gsd-pi 2.76.0-dev.fe143342a → 2.77.0-dev.538325aea

package/dist/resources/extensions/claude-code-cli/stream-adapter.js

@@ -365,8 +365,9 @@ async function promptElicitationWithDialogs(request, questions, ui, signal) {
     for (const question of questions) {
         const title = buildElicitationPromptTitle(request, question);
         if (question.allowMultiple) {
+            // allowMultiple removed from ExtensionUIDialogOptions in pi 0.67.2;
+            // fall back to single-select (result is wrapped in array below)
             const selected = await ui.select(title, question.options.map((option) => option.label), {
-                allowMultiple: true,
                 signal,
             });
             if (Array.isArray(selected)) {
@@ -484,25 +485,33 @@ export function makeAbortedMessage(model, lastTextContent) {
 /**
  * Resolve the Claude Code permission mode for the current run.
  *
- * GSD subagents run underneath a host Claude Code session the user has
- * already consented to, and their work (edits, shell inspection, MCP calls)
- * spans the full workflow toolset. Defaulting the inner SDK to
- * `bypassPermissions` avoids per-tool approval prompts that offer no
- * meaningful safety beyond what the host session and the subagent prompts
- * already enforce. `GSD_CLAUDE_CODE_PERMISSION_MODE` lets security-conscious
- * users opt into a stricter mode (`acceptEdits`, `default`, `plan`).
+ * Defaults to `acceptEdits`, which auto-approves file reads/edits but
+ * surfaces a permission dialog for dangerous operations (e.g. general Bash,
+ * Agent, WebFetch). This prevents tools outside the allowlist from being
+ * silently denied — the SDK emits an `extension_ui_request` event so the
+ * user sees a prompt instead of a silent refusal that Claude Code mistakes
+ * for user rejection (#4383).
+ *
+ * Set `GSD_CLAUDE_CODE_PERMISSION_MODE` to `bypassPermissions` to restore
+ * the old always-approve behaviour, or to `default` / `plan` for stricter
+ * modes.
  *
- * Tradeoff: bypass means a prompt-injection payload read from an untrusted
- * file could trigger tool calls without a second gate. Accepted for GSD
- * because the workflow is explicit user intent and the alternative
- * (#4099) is continuous approval fatigue that blocks real work.
+ * When `GSD_HEADLESS=1` is set (auto-mode / non-interactive runs), the
+ * default flips to `bypassPermissions` because there is no UI to approve
+ * permission dialogs — `acceptEdits` would hang verification commands like
+ * `npx tsc --noEmit` or `npx vitest run` indefinitely (#4657). Explicit
+ * overrides still win, so users can opt back into `acceptEdits` in headless.
  */
 export async function resolveClaudePermissionMode(env = process.env) {
     const override = env.GSD_CLAUDE_CODE_PERMISSION_MODE?.trim();
     if (override === "bypassPermissions" || override === "acceptEdits" || override === "default" || override === "plan") {
         return override;
     }
-    return "bypassPermissions";
+    if (env.GSD_HEADLESS === "1") {
+        console.warn("[claude-code-cli] Headless mode detected (GSD_HEADLESS=1): defaulting permissionMode to 'bypassPermissions' so verification Bash commands can run. Set GSD_CLAUDE_CODE_PERMISSION_MODE=acceptEdits to opt out.");
+        return "bypassPermissions";
+    }
+    return "acceptEdits";
 }
 // NOTE: These helpers intentionally mirror @gsd/pi-ai anthropic-shared
 // behavior so this extension remains typecheck-stable even when the published
@@ -554,7 +563,7 @@ function mapThinkingLevelToAnthropicEffort(level, modelId) {
 export function buildSdkOptions(modelId, prompt, overrides, extraOptions = {}) {
     const { reasoning, ...sdkExtraOptions } = extraOptions;
     const mcpServers = buildWorkflowMcpServers();
-    const permissionMode = overrides?.permissionMode ?? "bypassPermissions";
+    const permissionMode = overrides?.permissionMode ?? "acceptEdits";
     const disallowedTools = ["AskUserQuestion"];
     // Pre-authorize the safe built-ins and every registered workflow MCP
     // server's tools. `acceptEdits` mode (the interactive default) only
@@ -747,7 +756,8 @@ export function extractToolResultsFromSdkUserMessage(message) {
 /** Attach external tool results from the SDK synthetic user message to their corresponding tool-call blocks by ID. */
 function attachExternalResultsToToolBlocks(toolBlocks, toolResultsById) {
     for (const block of toolBlocks) {
-        if (block.type !== "toolCall" && block.type !== "serverToolUse")
+        // serverToolUse removed from AssistantMessage content in pi 0.67.2
+        if (block.type !== "toolCall")
             continue;
         const externalResult = toolResultsById.get(block.id);
         if (!externalResult)
@@ -900,8 +910,9 @@ async function pumpSdkMessages(model, context, options, stream) {
                             else if (block.type === "thinking" && block.thinking) {
                                 lastThinkingContent = block.thinking;
                             }
-                            else if (block.type === "toolCall" || block.type === "serverToolUse") {
+                            else if (block.type === "toolCall") {
                                 // Collect tool blocks for externalToolExecution rendering
+                                // (serverToolUse removed from pi-ai content types in pi 0.67.2)
                                 intermediateToolBlocks.push(block);
                             }
                         }
@@ -933,13 +944,7 @@ async function pumpSdkMessages(model, context, options, stream) {
                                     partial: builder.message,
                                 });
                             }
-                            else if (block.type === "serverToolUse") {
-                                stream.push({
-                                    type: "server_tool_use",
-                                    contentIndex,
-                                    partial: builder.message,
-                                });
-                            }
+                            // serverToolUse/server_tool_use removed from pi-ai in pi 0.67.2
                         }
                     }
                     builder = null;

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

@@ -30,6 +30,13 @@ function stuckStatePath(basePath) {
 function loadStuckState(basePath) {
     try {
         const data = JSON.parse(readFileSync(stuckStatePath(basePath), "utf-8"));
+        // Only load state written by a DIFFERENT process (real session restart).
+        // If the PID matches the current process, this state was written by an earlier
+        // autoLoop call in the same process (e.g., a test that completed before this
+        // one), not by a crashed session — skip it to prevent test state pollution.
+        if (data.pid === process.pid) {
+            return { recentUnits: [], stuckRecoveryAttempts: 0 };
+        }
         return {
             recentUnits: Array.isArray(data.recentUnits) ? data.recentUnits : [],
             stuckRecoveryAttempts: typeof data.stuckRecoveryAttempts === "number" ? data.stuckRecoveryAttempts : 0,
@@ -45,6 +52,7 @@ function saveStuckState(basePath, state) {
         const filePath = stuckStatePath(basePath);
         mkdirSync(join(gsdRoot(basePath), "runtime"), { recursive: true });
         writeFileSync(filePath, JSON.stringify({
+            pid: process.pid,
             recentUnits: state.recentUnits.slice(-20), // keep last 20 entries
             stuckRecoveryAttempts: state.stuckRecoveryAttempts,
             updatedAt: new Date().toISOString(),
@@ -497,6 +505,7 @@ export async function autoLoop(ctx, pi, s, deps, options) {
             consecutiveCooldowns = 0;
             recentErrorMessages.length = 0;
             deps.emitJournalEvent({ ts: new Date().toISOString(), flowId, seq: nextSeq(), eventType: "iteration-end", data: { iteration } });
+            saveStuckState(s.basePath, loopState); // persist across session restarts (#4382)
             debugLog("autoLoop", { phase: "iteration-complete", iteration });
             finishTurn("completed");
         }

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

@@ -30,7 +30,8 @@ import { isDbAvailable, getMilestoneSlices } from "../gsd-db.js";
 import { ensurePlanV2Graph } from "../uok/plan-v2.js";
 import { resolveUokFlags } from "../uok/flags.js";
 import { UokGateRunner } from "../uok/gate-runner.js";
-import { resetEvidence } from "../safety/evidence-collector.js";
+import { resetEvidence, loadEvidenceFromDisk } from "../safety/evidence-collector.js";
+import { parseUnitId } from "../unit-id.js";
 import { createCheckpoint, cleanupCheckpoint, rollbackToCheckpoint } from "../safety/git-checkpoint.js";
 import { resolveSafetyHarnessConfig } from "../safety/safety-harness.js";
 import { getWorkflowTransportSupportError, getRequiredWorkflowToolsForAutoUnit, supportsStructuredQuestions, } from "../workflow-mcp.js";
@@ -47,7 +48,11 @@ export function resetSessionTimeoutState() {
  * Exported for testing as _resolveReportBasePath.
  */
 export function _resolveReportBasePath(s) {
-    return s.originalBasePath || s.basePath;
+    // Strip /.gsd/worktrees/ suffix when basePath is itself a worktree path and
+    // originalBasePath is falsy — prevents reports landing in the worktree (#3729).
+    const resolved = s.originalBasePath || s.basePath;
+    const markerIdx = resolved.indexOf("/.gsd/worktrees/");
+    return markerIdx !== -1 ? resolved.slice(0, markerIdx) : resolved;
 }
 /**
  * Resolve the authoritative project base for dispatch guards.
@@ -55,7 +60,12 @@ export function _resolveReportBasePath(s) {
  * unit is running inside an auto worktree.
  */
 export function _resolveDispatchGuardBasePath(s) {
-    return s.originalBasePath || s.basePath;
+    // Strip /.gsd/worktrees/ suffix when basePath is itself a worktree path and
+    // originalBasePath is falsy — prevents guard checks running against the
+    // worktree instead of the project root (#3729).
+    const resolved = s.originalBasePath || s.basePath;
+    const markerIdx = resolved.indexOf("/.gsd/worktrees/");
+    return markerIdx !== -1 ? resolved.slice(0, markerIdx) : resolved;
 }
 const PLAN_V2_GATE_PHASES = new Set([
     "executing",
@@ -295,7 +305,7 @@ export async function runPreDispatch(ic, loopState) {
                 findings: reason,
                 milestoneId: state.activeMilestone?.id ?? undefined,
             });
-            ctx.ui.notify(`Plan gate failed-closed: ${reason}`, "error");
+            ctx.ui.notify(`Plan gate failed-closed: ${reason}\n\nIf this keeps happening, try: /gsd doctor heal`, "error");
             await deps.pauseAuto(ctx, pi);
             return { action: "break", reason: "plan-v2-gate-failed" };
         }
@@ -389,6 +399,8 @@ export async function runPreDispatch(ic, loopState) {
         loopState.recentUnits.length = 0;
         loopState.stuckRecoveryAttempts = 0;
         // Worktree lifecycle on milestone transition — merge current, enter next
+        // #2909: preflight — warn + stash dirty working tree before merge
+        const preflightTransition = deps.preflightCleanRoot(s.originalBasePath || s.basePath, s.currentMilestoneId, ctx.ui.notify.bind(ctx.ui));
         try {
             deps.resolver.mergeAndExit(s.currentMilestoneId, ctx.ui);
         }
@@ -405,6 +417,10 @@ export async function runPreDispatch(ic, loopState) {
             await deps.stopAuto(ctx, pi, `Merge error on milestone ${s.currentMilestoneId}: ${String(mergeErr)}`);
             return { action: "break", reason: "merge-failed" };
         }
+        // #2909: postflight — restore stashed changes after successful merge
+        if (preflightTransition.stashPushed) {
+            deps.postflightPopStash(s.originalBasePath || s.basePath, s.currentMilestoneId, ctx.ui.notify.bind(ctx.ui));
+        }
         // PR creation (auto_pr) is handled inside mergeMilestoneToMain (#2302)
         deps.invalidateAllCaches();
         state = await deps.deriveState(s.basePath);
@@ -459,6 +475,8 @@ export async function runPreDispatch(ic, loopState) {
         if (incomplete.length === 0 && state.registry.length > 0) {
             // All milestones complete — merge milestone branch before stopping
             if (s.currentMilestoneId) {
+                // #2909: preflight — warn + stash dirty working tree before merge
+                const preflightAllComplete = deps.preflightCleanRoot(s.originalBasePath || s.basePath, s.currentMilestoneId, ctx.ui.notify.bind(ctx.ui));
                 try {
                     deps.resolver.mergeAndExit(s.currentMilestoneId, ctx.ui);
                     // Prevent stopAuto from attempting the same merge (#2645)
@@ -475,6 +493,10 @@ export async function runPreDispatch(ic, loopState) {
                     await deps.stopAuto(ctx, pi, `Merge error on milestone ${s.currentMilestoneId}: ${String(mergeErr)}`);
                     return { action: "break", reason: "merge-failed" };
                 }
+                // #2909: postflight — restore stashed changes after successful merge
+                if (preflightAllComplete.stashPushed) {
+                    deps.postflightPopStash(s.originalBasePath || s.basePath, s.currentMilestoneId, ctx.ui.notify.bind(ctx.ui));
+                }
                 // PR creation (auto_pr) is handled inside mergeMilestoneToMain (#2302)
             }
             deps.sendDesktopNotification("GSD", "All milestones complete!", "success", "milestone", basename(s.originalBasePath || s.basePath));
@@ -539,6 +561,8 @@ export async function runPreDispatch(ic, loopState) {
     if (state.phase === "complete") {
         // Milestone merge on complete (before closeout so branch state is clean)
         if (s.currentMilestoneId) {
+            // #2909: preflight — warn + stash dirty working tree before merge
+            const preflightComplete = deps.preflightCleanRoot(s.originalBasePath || s.basePath, s.currentMilestoneId, ctx.ui.notify.bind(ctx.ui));
             try {
                 deps.resolver.mergeAndExit(s.currentMilestoneId, ctx.ui);
                 // Prevent stopAuto from attempting the same merge (#2645)
@@ -555,6 +579,10 @@ export async function runPreDispatch(ic, loopState) {
                 await deps.stopAuto(ctx, pi, `Merge error on milestone ${s.currentMilestoneId}: ${String(mergeErr)}`);
                 return { action: "break", reason: "merge-failed" };
             }
+            // #2909: postflight — restore stashed changes after successful merge
+            if (preflightComplete.stashPushed) {
+                deps.postflightPopStash(s.originalBasePath || s.basePath, s.currentMilestoneId, ctx.ui.notify.bind(ctx.ui));
+            }
             // PR creation (auto_pr) is handled inside mergeMilestoneToMain (#2302)
         }
         deps.sendDesktopNotification("GSD", `Milestone ${mid} complete!`, "success", "milestone", basename(s.originalBasePath || s.basePath));
@@ -1028,6 +1056,14 @@ export async function runUnitPhase(ic, iterData, loopState, sidecarItem) {
     const safetyConfig = resolveSafetyHarnessConfig(prefs?.safety_harness);
     if (safetyConfig.enabled && safetyConfig.evidence_collection) {
         resetEvidence();
+        // Restore persisted evidence so session-restart resumes don't produce
+        // false-positive "no bash calls" warnings (Bug #4385).
+        if (s.basePath && unitType === "execute-task") {
+            const { milestone: eMid, slice: eSid, task: eTid } = parseUnitId(unitId);
+            if (eMid && eSid && eTid) {
+                loadEvidenceFromDisk(s.basePath, eMid, eSid, eTid);
+            }
+        }
     }
     // Only checkpoint code-executing units (not lifecycle/planning units)
     if (safetyConfig.enabled && safetyConfig.checkpoints && unitType === "execute-task") {
@@ -1430,6 +1466,20 @@ export async function runFinalize(ic, iterData, loopState, sidecarItem) {
         }
         else {
             // s.pendingVerificationRetry was set by postUnitPreVerification.
+            // Emit a dedicated journal event so forensics can distinguish bounded
+            // verification retries from genuine stuck-loop dispatch repetitions (#4540).
+            const retryInfo = s.pendingVerificationRetry;
+            deps.emitJournalEvent({
+                ts: new Date().toISOString(),
+                flowId: ic.flowId,
+                seq: ic.nextSeq(),
+                eventType: "artifact-verification-retry",
+                data: {
+                    unitType: preUnitSnapshot?.type,
+                    unitId: retryInfo?.unitId,
+                    attempt: retryInfo?.attempt,
+                },
+            });
             // Continue the loop — next iteration will inject the retry context into the prompt.
             debugLog("autoLoop", { phase: "artifact-verification-retry", iteration: ic.iteration });
             return { action: "continue" };

package/dist/resources/extensions/gsd/auto/run-unit.js

@@ -26,15 +26,24 @@ export async function runUnit(ctx, pi, s, unitType, unitId, prompt) {
     let sessionResult;
     let sessionTimeoutHandle;
     const mySessionSwitchGeneration = ++sessionSwitchGeneration;
+    // #3731: Cancellation controller for newSession(). When the session-creation
+    // timeout fires, we abort this controller so that the still-in-flight
+    // newSession() discards itself after await this.abort() completes, preventing
+    // it from capturing the (now-root) process.cwd() and rebuilding the tool
+    // runtime with the wrong cwd.
+    const sessionAbortController = new AbortController();
     _setSessionSwitchInFlight(true);
     try {
-        const sessionPromise = s.cmdCtx.newSession().finally(() => {
+        const sessionPromise = s.cmdCtx.newSession({ abortSignal: sessionAbortController.signal }).finally(() => {
             if (sessionSwitchGeneration === mySessionSwitchGeneration) {
                 _setSessionSwitchInFlight(false);
             }
         });
         const timeoutPromise = new Promise((resolve) => {
-            sessionTimeoutHandle = setTimeout(() => resolve({ cancelled: true }), NEW_SESSION_TIMEOUT_MS);
+            sessionTimeoutHandle = setTimeout(() => {
+                sessionAbortController.abort();
+                resolve({ cancelled: true });
+            }, NEW_SESSION_TIMEOUT_MS);
         });
         sessionResult = await Promise.race([sessionPromise, timeoutPromise]);
     }
@@ -92,6 +101,33 @@ export async function runUnit(ctx, pi, s, unitType, unitId, prompt) {
     catch (e) {
         logWarning("engine", "Failed to chdir to basePath before dispatch", { basePath: s.basePath, error: String(e) });
     }
+    // ── Provider request-readiness pre-check (#4555) ──
+    // Verify the provider can accept requests before dispatching. If the token
+    // has expired since bootstrap, return cancelled immediately so the unit is
+    // not wasted on a guaranteed 401.
+    {
+        const provider = s.currentUnitModel?.provider ?? ctx.model?.provider;
+        const registry = ctx.modelRegistry;
+        if (provider && registry != null && typeof registry.isProviderRequestReady === "function") {
+            let ready = false;
+            try {
+                ready = registry.isProviderRequestReady(provider);
+            }
+            catch {
+                ready = false;
+            }
+            if (!ready) {
+                return {
+                    status: "cancelled",
+                    errorContext: {
+                        message: `Provider ${provider} is not request-ready (login/token expired)`,
+                        category: "provider",
+                        isTransient: false,
+                    },
+                };
+            }
+        }
+    }
     // ── Send the prompt ──
     debugLog("runUnit", { phase: "send-message", unitType, unitId });
     pi.sendMessage({ customType: "gsd-auto", content: prompt, display: s.verbose }, { triggerTurn: true });

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

@@ -79,6 +79,17 @@ export class AutoSession {
     lastStateRebuildAt = 0;
     // ── Sidecar queue ─────────────────────────────────────────────────────
     sidecarQueue = [];
+    // ── Pre-exec gate failure context (#4551) ───────────────────────────
+    /**
+     * Persisted when a pre-execution gate fails on a plan-slice or refine-slice
+     * unit. The planning → plan-slice dispatch rule reads this field and injects
+     * the failure details into the next re-dispatch prompt so the LLM can fix the
+     * specific issues instead of producing an identical plan.
+     *
+     * Cleared after it has been consumed (injected into the prompt) to avoid
+     * stale context bleeding into unrelated slices.
+     */
+    lastPreExecFailure = null;
     // ── Tool invocation errors (#2883) ──────────────────────────────────
     /** Set when a GSD tool execution ends with isError due to malformed/truncated
      *  JSON arguments. Checked by postUnitPreVerification to break retry loops. */
@@ -139,7 +150,12 @@ export class AutoSession {
         this.unitLifetimeDispatches.clear();
     }
     get lockBasePath() {
-        return this.originalBasePath || this.basePath;
+        // Prefer originalBasePath (project root); fall back to basePath.
+        // Strip /.gsd/worktrees/ suffix if basePath is itself a worktree path
+        // to avoid reading/writing the lock inside the worktree (#3729).
+        const resolved = this.originalBasePath || this.basePath;
+        const markerIdx = resolved.indexOf("/.gsd/worktrees/");
+        return markerIdx !== -1 ? resolved.slice(0, markerIdx) : resolved;
     }
     reset() {
         this.clearTimers();
@@ -199,6 +215,7 @@ export class AutoSession {
         this.sidecarQueue = [];
         this.rewriteAttemptCount = 0;
         this.consecutiveCompleteBootstraps = 0;
+        this.lastPreExecFailure = null;
         this.lastToolInvocationError = null;
         this.lastGitActionFailure = null;
         this.lastGitActionStatus = null;

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

@@ -17,7 +17,7 @@ import { parseRoadmap } from "./parsers-legacy.js";
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { logWarning, logError } from "./workflow-logger.js";
 import { join } from "node:path";
-import { hasImplementationArtifacts } from "./auto-recovery.js";
+import { hasImplementationArtifacts, classifyMilestoneSummaryContent } from "./auto-recovery.js";
 import { buildDiscussMilestonePrompt, buildResearchMilestonePrompt, buildPlanMilestonePrompt, buildResearchSlicePrompt, buildPlanSlicePrompt, buildRefineSlicePrompt, buildExecuteTaskPrompt, buildCompleteSlicePrompt, buildCompleteMilestonePrompt, buildValidateMilestonePrompt, buildReplanSlicePrompt, buildRunUatPrompt, buildReassessRoadmapPrompt, buildRewriteDocsPrompt, buildReactiveExecutePrompt, buildGateEvaluatePrompt, buildParallelResearchSlicesPrompt, checkNeedsReassessment, checkNeedsRunUat, } from "./auto-prompts.js";
 import { resolveModelWithFallbacksForUnit } from "./preferences-models.js";
 import { resolveUokFlags } from "./uok/flags.js";
@@ -451,18 +451,31 @@ export const DISPATCH_RULES = [
     },
     {
         name: "planning → plan-slice",
-        match: async ({ state, mid, midTitle, basePath, sessionContextWindow, modelRegistry }) => {
+        match: async ({ state, mid, midTitle, basePath, sessionContextWindow, modelRegistry, session }) => {
             if (state.phase !== "planning")
                 return null;
             if (!state.activeSlice)
                 return missingSliceStop(mid, state.phase);
             const sid = state.activeSlice.id;
             const sTitle = state.activeSlice.title;
+            // #4551: Consume any persisted pre-exec failure for this slice so the
+            // re-dispatched prompt includes the exact blocked references. Clear the
+            // field immediately after reading to prevent stale context leaking into
+            // a later, unrelated plan-slice run.
+            const unitId = `${mid}/${sid}`;
+            let priorPreExecFailure;
+            if (session?.lastPreExecFailure?.unitId === unitId) {
+                priorPreExecFailure = {
+                    blockingFindings: session.lastPreExecFailure.blockingFindings,
+                    verdictExcerpt: session.lastPreExecFailure.verdictExcerpt,
+                };
+                session.lastPreExecFailure = null;
+            }
             return {
                 action: "dispatch",
                 unitType: "plan-slice",
-                unitId: `${mid}/${sid}`,
-                prompt: await buildPlanSlicePrompt(mid, midTitle, sid, sTitle, basePath, undefined, { sessionContextWindow, modelRegistry }),
+                unitId,
+                prompt: await buildPlanSlicePrompt(mid, midTitle, sid, sTitle, basePath, undefined, { sessionContextWindow, modelRegistry, priorPreExecFailure }),
             };
         },
     },
@@ -689,22 +702,13 @@ export const DISPATCH_RULES = [
                     return { action: "skip" };
                 }
             }
-            // Reconciliation guard (#4324): when the SUMMARY file already exists
-            // on disk but the DB says the milestone is not complete, the DB is
-            // out of sync (e.g. journal reset, partial merge, crash recovery).
-            // Reconcile the DB status directly instead of re-dispatching the
-            // tool, which would overwrite the richer on-disk SUMMARY with a
-            // thinner regenerated version — causing silent data loss.
             const existingSummary = resolveMilestoneFile(basePath, mid, "SUMMARY");
+            let summaryOutcome = "unknown";
             if (existingSummary && isDbAvailable()) {
-                try {
-                    updateMilestoneStatus(mid, "complete", new Date().toISOString());
-                    logWarning("dispatch", `Milestone ${mid} has SUMMARY on disk but DB status was not complete — reconciled DB to complete (#4324)`);
-                }
-                catch (err) {
-                    logWarning("dispatch", `Failed to reconcile milestone ${mid} status: ${err instanceof Error ? err.message : String(err)}`);
+                const summaryContent = await loadFile(existingSummary);
+                if (summaryContent) {
+                    summaryOutcome = classifyMilestoneSummaryContent(summaryContent);
                 }
-                return { action: "skip" };
             }
             // Safety guard (#2675): block completion when VALIDATION verdict is
             // needs-remediation. The state machine treats needs-remediation as
@@ -782,6 +786,41 @@ export const DISPATCH_RULES = [
             catch (err) { /* fall through — don't block on DB errors */
                 logWarning("dispatch", `verification class check failed: ${err instanceof Error ? err.message : String(err)}`);
             }
+            // Disk/DB mismatch handling (#4658): SUMMARY presence alone is not enough.
+            // Apply post-gate policy:
+            // - success summary: reconcile DB and skip re-dispatch
+            // - failure summary: pause/fail-closed
+            // - unknown summary: pause/fail-closed
+            if (existingSummary && isDbAvailable()) {
+                const milestone = getMilestone(mid);
+                const status = milestone?.status ?? "missing";
+                if (summaryOutcome === "success") {
+                    try {
+                        updateMilestoneStatus(mid, "complete", new Date().toISOString());
+                        logWarning("dispatch", `Milestone ${mid} SUMMARY indicates completion while DB status was "${status}" — reconciled DB to complete (#4658)`);
+                        return { action: "skip" };
+                    }
+                    catch (err) {
+                        return {
+                            action: "stop",
+                            level: "warning",
+                            reason: `Milestone ${mid} SUMMARY indicates completion but DB reconciliation failed (${err instanceof Error ? err.message : String(err)}). Auto-mode paused for manual review.`,
+                        };
+                    }
+                }
+                if (summaryOutcome === "failure") {
+                    return {
+                        action: "stop",
+                        level: "warning",
+                        reason: `Milestone ${mid} has a failure-path SUMMARY while DB status is "${status}". Auto-mode will not promote completion from failure artifacts. Re-run complete-milestone only after blockers are resolved and verification passes.`,
+                    };
+                }
+                return {
+                    action: "stop",
+                    level: "warning",
+                    reason: `Milestone ${mid} has an ambiguous SUMMARY while DB status is "${status}". Auto-mode paused instead of promoting completion from file presence alone.`,
+                };
+            }
             return {
                 action: "dispatch",
                 unitType: "complete-milestone",

package/dist/resources/extensions/gsd/auto-model-selection.js

@@ -291,7 +291,7 @@ autoModeStartThinkingLevel) {
                 // ADR-005: Adjust active tool set for the selected model's provider capabilities.
                 // Hard-filter incompatible tools, then let extensions override via adjust_tool_set hook.
                 const activeToolNames = pi.getActiveTools();
-                const { toolNames: compatibleTools, removedTools } = adjustToolSet(activeToolNames, model.api);
+                const { toolNames: compatibleTools, removedTools } = adjustToolSet(activeToolNames, model.api, model.provider);
                 let finalToolNames = compatibleTools;
                 // Fire adjust_tool_set hook — extensions can override the filtered tool set
                 if (routingConfig.hooks !== false) {

package/dist/resources/extensions/gsd/auto-post-unit.js

@@ -30,7 +30,7 @@ import { checkPostUnitHooks, isRetryPending, consumeRetryTrigger, persistHookSta
 import { hasPendingCaptures, loadPendingCaptures, revertExecutorResolvedCaptures } from "./captures.js";
 import { debugLog } from "./debug-logger.js";
 import { runSafely } from "./auto-utils.js";
-import { getEvidence } from "./safety/evidence-collector.js";
+import { getEvidence, clearEvidenceFromDisk } from "./safety/evidence-collector.js";
 import { validateFileChanges } from "./safety/file-change-validator.js";
 // crossReferenceEvidence available for future use when verification_evidence is stored in DB
 // import { crossReferenceEvidence, type ClaimedEvidence } from "./safety/evidence-cross-ref.js";
@@ -45,8 +45,23 @@ import { ensureCodebaseMapFresh } from "./codebase-generator.js";
 import { resolveUokFlags } from "./uok/flags.js";
 import { UokGateRunner } from "./uok/gate-runner.js";
 import { writeTurnGitTransaction } from "./uok/gitops.js";
+import { isClosedStatus } from "./status-guards.js";
 /** Maximum verification retry attempts before escalating to blocker placeholder (#2653). */
 const MAX_VERIFICATION_RETRIES = 3;
+const COMPLETE_MILESTONE_DB_SETTLE_MS = 1500;
+const COMPLETE_MILESTONE_DB_SETTLE_POLL_MS = 100;
+async function waitForMilestoneDbClose(mid) {
+    const deadline = Date.now() + COMPLETE_MILESTONE_DB_SETTLE_MS;
+    while (Date.now() < deadline) {
+        if (!isDbAvailable())
+            return false;
+        const milestone = getMilestone(mid);
+        if (milestone && isClosedStatus(milestone.status))
+            return true;
+        await new Promise((resolve) => setTimeout(resolve, COMPLETE_MILESTONE_DB_SETTLE_POLL_MS));
+    }
+    return false;
+}
 /** Enqueue a sidecar item (hook, triage, or quick-task) for the main loop to
  *  drain via runUnit. Logs the enqueue event and notifies the UI. */
 function enqueueSidecar(s, ctx, entry, debugExtra, notification) {
@@ -535,7 +550,7 @@ export async function postUnitPreVerification(pctx, opts) {
                         if (taskRow) {
                             const expectedOutput = taskRow.expected_output ?? [];
                             const plannedFiles = taskRow.files ?? [];
-                            const audit = validateFileChanges(s.basePath, expectedOutput, plannedFiles);
+                            const audit = validateFileChanges(s.basePath, expectedOutput, plannedFiles, safetyConfig.file_change_allowlist);
                             if (audit && audit.violations.length > 0) {
                                 const warnings = audit.violations.filter(v => v.severity === "warning");
                                 for (const v of warnings) {
@@ -588,6 +603,16 @@ export async function postUnitPreVerification(pctx, opts) {
                         debugLog("postUnit", { phase: "safety-content-validation", error: String(e) });
                     }
                 }
+                // Clear persisted evidence file now that post-unit processing is complete
+                // (Bug #4385 — prevents stale evidence from affecting retries of same unit ID).
+                if (safetyConfig.evidence_collection && s.currentUnit.type === "execute-task" && sMid && sSid && sTid) {
+                    try {
+                        clearEvidenceFromDisk(s.basePath, sMid, sSid, sTid);
+                    }
+                    catch (e) {
+                        debugLog("postUnit", { phase: "safety-evidence-clear", error: String(e) });
+                    }
+                }
             }
         }
         catch (e) {
@@ -607,6 +632,25 @@ export async function postUnitPreVerification(pctx, opts) {
             }
             // If verification failed, attempt to regenerate missing projection files
             // from DB data before giving up (e.g. research-slice produces PLAN from engine).
+            if (!triggerArtifactVerified) {
+                if (s.currentUnit.type === "complete-milestone") {
+                    try {
+                        const { milestone: mid } = parseUnitId(s.currentUnit.id);
+                        if (mid) {
+                            const settled = await waitForMilestoneDbClose(mid);
+                            if (settled) {
+                                triggerArtifactVerified = verifyExpectedArtifact(s.currentUnit.type, s.currentUnit.id, s.basePath);
+                                if (triggerArtifactVerified) {
+                                    invalidateAllCaches();
+                                }
+                            }
+                        }
+                    }
+                    catch (e) {
+                        debugLog("postUnit", { phase: "artifact-verify-settle-db", error: String(e) });
+                    }
+                }
+            }
             if (!triggerArtifactVerified) {
                 try {
                     const { milestone: mid, slice: sid } = parseUnitId(s.currentUnit.id);
@@ -950,12 +994,25 @@ export async function postUnitPostVerification(pctx) {
                     const suffix = blockingChecks.length > 3 ? `\n  \u2022 ...and ${blockingChecks.length - 3} more` : "";
                     const evidenceNote = `\nSee ${sid}-PRE-EXEC-VERIFY.json for full details.`;
                     ctx.ui.notify(`Pre-execution checks failed: ${blockingCount} blocking issue${blockingCount === 1 ? "" : "s"} found\n${details}${suffix}${evidenceNote}`, "error");
+                    // Persist failure context so the next plan-slice re-dispatch can inject
+                    // it into the prompt and break the infinite loop (#4551).
+                    s.lastPreExecFailure = {
+                        unitId: currentUnit.id,
+                        blockingFindings: blockingChecks.map(c => `[${c.category}] ${c.target}: ${c.message}`),
+                        verdictExcerpt: `status=${result.status}; ${blockingCount} blocking issue${blockingCount === 1 ? "" : "s"} detected`,
+                    };
                     preExecPauseNeeded = true;
                 }
                 else if (result.status === "warn") {
                     ctx.ui.notify(`Pre-execution checks passed with warnings`, "warning");
                     // Strict mode: treat warnings as blocking
                     if (prefs?.enhanced_verification_strict === true) {
+                        const warnChecks = result.checks.filter(c => !c.passed);
+                        s.lastPreExecFailure = {
+                            unitId: currentUnit.id,
+                            blockingFindings: warnChecks.map(c => `[${c.category}] ${c.target}: ${c.message}`),
+                            verdictExcerpt: `status=${result.status} (strict mode); ${warnChecks.length} warning${warnChecks.length === 1 ? "" : "s"} treated as blocking`,
+                        };
                         preExecPauseNeeded = true;
                     }
                 }

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

@@ -1223,6 +1223,20 @@ export async function buildPlanSlicePrompt(mid, _midTitle, sid, sTitle, base, le
         prependBlocks.push(`## Prior Sketch Scope (soft hint — non-binding)\n\n${options.softScopeHint.trim()}\n\n` +
             `This scope was captured during an earlier progressive-planning pass that was later disabled. Treat it as context only — you may plan beyond it if the work genuinely requires more scope. Do NOT treat this as a hard boundary.`);
     }
+    // #4551: inject pre-exec failure context so the re-dispatched plan-slice
+    // addresses the exact blocked references rather than reproducing the same plan.
+    if (options?.priorPreExecFailure) {
+        const { blockingFindings, verdictExcerpt } = options.priorPreExecFailure;
+        const findingsList = blockingFindings.length > 0
+            ? blockingFindings.map(f => `- ${f}`).join("\n")
+            : "- (no specific findings recorded)";
+        prependBlocks.push(`## Fix these specific issues from the prior pre-exec check\n\n` +
+            `The previous plan-slice attempt was blocked by pre-execution validation.\n` +
+            `Gate verdict: ${verdictExcerpt}\n\n` +
+            `Blocked references that must be resolved in this plan:\n${findingsList}\n\n` +
+            `Revise the plan so that every reference listed above is satisfied before execution begins. ` +
+            `Do not reproduce the same file paths, package names, or task ordering that caused these failures.`);
+    }
     return renderSlicePrompt({
         mid, sid, sTitle, base,
         level: level ?? resolveInlineLevel(),