gsd-pi 2.80.0-dev.c5f2443b3 → 2.80.0-dev.e51d2c88c

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

@@ -1,3 +1,5 @@
+// GSD-2 + src/resources/extensions/gsd/bootstrap/agent-end-recovery.ts - Handles provider and agent-end recovery for GSD auto-mode.
+
 import type { ExtensionAPI, ExtensionContext } from "@gsd/pi-coding-agent";
 
 import { logWarning } from "../workflow-logger.js";
@@ -12,7 +14,7 @@ import { clearPathCache } from "../paths.js";
 import { getAutoDashboardData, getAutoModeStartModel, isAutoActive, pauseAuto, setCurrentDispatchedModelId } from "../auto.js";
 import { getNextFallbackModel, resolveModelWithFallbacksForUnit } from "../preferences.js";
 import { pauseAutoForProviderError } from "../provider-error-pause.js";
-import { isSessionSwitchInFlight, resolveAgentEnd } from "../auto/resolve.js";
+import { isSessionSwitchInFlight, resolveAgentEnd, resolveAgentEndCancelled } from "../auto/resolve.js";
 import { resolveModelId } from "../auto-model-selection.js";
 import { resolveProjectRoot } from "../worktree.js";
 import { clearDiscussionFlowState } from "./write-gate.js";
@@ -69,6 +71,11 @@ export function _buildAbortedPauseContext(lastMsg: { errorMessage?: unknown }):
   };
 }
 
+export function isUserInitiatedAbortMessage(message: string | undefined | null): boolean {
+  if (!message) return false;
+  return /\b(?:claude code process aborted by user|request aborted by user|process aborted by user)\b/i.test(message);
+}
+
 async function pauseTransientWithBackoff(
   cls: ErrorClass,
   pi: ExtensionAPI,
@@ -146,9 +153,29 @@ export async function handleAgentEnd(
   if (maybeHandleEmptyIntentTurn(event, isAutoActive())) return;
 
   if (!isAutoActive()) return;
-  if (isSessionSwitchInFlight()) return;
 
   const lastMsg = event.messages[event.messages.length - 1];
+  if (isSessionSwitchInFlight()) {
+    if (lastMsg && "stopReason" in lastMsg && lastMsg.stopReason === "error") {
+      const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+      if (isUserInitiatedAbortMessage(rawErrorMsg)) {
+        resolveAgentEndCancelled({
+          message: rawErrorMsg,
+          category: "aborted",
+          isTransient: false,
+        });
+      }
+    } else if (lastMsg && "stopReason" in lastMsg && lastMsg.stopReason === "aborted") {
+      const content = "content" in lastMsg ? lastMsg.content : undefined;
+      const hasEmptyContent = Array.isArray(content) && content.length === 0;
+      const hasErrorMessage = "errorMessage" in lastMsg && !!lastMsg.errorMessage;
+      if (!hasEmptyContent || hasErrorMessage) {
+        resolveAgentEndCancelled(_buildAbortedPauseContext(lastMsg as { errorMessage?: unknown }));
+      }
+    }
+    return;
+  }
+
   if (lastMsg && "stopReason" in lastMsg && lastMsg.stopReason === "aborted") {
     // Empty content with aborted stopReason is a non-fatal agent stop (the LLM
     // chose to end without producing output). Only pause on genuine fatal aborts
@@ -180,6 +207,14 @@ export async function handleAgentEnd(
     // is in the assistant message text content. Fall back to content when
     // errorMessage looks uninformative.
     const rawErrorMsg = ("errorMessage" in lastMsg && lastMsg.errorMessage) ? String(lastMsg.errorMessage) : "";
+    if (isUserInitiatedAbortMessage(rawErrorMsg)) {
+      resolveAgentEndCancelled({
+        message: rawErrorMsg,
+        category: "aborted",
+        isTransient: false,
+      });
+      return;
+    }
     const isUseless = !rawErrorMsg || /^(success|ok|true|error|unknown)$/i.test(rawErrorMsg.trim());
     // #3588: When errorMessage is uninformative, extract the real error from
     // the assistant message text content for display purposes only.

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

@@ -6,23 +6,36 @@
 import { Type } from "@sinclair/typebox";
 import type { ExtensionAPI } from "@gsd/pi-coding-agent";
 
+async function loadContextModePreferences(baseDir: string) {
+  const [{ loadEffectiveGSDPreferences }, { logWarning }] = await Promise.all([
+    import("../preferences.js"),
+    import("../workflow-logger.js"),
+  ]);
+  try {
+    return loadEffectiveGSDPreferences(baseDir)?.preferences ?? null;
+  } catch (err) {
+    logWarning("tool", `Context Mode tool could not load preferences: ${err instanceof Error ? err.message : String(err)}`);
+    return null;
+  }
+}
+
 export function registerExecTools(pi: ExtensionAPI): void {
   pi.registerTool({
     name: "gsd_exec",
     label: "Exec (Sandboxed)",
     description:
-      "Run a short script (bash/node/python) in a subprocess. Full stdout/stderr persist to " +
+      "Run a short script (bash/node/python) in a subprocess. Capped stdout/stderr and metadata persist to " +
       ".gsd/exec/<id>.{stdout,stderr,meta.json}; only a short digest returns in context. Use " +
       "this instead of reading many files or emitting large tool outputs — e.g. have the script " +
       "count/grep/summarize and log the finding. Enabled by default; opt out via " +
       "preferences.context_mode.enabled=false.",
     promptSnippet:
-      "Run a bash/node/python script in a sandbox; full output is saved to disk and only a digest returns",
+      "Run a bash/node/python script in a sandbox; capped output is saved to disk and only a digest returns",
     promptGuidelines: [
       "Prefer gsd_exec for analyses that would otherwise read >3 files or produce large tool output.",
       "Write scripts that log the finding (counts, matches, summaries) rather than raw dumps.",
       "The digest is the last ~300 chars of stdout — size your log output accordingly.",
-      "Need the full output? Read the stdout_path returned in details (file on local disk).",
+      "Need persisted output? Read the stdout_path returned in details (file on local disk).",
     ],
     parameters: Type.Object({
       runtime: Type.Union(
@@ -40,20 +53,11 @@ export function registerExecTools(pi: ExtensionAPI): void {
       ),
     }),
     async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
-      const [{ executeGsdExec }, { loadEffectiveGSDPreferences }, { logWarning }] = await Promise.all([
-        import("../tools/exec-tool.js"),
-        import("../preferences.js"),
-        import("../workflow-logger.js"),
-      ]);
-      let prefs: ReturnType<typeof loadEffectiveGSDPreferences> | null = null;
-      try {
-        prefs = loadEffectiveGSDPreferences();
-      } catch (err) {
-        logWarning("tool", `gsd_exec could not load preferences: ${err instanceof Error ? err.message : String(err)}`);
-      }
+      const { executeGsdExec } = await import("../tools/exec-tool.js");
+      const baseDir = process.cwd();
       return executeGsdExec(params as Parameters<typeof executeGsdExec>[0], {
-        baseDir: process.cwd(),
-        preferences: prefs?.preferences ?? null,
+        baseDir,
+        preferences: await loadContextModePreferences(baseDir),
       });
     },
   });
@@ -67,7 +71,7 @@ export function registerExecTools(pi: ExtensionAPI): void {
     promptSnippet: "Search prior gsd_exec runs by substring, runtime, or failing-only filter",
     promptGuidelines: [
       "Use this before re-running an expensive analysis — the prior run's stdout file may still answer.",
-      "The preview shows the trailing ~300 chars of stdout; read stdout_path for the full transcript.",
+      "The preview shows the trailing ~300 chars of stdout; read stdout_path for persisted output.",
     ],
     parameters: Type.Object({
       query: Type.Optional(Type.String({ description: "Substring matched against id and purpose (case-insensitive)." })),
@@ -81,8 +85,10 @@ export function registerExecTools(pi: ExtensionAPI): void {
     }),
     async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
       const { executeExecSearch } = await import("../tools/exec-search-tool.js");
+      const baseDir = process.cwd();
       return executeExecSearch(params as Parameters<typeof executeExecSearch>[0], {
-        baseDir: process.cwd(),
+        baseDir,
+        preferences: await loadContextModePreferences(baseDir),
       });
     },
   });
@@ -102,8 +108,10 @@ export function registerExecTools(pi: ExtensionAPI): void {
     parameters: Type.Object({}),
     async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
       const { executeResume } = await import("../tools/resume-tool.js");
+      const baseDir = process.cwd();
       return executeResume(params as Parameters<typeof executeResume>[0], {
-        baseDir: process.cwd(),
+        baseDir,
+        preferences: await loadContextModePreferences(baseDir),
       });
     },
   });

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

@@ -7,7 +7,7 @@ import type { GSDEcosystemBeforeAgentStartHandler } from "../ecosystem/gsd-exten
 import { updateSnapshot } from "../ecosystem/gsd-extension-api.js";
 
 import { buildMilestoneFileName, resolveMilestonePath, resolveSliceFile, resolveSlicePath } from "../paths.js";
-import { canonicalToolName, clearDiscussionFlowState, isDepthConfirmationAnswer, isQueuePhaseActive, markApprovalGateVerified, markDepthVerified, resetWriteGateState, shouldBlockContextWrite, shouldBlockPlanningUnit, shouldBlockQueueExecution, isGateQuestionId, setPendingGate, clearPendingGate, getPendingGate, shouldBlockPendingGate, shouldBlockPendingGateBash, extractDepthVerificationMilestoneId } from "./write-gate.js";
+import { canonicalToolName, clearDiscussionFlowState, isDepthConfirmationAnswer, isQueuePhaseActive, markApprovalGateVerified, markDepthVerified, resetWriteGateState, shouldBlockContextWrite, shouldBlockPlanningUnit, shouldBlockQueueExecution, shouldBlockWorktreeWrite, isGateQuestionId, setPendingGate, clearPendingGate, getPendingGate, shouldBlockPendingGate, shouldBlockPendingGateBash, extractDepthVerificationMilestoneId } from "./write-gate.js";
 import { resolveManifest } from "../unit-context-manifest.js";
 import { isBlockedStateFile, isBashWriteToStateFile, BLOCKED_WRITE_ERROR } from "../write-intercept.js";
 import { loadFile, saveFile, formatContinue } from "../files.js";
@@ -65,6 +65,26 @@ async function applyDisabledModelProviderPolicy(ctx: ExtensionContext): Promise<
   }
 }
 
+/**
+ * Bridge `context_management.compaction_threshold_percent` from GSD preferences
+ * into the agent's runtime compaction settings (#5475). The preference is
+ * validated to (0.5, 0.95) at load time, but defense-in-depth normalization
+ * here protects against a stale or hand-edited prefs file. Calling with
+ * `undefined` clears any prior override so a removed preference does not leak.
+ */
+async function applyCompactionThresholdOverride(ctx: ExtensionContext): Promise<void> {
+  try {
+    const { loadEffectiveGSDPreferences } = await import("../preferences.js");
+    const prefs = loadEffectiveGSDPreferences();
+    const raw = prefs?.preferences.context_management?.compaction_threshold_percent;
+    const value =
+      typeof raw === "number" && Number.isFinite(raw) && raw > 0 && raw < 1 ? raw : undefined;
+    ctx.setCompactionThresholdOverride(value);
+  } catch {
+    // Non-fatal: leave any existing override in place.
+  }
+}
+
 export function resolveNotificationStoreBasePath(cwd: string = process.cwd()): string {
   return resolveWorktreeProjectRoot(cwd);
 }
@@ -123,6 +143,7 @@ export function registerHooks(
     await resetAskUserQuestionsTurnCache();
     await syncServiceTierStatus(ctx);
     await applyDisabledModelProviderPolicy(ctx);
+    await applyCompactionThresholdOverride(ctx);
     // Skip MCP auto-prep when running inside an auto-worktree (see session_switch below).
     const { isInAutoWorktree } = await import("../auto-worktree.js");
     if (!isInAutoWorktree(process.cwd())) {
@@ -172,6 +193,7 @@ export function registerHooks(
     clearDiscussionFlowState(process.cwd());
     await syncServiceTierStatus(ctx);
     await applyDisabledModelProviderPolicy(ctx);
+    await applyCompactionThresholdOverride(ctx);
     // Skip MCP auto-prep when running inside an auto-worktree. The worktree
     // already has .mcp.json from createAutoWorktree, and re-running the writer
     // post-chdir rewrites the file mid-run (non-idempotent due to cwd-relative
@@ -482,6 +504,22 @@ export function registerHooks(
       }
     }
 
+    // ── Worktree-isolation write gate (#5199) ────────────────────────────
+    // Block planning-write tools from landing code at the project root when
+    // git.isolation=worktree but auto-mode hasn't created the milestone
+    // worktree yet. Without this, writes silently orphan outside git history.
+    if (isToolCallEventType("write", event) || isToolCallEventType("edit", event)) {
+      const wtBasePath = resolveWorktreeProjectRoot(dash.basePath ?? discussionBasePath);
+      const wtGuard = shouldBlockWorktreeWrite(
+        event.toolName,
+        event.input.path,
+        wtBasePath,
+        isAutoActive(),
+        dash.currentUnit?.type,
+      );
+      if (wtGuard.block) return wtGuard;
+    }
+
     // ── Single-writer engine: block direct writes to STATE.md ──────────
     // Covers write, edit, and bash tools to prevent bypass vectors.
     if (isToolCallEventType("write", event)) {

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

@@ -1,10 +1,12 @@
-import { copyFileSync, existsSync, mkdirSync, readFileSync, renameSync, unlinkSync, writeFileSync } from "node:fs";
+import { copyFileSync, existsSync, mkdirSync, readFileSync, realpathSync, renameSync, unlinkSync, writeFileSync } from "node:fs";
 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 { logWarning } from "../workflow-logger.js";
+import { isGsdWorktreePath, resolveWorktreeProjectRoot } from "../worktree-root.js";
 
 /**
  * Regex matching milestone CONTEXT.md file names in both legacy M001
@@ -882,3 +884,135 @@ export function shouldBlockPlanningUnit(
   // avoids breaking gsd_* MCP tools or future safe additions.
   return { block: false };
 }
+
+// ─── Worktree isolation write gate (#5199) ────────────────────────────────
+//
+// When `git.isolation: worktree` is configured, the per-unit commit pipeline
+// only runs inside the auto-mode loop (`auto-post-unit.ts`). If the LLM
+// authors code at the project root before auto-mode is started, those writes
+// land in the working tree but never reach a commit — they're silently
+// orphaned outside git history. This guard blocks those writes at the
+// tool_call seam so the agent receives a clear error instead.
+
+const WORKTREE_GATE_BOOTSTRAP_UNITS = new Set([
+  "discuss-milestone",
+  "plan-milestone",
+  "init",
+]);
+
+function realpathOrResolve(p: string): string {
+  const abs = resolve(p);
+  try {
+    return realpathSync(abs);
+  } catch {
+    // Path doesn't exist (yet) — realpath the deepest existing ancestor so
+    // platforms where /tmp -> /private/tmp don't break containment checks.
+    let dir = abs;
+    const tail: string[] = [];
+    while (dir && dir !== resolve(dir, "..")) {
+      try {
+        const real = realpathSync(dir);
+        return tail.length ? join(real, ...tail.reverse()) : real;
+      } catch {
+        const idx = dir.lastIndexOf(sep);
+        if (idx <= 0) break;
+        tail.push(dir.slice(idx + 1));
+        dir = dir.slice(0, idx) || sep;
+      }
+    }
+    return abs;
+  }
+}
+
+function isPathContained(target: string, container: string): boolean {
+  if (target === container) return true;
+  return target.startsWith(container.endsWith(sep) ? container : container + sep);
+}
+
+/**
+ * Block planning-write tool calls that would land code at the project root
+ * while `git.isolation: worktree` is in effect and auto-mode hasn't created
+ * (or flipped cwd into) the milestone worktree.
+ *
+ * Pure / unit-testable. Callers in `register-hooks.ts` supply the resolved
+ * project root and current auto liveness; this function does no I/O beyond
+ * realpath resolution.
+ *
+ * Allow rules (in order):
+ *   1. Tool isn't a planning-write (write/edit/multi_edit/notebook_edit).
+ *   2. `GSD_DISABLE_WORKTREE_WRITE_GUARD=1` self-hosting bypass.
+ *   3. Isolation mode is not "worktree".
+ *   4. Active unit is a bootstrap unit (discuss-milestone/plan-milestone/init).
+ *   5. Target is inside `<projectRoot>/.gsd/worktrees/` (a real worktree).
+ *   6. Target is inside `<projectRoot>/.gsd/` and isn't masquerading as a
+ *      worktrees sibling (rejects the `.gsd/worktrees-extra/…` prefix trick).
+ *   7. Auto is live AND `effectiveBasePath` is itself a `.gsd/worktrees/…` path.
+ *
+ * Otherwise: block with a message that points the agent at `/gsd` to start
+ * auto-mode.
+ */
+export function shouldBlockWorktreeWrite(
+  toolName: string,
+  targetPath: string,
+  effectiveBasePath: string,
+  isAutoLive: boolean,
+  currentUnitType?: string | null,
+): { block: boolean; reason?: string } {
+  const tool = canonicalToolName(toolName);
+  if (!PLANNING_WRITE_TOOLS.has(tool)) return { block: false };
+  if (process.env.GSD_DISABLE_WORKTREE_WRITE_GUARD === "1") return { block: false };
+  if (getIsolationMode(effectiveBasePath) !== "worktree") return { block: false };
+  if (currentUnitType && WORKTREE_GATE_BOOTSTRAP_UNITS.has(currentUnitType)) return { block: false };
+
+  if (!targetPath) {
+    return {
+      block: true,
+      reason: [
+        `HARD BLOCK: ${tool} called with empty path while \`git.isolation: worktree\` is configured`,
+        `and auto-mode is not active. Refusing to allow writes that cannot be located.`,
+      ].join(" "),
+    };
+  }
+
+  // Resolve the target relative to the project root, then realpath to defeat
+  // symlink-based escapes and prefix tricks (e.g. .gsd/worktrees-extra/).
+  const projectRoot = resolveWorktreeProjectRoot(effectiveBasePath);
+  const absTarget = isAbsolute(targetPath) ? targetPath : resolve(projectRoot, targetPath);
+  const realTarget = realpathOrResolve(absTarget);
+  const realRoot = realpathOrResolve(projectRoot);
+  const realGsd = realpathOrResolve(join(projectRoot, ".gsd"));
+  const realWorktreesDir = realpathOrResolve(join(projectRoot, ".gsd", "worktrees"));
+
+  // Allow writes inside the legitimate worktrees subtree.
+  if (isPathContained(realTarget, realWorktreesDir)) return { block: false };
+
+  // Allow writes to .gsd/ planning artifacts, but reject siblings whose name
+  // starts with "worktrees" (the worktrees-extra prefix trick — case 4).
+  if (isPathContained(realTarget, realGsd)) {
+    const rel = relative(realGsd, realTarget);
+    const firstSeg = rel.split(/[\/\\]/)[0] ?? "";
+    if (!firstSeg.startsWith("worktrees")) return { block: false };
+    // fall through: looks like worktrees<something> sibling — block
+  }
+
+  // Auto is live and the caller is operating inside a worktree path —
+  // host tool's write happens in worktree context; let it through.
+  if (isAutoLive && isGsdWorktreePath(effectiveBasePath)) return { block: false };
+
+  // Block. Provide enough context that the agent can self-correct.
+  const displayTarget = isPathContained(realTarget, realRoot)
+    ? relative(realRoot, realTarget) || "."
+    : realTarget;
+  return {
+    block: true,
+    reason: [
+      `HARD BLOCK: Worktree isolation is configured (\`git.isolation: worktree\`) but auto-mode is`,
+      `not running and the target "${displayTarget}" is not inside \`.gsd/worktrees/<MID>/\`.`,
+      `Code edits at the project root would be lost — only the auto-mode commit pipeline`,
+      `(auto-post-unit) commits work, and it never runs outside the loop.`,
+      `Required action: start auto-mode with \`/gsd\` so the milestone worktree is created,`,
+      `then write inside it. To disable this guard for self-hosting development, set`,
+      `GSD_DISABLE_WORKTREE_WRITE_GUARD=1.`,
+    ].join(" "),
+  };
+}

package/src/resources/extensions/gsd/clean-root-preflight.ts

@@ -21,10 +21,34 @@ import { nativeHasChanges } from "./native-git-bridge.js";
 export interface PreflightResult {
   /** true when a stash was pushed and postflightPopStash should be called */
   stashPushed: boolean;
+  /** Unique marker embedded in the stash message for targeted restoration */
+  stashMarker?: string;
   /** human-readable summary of what happened (empty string for clean trees) */
   summary: string;
 }
 
+function findPreflightStashRef(basePath: string, milestoneId: string, stashMarker?: string): string | null {
+  const markerPrefix = `gsd-preflight-stash:${milestoneId}:`;
+  let fallbackRef: string | null = null;
+  try {
+    const list = execFileSync("git", ["stash", "list", "--format=%gd%x00%s"], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+      encoding: "utf-8",
+      env: GIT_NO_PROMPT_ENV,
+    });
+    for (const line of list.split("\n")) {
+      const [ref, subject] = line.split("\x00");
+      if (!ref || !subject) continue;
+      if (stashMarker && subject.includes(stashMarker)) return ref;
+      if (!fallbackRef && subject.includes(markerPrefix)) fallbackRef = ref;
+    }
+  } catch (err) {
+    logWarning("preflight", `stash list failed before restore: ${err instanceof Error ? err.message : String(err)}`);
+  }
+  return fallbackRef;
+}
+
 /**
  * Check the working tree for dirty files before a milestone merge.
  *
@@ -62,7 +86,8 @@ export function preflightCleanRoot(
 
   // Push the stash
   try {
-    execFileSync("git", ["stash", "push", "--include-untracked", "-m", "gsd-preflight-stash"], {
+    const stashMarker = `gsd-preflight-stash:${milestoneId}:${process.pid}:${Date.now()}:${process.hrtime.bigint().toString(36)}`;
+    execFileSync("git", ["stash", "push", "--include-untracked", "-m", `gsd-preflight-stash [${stashMarker}]`], {
       cwd: basePath,
       stdio: ["ignore", "pipe", "pipe"],
       encoding: "utf-8",
@@ -70,6 +95,7 @@ export function preflightCleanRoot(
     });
     return {
       stashPushed: true,
+      stashMarker,
       summary: `Stashed uncommitted changes before merge (milestone ${milestoneId}).`,
     };
   } catch (err) {
@@ -91,10 +117,19 @@ export function preflightCleanRoot(
 export function postflightPopStash(
   basePath: string,
   milestoneId: string,
+  stashMarker: string | undefined,
   notify: (message: string, level: "info" | "warning" | "error") => void,
 ): void {
+  let stashRef: string | null = null;
   try {
-    execFileSync("git", ["stash", "pop"], {
+    stashRef = findPreflightStashRef(basePath, milestoneId, stashMarker);
+    if (!stashRef) {
+      const msg = `No matching GSD preflight stash found for milestone ${milestoneId}; leaving stash list untouched.`;
+      logWarning("preflight", msg);
+      notify(msg, "warning");
+      return;
+    }
+    execFileSync("git", ["stash", "pop", stashRef], {
       cwd: basePath,
       stdio: ["ignore", "pipe", "pipe"],
       encoding: "utf-8",
@@ -104,7 +139,10 @@ export function postflightPopStash(
   } catch (err) {
     // Pop conflicts mean the merged code collides with the stashed changes.
     // Log a warning — the user needs to resolve manually, but the merge succeeded.
-    const msg = `git stash pop failed after merge of milestone ${milestoneId}: ${err instanceof Error ? err.message : String(err)}. Run "git stash pop" manually to restore your changes.`;
+    const restoreHint = stashRef
+      ? `Run "git stash pop ${stashRef}" or "git stash apply ${stashRef}" manually to restore the correct stash.`
+      : `Run "git stash list" to find the matching GSD preflight stash before restoring manually.`;
+    const msg = `git stash pop ${stashRef ?? ""}`.trim() + ` failed after merge of milestone ${milestoneId}: ${err instanceof Error ? err.message : String(err)}. ${restoreHint}`;
     logWarning("preflight", msg);
     notify(msg, "warning");
   }

package/src/resources/extensions/gsd/commands/dispatcher.ts

@@ -45,5 +45,11 @@ export async function handleGSDCommand(
 
   if (handled) return;
 
+  if (trimmed.includes(" ")) {
+    const { handleDo } = await import("../commands-do.js");
+    await handleDo(trimmed, ctx, pi);
+    return;
+  }
+
   ctx.ui.notify(`Unknown: /gsd ${trimmed}. Run /gsd help for available commands.`, "warning");
 }

package/src/resources/extensions/gsd/commands-extract-learnings.ts

@@ -222,14 +222,19 @@ Using the \`write\` tool, persist the full structured report to
 LEARNINGS.md is the full, cited audit trail. Write it first — subsequent steps
 feed from its content.
 
-### Step 3 — Optionally pre-query the memory store for semantic duplicates
+### Step 3 — Run one bounded duplicate check
 
-Before persisting any extracted item in Steps 4–6, you may call
-\`memory_query\` with 2–3 keywords from the item to check whether the
-memory store already holds a semantically equivalent entry at high
-confidence. Skip those items in their respective steps. The memory store
-is the single source of truth for cross-session durable knowledge — no
-other persistence call is part of this flow.
+Before persisting extracted items in Steps 4–6, do at most one duplicate-check
+pass for the durable Decisions, Lessons, and Patterns from this milestone. Use
+the already-written LEARNINGS.md content and call \`memory_query\` once with a
+compact keyword summary for the whole batch. If that result clearly shows a
+semantically equivalent high-confidence memory for an item, mark only that item
+as already captured and skip it in its respective persistence step.
+
+Do not re-read milestone artefacts or repeat memory queries category-by-category
+after this point. The memory store is the single source of truth for
+cross-session durable knowledge — no other persistence call is part of this
+flow.
 
 ### Step 4 — Persist Patterns via \`capture_thought\`
 
@@ -268,11 +273,11 @@ later projection back to a human-visible decisions register stays lossless
 
 ### Step 7 — Deduplication rule (applies to Steps 4, 5, 6)
 
-Before each \`capture_thought\` call, optionally call \`memory_query\` with 2–3
-keywords from the entry. If a semantically equivalent memory is returned at
-high confidence, skip the capture entirely. Prefer skipping a near-duplicate
-over creating a second slightly-different row — redundancy degrades the
-signal.
+Use only the duplicate-check result from Step 3. If that bounded check returned
+a semantically equivalent memory at high confidence for an extracted item, skip
+the capture entirely. Otherwise, persist the item once via \`capture_thought\`.
+Prefer skipping a near-duplicate over creating a second slightly-different row
+— redundancy degrades the signal.
 
 ### Step 8 — Surprises stay only in LEARNINGS.md
 

package/src/resources/extensions/gsd/context-budget.ts

@@ -8,7 +8,12 @@
  * @see D001 (module location), D002 (200K fallback), D003 (section-boundary truncation)
  */
 
-import { type TokenProvider, getCharsPerToken } from "./token-counter.js";
+import {
+  type TokenProvider,
+  getCharsPerToken,
+  isAccurateCountingAvailable,
+  countTokensSync,
+} from "./token-counter.js";
 
 // ─── Budget ratio constants ──────────────────────────────────────────────────
 // Percentages of total context window allocated to each budget category.
@@ -32,6 +37,24 @@ const DEFAULT_CONTEXT_WINDOW = 200_000;
 /** Conservative effective context for Claude Code subscription routing (#4676) */
 const CLAUDE_CODE_EFFECTIVE_CONTEXT_WINDOW = 200_000;
 
+/**
+ * Cached empirical chars-per-token from a tiktoken probe, keyed by provider.
+ * countTokensSync's fallback path is provider-aware, so we cache per-provider
+ * to preserve that distinction once the encoder warms. The cl100k_base encoder
+ * itself gives a stable ratio for ASCII English so a single probe per provider
+ * key is sufficient. Empty map means "not yet probed" or "encoder unavailable".
+ */
+const _empiricalCharsPerTokenByProvider = new Map<string, number>();
+
+/**
+ * Test hook — clears the empirical chars-per-token cache so test cases that
+ * assert against the static char-ratio fallback aren't polluted by a prior
+ * tiktoken-warmed run in the same process. Production code must not call this.
+ */
+export function _resetEmpiricalCacheForTest(): void {
+  _empiricalCharsPerTokenByProvider.clear();
+}
+
 /** Percentage of context consumed before suggesting a continue-here checkpoint */
 const CONTINUE_THRESHOLD_PERCENT = 70;
 
@@ -101,7 +124,26 @@ export interface MinimalPreferences {
 export function computeBudgets(contextWindow: number, provider?: TokenProvider): BudgetAllocation {
   const effectiveWindow = contextWindow > 0 ? contextWindow : DEFAULT_CONTEXT_WINDOW;
   const charsPerToken = provider ? getCharsPerToken(provider) : CHARS_PER_TOKEN;
-  const totalChars = effectiveWindow * charsPerToken;
+
+  // Prefer the tiktoken encoder for total-char estimation when it has been
+  // warmed (initTokenCounter resolved). The cl100k_base ratio is stable for
+  // ASCII English, so probe once per provider and cache — computeBudgets is
+  // called multiple times per prompt build and the probe encode is otherwise
+  // wasted work.
+  let totalChars: number;
+  if (isAccurateCountingAvailable()) {
+    const providerKey = provider ?? "__default__";
+    let empirical = _empiricalCharsPerTokenByProvider.get(providerKey);
+    if (empirical === undefined) {
+      const probe = "the quick brown fox jumps over the lazy dog ".repeat(64);
+      const probeTokens = countTokensSync(probe, provider);
+      empirical = probeTokens > 0 ? probe.length / probeTokens : charsPerToken;
+      _empiricalCharsPerTokenByProvider.set(providerKey, empirical);
+    }
+    totalChars = effectiveWindow * empirical;
+  } else {
+    totalChars = effectiveWindow * charsPerToken;
+  }
 
   return {
     summaryBudgetChars: Math.floor(totalChars * SUMMARY_RATIO),