gsd-pi 2.82.0 → 3.0.0

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

@@ -32,6 +32,7 @@ import { shouldIgnoreAgentEndForActiveUnit } from "../auto/unit-runner-events.js
 import { resolveModelId } from "../auto-model-selection.js";
 import { resolveProjectRoot } from "../worktree.js";
 import { clearDiscussionFlowState } from "./write-gate.js";
+import { clearGuidedUnitContext } from "../guided-unit-context.js";
 import { resumeAutoAfterProviderDelay } from "./provider-error-resume.js";
 import {
   classifyError,
@@ -48,6 +49,11 @@ const MAX_NETWORK_RETRIES = 2;
 function isObjectRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === "object";
 }
+
+export function _hasEmptyAgentEndContent(content: unknown): boolean {
+  return content == null || (Array.isArray(content) && content.length === 0);
+}
+
 /**
  * Cap on auto-resume attempts for sustained transient-provider errors.
  *
@@ -94,6 +100,14 @@ export function isUserInitiatedAbortMessage(message: string | undefined | null):
   return /\b(?:claude code process aborted by user|request aborted by user|process aborted by user)\b/i.test(message);
 }
 
+export function shouldDeferTransientErrorToCoreRetry(
+  cls: ErrorClass,
+  rawErrorMsg: string,
+): boolean {
+  if (!isTransient(cls) || cls.kind === "rate-limit") return false;
+  return !/retry failed after \d+ attempts:/i.test(rawErrorMsg);
+}
+
 function isBareClaudeCodeSessionSwitchAbortMarker(message: string | undefined | null): boolean {
   if (!message) return false;
   const normalized = message.trim().replace(/\s+/g, " ").toLowerCase();
@@ -196,6 +210,14 @@ export function resolveAgentEndErrorDisplay(
   return rawErrorMsg;
 }
 
+export function isTerminalDeletedWorktreeProviderError(
+  message: string | undefined | null,
+): boolean {
+  if (!message) return false;
+  if (!/\bdoes not exist\b/i.test(message)) return false;
+  return /[/\\]\.gsd[/\\](?:projects[/\\][^/\\]+[/\\])?worktrees[/\\][^/\\\s"']+/i.test(message);
+}
+
 async function pauseTransientWithBackoff(
   cls: ErrorClass,
   pi: ExtensionAPI,
@@ -244,9 +266,11 @@ export async function handleAgentEnd(
   // falsely report files as missing — producing a spurious "ready signal
   // rejected" loop even though the files are on disk.
   clearPathCache();
+  const basePath = resolveAgentEndBasePath();
+  clearGuidedUnitContext(basePath);
 
   try {
-    if (await checkDeepProjectSetupAfterTurn(event, ctx, resolveAgentEndBasePath())) {
+    if (await checkDeepProjectSetupAfterTurn(event, ctx, basePath)) {
       return;
     }
   } catch (err) {
@@ -254,8 +278,8 @@ export async function handleAgentEnd(
     logWarning("bootstrap", `checkDeepProjectSetupAfterTurn failed: ${message}`);
   }
 
-  if (checkAutoStartAfterDiscuss()) {
-    clearDiscussionFlowState(resolveAgentEndBasePath() ?? process.cwd());
+  if (checkAutoStartAfterDiscuss(basePath)) {
+    clearDiscussionFlowState(basePath ?? process.cwd());
     return;
   }
 
@@ -263,14 +287,14 @@ export async function handleAgentEnd(
   // are missing, `checkAutoStartAfterDiscuss` returns false silently. Surface
   // that and nudge the LLM to complete the writes before the user hits the
   // downstream "All milestones complete" warning loop.
-  if (maybeHandleReadyPhraseWithoutFiles(event)) return;
+  if (maybeHandleReadyPhraseWithoutFiles(event, basePath)) return;
 
   // #4573 — Empty-turn recovery: if the LLM announced intent in prose but
   // emitted no tool calls, nudge it to execute. Fires only when auto-mode is
   // active or a discussion autostart is pending (non-auto interactive discuss
   // is user-driven). Runs before `isAutoActive` early return so pending
   // discussions (where isAutoActive may be false) still get recovered.
-  if (maybeHandleEmptyIntentTurn(event, isAutoActive())) return;
+  if (maybeHandleEmptyIntentTurn(event, isAutoActive(), basePath)) return;
 
   if (!isAutoActive()) return;
 
@@ -310,7 +334,7 @@ export async function handleAgentEnd(
     // that carry error context — e.g. errorMessage field or non-empty content
     // indicating a mid-stream failure. (#2695)
     const content = "content" in lastMsg ? lastMsg.content : undefined;
-    const hasEmptyContent = Array.isArray(content) && content.length === 0;
+    const hasEmptyContent = _hasEmptyAgentEndContent(content);
     const hasErrorMessage = "errorMessage" in lastMsg && !!lastMsg.errorMessage;
 
     if (hasEmptyContent && !hasErrorMessage) {
@@ -355,6 +379,17 @@ export async function handleAgentEnd(
       rawErrorMsg,
       "content" in lastMsg ? lastMsg.content : undefined,
     );
+    if (
+      isAutoCompletionStopInProgress() &&
+      isTerminalDeletedWorktreeProviderError(`${rawErrorMsg}\n${displayMsg}`)
+    ) {
+      resetRetryState(retryState);
+      logWarning(
+        "bootstrap",
+        `Ignoring stale deleted-worktree provider error during terminal completion reroot: ${displayMsg || rawErrorMsg}`,
+      );
+      return;
+    }
     const errorDetail = displayMsg ? `: ${displayMsg}` : "";
     const explicitRetryAfterMs = ("retryAfterMs" in lastMsg && typeof lastMsg.retryAfterMs === "number") ? lastMsg.retryAfterMs : undefined;
 
@@ -461,7 +496,7 @@ export async function handleAgentEnd(
     // Core retries transient failures in-session after this handler.
     // Keep that behavior for non-rate-limit classes to avoid pause/retry races,
     // but let rate-limit continue into model fallback logic below (#4373).
-    if (isTransient(cls) && cls.kind !== "rate-limit") {
+    if (shouldDeferTransientErrorToCoreRetry(cls, rawErrorMsg)) {
       return;
     }
 

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

@@ -487,17 +487,18 @@ export function registerDbTools(pi: ExtensionAPI): void {
     promptGuidelines: [
       "Use gsd_plan_milestone for milestone planning instead of writing ROADMAP.md directly.",
       "Keep parameters flat and provide the full milestone planning payload, including slices.",
+      "Milestone and slice titles must not contain forward slash (/), en dash, or em dash characters.",
       "The tool validates input, writes milestone and slice planning data transactionally, renders ROADMAP.md from DB, and clears both state and parse caches after success.",
       "Use the canonical name gsd_plan_milestone; gsd_milestone_plan is only an alias.",
     ],
     parameters: Type.Object({
       // ── Core identification + content (required) ──────────────────────
       milestoneId: Type.String({ description: "Milestone ID (e.g. M001)" }),
-      title: Type.String({ description: "Milestone title" }),
+      title: Type.String({ description: "Milestone title; must not contain forward slash (/), en dash, or em dash characters" }),
       vision: Type.String({ description: "Milestone vision" }),
       slices: Type.Array(Type.Object({
         sliceId: Type.String({ description: "Slice ID (e.g. S01)" }),
-        title: Type.String({ description: "Slice title" }),
+        title: Type.String({ description: "Slice title; must not contain forward slash (/), en dash, or em dash characters" }),
         risk: Type.String({ description: "Slice risk" }),
         depends: Type.Array(Type.String(), { description: "Slice dependency IDs" }),
         demo: Type.String({ description: "Roadmap demo text / After this" }),
@@ -570,10 +571,10 @@ export function registerDbTools(pi: ExtensionAPI): void {
         title: Type.String({ description: "Task title" }),
         description: Type.String({ description: "Task description / steps block" }),
         estimate: Type.String({ description: "Task estimate string" }),
-        files: Type.Array(Type.String(), { description: "Files likely touched" }),
+        files: Type.Array(Type.String(), { description: "Array<string> of files likely touched; pass [\"path\"] or [], never a single string" }),
         verify: Type.String({ description: "Verification command or block" }),
-        inputs: Type.Array(Type.String(), { description: "Input files or references" }),
-        expectedOutput: Type.Array(Type.String(), { description: "Expected output files or artifacts" }),
+        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 expected output files or artifacts; pass [\"path\"] or [], never a single string" }),
         observabilityImpact: Type.Optional(Type.String({ description: "Task observability impact" })),
       }), { description: "Planned tasks for the slice" }),
       // ── Enrichment metadata (optional — defaults to empty) ────────────
@@ -650,10 +651,10 @@ export function registerDbTools(pi: ExtensionAPI): void {
       title: Type.String({ description: "Task title" }),
       description: Type.String({ description: "Task description / steps block" }),
       estimate: Type.String({ description: "Task estimate string" }),
-      files: Type.Array(Type.String(), { description: "Files likely touched" }),
+      files: Type.Array(Type.String(), { description: "Array<string> of files likely touched; pass [\"path\"] or [], never a single string" }),
       verify: Type.String({ description: "Verification command or block" }),
-      inputs: Type.Array(Type.String(), { description: "Input files or references" }),
-      expectedOutput: Type.Array(Type.String(), { description: "Expected output files or artifacts" }),
+      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 expected output files or artifacts; pass [\"path\"] or [], never a single string" }),
       observabilityImpact: Type.Optional(Type.String({ description: "Task observability impact" })),
       // Single-writer v3 audit trail (Stream 2): caller-provided actor identity + causation.
       actorName: Type.Optional(Type.String({ description: "Caller-provided actor identity for the audit trail (e.g. 'executor-01', 'gsd-orchestrator')" })),
@@ -912,7 +913,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
     label: "Skip Slice",
     description:
       "Mark a slice as skipped so auto-mode advances past it without executing. " +
-      "Non-closed tasks within the slice are cascaded to skipped so milestone completion is not blocked by leftover pending tasks (#4375). " +
+      "Non-closed tasks within the slice are cascaded to skipped so milestone completion is not blocked by leftover pending tasks. " +
       "The slice data is preserved for reference. The state machine treats skipped slices like completed ones for dependency satisfaction.",
     promptSnippet: "Skip a GSD slice (mark as skipped, auto-mode will advance past it)",
     promptGuidelines: [

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

@@ -31,6 +31,7 @@ import { resolveWorktreeProjectRoot } from "../worktree-root.js";
 import { extractSubagentAgentClasses } from "./subagent-input.js";
 import { approvalGateIdForUnit, isExplicitApprovalResponse, shouldPauseForUserApprovalQuestion } from "../user-input-boundary.js";
 import { resolveSkillManifest } from "../skill-manifest.js";
+import { getGuidedUnitContext } from "../guided-unit-context.js";
 
 let approvalQuestionAbortInFlight = false;
 
@@ -423,6 +424,10 @@ export function registerHooks(
   pi: ExtensionAPI,
   ecosystemHandlers: GSDEcosystemBeforeAgentStartHandler[],
 ): void {
+  // ADR-005 Phase 3b: surface pi-ai ProviderSwitchReport via audit, notification, and counter.
+  // Idempotent — only the first registerHooks call installs.
+  void import("../provider-switch-observer.js").then((m) => m.installProviderSwitchObserver());
+
   pi.on("session_start", async (_event, ctx) => {
     const basePath = contextBasePath(ctx);
     initSessionNotifications(ctx);
@@ -480,6 +485,8 @@ export function registerHooks(
     }
     await loadToolApiKeysForSession();
     if (!isAutoActive()) {
+      ctx.ui.setWidget("gsd-progress", undefined);
+      ctx.ui.setWidget("gsd-outcome", undefined);
       const { initHealthWidget } = await import("../health-widget.js");
       initHealthWidget(ctx);
     } else {
@@ -766,7 +773,8 @@ export function registerHooks(
     // subagent dispatch. Closes the b23 bug class where a discuss-milestone
     // turn used the host Edit tool to modify user source files.
     const dash = getAutoRuntimeSnapshot();
-    const activeUnitType = dash.currentUnit?.type;
+    const guidedUnit = getGuidedUnitContext(discussionBasePath);
+    const activeUnitType = dash.currentUnit?.type ?? guidedUnit?.unitType;
     if (activeUnitType) {
       const manifest = resolveManifest(activeUnitType);
       if (manifest) {
@@ -785,7 +793,7 @@ export function registerHooks(
         const planningGuard = shouldBlockPlanningUnit(
           event.toolName,
           planningInput,
-          dash.basePath || discussionBasePath,
+          dash.basePath || guidedUnit?.basePath || discussionBasePath,
           activeUnitType,
           manifest.tools,
           agentClasses,

package/src/resources/extensions/gsd/bootstrap/subagent-input.ts

@@ -1,20 +1,32 @@
 export function extractSubagentAgentClasses(input: unknown): string[] {
   if (!input || typeof input !== "object") return [];
 
-  const record = input as Record<string, unknown>;
   const agentClasses: string[] = [];
+  const visited = new WeakSet<object>();
   const addAgentClass = (value: unknown): void => {
-    if (typeof value === "string" && value.trim().length > 0) agentClasses.push(value.trim());
+    if (typeof value !== "string") return;
+    const normalized = value.trim().replace(/\.md$/i, "");
+    if (normalized.length > 0) agentClasses.push(normalized);
   };
-  const addFromItems = (value: unknown): void => {
+
+  const visitItems = (value: unknown): void => {
     if (!Array.isArray(value)) return;
     for (const item of value) {
-      if (item && typeof item === "object") addAgentClass((item as Record<string, unknown>).agent);
+      visit(item);
     }
   };
 
-  addAgentClass(record.agent);
-  addFromItems(record.tasks);
-  addFromItems(record.chain);
+  const visit = (value: unknown): void => {
+    if (!value || typeof value !== "object") return;
+    if (visited.has(value)) return;
+    visited.add(value);
+    const record = value as Record<string, unknown>;
+    addAgentClass(record.agent);
+    visitItems(record.tasks);
+    visitItems(record.chain);
+    visitItems(record.parallel);
+  };
+
+  visit(input);
   return agentClasses;
 }

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

@@ -13,6 +13,7 @@ import { resolveAllSkillReferences, renderPreferencesForSystemPrompt, loadEffect
 import { resolveModelWithFallbacksForUnit } from "../preferences-models.js";
 import { resolveSkillReference } from "../preferences-skills.js";
 import { resolveGsdRootFile, resolveSliceFile, resolveSlicePath, resolveTaskFile, resolveTaskFiles, resolveTasksDir, relSliceFile, relSlicePath, relTaskFile } from "../paths.js";
+import { extractIntroAndRules } from "../knowledge-parser.js";
 import { ensureCodebaseMapFresh, readCodebaseMap } from "../codebase-generator.js";
 import { hasSkillSnapshot, detectNewSkills, formatSkillsXml } from "../skill-discovery.js";
 import { getActiveAutoWorktreeContext } from "../auto-worktree.js";
@@ -131,9 +132,14 @@ export async function buildBeforeAgentStartResult(
     logWarning("bootstrap", `cmux prompt setup skipped: ${(e as Error).message}`);
   }
 
+  const ctxProjectRoot = (ctx as { projectRoot?: unknown }).projectRoot;
+  const basePath = typeof ctxProjectRoot === "string" && ctxProjectRoot.length > 0
+    ? ctxProjectRoot
+    : process.cwd();
+
   let preferenceBlock = "";
   if (loadedPreferences) {
-    const cwd = process.cwd();
+    const cwd = basePath;
     const report = resolveAllSkillReferences(loadedPreferences.preferences, cwd);
     preferenceBlock = `\n\n${renderPreferencesForSystemPrompt(loadedPreferences.preferences, report.resolutions)}`;
     if (report.warnings.length > 0) {
@@ -144,14 +150,6 @@ export async function buildBeforeAgentStartResult(
     }
   }
 
-  const { block: knowledgeBlock, globalSizeKb } = loadKnowledgeBlock(gsdHome(), process.cwd());
-  if (globalSizeKb > 4) {
-    ctx.ui.notify(
-      `GSD: ~/.gsd/agent/KNOWLEDGE.md is ${globalSizeKb.toFixed(1)}KB — consider trimming to keep system prompt lean.`,
-      "warning",
-    );
-  }
-
   // ADR-013 step 5: opportunistic decisions->memories backfill. Idempotent
   // and best-effort — first run absorbs the existing decisions table into
   // the memory store; subsequent runs are a single sentinel SELECT.
@@ -159,12 +157,50 @@ export async function buildBeforeAgentStartResult(
     const { backfillDecisionsToMemories } = await import("../memory-backfill.js");
     const written = backfillDecisionsToMemories();
     if (written > 0) {
-      ctx.ui.notify(`GSD: backfilled ${written} decision${written === 1 ? "" : "s"} into the memory store (ADR-013).`, "info");
+      ctx.ui.notify(`GSD: backfilled ${written} decision${written === 1 ? "" : "s"} into the memory store.`, "info");
     }
   } catch (e) {
     logWarning("bootstrap", `decisions backfill failed: ${(e as Error).message}`);
   }
 
+  // ADR-013 Stage 2b: KNOWLEDGE.md Patterns + Lessons backfill, then
+  // re-render the hybrid projection (manual Rules + projected Patterns +
+  // projected Lessons). Both are idempotent and best-effort — failures here
+  // can't block agent startup.
+  try {
+    const { backfillKnowledgeToMemories } = await import("../knowledge-backfill.js");
+    const writtenK = backfillKnowledgeToMemories(basePath);
+    if (writtenK > 0) {
+      ctx.ui.notify(`GSD: backfilled ${writtenK} KNOWLEDGE.md row${writtenK === 1 ? "" : "s"} into the memory store.`, "info");
+    }
+  } catch (e) {
+    logWarning("bootstrap", `KNOWLEDGE.md backfill failed: ${(e as Error).message}`);
+  }
+  try {
+    const { renderKnowledgeProjection } = await import("../knowledge-projection.js");
+    renderKnowledgeProjection(basePath);
+  } catch (e) {
+    logWarning("bootstrap", `KNOWLEDGE.md projection render failed: ${(e as Error).message}`);
+  }
+
+  // ADR-013 step 6 preflight: warn when decisions / KNOWLEDGE.md rows are not
+  // yet in the memories table. Read-only; never throws. Runs after the two
+  // backfills above so the gap report reflects post-backfill state.
+  try {
+    const { reportConsolidationGaps } = await import("../memory-consolidation-scanner.js");
+    reportConsolidationGaps(basePath);
+  } catch (e) {
+    logWarning("bootstrap", `memory consolidation scan failed: ${(e as Error).message}`);
+  }
+
+  const { block: knowledgeBlock, globalSizeKb } = loadKnowledgeBlock(gsdHome(), basePath);
+  if (globalSizeKb > 4) {
+    ctx.ui.notify(
+      `GSD: ~/.gsd/agent/KNOWLEDGE.md is ${globalSizeKb.toFixed(1)}KB — consider trimming to keep system prompt lean.`,
+      "warning",
+    );
+  }
+
   let newSkillsBlock = "";
   if (hasSkillSnapshot()) {
     const newSkills = detectNewSkills();
@@ -370,7 +406,9 @@ export async function loadMemoryBlock(
 }
 
 export function loadKnowledgeBlock(gsdHomeDir: string, cwd: string): { block: string; globalSizeKb: number } {
-  // 1. Global knowledge (~/.gsd/agent/KNOWLEDGE.md) — cross-project, user-maintained
+  // 1. Global knowledge (~/.gsd/agent/KNOWLEDGE.md) — cross-project,
+  //    user-maintained. NOT migrated to memories (which are project-scoped),
+  //    so the full file is injected unchanged.
   let globalKnowledge = "";
   let globalSizeKb = 0;
   const globalKnowledgePath = join(gsdHomeDir, "agent", "KNOWLEDGE.md");
@@ -386,13 +424,18 @@ export function loadKnowledgeBlock(gsdHomeDir: string, cwd: string): { block: st
     }
   }
 
-  // 2. Project knowledge (.gsd/KNOWLEDGE.md) — project-specific
+  // 2. Project knowledge (.gsd/KNOWLEDGE.md) — project-specific.
+  //    ADR-013 Stage 2b: Patterns and Lessons are projected from the
+  //    memories table and already reach the LLM via loadMemoryBlock. Inject
+  //    only the intro prose + `## Rules` section here to avoid duplicating
+  //    Patterns/Lessons content in the prompt. Rules stay manual per
+  //    ADR-013 line 39 and have no memory equivalent.
   let projectKnowledge = "";
   const knowledgePath = resolveGsdRootFile(cwd, "KNOWLEDGE");
   if (existsSync(knowledgePath)) {
     try {
-      const content = readFileSync(knowledgePath, "utf-8").trim();
-      if (content) projectKnowledge = content;
+      const raw = readFileSync(knowledgePath, "utf-8").trim();
+      if (raw) projectKnowledge = extractIntroAndRules(raw).trim();
     } catch (e) {
       logWarning("bootstrap", `project knowledge file read failed: ${(e as Error).message}`);
     }
@@ -411,7 +454,7 @@ export function loadKnowledgeBlock(gsdHomeDir: string, cwd: string): { block: st
   }
   const body = limitKnowledgeBlock(parts.join("\n\n"), getKnowledgeCharLimit());
   return {
-    block: `\n\n[KNOWLEDGE — Rules, patterns, and lessons learned]\n\n${body}`,
+    block: `\n\n[KNOWLEDGE — Rules from KNOWLEDGE.md (Patterns and Lessons reach the LLM via the memory block)]\n\n${body}`,
     globalSizeKb,
   };
 }

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

@@ -5,7 +5,7 @@ import { isAbsolute, join, relative, resolve, sep } from "node:path";
 import { minimatch } from "minimatch";
 
 import { getIsolationMode } from "../preferences.js";
-import type { ToolsPolicy } from "../unit-context-manifest.js";
+import { compileSubagentPermissionContract, type ToolsPolicy } from "../unit-context-manifest.js";
 import { logWarning } from "../workflow-logger.js";
 import { isGsdWorktreePath, resolveWorktreeProjectRoot } from "../worktree-root.js";
 
@@ -65,6 +65,7 @@ const QUEUE_SAFE_TOOLS = new Set([
  *   true / false        — shell no-ops / test exit codes
  */
 const BASH_READ_ONLY_RE = /^\s*(cat|head|tail|less|more|wc|file|stat|du|df|which|type|echo|printf|ls|find|grep|rg|awk|sed\b(?!.*-i)|sort|uniq|diff|comm|tr|cut|tee\s+-a\s+\/dev\/null|git\s+(log|show|diff|status|branch|tag|remote|rev-parse|ls-files|blame|shortlog|describe|stash\s+list|config\s+--get|cat-file)|gh\s+(issue|pr|api|repo|release)\s+(view|list|diff|status|checks)|mkdir\s+-p\s+\.gsd|rtk\s|npm\s+run\s+(test|test:\w+|lint|lint:\w+|typecheck|type-check|type-check:\w+|check|verify|audit|outdated|format:check|ci|validate)\b|npm\s+(ls|list|info|view|show|outdated|audit|explain|doctor|ping|--version|-v)\b|npx\s|tsx\s|node\s+(--print|--version|-v\b)|python[23]?\s+(-c\s+'[^']*'|--version|-V\b|-m\s+(pip\s+show|pip\s+list|site))|pip[23]?\s+(show|list|freeze|check|index\s+versions)\b|jq\s|yq\s|curl\s+(-s\b|--silent\b)(?!\s+[^|>]*\s-[oO]\b)(?!\s+[^|>]*\s--output\b)[^|>]*$|openssl\s+(version|x509|s_client)|env\b|printenv\b|true\b|false\b)/;
+const BASH_VERIFICATION_RE = /^\s*(npm\s+(run\s+(build|test|test:\w+|lint|lint:\w+|typecheck|type-check|verify|ci|validate)\b|test\b)|pnpm\s+(build|test|lint|typecheck|verify)\b|yarn\s+(build|test|lint|typecheck|verify)\b|vitest\b|jest\b|go\s+test\b)/;
 
 interface InMemoryWriteGateState {
   verifiedDepthMilestones: Set<string>;
@@ -747,7 +748,7 @@ function matchesAllowedGlob(absPath: string, basePath: string, globs: readonly s
 function blockReason(unitType: string, mode: string, what: string): string {
   return [
     `HARD BLOCK: unit "${unitType}" runs under tools-policy "${mode}" — ${what}.`,
-    `This is a mechanical gate enforced by manifest.tools (#4934). You MUST NOT proceed,`,
+    `This is a mechanical gate enforced by manifest.tools. You MUST NOT proceed,`,
     `retry the same call, or rationalize past this block. If you need to write user source,`,
     `the work belongs in execute-task, not in a planning unit.`,
   ].join(" ");
@@ -767,6 +768,9 @@ function blockReason(unitType: string, mode: string, what: string): string {
  *                    and listed in the policy's allowedSubagents.
  *   - "docs"       → like "planning" but also allows writes to paths
  *                    matching `allowedPathGlobs` relative to basePath.
+ *   - "verification"
+ *                  → allows Bash for project verification commands, but keeps
+ *                    writes restricted to .gsd/ and blocks subagent dispatch.
  *
  * `pathOrCommand` is the file path for write/edit-shaped tools and the
  * shell command for bash. Other tools ignore this argument.
@@ -804,14 +808,15 @@ export function shouldBlockPlanningUnit(
     return { block: true, reason: blockReason(unitType, policy.mode, `tool "${tool}" is not on the read-only allowlist`) };
   }
 
-  // planning / planning-dispatch / docs modes share the same surface for safe tools, bash, and subagent.
+  // planning / planning-dispatch / docs / verification modes share the same surface for safe tools, bash, and subagent.
   if (PLANNING_SAFE_TOOLS.has(tool)) return { block: false };
   if (tool.startsWith("gsd_")) return { block: false };
 
   if (PLANNING_SUBAGENT_TOOLS.has(tool)) {
     if (policy.mode === "planning-dispatch") {
       const requested = (agentClasses ?? []).map(a => a.trim()).filter(Boolean);
-      const allowedSubagents = Array.isArray(policy.allowedSubagents) ? policy.allowedSubagents : [];
+      const dispatchContract = compileSubagentPermissionContract(policy);
+      const allowedSubagents = dispatchContract.allowedSubagents;
       const allowed = new Set(allowedSubagents);
       // When agentClasses is undefined, the caller has not been updated to extract
       // agent identities yet. Block and warn so stale callers surface in telemetry
@@ -861,6 +866,17 @@ export function shouldBlockPlanningUnit(
   }
 
   if (tool === "bash") {
+    if (policy.mode === "verification") {
+      if (BASH_VERIFICATION_RE.test(pathOrCommand) || BASH_READ_ONLY_RE.test(pathOrCommand)) return { block: false };
+      return {
+        block: true,
+        reason: blockReason(
+          unitType,
+          policy.mode,
+          `bash is restricted to build/test verification commands (npm run build, npm test, etc.); cannot run "${pathOrCommand.slice(0, 80)}${pathOrCommand.length > 80 ? "…" : ""}"`,
+        ),
+      };
+    }
     if (BASH_READ_ONLY_RE.test(pathOrCommand)) return { block: false };
     return {
       block: true,