@opengsd/gsd-pi 1.3.0-dev.65546769 → 1.3.0-dev.72e3af2a

package/src/resources/extensions/gsd/bootstrap/agent-end-recovery.ts

@@ -1,6 +1,8 @@
 // gsd-pi + src/resources/extensions/gsd/bootstrap/agent-end-recovery.ts - Handles provider and agent-end recovery for GSD auto-mode.
 
 import type { ExtensionAPI, ExtensionContext } from "@gsd/pi-coding-agent";
+import { mkdirSync, readdirSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
 
 import type { AgentEndEvent, ErrorContext } from "../auto/types.js";
 import { logWarning } from "../workflow-logger.js";
@@ -11,7 +13,7 @@ import {
   maybeHandleEmptyIntentTurn,
   resetEmptyTurnCounter,
 } from "../guided-flow.js";
-import { clearPathCache } from "../paths.js";
+import { clearPathCache, gsdRoot } from "../paths.js";
 import {
   getAutoDashboardData,
   getAutoModeStartModel,
@@ -34,7 +36,7 @@ import { resolveModelId } from "../auto-model-selection.js";
 import { resolveProjectRoot } from "../worktree.js";
 import { clearDiscussionFlowState } from "./write-gate.js";
 import { scheduleFallbackContinuation } from "./fallback-continuation.js";
-import { clearGuidedUnitContext } from "../guided-unit-context.js";
+import { clearGuidedUnitContext, getGuidedUnitContext, type GuidedUnitContext } from "../guided-unit-context.js";
 import { resumeAutoAfterProviderDelay } from "./provider-error-resume.js";
 import {
   classifyError,
@@ -116,9 +118,14 @@ export function isUserInitiatedAbortMessage(message: string | undefined | null):
 export function shouldDeferTransientErrorToCoreRetry(
   cls: ErrorClass,
   rawErrorMsg: string,
+  deferCheckMsg: string = rawErrorMsg,
 ): boolean {
   if (!isTransient(cls) || cls.kind === "rate-limit") return false;
-  return !/retry failed after \d+ attempts:/i.test(rawErrorMsg);
+  // Empty rawErrorMsg means the SDK terminated the session without providing an
+  // error string — core is done, not mid-retry.  GSD must schedule its own
+  // retry rather than silently deferring to a core that has already exited.
+  if (!rawErrorMsg) return false;
+  return !/retry failed after \d+ attempts:/i.test(deferCheckMsg);
 }
 
 type ProviderModelFallbackParams = {
@@ -388,6 +395,108 @@ export function suppressTerminalDeletedWorktreeMessageEnd(event: MessageEndLike)
   return true;
 }
 
+function modelLabel(ctx: ExtensionContext): string {
+  const provider = ctx.model?.provider;
+  const id = ctx.model?.id;
+  return provider && id ? `${provider}/${id}` : "unknown model";
+}
+
+function isFatalManualGuidedTerminalFailure(lastMsg: unknown): boolean {
+  if (!isObjectRecord(lastMsg) || !("stopReason" in lastMsg)) return false;
+  if (lastMsg.stopReason === "error") {
+    const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+    if (isUserInitiatedAbortMessage(rawErrorMsg)) return false;
+    return true;
+  }
+  if (lastMsg.stopReason !== "aborted") return false;
+
+  const content = "content" in lastMsg ? lastMsg.content : undefined;
+  const hasErrorMessage = "errorMessage" in lastMsg && !!lastMsg.errorMessage;
+  return hasErrorMessage || !_hasEmptyAgentEndContent(content);
+}
+
+function terminalFailureDetail(lastMsg: unknown): string {
+  if (!isObjectRecord(lastMsg)) return "Provider turn ended with an unknown terminal error.";
+  const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+  const content = "content" in lastMsg ? lastMsg.content : undefined;
+  const displayMsg = resolveAgentEndErrorDisplay(rawErrorMsg, content).replace(/\s+/g, " ").trim();
+  if (displayMsg) return displayMsg.length > 300 ? `${displayMsg.slice(0, 300)}...` : displayMsg;
+  return lastMsg.stopReason === "aborted"
+    ? "Provider turn aborted with error context."
+    : "Provider stream ended with stopReason=error.";
+}
+
+function nextManualActivitySequence(activityDir: string): string {
+  let maxSeq = 0;
+  try {
+    for (const file of readdirSync(activityDir)) {
+      const match = /^(\d+)-/.exec(file);
+      if (match) maxSeq = Math.max(maxSeq, Number.parseInt(match[1]!, 10));
+    }
+  } catch {
+    return "001";
+  }
+  return String(maxSeq + 1).padStart(3, "0");
+}
+
+function writeManualGuidedTerminalErrorActivity(
+  basePath: string,
+  unitType: string,
+  model: string,
+  detail: string,
+  stopReason: unknown,
+): void {
+  const activityDir = join(gsdRoot(basePath), "activity");
+  mkdirSync(activityDir, { recursive: true });
+  const seq = nextManualActivitySequence(activityDir);
+  const safeUnitType = unitType.replace(/[^a-z0-9_.-]+/gi, "-");
+  const markerPath = join(activityDir, `${seq}-${safeUnitType}-manual-terminal-provider-error.jsonl`);
+  const message = `Manual guided ${unitType} turn ended with provider ${String(stopReason)} on ${model}: ${detail}`;
+  writeFileSync(
+    markerPath,
+    JSON.stringify({
+      type: "message",
+      message: {
+        role: "toolResult",
+        toolCallId: "manual-guided-terminal-provider-error",
+        toolName: "provider",
+        isError: true,
+        content: [{ type: "text", text: message }],
+      },
+    }) + "\n",
+    "utf-8",
+  );
+}
+
+function observeManualDiscussTerminalError(
+  ctx: ExtensionContext,
+  lastMsg: unknown,
+  guidedUnit: GuidedUnitContext | null,
+): void {
+  if (!guidedUnit?.unitType.startsWith("discuss-")) return;
+  if (!isFatalManualGuidedTerminalFailure(lastMsg)) return;
+
+  const model = modelLabel(ctx);
+  const detail = terminalFailureDetail(lastMsg);
+  ctx.ui.notify(
+    `Manual /gsd discuss ${guidedUnit.unitType} ended with a provider error on ${model}: ${detail}`,
+    "warning",
+  );
+
+  try {
+    writeManualGuidedTerminalErrorActivity(
+      guidedUnit.basePath,
+      guidedUnit.unitType,
+      model,
+      detail,
+      isObjectRecord(lastMsg) ? lastMsg.stopReason : "unknown",
+    );
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err);
+    logWarning("bootstrap", `Failed to write manual guided terminal-error activity marker: ${message}`);
+  }
+}
+
 async function pauseTransientWithBackoff(
   cls: ErrorClass,
   pi: ExtensionAPI,
@@ -437,7 +546,9 @@ export async function handleAgentEnd(
   // rejected" loop even though the files are on disk.
   clearPathCache();
   const basePath = resolveAgentEndBasePath();
-  clearGuidedUnitContext(basePath);
+  const lastMsg = event.messages[event.messages.length - 1];
+  const guidedUnit = basePath ? getGuidedUnitContext(basePath) ?? getGuidedUnitContext() : getGuidedUnitContext();
+  clearGuidedUnitContext(guidedUnit?.basePath ?? basePath);
 
   try {
     if (await checkDeepProjectSetupAfterTurn(event, ctx, basePath)) {
@@ -466,13 +577,15 @@ export async function handleAgentEnd(
   // discussions (where isAutoActive may be false) still get recovered.
   if (maybeHandleEmptyIntentTurn(event, isAutoActive(), basePath)) return;
 
-  if (!isAutoActive()) return;
+  if (!isAutoActive()) {
+    observeManualDiscussTerminalError(ctx, lastMsg, guidedUnit);
+    return;
+  }
 
   if (shouldIgnoreAgentEndForActiveUnit(event)) {
     return;
   }
 
-  const lastMsg = event.messages[event.messages.length - 1];
   if (isSessionSwitchInFlight()) {
     _handleSessionSwitchAgentEnd(lastMsg, resolveAgentEndCancelled);
     return;
@@ -555,9 +668,9 @@ export async function handleAgentEnd(
       });
       return;
     }
-    // #3588: When errorMessage is uninformative, extract the real error from
-    // the assistant message text content for display purposes only.
-    // Classification still uses rawErrorMsg to avoid false positives from prose.
+    // #3588/#956: When errorMessage is uninformative, extract the real error
+    // from assistant text. Prefer rawErrorMsg for classification to avoid
+    // prose false-positives, but use display text when rawErrorMsg is empty.
     const displayMsg = resolveAgentEndErrorDisplay(
       rawErrorMsg,
       "content" in lastMsg ? lastMsg.content : undefined,
@@ -576,8 +689,8 @@ export async function handleAgentEnd(
     const errorDetail = displayMsg ? `: ${displayMsg}` : "";
     const explicitRetryAfterMs = ("retryAfterMs" in lastMsg && typeof lastMsg.retryAfterMs === "number") ? lastMsg.retryAfterMs : undefined;
 
-    // ── 1. Classify using rawErrorMsg to avoid prose false-positives ────
-    const cls = classifyError(rawErrorMsg, explicitRetryAfterMs);
+    // ── 1. Classify, preserving non-empty errorMessage precedence ──────
+    const cls = classifyError(rawErrorMsg || displayMsg, explicitRetryAfterMs);
 
     // ── 1a. Unsupported-model: provider rejected this model for the current
     //        account/plan at request time (#4513).  Persist a block so the
@@ -655,7 +768,7 @@ export async function handleAgentEnd(
     // Core retries transient failures in-session after this handler.
     // Keep that behavior for non-rate-limit classes to avoid pause/retry races,
     // but let rate-limit continue into model fallback logic below (#4373).
-    if (shouldDeferTransientErrorToCoreRetry(cls, rawErrorMsg)) {
+    if (shouldDeferTransientErrorToCoreRetry(cls, rawErrorMsg, rawErrorMsg || displayMsg)) {
       return;
     }
 

package/src/resources/extensions/gsd/bootstrap/core-session-tools.ts

@@ -0,0 +1,43 @@
+// Project/App: gsd-pi
+// File Purpose: Canonical core session tool surface shared by auto-mode scoping and loop guards.
+
+/** Base tools available during auto-mode execution before unit-specific GSD tools are added. */
+export const MINIMAL_AUTO_BASE_TOOL_NAMES = [
+  "ask_user_questions",
+  "bash",
+  "bg_shell",
+  "edit",
+  "find",
+  "glob",
+  "grep",
+  "fetch_page",
+  "search-the-web",
+  "ls",
+  "read",
+  "subagent",
+  "write",
+  "ToolSearch",
+] as const;
+
+/** Tools excluded from the high per-turn loop-guard cap (strict or orchestration). */
+const NON_REPEATABLE_FROM_MINIMAL_BASE = new Set<string>([
+  "ask_user_questions",
+  "subagent",
+  "ToolSearch",
+]);
+
+/** Additional core tools not in MINIMAL_AUTO_BASE but routinely multi-called per turn. */
+const EXTRA_INHERENTLY_REPEATABLE = [
+  "multi_edit",
+  "todo_write",
+  "notebook_edit",
+  "search_and_read",
+] as const;
+
+/** Core session tools that may be invoked many times per agent turn. */
+export const INHERENTLY_REPEATABLE_TOOL_NAMES = [
+  ...MINIMAL_AUTO_BASE_TOOL_NAMES.filter((name) => !NON_REPEATABLE_FROM_MINIMAL_BASE.has(name)),
+  ...EXTRA_INHERENTLY_REPEATABLE,
+] as const;
+
+export const INHERENTLY_REPEATABLE_TOOL_SET = new Set<string>(INHERENTLY_REPEATABLE_TOOL_NAMES);

package/src/resources/extensions/gsd/bootstrap/db-tools.ts

@@ -3,6 +3,7 @@
 import { Type, StringEnum } from "@gsd/pi-ai";
 import type { ExtensionAPI } from "@gsd/pi-coding-agent";
 import { Text } from "@gsd/pi-tui";
+import { SUMMARY_SAVE_CONTENT_MAX_LENGTH } from "@opengsd/contracts";
 
 import { loadEffectiveGSDPreferences } from "../preferences.js";
 import { ensureDbOpen, resolveCtxCwd, resolveWorkflowToolBasePath } from "./dynamic-tools.js";
@@ -392,13 +393,17 @@ export function registerDbTools(pi: ExtensionAPI): void {
       "Root-level artifact paths are PROJECT.md, PROJECT-DRAFT.md, REQUIREMENTS.md, and REQUIREMENTS-DRAFT.md.",
       "artifact_type must be one of: SUMMARY, RESEARCH, CONTEXT, ASSESSMENT, CONTEXT-DRAFT, PROJECT, PROJECT-DRAFT, REQUIREMENTS, REQUIREMENTS-DRAFT.",
       "Use CONTEXT-DRAFT for incremental draft persistence; use CONTEXT for the final milestone context after depth verification.",
+      `Keep each content payload under ${SUMMARY_SAVE_CONTENT_MAX_LENGTH} characters; save large context incrementally with CONTEXT-DRAFT/PROJECT-DRAFT/REQUIREMENTS-DRAFT instead of one oversized call.`,
     ],
     parameters: Type.Object({
       milestone_id: Type.Optional(Type.String({ description: "Milestone ID (e.g. M001). Omit only for root-level PROJECT/PROJECT-DRAFT/REQUIREMENTS/REQUIREMENTS-DRAFT artifacts." })),
       slice_id: Type.Optional(Type.String({ description: "Slice ID (e.g. S01)" })),
       task_id: Type.Optional(Type.String({ description: "Task ID (e.g. T01)" })),
       artifact_type: StringEnum(["SUMMARY", "RESEARCH", "CONTEXT", "ASSESSMENT", "CONTEXT-DRAFT", "PROJECT", "PROJECT-DRAFT", "REQUIREMENTS", "REQUIREMENTS-DRAFT"], { description: "Artifact type to save" }),
-      content: Type.String({ description: "The full markdown content of the artifact" }),
+      content: Type.String({
+        description: `The full markdown content of the artifact. Maximum ${SUMMARY_SAVE_CONTENT_MAX_LENGTH} characters per save.`,
+        maxLength: SUMMARY_SAVE_CONTENT_MAX_LENGTH,
+      }),
     }),
     execute: summarySaveExecute,
     renderCall(args: any, theme: any) {

package/src/resources/extensions/gsd/bootstrap/exec-tools.ts

@@ -72,6 +72,7 @@ export function registerExecTools(pi: ExtensionAPI): void {
       return executeUatExec(params as Parameters<typeof executeUatExec>[0], {
         baseDir,
         preferences: await loadContextModePreferences(baseDir),
+        signal: _signal,
       });
     },
   });
@@ -119,6 +120,7 @@ export function registerExecTools(pi: ExtensionAPI): void {
       return executeGsdExec(params as Parameters<typeof executeGsdExec>[0], {
         baseDir,
         preferences: await loadContextModePreferences(baseDir),
+        signal: _signal,
       });
     },
   });

package/src/resources/extensions/gsd/bootstrap/register-hooks.ts

@@ -13,7 +13,7 @@ import type { GSDEcosystemBeforeAgentStartHandler } from "../ecosystem/gsd-exten
 import { updateSnapshot } from "../ecosystem/gsd-extension-api.js";
 
 import { buildMilestoneFileName, canonicalPhaseDirName, clearPathCache, milestonesDir, legacyMilestonesDir, resolveMilestonePath, resolveSliceFile, resolveSlicePath } from "../paths.js";
-import { applyAskUserQuestionsGateResult, clearDiscussionFlowState, formatPendingAskUserQuestionsGateMessage, formatTimedOutAskUserQuestionsGateMessage, hostWriteGateAdapter, isApprovalGateVerifiedInSnapshot, isDepthConfirmationAnswer, isMilestoneDepthVerified, isMilestoneDepthVerifiedInSnapshot, isQueuePhaseActive, resetWriteGateState, shouldBlockContextWrite, shouldBlockPlanningUnit, shouldBlockQueueExecution, shouldBlockWorktreeBash, shouldBlockWorktreeWrite, isGateQuestionId, getPendingGate, shouldBlockPendingGate, shouldBlockPendingGateBash, extractDepthVerificationMilestoneId } from "./write-gate.js";
+import { applyAskUserQuestionsGateResult, clearDiscussionFlowState, currentWriteGateSnapshot, formatPendingAskUserQuestionsGateMessage, formatTimedOutAskUserQuestionsGateMessage, hostWriteGateAdapter, isApprovalGateVerifiedInSnapshot, isDepthConfirmationAnswer, isMilestoneDepthVerifiedInSnapshot, isQueuePhaseActive, resetWriteGateState, shouldBlockContextWrite, shouldBlockPlanningUnit, shouldBlockQueueExecution, shouldBlockWorktreeBash, shouldBlockWorktreeWrite, isGateQuestionId, getPendingGate, shouldBlockPendingGate, shouldBlockPendingGateBash, extractDepthVerificationMilestoneId, type WriteGateSnapshot } from "./write-gate.js";
 import { canonicalToolName } from "../engine-hook-contract.js";
 import { resolveManifest } from "../unit-context-manifest.js";
 import { isBlockedStateFile, isBashWriteToStateFile, BLOCKED_WRITE_ERROR } from "../write-intercept.js";
@@ -35,6 +35,7 @@ import {
 import { applyProviderPayloadPolicy } from "../provider-payload-policy.js";
 
 import { checkToolCallLoop, resetToolCallLoopGuard } from "./tool-call-loop-guard.js";
+import { MINIMAL_AUTO_BASE_TOOL_NAMES } from "./core-session-tools.js";
 import { maybePauseAutoForApprovalGate, resetPendingGatePauseGuard } from "./pending-gate-pause.js";
 import { saveActivityLog } from "../activity-log.js";
 import { recordToolCall as safetyRecordToolCall, recordToolResult as safetyRecordToolResult, saveEvidenceToDisk } from "../safety/evidence-collector.js";
@@ -178,22 +179,7 @@ export const MINIMAL_GSD_TOOL_NAMES = [
   "gsd_capture_thought",
 ] as const;
 
-export const MINIMAL_AUTO_BASE_TOOL_NAMES = [
-  "ask_user_questions",
-  "bash",
-  "bg_shell",
-  "edit",
-  "find",
-  "glob",
-  "grep",
-  "fetch_page",
-  "search-the-web",
-  "ls",
-  "read",
-  "subagent",
-  "write",
-  "ToolSearch",
-] as const;
+export { MINIMAL_AUTO_BASE_TOOL_NAMES } from "./core-session-tools.js";
 
 function withPreservedShimTools(toolNames: readonly string[]): string[] {
   return [...new Set([...toolNames, ...ALWAYS_PRESERVED_SHIM_TOOL_NAMES])];
@@ -583,6 +569,10 @@ function deferApprovalGate(gateId: string, basePath: string): void {
   // workflow MCP child already verified this gate, deferring would block
   // tools for a gate that can never legitimately arm.
   const snapshot = hostWriteGateAdapter.readState(basePath);
+  deferApprovalGateFromSnapshot(gateId, basePath, snapshot);
+}
+
+function deferApprovalGateFromSnapshot(gateId: string, basePath: string, snapshot: WriteGateSnapshot): void {
   if (isApprovalGateVerifiedInSnapshot(snapshot, gateId)) return;
   const milestoneId = extractDepthVerificationMilestoneId(gateId);
   if (milestoneId && isMilestoneDepthVerifiedInSnapshot(snapshot, milestoneId)) return;
@@ -1233,13 +1223,15 @@ export function registerHooks(
 
     const gateId = approvalGateIdForUnit(unitType, unitId);
     if (gateId) {
+      const basePath = contextBasePath(ctx);
+      const gateSnapshot = currentWriteGateSnapshot(basePath);
       // Skip the gate if this milestone is already depth-verified — the approval
       // pattern matched again on post-verification text (a false-positive re-trigger).
       // Without this guard, the second firing blocks gsd_plan_milestone in the same
       // turn and leaves CONTEXT.md on disk with no DB row (#discuss-milestone-no-db).
       const gateMilestoneId = extractDepthVerificationMilestoneId(gateId);
-      if (gateMilestoneId && isMilestoneDepthVerified(gateMilestoneId, contextBasePath(ctx))) return;
-      deferApprovalGate(gateId, contextBasePath(ctx));
+      if (gateMilestoneId && isMilestoneDepthVerifiedInSnapshot(gateSnapshot, gateMilestoneId)) return;
+      deferApprovalGateFromSnapshot(gateId, basePath, gateSnapshot);
     }
 
     approvalQuestionAbortInFlight = true;

package/src/resources/extensions/gsd/bootstrap/system-context.ts

@@ -28,9 +28,10 @@ const DEFAULT_CODEBASE_MAX_CHARS = 8_000;
 const MIN_CONTEXT_MESSAGE_MAX_CHARS = 1_000;
 const MIN_KNOWLEDGE_MAX_CHARS = 1_000;
 
-const contextMaintenanceCompletedForBasePath = new Set<string>();
-const contextMaintenanceInFlightByBasePath = new Map<string, Promise<boolean>>();
-const deferredContextMaintenanceByBasePath = new Map<string, Promise<void>>();
+const CONTEXT_MAINTENANCE_KEY_SEPARATOR = "\0";
+const contextMaintenanceCompletedForSession = new Set<string>();
+const contextMaintenanceInFlightBySession = new Map<string, Promise<boolean>>();
+const deferredContextMaintenanceBySession = new Map<string, Promise<void>>();
 
 /**
  * Bundled skill triggers — resolved dynamically at runtime instead of
@@ -114,7 +115,8 @@ async function runSessionStartupMaintenanceOnce(
   basePath: string,
   ctx: ExtensionContext,
 ): Promise<boolean> {
-  if (contextMaintenanceCompletedForBasePath.has(basePath)) {
+  const maintenanceKey = getContextMaintenanceKey(basePath, ctx);
+  if (contextMaintenanceCompletedForSession.has(maintenanceKey)) {
     // Backfills are session-once, but memory queries and other DB-backed
     // prompt assembly still need an active adapter on every turn.
     try {
@@ -126,16 +128,16 @@ async function runSessionStartupMaintenanceOnce(
     return false;
   }
 
-  const existing = contextMaintenanceInFlightByBasePath.get(basePath);
+  const existing = contextMaintenanceInFlightBySession.get(maintenanceKey);
   const isInitiator = !existing;
   // Use a definite Promise<boolean> so `await inFlight` has a known return type.
   let inFlight: Promise<boolean>;
   if (isInitiator) {
-    inFlight = performSessionStartupMaintenance(basePath, ctx);
-    contextMaintenanceInFlightByBasePath.set(basePath, inFlight);
+    inFlight = performSessionStartupMaintenance(basePath, ctx, maintenanceKey);
+    contextMaintenanceInFlightBySession.set(maintenanceKey, inFlight);
     void inFlight.finally(() => {
-      if (contextMaintenanceInFlightByBasePath.get(basePath) === inFlight) {
-        contextMaintenanceInFlightByBasePath.delete(basePath);
+      if (contextMaintenanceInFlightBySession.get(maintenanceKey) === inFlight) {
+        contextMaintenanceInFlightBySession.delete(maintenanceKey);
       }
     });
   } else {
@@ -149,6 +151,7 @@ async function runSessionStartupMaintenanceOnce(
 async function performSessionStartupMaintenance(
   basePath: string,
   ctx: ExtensionContext,
+  maintenanceKey: string,
 ): Promise<boolean> {
   // DB-backed memory backfills run below. On a cold session the database file
   // may exist without an active in-process adapter, so open the canonical
@@ -172,11 +175,40 @@ async function performSessionStartupMaintenance(
 
   // Mark session complete before scheduling deferred work so any concurrent
   // caller that observes the completed state does not re-enter maintenance.
-  contextMaintenanceCompletedForBasePath.add(basePath);
-  scheduleDeferredContextMaintenance(basePath);
+  contextMaintenanceCompletedForSession.add(maintenanceKey);
+  scheduleDeferredContextMaintenance(basePath, maintenanceKey);
   return true;
 }
 
+function getContextMaintenanceKey(basePath: string, ctx: ExtensionContext): string {
+  return `${basePath}${CONTEXT_MAINTENANCE_KEY_SEPARATOR}${getContextSessionPart(ctx)}`;
+}
+
+function getContextSessionPart(ctx: ExtensionContext): string {
+  const sessionManager = (ctx as {
+    sessionManager?: {
+      getSessionId?: () => string | undefined;
+      getSessionFile?: () => string | undefined;
+    };
+  }).sessionManager;
+
+  try {
+    const sessionId = sessionManager?.getSessionId?.();
+    if (typeof sessionId === "string" && sessionId.length > 0) return `id:${sessionId}`;
+  } catch (e) {
+    logWarning("bootstrap", `session-id fetch failed: ${(e as Error).message}`);
+  }
+
+  try {
+    const sessionFile = sessionManager?.getSessionFile?.();
+    if (typeof sessionFile === "string" && sessionFile.length > 0) return `file:${sessionFile}`;
+  } catch (e) {
+    logWarning("bootstrap", `session-file fetch failed: ${(e as Error).message}`);
+  }
+
+  return "process";
+}
+
 async function runDecisionsMemoryBackfill(ctx: ExtensionContext): Promise<void> {
   // ADR-013 step 5: opportunistic decisions->memories backfill. Idempotent
   // and best-effort — first run absorbs the existing decisions table into
@@ -210,8 +242,8 @@ async function runKnowledgeMemoryBackfill(
   }
 }
 
-function scheduleDeferredContextMaintenance(basePath: string): void {
-  if (deferredContextMaintenanceByBasePath.has(basePath)) return;
+function scheduleDeferredContextMaintenance(basePath: string, maintenanceKey: string): void {
+  if (deferredContextMaintenanceBySession.has(maintenanceKey)) return;
 
   const task = new Promise<void>((resolve) => {
     setTimeout(() => {
@@ -219,10 +251,10 @@ function scheduleDeferredContextMaintenance(basePath: string): void {
     }, 0);
   });
 
-  deferredContextMaintenanceByBasePath.set(basePath, task);
+  deferredContextMaintenanceBySession.set(maintenanceKey, task);
   void task.finally(() => {
-    if (deferredContextMaintenanceByBasePath.get(basePath) === task) {
-      deferredContextMaintenanceByBasePath.delete(basePath);
+    if (deferredContextMaintenanceBySession.get(maintenanceKey) === task) {
+      deferredContextMaintenanceBySession.delete(maintenanceKey);
     }
   });
 }
@@ -257,8 +289,10 @@ async function reportConsolidationGapsDeferred(basePath: string): Promise<void>
 
 export async function _flushDeferredContextMaintenanceForTest(basePath?: string): Promise<void> {
   const tasks = basePath
-    ? [deferredContextMaintenanceByBasePath.get(basePath)].filter((task): task is Promise<void> => Boolean(task))
-    : [...deferredContextMaintenanceByBasePath.values()];
+    ? [...deferredContextMaintenanceBySession.entries()]
+        .filter(([key]) => key.startsWith(`${basePath}${CONTEXT_MAINTENANCE_KEY_SEPARATOR}`))
+        .map(([, task]) => task)
+    : [...deferredContextMaintenanceBySession.values()];
   await Promise.allSettled(tasks);
 }
 

package/src/resources/extensions/gsd/bootstrap/tool-call-loop-guard.ts

@@ -1,18 +1,28 @@
 /**
  * Tool-call loop guard.
  *
- * Detects when a model calls the same tool with identical arguments
- * repeatedly within a single agent turn. Works in both auto-mode and
- * interactive sessions by hooking into the `tool_call` event, which
- * fires before execution and can block the call.
+ * Detects when a model repeats tool calls within a single Agent Turn.
+ * Works in both auto-mode and interactive sessions by hooking into the
+ * native engine's `tool_call` event, which fires before execution and can
+ * block the call.
  *
- * The guard uses a sliding window: it tracks the last N tool signatures
- * and blocks when the same signature appears more than MAX_CONSECUTIVE
- * times in a row. Resets on each agent turn (session_start, agent_end)
- * and when a different tool call breaks the streak.
+ * The guard has two independent checks: a sliding window for identical
+ * tool signatures, and a per-tool-name cap for repeated calls with varied
+ * arguments. State resets at Agent Turn boundaries (session_start,
+ * agent_end) and the identical-signature streak also resets when a
+ * different tool call breaks the streak. Block messages instruct the model
+ * to stop tooling for the rest of that turn and answer in text.
+ *
+ * The per-tool-name check (#783 Brief C) tracks call counts within a
+ * turn regardless of args. This catches improvisation
+ * loops where the model attempts the same missing workflow tool through
+ * varied surfaces (bash → `node -e` → CLI), each with a different
+ * signature, so the identical-args streak never trips. Whichever guard
+ * trips first blocks.
  */
 
 import { createHash } from "node:crypto";
+import { INHERENTLY_REPEATABLE_TOOL_SET } from "./core-session-tools.js";
 
 const MAX_CONSECUTIVE_IDENTICAL_CALLS = 4;
 
@@ -20,11 +30,27 @@ const MAX_CONSECUTIVE_IDENTICAL_CALLS = 4;
 const STRICT_LOOP_TOOLS = new Set(["ask_user_questions"]);
 const MAX_CONSECUTIVE_STRICT = 1;
 
+/**
+ * Per-turn cap on calls to the SAME tool name, regardless of args (#783).
+ *
+ * General-purpose execution tools are routinely called many times per turn
+ * (touching multiple files, running several commands), so they get a higher
+ * ceiling. Everything else — workflow one-shot tools (e.g. gsd_complete_milestone)
+ * and any non-allowlisted tool — gets the default cap. The default is generous
+ * enough to absorb legitimate retries but catches the reported improvisation
+ * loop (~51 calls) well before a cost spike.
+ */
+const PER_TOOL_DEFAULT_CAP = 6;
+const PER_TOOL_REPEATABLE_CAP = 15;
+
 let consecutiveCount = 0;
 let lastSignature = "";
 let lastToolName = "";
 let enabled = true;
 
+/** Per-tool-name call counts within the current turn (#783 Brief C). */
+const perToolCounts = new Map<string, number>();
+
 /** Hash tool name + args into a compact signature for comparison. */
 function hashToolCall(toolName: string, args: Record<string, unknown>): string {
   const h = createHash("sha256");
@@ -46,6 +72,12 @@ function hashToolCall(toolName: string, args: Record<string, unknown>): string {
  *
  * Returns `{ block: false }` for allowed calls.
  * Returns `{ block: true, reason }` when the loop threshold is exceeded.
+ *
+ * Two independent guards run; whichever trips first blocks:
+ *  1. Identical-signature streak (MAX_CONSECUTIVE_IDENTICAL_CALLS, strict for
+ *     ask_user_questions).
+ *  2. Per-tool-name cap (PER_TOOL_DEFAULT_CAP / PER_TOOL_REPEATABLE_CAP),
+ *     independent of args — catches improvisation loops (#783).
  */
 export function checkToolCallLoop(
   toolName: string,
@@ -63,6 +95,7 @@ export function checkToolCallLoop(
     lastToolName = toolName;
   }
 
+  // ── Guard 1: identical-signature streak ──
   const threshold = STRICT_LOOP_TOOLS.has(toolName)
     ? MAX_CONSECUTIVE_STRICT
     : MAX_CONSECUTIVE_IDENTICAL_CALLS;
@@ -71,13 +104,34 @@ export function checkToolCallLoop(
     return {
       block: true,
       reason:
-        `Tool loop detected: ${toolName} called ${consecutiveCount} times ` +
+        `Tool loop detected (identical args): ${toolName} called ${consecutiveCount} times ` +
         `with identical arguments. Blocking to prevent infinite loop. ` +
-        `Try a different approach or modify your arguments.`,
+        `Do not retry this tool or call other tools this turn — stop and respond to the user in text.`,
       count: consecutiveCount,
     };
   }
 
+  // ── Guard 2: per-tool-name cap, independent of args (#783 Brief C) ──
+  // Catches improvisation loops where the same tool is invoked many times with
+  // varied args (e.g. retrying a missing workflow tool via bash/node -e/CLI).
+  const perToolCount = (perToolCounts.get(toolName) ?? 0) + 1;
+  perToolCounts.set(toolName, perToolCount);
+  const perToolCap = INHERENTLY_REPEATABLE_TOOL_SET.has(toolName)
+    ? PER_TOOL_REPEATABLE_CAP
+    : PER_TOOL_DEFAULT_CAP;
+
+  if (perToolCount > perToolCap) {
+    return {
+      block: true,
+      reason:
+        `Tool loop detected (repeated tool): ${toolName} called ${perToolCount} times ` +
+        `this turn (cap ${perToolCap}). Blocking to prevent infinite loop. ` +
+        `The tool may be unavailable or failing repeatedly. ` +
+        `Do not retry this tool or pivot to other tools this turn — stop and respond to the user in text.`,
+      count: perToolCount,
+    };
+  }
+
   return { block: false, count: consecutiveCount };
 }
 
@@ -87,6 +141,7 @@ export function resetToolCallLoopGuard(): void {
   lastSignature = "";
   lastToolName = "";
   enabled = true;
+  perToolCounts.clear();
 }
 
 /** Disable the guard (e.g. during shutdown). */
@@ -95,9 +150,18 @@ export function disableToolCallLoopGuard(): void {
   consecutiveCount = 0;
   lastSignature = "";
   lastToolName = "";
+  perToolCounts.clear();
 }
 
 /** Get current consecutive count for diagnostics. */
 export function getToolCallLoopCount(): number {
   return consecutiveCount;
 }
+
+/**
+ * Get the per-tool-name call count for the current turn (#783 Brief C).
+ * Returns 0 for tools not yet called. Diagnostic only.
+ */
+export function getToolCallCountForTool(toolName: string): number {
+  return perToolCounts.get(toolName) ?? 0;
+}

package/src/resources/extensions/gsd/bootstrap/write-gate.ts

@@ -180,7 +180,7 @@ function ensureWriteGateSnapshotDirectory(basePath: string): void {
   mkdirSync(join(gsdPath, "runtime"), { recursive: true });
 }
 
-function currentWriteGateSnapshot(basePath: string = process.cwd()): WriteGateSnapshot {
+export function currentWriteGateSnapshot(basePath: string = process.cwd()): WriteGateSnapshot {
   const state = getWriteGateState(basePath);
   return {
     verifiedDepthMilestones: [...state.verifiedDepthMilestones].sort(),