gsd-pi 2.80.0-dev.c5f2443b3 → 2.80.0-dev.d4fc28e6b

package/src/resources/extensions/gsd/native-git-bridge.ts

@@ -1183,6 +1183,33 @@ export function nativeRmForce(basePath: string, paths: string[]): void {
   }
 }
 
+function runGitWorktreeAdd(
+  basePath: string,
+  wtPath: string,
+  branch: string,
+  createBranch?: boolean,
+  startPoint?: string,
+): void {
+  if (createBranch) {
+    const branchRef = gitExec(basePath, ["show-ref", "--verify", `refs/heads/${branch}`], true);
+    if (branchRef) {
+      gitExec(basePath, ["worktree", "add", wtPath, branch]);
+      return;
+    }
+    gitExec(basePath, ["worktree", "add", "-b", branch, wtPath, startPoint ?? "HEAD"]);
+  } else {
+    gitExec(basePath, ["worktree", "add", wtPath, branch]);
+  }
+}
+
+export function assertWorktreeMaterialized(wtPath: string): void {
+  if (existsSync(join(wtPath, ".git"))) return;
+  throw new GSDError(
+    GSD_GIT_ERROR,
+    `git worktree add did not materialize a valid worktree at ${wtPath}: missing .git file`,
+  );
+}
+
 /**
  * Add a new git worktree.
  * Native: libgit2 worktree API.
@@ -1198,14 +1225,20 @@ export function nativeWorktreeAdd(
   const native = loadNative();
   if (native) {
     native.gitWorktreeAdd(basePath, wtPath, branch, createBranch, startPoint);
-    return;
+    try {
+      assertWorktreeMaterialized(wtPath);
+      return;
+    } catch {
+      rmSync(wtPath, { recursive: true, force: true });
+      gitExec(basePath, ["worktree", "prune"], true);
+      runGitWorktreeAdd(basePath, wtPath, branch, createBranch, startPoint);
+      assertWorktreeMaterialized(wtPath);
+      return;
+    }
   }
 
-  if (createBranch) {
-    gitExec(basePath, ["worktree", "add", "-b", branch, wtPath, startPoint ?? "HEAD"]);
-  } else {
-    gitExec(basePath, ["worktree", "add", wtPath, branch]);
-  }
+  runGitWorktreeAdd(basePath, wtPath, branch, createBranch, startPoint);
+  assertWorktreeMaterialized(wtPath);
 }
 
 /**

package/src/resources/extensions/gsd/orphan-stash-audit.ts

@@ -0,0 +1,117 @@
+// GSD-2 + src/resources/extensions/gsd/orphan-stash-audit.ts
+// Startup sweep for orphaned gsd-preflight-stash entries left behind by
+// interrupted milestone merges (#5538-followup).
+
+import { execFileSync } from "node:child_process";
+import { GIT_NO_PROMPT_ENV } from "./git-constants.js";
+
+export interface OrphanPreflightStashAuditResult {
+  applied: Array<{ milestoneId: string; stashRef: string }>;
+  warnings: string[];
+}
+
+/**
+ * Recognize the "already restored" failure mode of `git stash apply`.
+ *
+ * When a preflight stash captured untracked files via `--include-untracked`
+ * and those files are now present in the working tree (e.g. a prior audit
+ * run already applied this stash), `git stash apply` aborts with
+ * `<path> already exists, no checkout` and exits non-zero. That is the
+ * idempotent steady state for this audit, not a recovery failure — treat
+ * it as a no-op so repeated GSD startups stop spamming the user with
+ * warnings about stashes that have already been restored (#5538-followup
+ * peer-review feedback).
+ */
+function _isAlreadyRestoredApplyError(err: unknown): boolean {
+  if (!err || typeof err !== "object") return false;
+  const stderr = (err as { stderr?: unknown }).stderr;
+  const stderrText = typeof stderr === "string" ? stderr : stderr instanceof Uint8Array ? Buffer.from(stderr).toString("utf-8") : "";
+  if (stderrText && /already exists, no checkout/i.test(stderrText)) return true;
+  const message = err instanceof Error ? err.message : String(err);
+  return /already exists, no checkout/i.test(message);
+}
+
+export { _isAlreadyRestoredApplyError };
+
+/**
+ * Audit `git stash list` for orphaned `gsd-preflight-stash:M00x:*` entries.
+ *
+ * The matching merge code in `phases.ts` previously skipped the postflight
+ * pop whenever `mergeAndExit` threw, leaking the user's pre-merge working
+ * tree into the stash list. For every preflight-stash entry whose milestone
+ * is now complete (per the supplied callback), `git stash apply` is invoked —
+ * NOT `pop`. The stash entry stays in the list so the user retains a backup
+ * if the apply produces unexpected merge results. Idempotent across repeated
+ * startup runs.
+ *
+ * Failures are best-effort: a list error (no repo, git unavailable) returns
+ * an empty result. An apply error becomes a warning the user sees alongside
+ * the existing orphan-branch audit messages — startup continues.
+ */
+export function auditOrphanedPreflightStashes(
+  basePath: string,
+  isMilestoneComplete: (milestoneId: string) => boolean,
+): OrphanPreflightStashAuditResult {
+  const result: OrphanPreflightStashAuditResult = { applied: [], warnings: [] };
+
+  let listOutput: string;
+  try {
+    listOutput = execFileSync(
+      "git",
+      ["stash", "list", "--format=%gd%x00%s"],
+      {
+        cwd: basePath,
+        stdio: ["ignore", "pipe", "pipe"],
+        encoding: "utf-8",
+        env: GIT_NO_PROMPT_ENV,
+      },
+    );
+  } catch {
+    return result;
+  }
+
+  const MARKER_RE = /\bgsd-preflight-stash:([A-Za-z0-9_-]+):/;
+  for (const line of listOutput.split("\n")) {
+    const sep = line.indexOf("\x00");
+    if (sep < 0) continue;
+    const ref = line.slice(0, sep);
+    const subject = line.slice(sep + 1);
+    if (!ref || !subject) continue;
+
+    const match = MARKER_RE.exec(subject);
+    if (!match) continue;
+    const milestoneId = match[1];
+
+    let complete = false;
+    try {
+      complete = isMilestoneComplete(milestoneId);
+    } catch (err) {
+      result.warnings.push(
+        `Could not determine completion status for ${milestoneId} during preflight-stash audit: ${err instanceof Error ? err.message : String(err)}.`,
+      );
+      continue;
+    }
+    if (!complete) continue;
+
+    try {
+      execFileSync("git", ["stash", "apply", "--quiet", ref], {
+        cwd: basePath,
+        stdio: ["ignore", "pipe", "pipe"],
+        encoding: "utf-8",
+        env: GIT_NO_PROMPT_ENV,
+      });
+      result.applied.push({ milestoneId, stashRef: ref });
+    } catch (err) {
+      // Idempotent steady state: stash was already applied in a prior audit
+      // run; the files exist and `git stash apply` refuses to overwrite.
+      // Skip silently so repeat runs are no-ops.
+      if (_isAlreadyRestoredApplyError(err)) continue;
+      result.warnings.push(
+        `Could not apply orphaned preflight stash ${ref} (milestone ${milestoneId}): ${err instanceof Error ? err.message : String(err)}. ` +
+          `Run \`git stash apply ${ref}\` manually to restore your pre-merge changes.`,
+      );
+    }
+  }
+
+  return result;
+}

package/src/resources/extensions/gsd/parallel-orchestrator.ts

@@ -21,7 +21,7 @@ import { join, dirname } from "node:path";
 import { fileURLToPath } from "node:url";
 import { gsdRoot } from "./paths.js";
 import { createWorktree, worktreePath } from "./worktree-manager.js";
-import { autoWorktreeBranch, runWorktreePostCreateHook, syncGsdStateToWorktree } from "./auto-worktree.js";
+import { autoWorktreeBranch, fastForwardReusedMilestoneBranchIfSafe, runWorktreePostCreateHook, syncGsdStateToWorktree } from "./auto-worktree.js";
 import { nativeBranchExists } from "./native-git-bridge.js";
 import { readIntegrationBranch } from "./git-service.js";
 import { resolveParallelConfig } from "./preferences.js";
@@ -471,7 +471,7 @@ export async function startParallel(
       // Create the worktree (without chdir — coordinator stays in project root)
       let wtPath: string;
       try {
-        wtPath = createMilestoneWorktree(basePath, mid);
+        wtPath = _createMilestoneWorktree(basePath, mid);
       } catch (e) {
         logWarning("parallel", `createMilestoneWorktree fallback for ${mid}: ${(e as Error).message}`);
         wtPath = worktreePath(basePath, mid);
@@ -532,13 +532,23 @@ export async function startParallel(
 /**
  * Create a git worktree for a milestone without changing the coordinator's cwd.
  * Uses milestone/<MID> branch naming (same as auto-worktree.ts).
+ *
+ * Exported with the `_` prefix purely for tests — production callers stay on
+ * the closure-private name `createMilestoneWorktree` below.
  */
-function createMilestoneWorktree(basePath: string, milestoneId: string): string {
+export function _createMilestoneWorktree(basePath: string, milestoneId: string): string {
   const branch = autoWorktreeBranch(milestoneId);
   const branchExists = nativeBranchExists(basePath, branch);
 
   let info: { name: string; path: string; branch: string; exists: boolean };
   if (branchExists) {
+    // #5549 post-merge audit (R3): match the fast-forward behavior added to
+    // `createAutoWorktree` in commit 8996cb68e. When a worker reuses an
+    // existing milestone branch, fast-forward it onto the integration branch
+    // when safe so per-worker worktrees don't fork from a stale base after a
+    // sibling milestone has merged. Same `nativeIsAncestor` + worktree-list
+    // safety guards apply.
+    fastForwardReusedMilestoneBranchIfSafe(basePath, milestoneId, branch);
     info = createWorktree(basePath, milestoneId, { branch, reuseExistingBranch: true });
   } else {
     const integrationBranch = readIntegrationBranch(basePath, milestoneId) ?? undefined;

package/src/resources/extensions/gsd/planning-path-scope.ts

@@ -0,0 +1,35 @@
+import { isAbsolute, relative, resolve } from "node:path";
+import { normalizePlannedFileReference } from "./files.js";
+
+export interface PlanningPathScopeField {
+  field: string;
+  values: string[];
+}
+
+function isInsideBase(basePath: string, candidate: string): boolean {
+  const base = resolve(basePath);
+  const abs = resolve(candidate);
+  const rel = relative(base, abs);
+  return rel === "" || (!!rel && !rel.startsWith("..") && !isAbsolute(rel));
+}
+
+/**
+ * Planning IO fields are execution contracts. Absolute paths are only safe when
+ * they stay inside the active working directory; in worktree mode, an absolute
+ * path to the original checkout makes executors edit the wrong tree.
+ */
+export function validatePlanningPathScope(
+  basePath: string,
+  fields: PlanningPathScopeField[],
+): string | null {
+  for (const { field, values } of fields) {
+    for (const raw of values) {
+      const candidate = normalizePlannedFileReference(raw);
+      if (!isAbsolute(candidate)) continue;
+      if (isInsideBase(basePath, candidate)) continue;
+      return `${field} contains absolute path outside working directory: ${candidate}. Use a path relative to ${basePath}.`;
+    }
+  }
+
+  return null;
+}

package/src/resources/extensions/gsd/pr-evidence.ts

@@ -11,6 +11,7 @@ export interface PrEvidenceInput {
   changeType?: PrChangeType;
   linkedIssue?: string;
   summaries?: string[];
+  blockers?: string[];
   roadmapItems?: string[];
   metrics?: string[];
   testsRun?: string[];
@@ -34,8 +35,54 @@ const CHANGE_TYPE_LABELS: Record<PrChangeType, string> = {
   chore: "Build, CI, or tooling changes",
 };
 
+// Per-item cap for user-supplied content. 2 KB gives slice titles plus
+// descriptions room while still bounding malicious DoS-via-PR-body input.
+const USER_CONTENT_CAP_BYTES = 2048;
+const TRUNCATION_SUFFIX = " … [truncated]";
+
+// Strips HTML comments, fake commit trailers (Co-Authored-By, Signed-off-by —
+// case-insensitive on the trailer name), and caps total length. Designed to
+// be a no-op for well-formed input; golden fixtures must remain byte-stable.
+// Trailer lines are removed (not rejected) so that a single bad line in an
+// otherwise legitimate description does not block the entire PR.
+function sanitizeUserContent(s: string): string {
+  if (!s) return s;
+  // Strip HTML comments greedily across newlines.
+  let out = s.replace(/<!--[\s\S]*?-->/g, "");
+  // Drop lines that look like commit trailers we do not want forged.
+  out = out
+    .split("\n")
+    .filter((line) => !/^\s*(co-authored-by|signed-off-by)\s*:/i.test(line))
+    .join("\n");
+  if (Buffer.byteLength(out, "utf8") > USER_CONTENT_CAP_BYTES) {
+    const budget = USER_CONTENT_CAP_BYTES - Buffer.byteLength(TRUNCATION_SUFFIX, "utf8");
+    // Truncate by code units to stay safely under the byte budget for ASCII;
+    // for multibyte content we conservatively walk back until under budget.
+    let sliced = out.slice(0, Math.max(0, budget));
+    while (Buffer.byteLength(sliced, "utf8") > budget && sliced.length > 0) {
+      sliced = sliced.slice(0, -1);
+    }
+    out = sliced + TRUNCATION_SUFFIX;
+  }
+  return out;
+}
+
+// Strips HTML comments and fake trailers without applying the length cap.
+// Used for short fields like linkedIssue where truncation would be confusing.
+function sanitizeIssueRef(s: string): string {
+  if (!s) return s;
+  let out = s.replace(/<!--[\s\S]*?-->/g, "");
+  out = out
+    .split("\n")
+    .filter((line) => !/^\s*(co-authored-by|signed-off-by)\s*:/i.test(line))
+    .join("\n");
+  return out;
+}
+
 function normalizeList(values: readonly string[] | undefined): string[] {
-  return (values ?? []).map((value) => value.trim()).filter(Boolean);
+  return (values ?? [])
+    .map((value) => sanitizeUserContent(value).trim())
+    .filter(Boolean);
 }
 
 function changeTypeChecklist(selected: PrChangeType): string[] {
@@ -56,13 +103,17 @@ export function buildPrEvidence(input: PrEvidenceInput): PrEvidence {
   const subjectTitle = input.milestoneTitle?.trim() || subjectId;
   const changeType = input.changeType ?? "feat";
   const summaries = normalizeList(input.summaries);
+  const blockers = normalizeList(input.blockers);
   const roadmapItems = normalizeList(input.roadmapItems);
   const metrics = normalizeList(input.metrics);
   const testsRun = normalizeList(input.testsRun);
   const rollbackNotes = normalizeList(input.rollbackNotes);
-  const linkedIssue = input.linkedIssue?.trim() || "Not specified. Add an issue link before marking this PR ready if CONTRIBUTING.md requires one.";
-  const why = input.why?.trim() || `${capitalize(subjectKind)} work is complete and ready for review.`;
-  const how = input.how?.trim() || "Generated from GSD evidence and local workflow artifacts.";
+  // linkedIssue is sanitized but not length-capped: legitimate issue refs
+  // are short by nature, and truncating "Closes #123" would be unhelpful.
+  const linkedIssueRaw = input.linkedIssue ? sanitizeIssueRef(input.linkedIssue).trim() : "";
+  const linkedIssue = linkedIssueRaw || "Not specified. Add an issue link before marking this PR ready if CONTRIBUTING.md requires one.";
+  const why = (input.why ? sanitizeUserContent(input.why).trim() : "") || `${capitalize(subjectKind)} work is complete and ready for review.`;
+  const how = (input.how ? sanitizeUserContent(input.how).trim() : "") || "Generated from GSD evidence and local workflow artifacts.";
   const title = `${changeType}: ${subjectTitle}`;
 
   const sections: string[] = [
@@ -75,6 +126,13 @@ export function buildPrEvidence(input: PrEvidenceInput): PrEvidence {
     "## What",
     "",
     summaries.length > 0 ? summaries.join("\n\n") : `${capitalize(subjectKind)} ${subjectId} completed.`,
+  ];
+
+  if (blockers.length > 0) {
+    sections.push("", "## Blockers", "", blockers.map((blocker) => `- ${blocker}`).join("\n"));
+  }
+
+  sections.push(
     "",
     "## Why",
     "",
@@ -87,7 +145,7 @@ export function buildPrEvidence(input: PrEvidenceInput): PrEvidence {
     "## Linked Issue",
     "",
     linkedIssue,
-  ];
+  );
 
   if (roadmapItems.length > 0) {
     sections.push("", "## Roadmap", "", roadmapItems.join("\n"));

package/src/resources/extensions/gsd/pre-execution-checks.ts

@@ -380,6 +380,8 @@ function shouldValidateInputAsPath(raw: string): boolean {
   const trimmed = raw.trim();
   if (!trimmed) return false;
 
+  if (isRuntimeOnlyInput(trimmed)) return false;
+
   const candidate = extractPathFromAnnotation(trimmed);
   if (!candidate) return false;
 
@@ -405,6 +407,10 @@ function shouldValidateInputAsPath(raw: string): boolean {
   );
 }
 
+function isRuntimeOnlyInput(raw: string): boolean {
+  return /\(\s*runtime\s*\)/i.test(raw);
+}
+
 function containsGlobPattern(candidate: string): boolean {
   return ["*", "?", "[", "]", "{", "}"].some((char) => candidate.includes(char));
 }
@@ -448,6 +454,16 @@ export function checkFilePathConsistency(
 ): PreExecutionCheckJSON[] {
   const results: PreExecutionCheckJSON[] = [];
 
+  // Build a set of all files created by any task at any position (normalized).
+  // Used to suppress consistency errors for files that will be caught with a
+  // more precise message by checkTaskOrdering (sequence violation).
+  const allTaskOutputs = new Set<string>();
+  for (const t of tasks) {
+    for (const f of t.expected_output) {
+      allTaskOutputs.add(normalizeFilePath(f));
+    }
+  }
+
   for (let i = 0; i < tasks.length; i++) {
     const task = tasks[i];
     const priorOutputs = getExpectedOutputsUpTo(tasks, i);
@@ -481,6 +497,12 @@ export function checkFilePathConsistency(
       }
 
       if (!existsOnDisk && !inPriorOutputs && !inOwnOutputs && !directorySatisfied) {
+        // If a later task claims to create this file, the ordering check will
+        // fire a more precise "sequence violation" error for the same file.
+        // Suppress the consistency error here to avoid duplicate noise.
+        if (allTaskOutputs.has(normalizedFile) && !ownOutputs.has(normalizedFile)) {
+          continue;
+        }
         results.push({
           category: "file",
           target: file,
@@ -535,6 +557,7 @@ export function checkTaskOrdering(
     const filesToCheck = [...task.inputs];
 
     for (const file of filesToCheck) {
+      if (isRuntimeOnlyInput(file)) continue;
       if (!shouldValidateInputAsPath(file)) continue;
 
       const normalizedFile = normalizeFilePath(file);

package/src/resources/extensions/gsd/preferences-types.ts

@@ -32,7 +32,7 @@ export interface ContextManagementConfig {
 /**
  * Opt-in tool-output sandboxing for sub-sessions. When enabled, the gsd_exec
  * MCP tool runs scripts in an isolated subprocess and returns only a short
- * digest to the calling agent's context window; full stdout/stderr persist
+ * digest to the calling agent's context window; capped stdout/stderr persist
  * in the project memory store and can be retrieved by id later.
  *
  * Inspired by mksglu/context-mode (Elastic License 2.0). This is an

package/src/resources/extensions/gsd/prompt-loader.ts

@@ -123,17 +123,42 @@ function warmCache(): void {
 }
 
 let warmCacheScheduled = false;
+let warmCacheRan = false;
+let warmCacheTimer: ReturnType<typeof setTimeout> | undefined;
+
+/**
+ * Synchronously snapshot the prompt/template tree into the cache.
+ *
+ * Safe to call immediately after `initResources()` because that function uses
+ * synchronous fs APIs — there is no race window that requires deferring the
+ * cache warm via setTimeout. Idempotent: subsequent calls are no-ops.
+ *
+ * Cancels the fallback `scheduleWarmCache` timer if it is still pending.
+ */
+export function primeCache(): void {
+  if (warmCacheRan) return;
+  if (warmCacheTimer) {
+    clearTimeout(warmCacheTimer);
+    warmCacheTimer = undefined;
+  }
+  warmCacheScheduled = true;
+  warmCacheRan = true;
+  warmCache();
+}
 
 function scheduleWarmCache(): void {
   if (warmCacheScheduled) return;
   warmCacheScheduled = true;
 
   const run = () => {
+    warmCacheTimer = undefined;
+    if (warmCacheRan) return;
+    warmCacheRan = true;
     warmCache();
   };
 
-  const timer = setTimeout(run, 1000);
-  timer.unref?.();
+  warmCacheTimer = setTimeout(run, 1000);
+  warmCacheTimer.unref?.();
 }
 
 // Snapshot the full prompt/template tree after import so extension startup only

package/src/resources/extensions/gsd/prompts/complete-milestone.md

@@ -6,6 +6,8 @@ You are executing GSD auto-mode.
 
 Your working directory is `{{workingDirectory}}`. All file reads, writes, and shell commands MUST operate relative to this directory. Do NOT `cd` to any other directory.
 
+If any inlined plan, summary, verification command, or prior artifact names an absolute path outside `{{workingDirectory}}`, treat that path as stale context. Convert it to the equivalent relative path under `{{workingDirectory}}` before reading, writing, or executing. If no equivalent path exists under `{{workingDirectory}}`, record a verification failure and stop; do not edit or run commands in another checkout.
+
 ## Mission
 
 All slices are complete. Verify the integrated work, persist milestone completion, refresh project state, and write the final record future milestones will rely on.
@@ -16,28 +18,28 @@ Start with what the excerpts give you. Read full files when the section heads si
 
 **On-demand Read ordering:** Complete all slice SUMMARY Reads you need for cross-slice synthesis, the Decision Re-evaluation table, and LEARNINGS **before** calling `gsd_complete_milestone` (step 12). Once that tool runs, the milestone is marked complete in the DB, so it must be the final persistent milestone-closeout write.
 
-### Delegate Review Work
+### Closeout Review Mode
 
-Use `subagent` for review work needing fresh context, before drafting LEARNINGS:
+The inlined context includes a validation status block.
 
-- Cross-slice integrations or new public APIs -> **reviewer** with milestone diff and roadmap.
-- Auth, network, parsing, file IO, shell exec, or crypto -> **security** audit.
-- Significant tests added or changed -> **tester** coverage check against success criteria.
+- If it says a passing validation artifact is present, treat that artifact as authoritative for success criteria, requirement coverage, verification classes, and cross-slice integration. Do not delegate fresh reviewer/security/tester audits unless the validation artifact is internally inconsistent with the inlined summaries.
+- If validation is missing, stale, non-pass, or internally inconsistent, use `subagent` for review work needing fresh context before drafting LEARNINGS: cross-slice integrations or new public APIs -> **reviewer**; auth, network, parsing, file IO, shell exec, or crypto -> **security**; significant tests added or changed -> **tester**.
 
-Subagents report only; they do not write user source. Fold findings into Decision Re-evaluation and LEARNINGS before completion.
+Subagents report only; they do not write user source. Fold any findings into Decision Re-evaluation and LEARNINGS before completion.
 
 {{inlinedContext}}
 
 ## Steps
 
-1. Use the **Milestone Summary** output template from the inlined context above
-2. {{skillActivation}}
-3. **Verify code changes exist.** Compare milestone work against the integration branch (`main`, `master`, or recorded branch), using merge-base as older revision and `HEAD` as newer. If the diff lists non-`.gsd/` files, pass. If `HEAD` equals the integration branch/merge-base, treat it as a self-diff retry: inspect milestone-scoped commit evidence (`GSD-Unit: {{milestoneId}}` or production `GSD-Task: Sxx/Tyy` trailers touching `.gsd/milestones/{{milestoneId}}/`) and verify those commits touched non-`.gsd/` files. Record **verification failure** only when neither source shows implementation files.
-4. Verify every **success criterion** from `{{roadmapPath}}` with evidence from summaries, tests, or observable behavior. Record unmet criteria as **verification failure**.
-5. Verify **definition of done**: all slices `[x]`, summaries exist, and integrations work. Record unmet items as **verification failure**.
-6. If the roadmap includes a **Horizontal Checklist**, verify each item and note unchecked items in the summary.
-7. Fill the **Decision Re-evaluation** table: compare each key `.gsd/DECISIONS.md` decision from this milestone with what shipped, and flag decisions to revisit.
-8. Validate **requirement status transitions**. For each changed requirement, confirm evidence supports the new status. Requirements may move between Active, Validated, Deferred, Blocked, or Out of Scope only with proof.
+1. **Duplicate completion guard:** Call `gsd_milestone_status` for `{{milestoneId}}` before any durable writes. If the returned milestone **status is `complete`**, this is a stale or duplicate closeout turn: do NOT mutate requirements, do NOT refresh the project document, do NOT write LEARNINGS, and do NOT persist milestone completion again. Say: "Milestone {{milestoneId}} is already complete." and stop.
+2. Use the **Milestone Summary** output template from the inlined context above
+3. {{skillActivation}}
+4. **Verify code changes exist.** Compare milestone work against the integration branch (`main`, `master`, or recorded branch), using merge-base as older revision and `HEAD` as newer. If the diff lists non-`.gsd/` files, pass. If `HEAD` equals the integration branch/merge-base, treat it as a self-diff retry: inspect milestone-scoped commit evidence (`GSD-Unit: {{milestoneId}}` or production `GSD-Task: Sxx/Tyy` trailers touching `.gsd/milestones/{{milestoneId}}/`) and verify those commits touched non-`.gsd/` files. Record **verification failure** only when neither source shows implementation files.
+5. Verify every **success criterion** from `{{roadmapPath}}`. If passing validation is present, summarize the validation evidence instead of re-auditing it; otherwise verify with evidence from summaries, tests, or observable behavior. Record unmet criteria as **verification failure**.
+6. Verify **definition of done**: all slices `[x]`, summaries exist, and integrations work. If passing validation is present, trust its integration/verification verdict unless inconsistent with current artifacts. Record unmet items as **verification failure**.
+7. If the roadmap includes a **Horizontal Checklist**, verify each item and note unchecked items in the summary.
+8. Fill the **Decision Re-evaluation** table: compare each key `.gsd/DECISIONS.md` decision from this milestone with what shipped, and flag decisions to revisit.
+9. Validate **requirement status transitions**. For each changed requirement, confirm evidence supports the new status. Requirements may move between Active, Validated, Deferred, Blocked, or Out of Scope only with proof.
 
 **DB access safety:** Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')`; the engine owns the WAL connection. Use `gsd_milestone_status`, inlined context, or `gsd_*` tools; never direct SQL.
 
@@ -54,13 +56,13 @@ Subagents report only; they do not write user source. Fold findings into Decisio
 
 **Success path** (all verifications passed):
 
-9. For each requirement whose status changed in step 8, call `gsd_requirement_update` with the requirement ID and updated `status` and `validation` fields — the tool regenerates `.gsd/REQUIREMENTS.md` automatically. Do this BEFORE completing the milestone so requirement updates are persisted.
-10. Update `.gsd/PROJECT.md`: use the `write` tool with `path: ".gsd/PROJECT.md"` and `content` containing the full updated document reflecting milestone completion and current project state. Do NOT use the `edit` tool for this — PROJECT.md is a full-document refresh.
-11. Extract structured learnings from this milestone and persist them to the GSD memory store. Follow the procedure block immediately below — it writes `{{milestoneId}}-LEARNINGS.md` as the audit trail and persists Patterns, Lessons, and Decisions via `capture_thought` (categories: pattern, gotcha/convention, architecture). The memory store is the single source of truth for cross-session durable knowledge (ADR-013).
+10. For each requirement whose status changed in step 9, call `gsd_requirement_update` with the requirement ID and updated `status` and `validation` fields — the tool regenerates `.gsd/REQUIREMENTS.md` automatically. Do this BEFORE completing the milestone so requirement updates are persisted.
+11. Update `.gsd/PROJECT.md`: use the `write` tool with `path: ".gsd/PROJECT.md"` and `content` containing the full updated document reflecting milestone completion and current project state. Do NOT use the `edit` tool for this — PROJECT.md is a full-document refresh.
+12. Extract structured learnings from this milestone and persist them to the GSD memory store. Follow the procedure block immediately below — it writes `{{milestoneId}}-LEARNINGS.md` as the audit trail and persists Patterns, Lessons, and Decisions via `capture_thought` (categories: pattern, gotcha/convention, architecture). The memory store is the single source of truth for cross-session durable knowledge (ADR-013).
 
 {{extractLearningsSteps}}
 
-12. **Persist completion through `gsd_complete_milestone`.** Call it with the parameters below. This must be the final persistent write in the unit. The tool updates the milestone status in the DB, renders `{{milestoneSummaryPath}}`, and validates all slices are complete.
+13. **Persist completion through `gsd_complete_milestone`.** Call it with the parameters below. This must be the final persistent write in the unit. The tool updates the milestone status in the DB, renders `{{milestoneSummaryPath}}`, and validates all slices are complete.
 
    **Required parameters:**
    - `milestoneId` (string) — Milestone ID (e.g. M001)
@@ -79,7 +81,7 @@ Subagents report only; they do not write user source. Fold findings into Decisio
    - `followUps` (string) — Follow-up items for future milestones
    - `deviations` (string) — Deviations from the original plan
 
-13. Do not commit manually — the system auto-commits your changes after this unit completes.
+14. Do not commit manually — the system auto-commits your changes after this unit completes.
 - Say: "Milestone {{milestoneId}} complete."
 
 **Important:** Do NOT skip code-change, success-criteria, or definition-of-done verification (steps 3-5). The summary must reflect verified outcomes. Verification failures block completion; there is no override. If a verification tool fails, errors, or returns unexpected output, treat it as failure.

package/src/resources/extensions/gsd/prompts/complete-slice.md

@@ -29,7 +29,7 @@ Use `subagent` only for fresh-context review when useful: reviewer for cross-cut
 7. If requirement status changed, call `gsd_requirement_update`; do not write `.gsd/REQUIREMENTS.md` directly.
 8. Prepare `gsd_slice_complete` content with camelCase fields `milestoneId`, `sliceId`, `sliceTitle`, `oneLiner`, `narrative`, `verification`, and `uatContent`.
 9. Draft concrete UAT with preconditions, numbered steps, expected outcomes, edge cases, UAT Type, and Not Proven By This UAT.
-10. Review task summaries for DECISIONS.md and KNOWLEDGE.md-worthy decisions, patterns, and gotchas. Capture significant items with `capture_thought`; do not append knowledge files directly.
+10. Review the inlined task-summary excerpts for DECISIONS.md and KNOWLEDGE.md-worthy decisions, patterns, and gotchas. Read full `*-SUMMARY.md` files only when an excerpt is absent, truncated, or lacks the specific evidence needed for the slice narrative. Capture significant items with `capture_thought`; do not append knowledge files directly.
 11. Call `gsd_slice_complete`. The DB-backed tool is the canonical write path. Do **not** manually write `{{sliceSummaryPath}}`. Do **not** manually write `{{sliceUatPath}}`. Do not edit roadmap checkboxes; the tool renders files and updates projections.
 12. Do not run git commands.
 13. Update `.gsd/PROJECT.md` with a full `write` only if the current project state needs refresh.

package/src/resources/extensions/gsd/prompts/execute-task.md

@@ -33,7 +33,7 @@ You execute. The inlined task plan is authoritative. Verify referenced files and
 ## Execution Rules
 
 1. Tersely narrate transitions, decisions, and verification outcomes between tool-call clusters.
-2. Call `memory_query` with 2-4 keywords from the task title and touched files unless this is purely mechanical.
+2. Use the injected memory/context blocks first. Call `memory_query` with 2-4 keywords from the task title and touched files only when no injected memory block exists or the inlined memory/context is insufficient for this task.
 3. {{skillActivation}} Follow activated skills before code edits.
 4. Execute the task plan. Before any `Write` that creates an artifact or output file, check whether that path already exists. If it does, read it first and decide whether the work is already done, should be extended, or truly needs replacement.
 5. Build real behavior through the intended surface; stubs/mocks belong in tests only.
@@ -59,11 +59,13 @@ Keep about **{{verificationBudget}}** for verification and summary. If context i
 - If the plan is fundamentally invalid, set `blocker_discovered: true` in the summary and explain.
 - For downstream-impacting ambiguity that cannot be resolved from code, plans, or decisions, include an `escalation` object with question, options, recommendation, rationale, and `continueWithDefault`.
 - Capture meaningful architecture/pattern/observability decisions with `capture_thought`; capture non-obvious gotchas or conventions only when they save future investigation.
-- Read the template at `{{taskSummaryTemplatePath}}`.
+- Use the inlined Task Summary template below. Read `{{taskSummaryTemplatePath}}` only if the inlined template is absent or visibly truncated.
 - Call `gsd_task_complete` with camelCase fields `milestoneId`, `sliceId`, `taskId`, `oneLiner`, `narrative`, `verification`, and `verificationEvidence`.
 - The DB-backed tool is the canonical write path. Do **not** manually write `{{taskSummaryPath}}` or edit PLAN.md checkboxes; the tool renders the summary and updates state.
 - Do not run git commands; the system commits from your summary.
 
+{{inlinedTemplates}}
+
 **Autonomous execution:** no human is available. Do not call `ask_user_questions` or `secure_env_collect`; make reasonable assumptions and document them.
 
 **You MUST call `gsd_task_complete` before finishing.**

package/src/resources/extensions/gsd/prompts/parallel-research-slices.md

@@ -16,7 +16,7 @@ Dispatch ALL slices simultaneously using the `subagent` tool in **parallel mode*
 
 1. Call `subagent` with `tasks: [...]` containing one entry per slice below
 2. Wait for ALL subagents to complete
-3. Verify each slice's RESEARCH file was written (check the `.gsd/{{mid}}/` directory)
+3. Verify each slice's RESEARCH file was written (check `.gsd/milestones/{{mid}}/slices/<slice-id>/`)
 4. If a subagent failed to write its RESEARCH file, retry it **once** individually
 5. If it fails a second time, write a partial RESEARCH file for that slice with a `## BLOCKER` section explaining the failure — do NOT retry again
 6. Report which slices completed research and which (if any) needed a blocker note

package/src/resources/extensions/gsd/prompts/plan-milestone.md

@@ -48,7 +48,7 @@ Narrate decomposition reasoning in complete sentences: grouping, risk order, ver
 Then:
 1. Use the **Roadmap** output template from the inlined context above
 2. {{skillActivation}}
-3. Create only as many demoable vertical slices as the work genuinely needs.
+3. Create only as many demoable vertical slices as the work genuinely needs. Use 1-10 slices, sized to the work; tiny/single-file/static work should usually be one slice.
 4. Order by risk, high-risk first.
 5. Call `gsd_plan_milestone` to persist milestone fields, slice rows, and **Horizontal Checklist** through the DB-backed path. Fill checklist concerns considered during planning: requirements, decisions, shutdown, revenue, auth, shared resources, reconnection. Omit for trivial milestones. Do **not** write `{{outputPath}}`, `ROADMAP.md`, or other planning artifacts manually; the tool owns rendering and persistence.
 6. If planning produced structural decisions (slice ordering, technology choices, scope exclusions), call `gsd_decision_save` for each; the tool assigns IDs and regenerates `.gsd/DECISIONS.md`.
@@ -78,6 +78,8 @@ Apply these when decomposing and ordering slices:
 - Ship features, not proofs; use clearly marked realistic stubs only when necessary.
 - **Dependency format is comma-separated, never range syntax.** Write `depends:[S01,S02,S03]`, not `depends:[S01-S03]`.
 - Roadmap ambition must match the milestone; right-size decomposition.
+- Missing ecosystem markers are not a reason to over-plan. If Project Classification says `untyped-existing`, treat the listed content files as the project surface and use generic file-level workflow guidance.
+- For `untyped-existing` projects with 1-2 content files, prefer exactly one slice unless the request clearly spans multiple independent user-visible capabilities. For 3-5 content files, prefer 1-2 slices.
 
 ## Progressive Planning (ADR-011)
 

package/src/resources/extensions/gsd/prompts/plan-slice.md

@@ -39,7 +39,7 @@ If slice research is inlined, trust it. Explore enough code to confirm paths, bo
 5. Define slice verification before tasks. Non-trivial slices need real tests or executable assertions; boundary contracts need contract-exercising checks. Tests must not read .gitignore/gitignored paths such as `.gsd/`, `.planning/`, or `.audits/`.
 6. Include Threat Surface (Q3), Requirement Impact (Q4), proof level, observability, integration closure, Failure Modes (Q5), Load Profile (Q6), and Negative Tests (Q7) only where applicable.
 7. Right-size tasks. Simple slices can be one task; split only when context, ownership, or verification boundaries justify it.
-8. Each task needs a concrete title, Why / Files / Do / Verify / Done when, plus task-plan description, steps, must-haves, verification, inputs, and expected output. Inputs and Expected Output must include concrete backtick-wrapped paths; each task needs at least one output path.
+8. Each task needs a concrete title, Why / Files / Do / Verify / Done when, plus task-plan description, steps, must-haves, verification, inputs, and expected output. Inputs and Expected Output must include concrete backtick-wrapped paths. Use paths relative to `{{workingDirectory}}`; do not put absolute paths to the original checkout or any directory outside `{{workingDirectory}}` in `files`, `inputs`, `expectedOutput`, or verification commands. **`expected_output` must only list files the task actually creates or overwrites on disk.** Do NOT include files the task merely reads, verifies, or tests — those belong only in `inputs`. If a task is a pure verification or test task that produces no new files, its `expected_output` may be empty or limited to test-result artifacts (e.g. a log or assertion output). A file that does not yet exist on disk and is needed as an `input` must be produced by an earlier task's `expected_output` — if no prior task creates it, add a task before this one that does.
 9. Persist with `gsd_plan_slice` using goal, successCriteria, optional proofLevel/integrationClosure/observabilityImpact, and tasks. `gsd_plan_slice` handles task persistence transactionally and renders `{{outputPath}}` plus task plans; do not call `gsd_plan_task`. The DB-backed tool is the canonical write path. Do **not** rely on direct `PLAN.md` writes as the source of truth.
 10. Self-audit before finishing: goal/demo closure, requirement coverage, locked decisions, concrete paths, dependency order, wiring, scope size, proof truthfulness, feature completeness, and quality gates. Quality gates: non-trivial slices/tasks include specific Q3-Q7 coverage where applicable.
 11. If planning creates structural decisions, append them to `.gsd/DECISIONS.md`.