@opengsd/gsd-pi 1.3.0-dev.65546769 → 1.3.0-dev.738160b9

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

@@ -27,7 +27,7 @@ import { writeLock, clearLock, clearStaleWorkerLock, readCrashLock, isLockProces
 import { acquireSessionLock, getSessionLockStatus, releaseSessionLock, updateSessionLock, } from "./session-lock.js";
 import { resolveAutoSupervisorConfig, loadEffectiveGSDPreferences, loadEffectiveGSDPreferencesWithRegistry, resolveProfileAnchorProvider, getIsolationMode, resolveEffectiveUnitIsolationMode, } from "./preferences.js";
 import { playNotificationBell, sendDesktopNotification } from "./notifications.js";
-import { getBudgetAlertLevel, getNewBudgetAlertLevel, getBudgetEnforcementAction, resolveCompactionThresholdPercent, shouldRerootStepSessionForContext, } from "./auto-budget.js";
+import { getBudgetAlertLevel, getNewBudgetAlertLevel, getBudgetEnforcementAction, HARD_CONTEXT_REROOT_THRESHOLD_PERCENT, resolveCompactionThresholdPercent, shouldRerootStepSessionForContext, shouldWarnStepSessionForContext, } from "./auto-budget.js";
 import { markToolStart as _markToolStart, markToolEnd as _markToolEnd, getOldestInFlightToolAgeMs as _getOldestInFlightToolAgeMs, clearInFlightTools, isToolInvocationError, isQueuedUserMessageSkip, isDeterministicPolicyError, } from "./auto-tool-tracking.js";
 import { closeoutUnit } from "./auto-unit-closeout.js";
 import { selectAndApplyModel, resolveModelId, clearToolBaseline } from "./auto-model-selection.js";
@@ -369,7 +369,7 @@ function getEffectiveUnitIsolationMode(basePath) {
  * running suspiciously long (e.g., a Bash command hung because `&` kept stdout open).
  */
 // Re-export budget utilities for external consumers
-export { getBudgetAlertLevel, getNewBudgetAlertLevel, getBudgetEnforcementAction, getContextPauseAction, resolveCompactionThresholdPercent, shouldRerootStepSessionForContext, } from "./auto-budget.js";
+export { getBudgetAlertLevel, getNewBudgetAlertLevel, getBudgetEnforcementAction, getContextPauseAction, resolveCompactionThresholdPercent, shouldRerootStepSessionForContext, shouldWarnStepSessionForContext, } from "./auto-budget.js";
 function closeOutSignalInterruptedUnit(currentBasePath) {
     const currentUnit = s.currentUnit;
     if (!currentUnit)
@@ -904,9 +904,10 @@ export async function rerootCommandSession(cmdCtx, workspaceRoot) {
     }
 }
 /**
- * After a completed step-mode unit, reclaim context headroom when usage is at
- * or above the compaction threshold. Writes a context-mode snapshot first, then
- * re-roots the visible command session while leaving the NEXT progress surface.
+ * After a completed step-mode unit, warn when usage reaches the configurable
+ * soft context threshold. At the fixed hard boundary, write a context-mode
+ * snapshot and re-root the visible command session while leaving the NEXT
+ * progress surface.
  */
 export async function maybeRerootStepSessionForHighContext(ctx, workspaceRoot, cmdCtx, originalBasePath) {
     if (!cmdCtx)
@@ -915,18 +916,24 @@ export async function maybeRerootStepSessionForHighContext(ctx, workspaceRoot, c
     const { loadEffectiveGSDPreferences } = await import("./preferences.js");
     const prefs = loadEffectiveGSDPreferences(workspaceRoot);
     const compactionThreshold = prefs?.preferences.context_management?.compaction_threshold_percent;
-    if (!shouldRerootStepSessionForContext(contextPercent, compactionThreshold)) {
+    const shouldWarn = shouldWarnStepSessionForContext(contextPercent, compactionThreshold);
+    const shouldReroot = shouldRerootStepSessionForContext(contextPercent);
+    if (!shouldWarn && !shouldReroot) {
+        return { rerooted: false };
+    }
+    const displayPercent = contextPercent.toFixed(1);
+    const thresholdDisplay = resolveCompactionThresholdPercent(compactionThreshold).toFixed(0);
+    if (!shouldReroot) {
+        ctx.ui.notify(`Step complete — context at ${displayPercent}% (soft threshold: ${thresholdDisplay}%). Use /compact when convenient; continuing in this session is still safe.`, "info");
         return { rerooted: false };
     }
     const { resolveWorktreeProjectRoot } = await import("./worktree-root.js");
     const snapshotBase = resolveWorktreeProjectRoot(workspaceRoot, originalBasePath);
     const { writeContextModeCompactionSnapshot } = await import("./context-mode-snapshot.js");
     await writeContextModeCompactionSnapshot(snapshotBase);
-    const displayPercent = contextPercent.toFixed(1);
-    const thresholdDisplay = resolveCompactionThresholdPercent(compactionThreshold).toFixed(0);
     const result = await rerootCommandSession(cmdCtx, workspaceRoot);
     if (result.status === "ok") {
-        ctx.ui.notify(`Step complete — context at ${displayPercent}% (threshold: ${thresholdDisplay}%). Fresh session ready for /gsd next.`, "info");
+        ctx.ui.notify(`Step complete — context at ${displayPercent}% (hard threshold: ${HARD_CONTEXT_REROOT_THRESHOLD_PERCENT}%). Fresh session ready for /gsd next.`, "info");
         return { rerooted: true };
     }
     if (result.status === "cancelled") {
@@ -1545,7 +1552,7 @@ export function _selectStopAutoWorktreeExit(args) {
 /**
  * Pause auto-mode without destroying state. Context is preserved.
  * The user can interact with the agent, then `/gsd auto` resumes
- * from disk state. Called when the user presses Escape during auto-mode.
+ * from disk state. Called when the user presses Escape or runs `/gsd pause`.
  */
 export async function pauseAuto(ctx, _pi, _errorContext, options = {}) {
     if (!s.active)
@@ -1564,6 +1571,8 @@ export async function pauseAuto(ctx, _pi, _errorContext, options = {}) {
     ].filter((path) => Boolean(path)));
     s.active = false;
     s.paused = true;
+    if (options.abortActiveTurn && ctx && !ctx.isIdle())
+        ctx.abort();
     clearUnitTimeout();
     stopAutoCommandPolling();
     // Flush queued follow-up messages (#3512).

package/dist/resources/extensions/gsd/bootstrap/agent-end-recovery.js

@@ -1,7 +1,9 @@
 // gsd-pi + src/resources/extensions/gsd/bootstrap/agent-end-recovery.ts - Handles provider and agent-end recovery for GSD auto-mode.
+import { mkdirSync, readdirSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
 import { logWarning } from "../workflow-logger.js";
 import { checkDeepProjectSetupAfterTurn, checkAutoStartAfterDiscuss, maybeHandleReadyPhraseWithoutFiles, maybeHandleEmptyIntentTurn, resetEmptyTurnCounter, } from "../guided-flow.js";
-import { clearPathCache } from "../paths.js";
+import { clearPathCache, gsdRoot } from "../paths.js";
 import { getAutoDashboardData, getAutoModeStartModel, isAutoActive, isAutoCompletionStopInProgress, pauseAuto, setCurrentDispatchedModelId, setCurrentUnitModelForRecovery, } from "../auto.js";
 import { getNextFallbackModel, resolveModelWithFallbacksForUnit } from "../preferences.js";
 import { pauseAutoForProviderError } from "../provider-error-pause.js";
@@ -11,7 +13,7 @@ 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 } from "../guided-unit-context.js";
 import { resumeAutoAfterProviderDelay } from "./provider-error-resume.js";
 import { classifyError, createRetryState, resetRetryState, isTransient, } from "../error-classifier.js";
 import { blockModel, blockModelUntil, isModelBlocked, isModelTemporarilyUnavailable } from "../blocked-models.js";
@@ -78,10 +80,15 @@ export function isUserInitiatedAbortMessage(message) {
         return false;
     return /\b(?:claude code process aborted by user|request aborted by user|process aborted by user)\b/i.test(message);
 }
-export function shouldDeferTransientErrorToCoreRetry(cls, rawErrorMsg) {
+export function shouldDeferTransientErrorToCoreRetry(cls, rawErrorMsg, deferCheckMsg = rawErrorMsg) {
     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);
 }
 /** Try configured fallbacks, then the auto-mode start model. Returns true if switched. */
 async function tryProviderModelFallback(params) {
@@ -277,6 +284,86 @@ export function suppressTerminalDeletedWorktreeMessageEnd(event) {
     logWarning("bootstrap", `Suppressing stale deleted-worktree provider error during terminal completion reroot: ${displayMsg || rawErrorMsg}`);
     return true;
 }
+function modelLabel(ctx) {
+    const provider = ctx.model?.provider;
+    const id = ctx.model?.id;
+    return provider && id ? `${provider}/${id}` : "unknown model";
+}
+function isFatalManualGuidedTerminalFailure(lastMsg) {
+    if (!isObjectRecord(lastMsg) || !("stopReason" in lastMsg))
+        return false;
+    if (lastMsg.stopReason === "error") {
+        const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+        if (isUserInitiatedAbortMessage(rawErrorMsg))
+            return false;
+        return true;
+    }
+    if (lastMsg.stopReason !== "aborted")
+        return false;
+    const content = "content" in lastMsg ? lastMsg.content : undefined;
+    const hasErrorMessage = "errorMessage" in lastMsg && !!lastMsg.errorMessage;
+    return hasErrorMessage || !_hasEmptyAgentEndContent(content);
+}
+function terminalFailureDetail(lastMsg) {
+    if (!isObjectRecord(lastMsg))
+        return "Provider turn ended with an unknown terminal error.";
+    const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+    const content = "content" in lastMsg ? lastMsg.content : undefined;
+    const displayMsg = resolveAgentEndErrorDisplay(rawErrorMsg, content).replace(/\s+/g, " ").trim();
+    if (displayMsg)
+        return displayMsg.length > 300 ? `${displayMsg.slice(0, 300)}...` : displayMsg;
+    return lastMsg.stopReason === "aborted"
+        ? "Provider turn aborted with error context."
+        : "Provider stream ended with stopReason=error.";
+}
+function nextManualActivitySequence(activityDir) {
+    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, unitType, model, detail, stopReason) {
+    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, lastMsg, guidedUnit) {
+    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, pi, ctx, errorDetail, isRateLimit) {
     retryState.consecutiveTransientCount += 1;
     const baseRetryAfterMs = "retryAfterMs" in cls ? cls.retryAfterMs : 15_000;
@@ -315,7 +402,9 @@ export async function handleAgentEnd(pi, event, ctx) {
     // 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)) {
             return;
@@ -342,12 +431,13 @@ export async function handleAgentEnd(pi, event, ctx) {
     // discussions (where isAutoActive may be false) still get recovered.
     if (maybeHandleEmptyIntentTurn(event, isAutoActive(), basePath))
         return;
-    if (!isAutoActive())
+    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;
@@ -435,9 +525,9 @@ export async function handleAgentEnd(pi, event, ctx) {
             });
             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);
         if (isAutoCompletionStopInProgress() &&
             isTerminalDeletedWorktreeProviderError(`${rawErrorMsg}\n${displayMsg}`)) {
@@ -447,8 +537,8 @@ export async function handleAgentEnd(pi, event, ctx) {
         }
         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
         //        same dead model isn't reselected on the next /gsd auto restart,
@@ -517,7 +607,7 @@ export async function handleAgentEnd(pi, event, ctx) {
         // 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;
         }
         // ── Tool-schema overload: the active model repeatedly emitted tool-call

package/dist/resources/extensions/gsd/bootstrap/core-session-tools.js

@@ -0,0 +1,39 @@
+// 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",
+];
+/** Tools excluded from the high per-turn loop-guard cap (strict or orchestration). */
+const NON_REPEATABLE_FROM_MINIMAL_BASE = new Set([
+    "ask_user_questions",
+    "subagent",
+    "ToolSearch",
+]);
+/** Additional core tools not in MINIMAL_AUTO_BASE but routinely multi-called per turn. */
+const EXTRA_INHERENTLY_REPEATABLE = [
+    "gsd_exec",
+    "multi_edit",
+    "todo_write",
+    "notebook_edit",
+    "search_and_read",
+];
+/** 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,
+];
+export const INHERENTLY_REPEATABLE_TOOL_SET = new Set(INHERENTLY_REPEATABLE_TOOL_NAMES);

package/dist/resources/extensions/gsd/bootstrap/db-tools.js

@@ -2,6 +2,7 @@
 // File Purpose: Registers DB-backed GSD workflow tools and compatibility aliases.
 import { Type, StringEnum } from "@gsd/pi-ai";
 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";
 import { importWorkflowExecutorsModule } from "../workflow-mcp.js";
@@ -373,13 +374,17 @@ export function registerDbTools(pi) {
             "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, theme) {
@@ -649,8 +654,9 @@ export function registerDbTools(pi) {
         promptSnippet: "Plan a slice via DB write + PLAN render + cache invalidation",
         promptGuidelines: [
             "Use gsd_plan_slice for slice planning instead of writing S##-PLAN.md or task PLAN files directly.",
-            "Keep parameters flat and provide the full slice planning payload, including tasks.",
-            "The tool validates input, requires an existing parent slice, writes slice/task planning data, renders PLAN.md and task plan files from DB, and clears both state and parse caches after success.",
+            "For incremental planning, call gsd_plan_slice with slice metadata only, then call gsd_plan_task once per task.",
+            "When tasks is omitted or empty, the tool writes only slice planning metadata and preserves existing tasks.",
+            "When tasks is non-empty, the tool validates input, requires an existing parent slice, writes slice/task planning data, renders PLAN.md and task plan files from DB, and clears both state and parse caches after success.",
             "Use the canonical name gsd_plan_slice; gsd_slice_plan is only an alias.",
         ],
         parameters: Type.Object({
@@ -658,7 +664,7 @@ export function registerDbTools(pi) {
             milestoneId: Type.String({ description: "Milestone ID (e.g. M001)" }),
             sliceId: Type.String({ description: "Slice ID (e.g. S01)" }),
             goal: Type.String({ description: "Slice goal" }),
-            tasks: Type.Array(Type.Object({
+            tasks: Type.Optional(Type.Array(Type.Object({
                 taskId: Type.String({ description: "Task ID (e.g. T01)" }),
                 title: Type.String({ description: "Task title" }),
                 description: Type.String({ description: "Task description / steps block" }),
@@ -668,7 +674,7 @@ export function registerDbTools(pi) {
                 inputs: Type.Array(Type.String(), { description: "Array<string> of input files or references; pass [\"path\"] or [], never a single string" }),
                 expectedOutput: Type.Array(Type.String(), { description: "Array<string> of files this task creates or overwrites; pass [\"path\"] or [], never prose or a single string" }),
                 observabilityImpact: Type.Optional(Type.String({ description: "Task observability impact" })),
-            }), { description: "Planned tasks for the slice" }),
+            }), { description: "Optional full task replacement for the slice. Omit for incremental planning, then call gsd_plan_task once per task." })),
             // ── Enrichment metadata (optional — defaults to empty) ────────────
             successCriteria: Type.Optional(Type.String({ description: "Slice success criteria block" })),
             proofLevel: Type.Optional(Type.String({ description: "Slice proof level" })),

package/dist/resources/extensions/gsd/bootstrap/exec-tools.js

@@ -60,6 +60,7 @@ export function registerExecTools(pi) {
             return executeUatExec(params, {
                 baseDir,
                 preferences: await loadContextModePreferences(baseDir),
+                signal: _signal,
             });
         },
     });
@@ -99,6 +100,7 @@ export function registerExecTools(pi) {
             return executeGsdExec(params, {
                 baseDir,
                 preferences: await loadContextModePreferences(baseDir),
+                signal: _signal,
             });
         },
     });

package/dist/resources/extensions/gsd/bootstrap/register-extension.js

@@ -14,6 +14,7 @@ import { registerShortcuts } from "./register-shortcuts.js";
 import { writeCrashLog } from "./crash-log.js";
 import { logWarning, isGsdExtensionStderrEnabled } from "../workflow-logger.js";
 import { UNIT_TOOL_CONTRACTS } from "../unit-tool-contracts.js";
+import { installManifestFlushOnProcessTeardown } from "../workflow-manifest.js";
 // Static import so cmux event listeners are registered synchronously during
 // extension bootstrap. Prior implementation used `void import().then()` which
 // queued listener registration as a microtask — any CMUX_CHANNELS emit fired
@@ -177,6 +178,7 @@ export function registerGsdExtension(pi) {
         process.stderr.write(`[gsd] Failed to bootstrap hook-emitter bridge: ${err instanceof Error ? err.stack ?? err.message : String(err)}\n`);
     });
     installEpipeGuard();
+    installManifestFlushOnProcessTeardown();
     // Ecosystem handlers captured by the GSDExtensionAPI wrapper for the
     // GSD-owned `before_agent_start` dispatch step (#3338).
     const ecosystemHandlers = [];

package/dist/resources/extensions/gsd/bootstrap/register-hooks.js

@@ -7,14 +7,15 @@ import { isToolCallEventType } from "@gsd/pi-coding-agent";
 import { ALWAYS_PRESERVED_SHIM_TOOL_NAMES } from "@gsd/pi-ai";
 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 } 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";
 import { loadFile, saveFile, formatContinue } from "../files.js";
 import { clearAutoCompletionStopInProgress, clearToolInvocationError, getAutoRuntimeSnapshot, getSourceObservationStore, isAutoActive, isAutoCompletionStopInProgress, isAutoPaused, isInteractiveElicitationInFlight, markToolEnd, markToolStart, recordAutoToolSurfaceSnapshot, recordToolInvocationError, } from "../auto-runtime-state.js";
 import { applyProviderPayloadPolicy } from "../provider-payload-policy.js";
-import { checkToolCallLoop, resetToolCallLoopGuard } from "./tool-call-loop-guard.js";
+import { checkToolCallLoop, recordToolCallLoopMutation, 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";
@@ -42,6 +43,7 @@ import { supportsSourceObservationsForUnit } from "../source-observations.js";
 import { clearPendingAutoStart } from "../pending-auto-start.js";
 import { resolveWorkflowToolBasePath } from "./dynamic-tools.js";
 import { getRequiredWorkflowToolsForUnit } from "../unit-tool-contracts.js";
+import { flushAllManifests } from "../workflow-manifest.js";
 let approvalQuestionAbortInFlight = false;
 async function loadWelcomeScreenModule() {
     const candidates = [];
@@ -138,22 +140,7 @@ export const MINIMAL_GSD_TOOL_NAMES = [
     "capture_thought",
     "gsd_capture_thought",
 ];
-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",
-];
+export { MINIMAL_AUTO_BASE_TOOL_NAMES } from "./core-session-tools.js";
 function withPreservedShimTools(toolNames) {
     return [...new Set([...toolNames, ...ALWAYS_PRESERVED_SHIM_TOOL_NAMES])];
 }
@@ -436,6 +423,9 @@ function deferApprovalGate(gateId, basePath) {
     // 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, basePath, snapshot) {
     if (isApprovalGateVerifiedInSnapshot(snapshot, gateId))
         return;
     const milestoneId = extractDepthVerificationMilestoneId(gateId);
@@ -446,6 +436,16 @@ function deferApprovalGate(gateId, basePath) {
 function contextBasePath(ctx) {
     return typeof ctx?.cwd === "string" ? ctx.cwd : process.cwd();
 }
+const LOOP_GUARD_INTERACTIVE_INSTRUCTIONS = [
+    "Do not retry this tool or call other tools this turn — stop and respond to the user in text.",
+    "Do not retry this tool or pivot to other tools this turn — stop and respond to the user in text.",
+];
+const LOOP_GUARD_AUTO_INSTRUCTION = "Do not re-issue this blocked tool. In /gsd auto, stop tool calls for this turn and return control to the auto-mode recovery/replan path.";
+function formatLoopGuardBlockReason(reason) {
+    if (!reason || !getAutoRuntimeSnapshot().active)
+        return reason;
+    return LOOP_GUARD_INTERACTIVE_INSTRUCTIONS.reduce((formatted, instruction) => formatted.replace(instruction, LOOP_GUARD_AUTO_INSTRUCTION), reason);
+}
 function beginSourceObservationStoreForCurrentUnit(ctx) {
     if (!isAutoActive())
         return null;
@@ -714,6 +714,12 @@ export function registerHooks(pi, ecosystemHandlers) {
                     }
                 }
             }
+            const projectRoot = resolveWorktreeProjectRoot(basePath);
+            const { pruneStaleFlatPhaseBackups } = await import("../flat-phase-migration.js");
+            const pruned = pruneStaleFlatPhaseBackups(projectRoot);
+            if (pruned > 0) {
+                safetyLogWarning("bootstrap", `pruned ${pruned} stale flat-phase migration backup(s) from .gsd-backups/ (retention exceeded)`);
+            }
         }
         catch (err) {
             safetyLogWarning("bootstrap", `flat-phase migration: ${err instanceof Error ? err.message : String(err)}`);
@@ -871,6 +877,17 @@ export function registerHooks(pi, ecosystemHandlers) {
         const { handleAgentEnd } = await import("./agent-end-recovery.js");
         const agentEndBasePath = contextBasePath(ctx);
         try {
+            // The manifest is a non-critical projection (the append-only event log is
+            // the authoritative recovery source), so a flush failure must not skip
+            // agent_end recovery. drainManifestWrites still propagates write failures
+            // to explicit flush callers; here we deliberately log and continue so
+            // handleAgentEnd (and the finally below) always run.
+            try {
+                await flushAllManifests();
+            }
+            catch (err) {
+                safetyLogWarning("manifest", `flushAllManifests on agent_end failed: ${err instanceof Error ? err.message : String(err)}`);
+            }
             await handleAgentEnd(pi, event, ctx);
         }
         finally {
@@ -1009,14 +1026,16 @@ export function registerHooks(pi, ecosystemHandlers) {
             return;
         const gateId = approvalGateIdForUnit(unitType, unitId);
         if (gateId) {
+            const basePath = contextBasePath(ctx);
+            const gateSnapshot = currentWriteGateSnapshot(basePath);
             // Skip the gate if this milestone is already depth-verified — the approval
             // pattern matched again on post-verification text (a false-positive re-trigger).
             // Without this guard, the second firing blocks gsd_plan_milestone in the same
             // turn and leaves CONTEXT.md on disk with no DB row (#discuss-milestone-no-db).
             const gateMilestoneId = extractDepthVerificationMilestoneId(gateId);
-            if (gateMilestoneId && isMilestoneDepthVerified(gateMilestoneId, contextBasePath(ctx)))
+            if (gateMilestoneId && isMilestoneDepthVerifiedInSnapshot(gateSnapshot, gateMilestoneId))
                 return;
-            deferApprovalGate(gateId, contextBasePath(ctx));
+            deferApprovalGateFromSnapshot(gateId, basePath, gateSnapshot);
         }
         approvalQuestionAbortInFlight = true;
         ctx.ui.notify(`${unitType ?? "The discussion"}${unitId ? ` ${unitId}` : ""} is waiting for your approval - pausing before more tool calls run.`, "info");
@@ -1058,7 +1077,7 @@ export function registerHooks(pi, ecosystemHandlers) {
         // ── Loop guard: block repeated identical tool calls ──
         const loopCheck = checkToolCallLoop(toolName, event.input);
         if (loopCheck.block) {
-            return { block: true, reason: loopCheck.reason };
+            return { block: true, reason: formatLoopGuardBlockReason(loopCheck.reason) };
         }
         const deferredGateGuard = shouldBlockDeferredApprovalTool(toolName, event.input, discussionBasePath);
         if (deferredGateGuard.block) {
@@ -1264,6 +1283,7 @@ export function registerHooks(pi, ecosystemHandlers) {
             }
         }
         if (!event.isError) {
+            recordToolCallLoopMutation(toolName);
             refreshSourceObservationAfterMutation(toolName, event.input, ctx);
             clearSourceObservationsAfterShell(toolName);
         }