSync, mkdirSync, readdirSync, readFileSync } from "node:fs";
+import {
+  appendFileSync,
+  closeSync,
+  existsSync,
+  mkdirSync,
+  openSync,
+  readdirSync,
+  readFileSync,
+} from "node:fs";
 import { join } from "node:path";
+import { isStaleWrite } from "./auto/turn-epoch.js";
+import { withFileLockSync } from "./file-lock.js";
 import { gsdRoot } from "./paths.js";
 import { buildAuditEnvelope, emitUokAuditEvent } from "./uok/audit.js";
 import { isUnifiedAuditEnabled } from "./uok/audit-toggle.js";
@@ -39,7 +49,8 @@ export type JournalEventType =
   | "worktree-create-failed"
   | "worktree-skip"
   | "worktree-merge-start"
-  | "worktree-merge-failed";
+  | "worktree-merge-failed"
+  | "artifact-verification-retry";
 
 /** A single structured event in the journal. */
 export interface JournalEntry {
@@ -83,12 +94,27 @@ export interface JournalQueryFilters {
  * Never throws — all errors are silently caught.
  */
 export function emitJournalEvent(basePath: string, entry: JournalEntry): void {
+  // Drop writes from a turn superseded by timeout recovery / cancellation.
+  // See auto/turn-epoch.ts for the full rationale.
+  if (isStaleWrite("journal")) return;
   try {
     const journalDir = join(gsdRoot(basePath), "journal");
     mkdirSync(journalDir, { recursive: true });
     const dateStr = entry.ts.slice(0, 10);
     const filePath = join(journalDir, `${dateStr}.jsonl`);
-    appendFileSync(filePath, JSON.stringify(entry) + "\n");
+    // Ensure file exists so proper-lockfile can acquire a lock against it.
+    if (!existsSync(filePath)) closeSync(openSync(filePath, "a"));
+    // onLocked: "skip" — journal writes are best-effort. POSIX O_APPEND
+    // atomicity still protects small entries; the lock mainly serializes
+    // larger writes and gives cross-process exclusivity on platforms where
+    // O_APPEND semantics are weaker (Windows).
+    withFileLockSync(
+      filePath,
+      () => {
+        appendFileSync(filePath, JSON.stringify(entry) + "\n");
+      },
+      { onLocked: "skip" },
+    );
   } catch {
     // Silent failure — journal must never break auto-mode
   }

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

@@ -51,6 +51,8 @@ export const PROVIDER_REGISTRY: ProviderInfo[] = [
   { id: "xai",              label: "xAI (Grok)",              category: "llm", envVar: "XAI_API_KEY",            dashboardUrl: "console.x.ai" },
   { id: "openrouter",       label: "OpenRouter",              category: "llm", envVar: "OPENROUTER_API_KEY",     dashboardUrl: "openrouter.ai/keys" },
   { id: "mistral",          label: "Mistral",                 category: "llm", envVar: "MISTRAL_API_KEY",        dashboardUrl: "console.mistral.ai" },
+  { id: "minimax",          label: "MiniMax",                 category: "llm", envVar: "MINIMAX_API_KEY",        dashboardUrl: "platform.minimax.io" },
+  { id: "minimax-cn",       label: "MiniMax CN",              category: "llm", envVar: "MINIMAX_CN_API_KEY",     dashboardUrl: "platform.minimax.io" },
   { id: "ollama-cloud",     label: "Ollama Cloud",            category: "llm", envVar: "OLLAMA_API_KEY" },
   { id: "custom-openai",    label: "Custom (OpenAI-compat)",  category: "llm", envVar: "CUSTOM_OPENAI_API_KEY" },
   { id: "cerebras",         label: "Cerebras",                category: "llm", envVar: "CEREBRAS_API_KEY" },
@@ -486,6 +488,26 @@ const TEST_ENDPOINTS: Record<string, { url: string; method?: string; headers?: (
     url: "https://api.mistral.ai/v1/models",
     headers: (key) => ({ Authorization: `Bearer ${key}` }),
   },
+  minimax: {
+    url: "https://api.minimax.io/anthropic/v1/messages",
+    method: "POST",
+    headers: (key) => ({
+      "x-api-key": key,
+      "anthropic-version": "2023-06-01",
+      "content-type": "application/json",
+    }),
+    body: JSON.stringify({ model: "MiniMax-M2.7", max_tokens: 1, messages: [{ role: "user", content: "hi" }] }),
+  },
+  "minimax-cn": {
+    url: "https://api.minimaxi.com/anthropic/v1/messages",
+    method: "POST",
+    headers: (key) => ({
+      "x-api-key": key,
+      "anthropic-version": "2023-06-01",
+      "content-type": "application/json",
+    }),
+    body: JSON.stringify({ model: "MiniMax-M2.7", max_tokens: 1, messages: [{ role: "user", content: "hi" }] }),
+  },
   openrouter: {
     url: "https://openrouter.ai/api/v1/models",
     headers: (key) => ({ Authorization: `Bearer ${key}` }),

package/src/resources/extensions/gsd/milestone-actions.ts

@@ -23,6 +23,21 @@ import { loadQueueOrder, saveQueueOrder } from "./queue-order.js";
 import { deleteMilestone, getMilestone, isDbAvailable, updateMilestoneStatus } from "./gsd-db.js";
 import { removeWorktree } from "./worktree-manager.js";
 import { logWarning } from "./workflow-logger.js";
+import { isAutoActive } from "./auto.js";
+
+/**
+ * Writer-side assert for mutations that race with auto-mode's squash merge (#4704).
+ * Auto-mode is confirmed not to call parkMilestone/discardMilestone/unparkMilestone
+ * internally — these throws only surface invariant violations from new or forgotten
+ * call sites, which is the correct failure mode to catch loudly.
+ */
+function assertNotAutoActive(action: string): void {
+  if (isAutoActive()) {
+    throw new Error(
+      `${action} cannot run while auto-mode is active. Stop auto-mode first with /gsd stop.`,
+    );
+  }
+}
 
 // ─── Park ──────────────────────────────────────────────────────────────────
 
@@ -32,6 +47,7 @@ import { logWarning } from "./workflow-logger.js";
  * Returns true if successfully parked, false if milestone not found or already parked.
  */
 export function parkMilestone(basePath: string, milestoneId: string, reason: string): boolean {
+  assertNotAutoActive("park milestone");
   const mDir = resolveMilestonePath(basePath, milestoneId);
   if (!mDir || !existsSync(mDir)) return false;
 
@@ -74,6 +90,7 @@ export function parkMilestone(basePath: string, milestoneId: string, reason: str
  * Returns true if successfully unparked, false if milestone not found or not parked.
  */
 export function unparkMilestone(basePath: string, milestoneId: string): boolean {
+  assertNotAutoActive("unpark milestone");
   const mDir = resolveMilestonePath(basePath, milestoneId);
   if (!mDir || !existsSync(mDir)) return false;
 
@@ -108,6 +125,7 @@ export function unparkMilestone(basePath: string, milestoneId: string): boolean
  * Returns true if successfully discarded, false if milestone not found.
  */
 export function discardMilestone(basePath: string, milestoneId: string): boolean {
+  assertNotAutoActive("discard milestone");
   const mDir = resolveMilestonePath(basePath, milestoneId);
   if (!mDir || !existsSync(mDir)) return false;
 

package/src/resources/extensions/gsd/milestone-summary-classifier.ts

@@ -0,0 +1,42 @@
+/**
+ * Shared milestone SUMMARY classifier.
+ *
+ * SUMMARY presence alone is not enough to prove milestone completion: recovery
+ * and blocker paths also write SUMMARY files. Keep this leaf module free of
+ * state/auto imports so state derivation, dispatch guards, and recovery can
+ * share one definition without cycles.
+ */
+
+import { splitFrontmatter, parseFrontmatterMap } from "../shared/frontmatter.js";
+import { isClosedStatus } from "./status-guards.js";
+
+export type MilestoneSummaryOutcome = "success" | "failure" | "unknown";
+
+export function classifyMilestoneSummaryContent(content: string): MilestoneSummaryOutcome {
+  const [fmLines] = splitFrontmatter(content);
+  const fm = fmLines ? parseFrontmatterMap(fmLines) : null;
+  const rawStatus = typeof fm?.status === "string" ? fm.status.trim().toLowerCase() : "";
+  if (rawStatus) {
+    if (isClosedStatus(rawStatus)) return "success";
+    if (["active", "pending", "blocked", "failed", "failure", "incomplete"].includes(rawStatus)) {
+      return "failure";
+    }
+  }
+
+  const failureSignal =
+    /(?:^|\n)\s*#\s*BLOCKER\b/i.test(content)
+    || /auto-mode recovery failed/i.test(content)
+    || /verification\s+failed/i.test(content)
+    || /(?:^|\n)\s*(?:status|verdict|outcome|result)\s*[:=-]\s*not complete\b/i.test(content);
+  if (failureSignal) return "failure";
+  return "unknown";
+}
+
+/**
+ * Legacy-compatible terminal check for state derivation.
+ * Unknown summaries remain terminal to preserve old handwritten SUMMARY files;
+ * explicit failure summaries do not.
+ */
+export function isTerminalMilestoneSummaryContent(content: string): boolean {
+  return classifyMilestoneSummaryContent(content) !== "failure";
+}

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 };
 }