gsd-pi 2.80.0-dev.c5c38454b → 2.80.0-dev.cf9433f56

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

@@ -6,7 +6,7 @@
  * utility.
  */
 import { loadFile, parseContinue, parseSummary, loadActiveOverrides, formatOverridesSection, parseTaskPlanFile } from "./files.js";
-import { hasVerdict, getUatType } from "./verdict-parser.js";
+import { hasVerdict, getUatType, extractVerdict } from "./verdict-parser.js";
 import { loadPrompt, inlineTemplate } from "./prompt-loader.js";
 import { resolveMilestoneFile, resolveSliceFile, resolveSlicePath, resolveTasksDir, resolveTaskFiles, resolveTaskFile, relMilestoneFile, relSliceFile, relSlicePath, relMilestonePath, resolveGsdRootFile, relGsdRootFile, resolveRuntimeFile, } from "./paths.js";
 import { resolveSkillDiscoveryMode, resolveInlineLevel, loadEffectiveGSDPreferences, resolveAllSkillReferences } from "./preferences.js";
@@ -21,10 +21,12 @@ import { assertGateCoverage, getGatesForTurn, } from "./gate-registry.js";
 import { formatDecisionsCompact, formatRequirementsCompact } from "./structured-data-formatter.js";
 import { readPhaseAnchor, formatAnchorForPrompt } from "./phase-anchor.js";
 import { composeContextModeInstructions, composeInlinedContext } from "./unit-context-composer.js";
+import { readCompactionSnapshot } from "./compaction-snapshot.js";
 import { logWarning } from "./workflow-logger.js";
 import { inlineGraphSubgraph } from "./graph-context.js";
 import { buildExtractionStepsBlock } from "./commands-extract-learnings.js";
 import { resolveSkillManifest, warnIfManifestHasMissingSkills } from "./skill-manifest.js";
+import { classifyProject } from "./detection.js";
 // ─── Preamble Cap ─────────────────────────────────────────────────────────────
 /**
  * Historical static ceiling for the preamble cap. Kept as an upper bound even
@@ -62,6 +64,104 @@ function resolvePromptBudgets() {
 function resolveSummaryBudgetChars() {
     return resolvePromptBudgets().summaryBudgetChars;
 }
+function formatProjectClassificationForPlanning(classification) {
+    const sampleFiles = classification.contentFiles.slice(0, 8);
+    const sample = sampleFiles.length > 0 ? sampleFiles.map((file) => `\`${file}\``).join(", ") : "(none)";
+    const lines = [
+        "### Project Classification",
+        "",
+        `- **Kind:** ${classification.kind}`,
+        `- **Content files:** ${classification.contentFiles.length}`,
+        `- **Sample files:** ${sample}`,
+        `- **Reason:** ${classification.reason}`,
+        "",
+    ];
+    if (classification.kind === "untyped-existing") {
+        if (classification.contentFiles.length <= 2) {
+            lines.push("**Workflow sizing:** This is a tiny existing untyped project. Prefer exactly one slice unless the milestone request clearly spans multiple independent user-visible capabilities.");
+        }
+        else if (classification.contentFiles.length <= 5) {
+            lines.push("**Workflow sizing:** This is a small existing untyped project. Prefer 1-2 slices unless the milestone request clearly spans multiple independent user-visible capabilities.");
+        }
+        else {
+            lines.push("**Workflow sizing:** Existing untyped project. Use generic file-level workflow guidance and size slices by real capability boundaries, not by missing tooling markers.");
+        }
+    }
+    else if (classification.kind === "greenfield") {
+        lines.push("**Workflow sizing:** No project content exists yet. Use normal greenfield sizing for the requested scope.");
+    }
+    else if (classification.kind === "typed-existing") {
+        lines.push("**Workflow sizing:** Known project markers exist. Use normal ecosystem-aware planning guidance.");
+    }
+    else {
+        lines.push("**Workflow sizing:** Invalid repository state. Planning should surface this as a blocker rather than inventing project structure.");
+    }
+    return lines.join("\n");
+}
+function normalizeArtifactRef(value) {
+    return value.trim().replace(/^[-\s]+/, "").replace(/^["'`]+|["'`]+$/g, "").replaceAll("\\", "/").replace(/^\.\//, "");
+}
+function parseCoveredArtifacts(validationContent) {
+    const covered = new Set();
+    const lines = validationContent.split(/\r?\n/);
+    let inCoveredArtifacts = false;
+    for (const line of lines) {
+        if (/^\s*covered[-_]?artifacts\s*:/i.test(line)) {
+            inCoveredArtifacts = true;
+            const inline = line.split(/covered[-_]?artifacts\s*:/i)[1]?.trim();
+            if (inline && inline !== "[]") {
+                inline.replace(/^\[|\]$/g, "").split(",").map(normalizeArtifactRef).filter(Boolean).forEach((item) => covered.add(item));
+            }
+            continue;
+        }
+        if (!inCoveredArtifacts)
+            continue;
+        if (/^\S/.test(line) && !/^\s*-/.test(line))
+            break;
+        const item = line.match(/^\s*-\s*(.+)$/)?.[1];
+        if (item)
+            covered.add(normalizeArtifactRef(item));
+    }
+    return covered;
+}
+function isValidationFreshOrApplicable(validationContent, currentArtifacts) {
+    if (!validationContent)
+        return false;
+    if (!/validation_metadata:/i.test(validationContent))
+        return false;
+    const coveredArtifacts = parseCoveredArtifacts(validationContent);
+    if (coveredArtifacts.size === 0)
+        return false;
+    return currentArtifacts
+        .map(normalizeArtifactRef)
+        .filter(Boolean)
+        .every((artifact) => coveredArtifacts.has(artifact));
+}
+function formatCloseoutReviewInstructions(validationContent, validationRel, currentArtifacts) {
+    const verdict = validationContent ? extractVerdict(validationContent) : null;
+    const validationFresh = isValidationFreshOrApplicable(validationContent, currentArtifacts);
+    if (verdict === "pass" && validationFresh) {
+        return [
+            "### Passing Validation Artifact",
+            "",
+            `A passing validation artifact is present at \`${validationRel}\`. Treat it as authoritative for success criteria, requirement coverage, verification classes, and cross-slice integration.`,
+            "",
+            "Do not delegate fresh reviewer/security/tester audits and do not redo the validation evidence review unless the artifact is internally inconsistent with the inlined summaries. Focus this unit on final milestone narrative, learnings, PROJECT/requirements updates, and `gsd_complete_milestone`.",
+        ].join("\n");
+    }
+    if (verdict) {
+        return [
+            "### Validation Requires Attention",
+            "",
+            `A validation artifact is present at \`${validationRel}\` with verdict \`${verdict}\`, but it is missing freshness metadata or does not cover current milestone artifacts. Do not treat the milestone as complete unless the issues are resolved and evidence supports completion.`,
+        ].join("\n");
+    }
+    return [
+        "### No Passing Validation Artifact",
+        "",
+        `No passing validation artifact was found at \`${validationRel}\`. Use the full closeout review path before completion.`,
+    ].join("\n");
+}
 function capPreamble(preamble) {
     // Cap inlined context at min(historical 30K ceiling, scaled inline budget).
     // The ceiling preserves pre-fix behavior for large-window users; the scaled
@@ -79,8 +179,19 @@ function renderContextModeForPrompt(unitType, base, renderMode = "standalone") {
         renderMode,
     });
 }
-function prependContextModeToBlock(unitType, base, block, renderMode = "standalone") {
+function renderContextModeBlockForPrompt(unitType, base, renderMode = "standalone") {
     const contextMode = renderContextModeForPrompt(unitType, base, renderMode);
+    if (!contextMode)
+        return "";
+    if (renderMode === "nested")
+        return contextMode;
+    const snapshot = readCompactionSnapshot(base);
+    if (!snapshot?.trim())
+        return contextMode;
+    return `${contextMode}\n\n## Context Snapshot\nSource: \`.gsd/last-snapshot.md\`\n\n${snapshot.trimEnd()}`;
+}
+function prependContextModeToBlock(unitType, base, block, renderMode = "standalone") {
+    const contextMode = renderContextModeBlockForPrompt(unitType, base, renderMode);
     if (!contextMode)
         return block;
     if (!block.trim())
@@ -1465,6 +1576,7 @@ export async function buildPlanMilestonePrompt(mid, midTitle, base, level) {
     const researchAnchor = readPhaseAnchor(base, mid, "research-milestone");
     if (researchAnchor)
         inlined.push(formatAnchorForPrompt(researchAnchor));
+    inlined.push(formatProjectClassificationForPlanning(classifyProject(base)));
     inlined.push(await inlineFile(contextPath, contextRel, "Milestone Context"));
     const researchInline = await inlineFileOptional(researchPath, researchRel, "Milestone Research");
     if (researchInline)
@@ -2017,6 +2129,9 @@ export async function buildCompleteMilestonePrompt(mid, midTitle, base, level) {
     const inlineLevel = level ?? resolveInlineLevel();
     const roadmapPath = resolveMilestoneFile(base, mid, "ROADMAP");
     const roadmapRel = relMilestoneFile(base, mid, "ROADMAP");
+    const validationPath = resolveMilestoneFile(base, mid, "VALIDATION");
+    const validationRel = relMilestoneFile(base, mid, "VALIDATION");
+    const validationContent = validationPath ? await loadFile(validationPath) : null;
     const inlined = [];
     inlined.push(await inlineFile(roadmapPath, roadmapRel, "Milestone Roadmap"));
     // Inline all slice summaries (deduplicated by slice ID)
@@ -2056,6 +2171,13 @@ export async function buildCompleteMilestonePrompt(mid, midTitle, base, level) {
         const pathList = summaryRelPaths.map(p => `- \`${p}\``).join("\n");
         inlined.push(`### On-demand Slice Summaries\n\nExcerpted above. Read the full file for any slice when the excerpt's section heads don't carry enough narrative for the milestone summary you're drafting:\n\n${pathList}`);
     }
+    const validationContext = [
+        formatCloseoutReviewInstructions(validationContent, validationRel, [validationRel, roadmapRel, ...summaryRelPaths]),
+    ];
+    if (validationContent) {
+        validationContext.push(`### Milestone Validation\nSource: \`${validationRel}\`\n\n${validationContent.trim()}`);
+    }
+    inlined.unshift(...validationContext);
     // Inline root GSD files (skip for minimal — completion can read these if needed)
     if (inlineLevel !== "minimal") {
         const requirementsInline = await inlineRequirementsFromDb(base, mid, undefined, inlineLevel);

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

@@ -30,7 +30,49 @@ import { getProjectResearchStatus } from "./project-research-policy.js";
 // Re-export so existing consumers of auto-recovery.ts keep working.
 export { resolveExpectedArtifactPath, diagnoseExpectedArtifact };
 export { classifyMilestoneSummaryContent, } from "./milestone-summary-classifier.js";
-// ─── Artifact Resolution & Verification ───────────────────────────────────────
+export function refreshRecoveryDbForArtifact(unitType, unitId) {
+    if (unitType !== "plan-slice" && unitType !== "execute-task")
+        return { ok: true };
+    if (!isDbAvailable())
+        return { ok: true };
+    if (!refreshOpenDatabaseFromDisk()) {
+        return {
+            ok: false,
+            fatal: unitType === "execute-task",
+            reason: `${unitType}-db-refresh-failed`,
+            message: `Stuck recovery found ${unitType} ${unitId} artifacts, but the DB refresh failed.`,
+        };
+    }
+    if (unitType !== "execute-task")
+        return { ok: true };
+    const { milestone: mid, slice: sid, task: tid } = parseUnitId(unitId);
+    if (!mid || !sid || !tid) {
+        return {
+            ok: false,
+            fatal: true,
+            reason: "execute-task-invalid-unit-id",
+            message: `Stuck recovery found execute-task ${unitId} artifacts, but the unit id could not be parsed for DB verification.`,
+        };
+    }
+    const task = getTask(mid, sid, tid);
+    if (!task) {
+        return {
+            ok: false,
+            fatal: true,
+            reason: "execute-task-artifact-db-missing",
+            message: `Stuck recovery found execute-task ${unitId} artifacts, but no matching DB task row exists after refresh.`,
+        };
+    }
+    if (!isClosedStatus(task.status)) {
+        return {
+            ok: false,
+            fatal: true,
+            reason: "execute-task-artifact-db-mismatch",
+            message: `Stuck recovery found execute-task ${unitId} artifacts, but the DB task status is still '${task.status}' after refresh.`,
+        };
+    }
+    return { ok: true };
+}
 function hasCapturedWorkflowPrefs(base) {
     const prefsPath = resolveExpectedArtifactPath("workflow-preferences", "WORKFLOW-PREFS", base);
     if (!prefsPath || !existsSync(prefsPath))

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

@@ -24,7 +24,7 @@ let _currentSigtermHandler = null;
  *
  * Returns the new handler so the caller can store and deregister it later.
  */
-export function registerSigtermHandler(currentBasePath, previousHandler) {
+export function registerSigtermHandler(currentBasePath, previousHandler, onSignalCleanup) {
     // Remove the explicitly-passed previous handler
     if (previousHandler) {
         for (const sig of CLEANUP_SIGNALS)
@@ -37,6 +37,13 @@ export function registerSigtermHandler(currentBasePath, previousHandler) {
             process.off(sig, _currentSigtermHandler);
     }
     const handler = () => {
+        try {
+            onSignalCleanup?.();
+        }
+        catch {
+            void 0;
+            // Signal cleanup is best-effort; lock cleanup and process exit still run.
+        }
         clearLock(currentBasePath);
         releaseSessionLock(currentBasePath);
         process.exit(0);

package/dist/resources/extensions/gsd/auto-timeout-recovery.js

@@ -35,13 +35,13 @@ export async function recoverTimedOutUnit(ctx, pi, unitType, unitId, reason, rct
         writeUnitRuntimeRecord(basePath, unitType, unitId, currentUnitStartedAt, {
             recovery: status,
         });
-        const durableComplete = status.summaryExists && status.taskChecked && status.nextActionAdvanced;
+        const durableComplete = status.dbComplete || (status.summaryExists && status.taskChecked && status.nextActionAdvanced);
         if (durableComplete) {
             writeUnitRuntimeRecord(basePath, unitType, unitId, currentUnitStartedAt, {
                 phase: "finalized",
                 recovery: status,
             });
-            ctx.ui.notify(`${reason === "idle" ? "Idle" : "Timeout"} recovery: ${unitType} ${unitId} already completed on disk. Continuing auto-mode. (attempt ${attemptNumber})`, "info");
+            ctx.ui.notify(`${reason === "idle" ? "Idle" : "Timeout"} recovery: ${unitType} ${unitId} already completed. Continuing auto-mode. (attempt ${attemptNumber})`, "info");
             unitRecoveryCount.delete(recoveryKey);
             bumpAndResolveSynthetic(`timeout-recovery:${reason}:${unitType}/${unitId}`);
             return "recovered";

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

@@ -20,7 +20,7 @@ import { gsdRoot, resolveMilestoneFile, resolveMilestonePath, resolveDir, milest
 import { invalidateAllCaches } from "./cache.js";
 import { clearActivityLogState } from "./activity-log.js";
 import { synthesizeCrashRecovery, getDeepDiagnostic, readActiveMilestoneId, } from "./session-forensics.js";
-import { writeLock, clearLock, readCrashLock, isLockProcessAlive, formatCrashInfo, emitCrashRecoveredUnitEnd, } from "./crash-recovery.js";
+import { writeLock, clearLock, readCrashLock, isLockProcessAlive, formatCrashInfo, emitCrashRecoveredUnitEnd, emitOpenUnitEndForUnit, } from "./crash-recovery.js";
 import { acquireSessionLock, getSessionLockStatus, releaseSessionLock, updateSessionLock, } from "./session-lock.js";
 import { resolveAutoSupervisorConfig, loadEffectiveGSDPreferences, getIsolationMode, } from "./preferences.js";
 import { sendDesktopNotification } from "./notifications.js";
@@ -61,6 +61,8 @@ import { isClosedStatus } from "./status-guards.js";
 import { updateProgressWidget as _updateProgressWidget, updateSliceProgressCache, clearSliceProgressCache, } from "./auto-dashboard.js";
 import { registerSigtermHandler as _registerSigtermHandler, deregisterSigtermHandler as _deregisterSigtermHandler, } from "./auto-supervisor.js";
 import { isDbAvailable, getMilestone } from "./gsd-db.js";
+import { markLatestActiveForWorkerCanceled } from "./db/unit-dispatches.js";
+import { writeUnitRuntimeRecord } from "./unit-runtime.js";
 import { countPendingCaptures } from "./captures.js";
 import { CMUX_CHANNELS } from "../shared/cmux-events.js";
 import { ensureDbOpen } from "./bootstrap/dynamic-tools.js";
@@ -85,6 +87,16 @@ import { validateDirectory } from "./validate-directory.js";
 import { createAutoOrchestrator } from "./auto/orchestrator.js";
 import { WorktreeResolver, } from "./worktree-resolver.js";
 import { reorderForCaching } from "./prompt-ordering.js";
+import { initTokenCounter } from "./token-counter.js";
+// Warm the tiktoken encoder at extension startup so context-budget computations
+// can use accurate token counts via countTokensSync without paying the load
+// cost mid-prompt-build. Fire-and-forget — failure falls back to the
+// provider-aware char-ratio estimator already used by getCharsPerToken().
+// Catch rejections explicitly: an unhandled rejection at module-import time
+// can destabilize startup before the engine logger is configured.
+void initTokenCounter().catch((err) => {
+    logWarning("engine", `token counter warm-up failed: ${err instanceof Error ? err.message : String(err)}`);
+});
 export { STUB_RECOVERY_THRESHOLD, NEW_SESSION_TIMEOUT_MS, } from "./auto/session.js";
 import { autoSession as s } from "./auto-runtime-state.js";
 import { gsdHome } from "./gsd-home.js";
@@ -112,11 +124,12 @@ const STATE_REBUILD_MIN_INTERVAL_MS = 30_000;
  * the DB is unavailable (e.g. fresh project before init) we skip registration
  * silently rather than blocking session start.
  */
-function registerAutoWorkerForSession(session) {
+function registerAutoWorkerForSession(session, projectRootOverride) {
     if (session.workerId)
         return; // already registered (e.g. resume re-runs)
     try {
-        const projectRootRealpath = normalizeRealPath(session.scope?.workspace.projectRoot
+        const projectRootRealpath = normalizeRealPath(projectRootOverride
+            ?? session.scope?.workspace.projectRoot
             ?? (session.originalBasePath || session.basePath));
         session.workerId = registerAutoWorker({ projectRootRealpath });
     }
@@ -225,8 +238,16 @@ function synthesizePausedSessionRecovery(basePath, unitType, unitId, sessionFile
 export function _synthesizePausedSessionRecoveryForTest(basePath, unitType, unitId, sessionFile) {
     return synthesizePausedSessionRecovery(basePath, unitType, unitId, sessionFile);
 }
+const DETACHED_AUTO_KEEPALIVE_INTERVAL_MS = 30_000;
+function withDetachedAutoKeepalive(run) {
+    const keepAlive = setInterval(() => { }, DETACHED_AUTO_KEEPALIVE_INTERVAL_MS);
+    return run.finally(() => {
+        clearInterval(keepAlive);
+    });
+}
+export const _withDetachedAutoKeepaliveForTest = withDetachedAutoKeepalive;
 export function startAutoDetached(ctx, pi, base, verboseMode, options) {
-    void startAuto(ctx, pi, base, verboseMode, options).catch((err) => {
+    void withDetachedAutoKeepalive(startAuto(ctx, pi, base, verboseMode, options)).catch((err) => {
         const message = getErrorMessage(err);
         ctx.ui.notify(`Auto-start failed: ${message}`, "error");
         logWarning("engine", `auto start error: ${message}`, { file: "auto.ts" });
@@ -264,9 +285,50 @@ export function shouldUseWorktreeIsolation(basePath) {
  */
 // Re-export budget utilities for external consumers
 export { getBudgetAlertLevel, getNewBudgetAlertLevel, getBudgetEnforcementAction, } from "./auto-budget.js";
+function closeOutSignalInterruptedUnit(currentBasePath) {
+    const currentUnit = s.currentUnit;
+    if (!currentUnit)
+        return;
+    const reason = "Auto-mode process received a termination signal";
+    const errorContext = {
+        message: reason,
+        category: "aborted",
+        isTransient: false,
+    };
+    const basePath = s.basePath || currentBasePath;
+    try {
+        emitOpenUnitEndForUnit(basePath, currentUnit.type, currentUnit.id, "cancelled", errorContext);
+    }
+    catch (err) {
+        logWarning("engine", `signal unit-end cleanup failed: ${getErrorMessage(err)}`, { file: "auto.ts" });
+    }
+    try {
+        writeUnitRuntimeRecord(basePath, currentUnit.type, currentUnit.id, currentUnit.startedAt, {
+            phase: "crashed",
+            lastProgressAt: Date.now(),
+            lastProgressKind: "signal",
+        });
+    }
+    catch (err) {
+        logWarning("engine", `signal runtime cleanup failed: ${getErrorMessage(err)}`, { file: "auto.ts" });
+    }
+    try {
+        if (s.workerId)
+            markLatestActiveForWorkerCanceled(s.workerId, "signal-exit");
+    }
+    catch (err) {
+        logWarning("engine", `signal dispatch cleanup failed: ${getErrorMessage(err)}`, { file: "auto.ts" });
+    }
+    try {
+        resolveAgentEndCancelled(errorContext);
+    }
+    catch (err) {
+        logWarning("engine", `signal resolve cleanup failed: ${getErrorMessage(err)}`, { file: "auto.ts" });
+    }
+}
 /** Wrapper: register SIGTERM handler and store reference. */
 function registerSigtermHandler(currentBasePath) {
-    s.sigtermHandler = _registerSigtermHandler(currentBasePath, s.sigtermHandler);
+    s.sigtermHandler = _registerSigtermHandler(currentBasePath, s.sigtermHandler, () => closeOutSignalInterruptedUnit(currentBasePath));
 }
 /** Wrapper: deregister SIGTERM handler and clear reference. */
 function deregisterSigtermHandler() {
@@ -658,6 +720,8 @@ export async function stopAuto(ctx, pi, reason) {
             if (s.workerId) {
                 markWorkerStopping(s.workerId);
             }
+            s.workerId = null;
+            s.milestoneLeaseToken = null;
         }
         catch (e) {
             debugLog("stop-cleanup-coordination", { error: e instanceof Error ? e.message : String(e) });
@@ -1558,6 +1622,10 @@ export async function startAuto(ctx, pi, base, verboseMode, options) {
             : new URL("../../../resource-loader.js", import.meta.url).href;
         const { initResources } = await import(resourceLoaderPath);
         initResources(agentDir);
+        // initResources() uses synchronous fs APIs, so the prompt-template cache
+        // can be primed immediately — no need for the legacy 1s setTimeout deferral.
+        const { primeCache } = await import("./prompt-loader.js");
+        primeCache();
         // Open the project DB before rebuild/derive so resume uses DB-backed
         // state instead of falling back to stale markdown parsing (#2940).
         await openProjectDbIfPresent(s.basePath);
@@ -1623,6 +1691,10 @@ export async function startAuto(ctx, pi, base, verboseMode, options) {
         lockBase,
         buildResolver,
     };
+    // Register the worker before bootstrap enters a milestone worktree.
+    // This ensures enterMilestone can claim a lease and seed dispatch claims
+    // for crash-recovery fidelity (#5405).
+    registerAutoWorkerForSession(s, base);
     const ready = await bootstrapAutoSession(s, ctx, pi, base, verboseMode, requestedStepMode, bootstrapDeps, freshStartAssessment);
     if (!ready)
         return;

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

@@ -118,9 +118,28 @@ export async function handleAgentEnd(pi, event, ctx) {
         return;
     if (!isAutoActive())
         return;
-    if (isSessionSwitchInFlight())
-        return;
     const lastMsg = event.messages[event.messages.length - 1];
+    if (isSessionSwitchInFlight()) {
+        if (lastMsg && "stopReason" in lastMsg && lastMsg.stopReason === "error") {
+            const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+            if (isUserInitiatedAbortMessage(rawErrorMsg)) {
+                resolveAgentEndCancelled({
+                    message: rawErrorMsg,
+                    category: "aborted",
+                    isTransient: false,
+                });
+            }
+        }
+        else if (lastMsg && "stopReason" in lastMsg && lastMsg.stopReason === "aborted") {
+            const content = "content" in lastMsg ? lastMsg.content : undefined;
+            const hasEmptyContent = Array.isArray(content) && content.length === 0;
+            const hasErrorMessage = "errorMessage" in lastMsg && !!lastMsg.errorMessage;
+            if (!hasEmptyContent || hasErrorMessage) {
+                resolveAgentEndCancelled(_buildAbortedPauseContext(lastMsg));
+            }
+        }
+        return;
+    }
     if (lastMsg && "stopReason" in lastMsg && lastMsg.stopReason === "aborted") {
         // Empty content with aborted stopReason is a non-fatal agent stop (the LLM
         // chose to end without producing output). Only pause on genuine fatal aborts

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

@@ -3,21 +3,34 @@
 // Exposes the Context Mode runtime tools in-process. Default-on; opt out with
 // `context_mode.enabled: false` in preferences.
 import { Type } from "@sinclair/typebox";
+async function loadContextModePreferences(baseDir) {
+    const [{ loadEffectiveGSDPreferences }, { logWarning }] = await Promise.all([
+        import("../preferences.js"),
+        import("../workflow-logger.js"),
+    ]);
+    try {
+        return loadEffectiveGSDPreferences(baseDir)?.preferences ?? null;
+    }
+    catch (err) {
+        logWarning("tool", `Context Mode tool could not load preferences: ${err instanceof Error ? err.message : String(err)}`);
+        return null;
+    }
+}
 export function registerExecTools(pi) {
     pi.registerTool({
         name: "gsd_exec",
         label: "Exec (Sandboxed)",
-        description: "Run a short script (bash/node/python) in a subprocess. Full stdout/stderr persist to " +
+        description: "Run a short script (bash/node/python) in a subprocess. Capped stdout/stderr and metadata persist to " +
             ".gsd/exec/<id>.{stdout,stderr,meta.json}; only a short digest returns in context. Use " +
             "this instead of reading many files or emitting large tool outputs — e.g. have the script " +
             "count/grep/summarize and log the finding. Enabled by default; opt out via " +
             "preferences.context_mode.enabled=false.",
-        promptSnippet: "Run a bash/node/python script in a sandbox; full output is saved to disk and only a digest returns",
+        promptSnippet: "Run a bash/node/python script in a sandbox; capped output is saved to disk and only a digest returns",
         promptGuidelines: [
             "Prefer gsd_exec for analyses that would otherwise read >3 files or produce large tool output.",
             "Write scripts that log the finding (counts, matches, summaries) rather than raw dumps.",
             "The digest is the last ~300 chars of stdout — size your log output accordingly.",
-            "Need the full output? Read the stdout_path returned in details (file on local disk).",
+            "Need persisted output? Read the stdout_path returned in details (file on local disk).",
         ],
         parameters: Type.Object({
             runtime: Type.Union([Type.Literal("bash"), Type.Literal("node"), Type.Literal("python")], { description: "Interpreter: bash (-c), node (-e), or python3 (-c)." }),
@@ -30,21 +43,11 @@ export function registerExecTools(pi) {
             })),
         }),
         async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
-            const [{ executeGsdExec }, { loadEffectiveGSDPreferences }, { logWarning }] = await Promise.all([
-                import("../tools/exec-tool.js"),
-                import("../preferences.js"),
-                import("../workflow-logger.js"),
-            ]);
-            let prefs = null;
-            try {
-                prefs = loadEffectiveGSDPreferences();
-            }
-            catch (err) {
-                logWarning("tool", `gsd_exec could not load preferences: ${err instanceof Error ? err.message : String(err)}`);
-            }
+            const { executeGsdExec } = await import("../tools/exec-tool.js");
+            const baseDir = process.cwd();
             return executeGsdExec(params, {
-                baseDir: process.cwd(),
-                preferences: prefs?.preferences ?? null,
+                baseDir,
+                preferences: await loadContextModePreferences(baseDir),
             });
         },
     });
@@ -56,7 +59,7 @@ export function registerExecTools(pi) {
         promptSnippet: "Search prior gsd_exec runs by substring, runtime, or failing-only filter",
         promptGuidelines: [
             "Use this before re-running an expensive analysis — the prior run's stdout file may still answer.",
-            "The preview shows the trailing ~300 chars of stdout; read stdout_path for the full transcript.",
+            "The preview shows the trailing ~300 chars of stdout; read stdout_path for persisted output.",
         ],
         parameters: Type.Object({
             query: Type.Optional(Type.String({ description: "Substring matched against id and purpose (case-insensitive)." })),
@@ -68,8 +71,10 @@ export function registerExecTools(pi) {
         }),
         async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
             const { executeExecSearch } = await import("../tools/exec-search-tool.js");
+            const baseDir = process.cwd();
             return executeExecSearch(params, {
-                baseDir: process.cwd(),
+                baseDir,
+                preferences: await loadContextModePreferences(baseDir),
             });
         },
     });
@@ -87,8 +92,10 @@ export function registerExecTools(pi) {
         parameters: Type.Object({}),
         async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
             const { executeResume } = await import("../tools/resume-tool.js");
+            const baseDir = process.cwd();
             return executeResume(params, {
-                baseDir: process.cwd(),
+                baseDir,
+                preferences: await loadContextModePreferences(baseDir),
             });
         },
     });

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

@@ -52,6 +52,25 @@ async function applyDisabledModelProviderPolicy(ctx) {
         // Non-fatal: keep default provider visibility if preferences cannot be loaded.
     }
 }
+/**
+ * Bridge `context_management.compaction_threshold_percent` from GSD preferences
+ * into the agent's runtime compaction settings (#5475). The preference is
+ * validated to (0.5, 0.95) at load time, but defense-in-depth normalization
+ * here protects against a stale or hand-edited prefs file. Calling with
+ * `undefined` clears any prior override so a removed preference does not leak.
+ */
+async function applyCompactionThresholdOverride(ctx) {
+    try {
+        const { loadEffectiveGSDPreferences } = await import("../preferences.js");
+        const prefs = loadEffectiveGSDPreferences();
+        const raw = prefs?.preferences.context_management?.compaction_threshold_percent;
+        const value = typeof raw === "number" && Number.isFinite(raw) && raw > 0 && raw < 1 ? raw : undefined;
+        ctx.setCompactionThresholdOverride(value);
+    }
+    catch {
+        // Non-fatal: leave any existing override in place.
+    }
+}
 export function resolveNotificationStoreBasePath(cwd = process.cwd()) {
     return resolveWorktreeProjectRoot(cwd);
 }
@@ -101,6 +120,7 @@ export function registerHooks(pi, ecosystemHandlers) {
         await resetAskUserQuestionsTurnCache();
         await syncServiceTierStatus(ctx);
         await applyDisabledModelProviderPolicy(ctx);
+        await applyCompactionThresholdOverride(ctx);
         // Skip MCP auto-prep when running inside an auto-worktree (see session_switch below).
         const { isInAutoWorktree } = await import("../auto-worktree.js");
         if (!isInAutoWorktree(process.cwd())) {
@@ -149,6 +169,7 @@ export function registerHooks(pi, ecosystemHandlers) {
         clearDiscussionFlowState(process.cwd());
         await syncServiceTierStatus(ctx);
         await applyDisabledModelProviderPolicy(ctx);
+        await applyCompactionThresholdOverride(ctx);
         // Skip MCP auto-prep when running inside an auto-worktree. The worktree
         // already has .mcp.json from createAutoWorktree, and re-running the writer
         // post-chdir rewrites the file mid-run (non-idempotent due to cwd-relative