gsd-pi 2.76.0 → 2.77.0-dev.1d17f366c

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

@@ -1,7 +1,13 @@
 import type { ExtensionAPI, ExtensionContext } from "@gsd/pi-coding-agent";
 
 import { logWarning } from "../workflow-logger.js";
-import { checkAutoStartAfterDiscuss } from "../guided-flow.js";
+import {
+  checkAutoStartAfterDiscuss,
+  maybeHandleReadyPhraseWithoutFiles,
+  maybeHandleEmptyIntentTurn,
+  resetEmptyTurnCounter,
+} from "../guided-flow.js";
+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";
@@ -16,6 +22,7 @@ import {
   isTransient,
   type ErrorClass,
 } from "../error-classifier.js";
+import { blockModel, isModelBlocked } from "../blocked-models.js";
 
 const retryState = createRetryState();
 const MAX_NETWORK_RETRIES = 2;
@@ -70,10 +77,34 @@ export async function handleAgentEnd(
   event: { messages: any[] },
   ctx: ExtensionContext,
 ): Promise<void> {
+  // #4648 — Invalidate the directory-listing cache before any artifact-existence
+  // checks. The LLM may have written milestone files (CONTEXT.md, ROADMAP.md,
+  // PROJECT.md, REQUIREMENTS.md) via tool calls during the turn that just
+  // ended. `paths.ts` caches readdir() results without a TTL, so without this
+  // flush, `resolveMilestoneFile` returns the pre-write listing and the guards
+  // below (`checkAutoStartAfterDiscuss` and `maybeHandleReadyPhraseWithoutFiles`)
+  // falsely report files as missing — producing a spurious "ready signal
+  // rejected" loop even though the files are on disk.
+  clearPathCache();
+
   if (checkAutoStartAfterDiscuss()) {
     clearDiscussionFlowState();
     return;
   }
+
+  // #4573 — When the LLM emits "Milestone X ready." but the required files
+  // 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;
+
+  // #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 (!isAutoActive()) return;
   if (isSessionSwitchInFlight()) return;
 
@@ -124,6 +155,102 @@ export async function handleAgentEnd(
     // ── 1. Classify using rawErrorMsg to avoid prose false-positives ────
     const cls = classifyError(rawErrorMsg, explicitRetryAfterMs);
 
+    // ── 1a. Unsupported-model: provider rejected this model for the current
+    //        account/plan at request time (#4513).  Persist a block so the
+    //        same dead model isn't reselected on the next /gsd auto restart,
+    //        then try a fallback before pausing.
+    if (cls.kind === "unsupported-model") {
+      const dash = getAutoDashboardData();
+      const rejectedProvider = ctx.model?.provider;
+      const rejectedId = ctx.model?.id;
+      if (dash.basePath && rejectedProvider && rejectedId) {
+        try {
+          blockModel(dash.basePath, rejectedProvider, rejectedId, rawErrorMsg || "unsupported for account");
+          ctx.ui.notify(
+            `Blocked ${rejectedProvider}/${rejectedId} for this project — provider rejected it for the current account.`,
+            "warning",
+          );
+        } catch (err) {
+          const m = err instanceof Error ? err.message : String(err);
+          logWarning("bootstrap", `Failed to persist blocked model: ${m}`);
+        }
+      }
+
+      // Try configured fallback chain, skipping anything already blocked.
+      if (dash.currentUnit && dash.basePath) {
+        const modelConfig = resolveModelWithFallbacksForUnit(dash.currentUnit.type);
+        if (modelConfig && modelConfig.fallbacks.length > 0) {
+          const availableModels = ctx.modelRegistry.getAvailable();
+          let cursorModelId: string | undefined = ctx.model?.id;
+          while (true) {
+            const nextModelId = getNextFallbackModel(cursorModelId, modelConfig);
+            if (!nextModelId) break;
+            const candidate = resolveModelId(nextModelId, availableModels, ctx.model?.provider);
+            if (candidate && !isModelBlocked(dash.basePath, candidate.provider, candidate.id)) {
+              const ok = await pi.setModel(candidate, { persist: false });
+              if (ok) {
+                setCurrentDispatchedModelId({ provider: candidate.provider, id: candidate.id });
+                ctx.ui.notify(
+                  `Switched to fallback ${candidate.provider}/${candidate.id} after account entitlement rejection.`,
+                  "warning",
+                );
+                pi.sendMessage(
+                  { customType: "gsd-auto-timeout-recovery", content: "Continue execution.", display: false },
+                  { triggerTurn: true },
+                );
+                return;
+              }
+            }
+            cursorModelId = nextModelId;
+          }
+        }
+
+        // Fallback chain exhausted — try the auto-mode start model if it isn't
+        // the same one we just blocked and isn't itself blocked.
+        const sessionModel = getAutoModeStartModel();
+        if (
+          sessionModel &&
+          !(sessionModel.provider === rejectedProvider && sessionModel.id === rejectedId) &&
+          !isModelBlocked(dash.basePath, sessionModel.provider, sessionModel.id)
+        ) {
+          const startModel = ctx.modelRegistry
+            .getAvailable()
+            .find((m) => m.provider === sessionModel.provider && m.id === sessionModel.id);
+          if (startModel) {
+            const ok = await pi.setModel(startModel, { persist: false });
+            if (ok) {
+              setCurrentDispatchedModelId({ provider: startModel.provider, id: startModel.id });
+              ctx.ui.notify(
+                `Restored auto-mode start model ${startModel.provider}/${startModel.id} after entitlement rejection.`,
+                "warning",
+              );
+              pi.sendMessage(
+                { customType: "gsd-auto-timeout-recovery", content: "Continue execution.", display: false },
+                { triggerTurn: true },
+              );
+              return;
+            }
+          }
+        }
+      }
+
+      // No usable fallback — pause with a clearly named message.
+      const blockedLabel = rejectedProvider && rejectedId ? `${rejectedProvider}/${rejectedId}` : "current model";
+      const pauseDetail = `Model ${blockedLabel} blocked for this account${errorDetail}. Configure a different model and restart /gsd auto.`;
+      await pauseAutoForProviderError(ctx.ui, pauseDetail, () =>
+        pauseAuto(ctx, pi, {
+          message: pauseDetail,
+          category: "provider",
+          isTransient: false,
+        }),
+      {
+        isRateLimit: false,
+        isTransient: false,
+        retryAfterMs: 0,
+      });
+      return;
+    }
+
     // ── 1b. Defer to Core RetryHandler for most transient errors ────────
     // Core retries transient failures in-session after this handler.
     // Keep that behavior for non-rate-limit classes to avoid pause/retry races,
@@ -239,6 +366,9 @@ export async function handleAgentEnd(
   // ── Success path ─────────────────────────────────────────────────────────
   try {
     resetRetryState(retryState);
+    // #4573 — Reset the empty-turn counter on any successful agent turn so
+    // transient stalls don't accumulate across independent units.
+    resetEmptyTurnCounter();
     resolveAgentEnd(event);
   } catch (err) {
     const message = err instanceof Error ? err.message : String(err);

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

@@ -35,6 +35,19 @@ function registerAlias(pi: ExtensionAPI, toolDef: any, aliasName: string, canoni
   });
 }
 
+/**
+ * Read a tool result's structured payload, accommodating MCP's `details` →
+ * `structuredContent` rename (#4472, #4477). In-process executions still
+ * deliver the payload on `result.details`; MCP-routed executions deliver it
+ * on `result.structuredContent` (post `adaptExecutorResult` transform). All
+ * `renderResult` callbacks in this file route through this helper so a future
+ * field rename only needs to be applied in one place.
+ */
+// eslint-disable-next-line @typescript-eslint/no-explicit-any -- result shape varies by tool
+function readDetails(result: any): any {
+  return result?.details ?? result?.structuredContent;
+}
+
 export function registerDbTools(pi: ExtensionAPI): void {
   // ─── gsd_decision_save (formerly gsd_save_decision) ─────────────────────
 
@@ -110,7 +123,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
       return new Text(text, 0, 0);
     },
     renderResult(result: any, _options: any, theme: any) {
-      const d = result.details;
+      const d = readDetails(result);
       if (result.isError || d?.error) {
         return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
       }
@@ -188,7 +201,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
       return new Text(text, 0, 0);
     },
     renderResult(result: any, _options: any, theme: any) {
-      const d = result.details;
+      const d = readDetails(result);
       if (result.isError || d?.error) {
         return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
       }
@@ -273,7 +286,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
       return new Text(text, 0, 0);
     },
     renderResult(result: any, _options: any, theme: any) {
-      const d = result.details;
+      const d = readDetails(result);
       if (result.isError || d?.error) {
         return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
       }
@@ -322,7 +335,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
       return new Text(text, 0, 0);
     },
     renderResult(result: any, _options: any, theme: any) {
-      const d = result.details;
+      const d = readDetails(result);
       if (result.isError || d?.error) {
         return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
       }
@@ -406,7 +419,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
       return new Text(theme.fg("toolTitle", theme.bold("milestone_generate_id")), 0, 0);
     },
     renderResult(result: any, _options: any, theme: any) {
-      const d = result.details;
+      const d = readDetails(result);
       if (result.isError || d?.error) {
         return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
       }
@@ -1074,13 +1087,31 @@ export function registerDbTools(pi: ExtensionAPI): void {
       text += theme.fg("dim", ` → ${args.verdict ?? ""}`);
       return new Text(text, 0, 0);
     },
+    /**
+     * Render the save_gate_result tool output for the TUI.
+     *
+     * Prefers structured fields, but falls back to `content[0].text` when the
+     * structured payload is empty. Defensive: the structural fix on this
+     * branch plumbs `details` through MCP via `structuredContent`, but older
+     * hosts, a future handler that forgets `structuredContent`, or any drop
+     * of non-standard return fields would otherwise render as
+     * "undefined: undefined". Same fallback applies to error rendering, and
+     * we strip a leading `Error:` from the fallback text to avoid producing
+     * `Error: Error: ...`.
+     */
     renderResult(result: any, _options: any, theme: any) {
-      const d = result.details;
+      const d = readDetails(result);
       if (result.isError || d?.error) {
-        return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
+        const rawMsg = d?.error ?? result.content?.[0]?.text ?? "unknown";
+        const msg = rawMsg.replace(/^\s*Error:\s*/i, "");
+        return new Text(theme.fg("error", `Error: ${msg}`), 0, 0);
+      }
+      if (!d?.gateId || !d?.verdict) {
+        const text = result.content?.[0]?.text ?? "Gate result saved";
+        return new Text(theme.fg("success", text), 0, 0);
       }
-      const color = d?.verdict === "flag" ? "warning" : "success";
-      return new Text(theme.fg(color, `${d?.gateId}: ${d?.verdict}`), 0, 0);
+      const color = d.verdict === "flag" ? "warning" : "success";
+      return new Text(theme.fg(color, `${d.gateId}: ${d.verdict}`), 0, 0);
     },
   };
 

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

@@ -0,0 +1,109 @@
+// GSD2 — Exec (context-mode) tool registration.
+//
+// Exposes the `gsd_exec` tool over MCP. Opt-in: disabled unless
+// `context_mode.enabled: true` is set in preferences.
+
+import { Type } from "@sinclair/typebox";
+import type { ExtensionAPI } from "@gsd/pi-coding-agent";
+
+import { executeGsdExec } from "../tools/exec-tool.js";
+import { executeExecSearch } from "../tools/exec-search-tool.js";
+import { executeResume } from "../tools/resume-tool.js";
+import { loadEffectiveGSDPreferences } from "../preferences.js";
+import { logWarning } from "../workflow-logger.js";
+
+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 " +
+      ".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",
+    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).",
+    ],
+    parameters: Type.Object({
+      runtime: Type.Union(
+        [Type.Literal("bash"), Type.Literal("node"), Type.Literal("python")],
+        { description: "Interpreter: bash (-c), node (-e), or python3 (-c)." },
+      ),
+      script: Type.String({ description: "Script body. Keep output small (log the finding, not the data)." }),
+      purpose: Type.Optional(Type.String({ description: "Short label recorded in meta.json for later review." })),
+      timeout_ms: Type.Optional(
+        Type.Number({
+          description: "Per-invocation timeout (ms). Capped at 600000. Default from preferences.",
+          minimum: 1_000,
+          maximum: 600_000,
+        }),
+      ),
+    }),
+    async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
+      let prefs: Awaited<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)}`);
+      }
+      return executeGsdExec(params as Parameters<typeof executeGsdExec>[0], {
+        baseDir: process.cwd(),
+        preferences: prefs?.preferences ?? null,
+      });
+    },
+  });
+
+  pi.registerTool({
+    name: "gsd_exec_search",
+    label: "Search gsd_exec History",
+    description:
+      "List prior gsd_exec runs (most recent first) from .gsd/exec/*.meta.json. Useful for " +
+      "rediscovering the stdout_path of an earlier run without re-executing it. Read-only.",
+    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.",
+    ],
+    parameters: Type.Object({
+      query: Type.Optional(Type.String({ description: "Substring matched against id and purpose (case-insensitive)." })),
+      runtime: Type.Optional(
+        Type.Union([Type.Literal("bash"), Type.Literal("node"), Type.Literal("python")], {
+          description: "Restrict to one runtime.",
+        }),
+      ),
+      failing_only: Type.Optional(Type.Boolean({ description: "Only non-zero exit codes and timeouts." })),
+      limit: Type.Optional(Type.Number({ description: "Max results (default 20, cap 200)", minimum: 1, maximum: 200 })),
+    }),
+    async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
+      return executeExecSearch(params as Parameters<typeof executeExecSearch>[0], {
+        baseDir: process.cwd(),
+      });
+    },
+  });
+
+  pi.registerTool({
+    name: "gsd_resume",
+    label: "Resume (Read Snapshot)",
+    description:
+      "Return the contents of .gsd/last-snapshot.md — a ≤2 KB digest of top memories, recent " +
+      "gsd_exec runs, and active context, written automatically on session_before_compact. Use " +
+      "this after compaction or session resume to re-orient quickly.",
+    promptSnippet: "Read the pre-compaction snapshot to re-orient after context loss",
+    promptGuidelines: [
+      "Call this right after a session resumes if you feel you've lost durable context.",
+      "The snapshot is a summary — use memory_query or gsd_exec_search for detail.",
+    ],
+    parameters: Type.Object({}),
+    async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
+      return executeResume(params as Parameters<typeof executeResume>[0], {
+        baseDir: process.cwd(),
+      });
+    },
+  });
+}

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

@@ -50,6 +50,11 @@ export function registerMemoryTools(pi: ExtensionAPI): void {
       ),
       tags: Type.Optional(Type.Array(Type.String(), { description: "Free-form tags (reserved for future use)" })),
       scope: Type.Optional(Type.String({ description: "Scope name (reserved for future use; defaults to project)" })),
+      structuredFields: Type.Optional(
+        Type.Record(Type.String(), Type.Unknown(), {
+          description: "Optional structured payload preserved alongside content (ADR-013). Use for decisions to retain scope/decision/choice/rationale/made_by/revisable. Omit for plain captures.",
+        }),
+      ),
     }),
     async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
       const ok = await ensureDbOpen();

package/src/resources/extensions/gsd/bootstrap/provider-error-resume.ts

@@ -12,6 +12,8 @@ type AutoResumeSnapshot = Pick<AutoDashboardData, "active" | "paused" | "stepMod
 
 export interface ProviderErrorResumeDeps {
   getSnapshot(): AutoResumeSnapshot;
+  resetTransientRetryState(): void;
+  resetSessionTimeoutState(): void;
   startAuto(
     ctx: ExtensionCommandContext,
     pi: ExtensionAPI,
@@ -23,6 +25,8 @@ export interface ProviderErrorResumeDeps {
 
 const defaultDeps: ProviderErrorResumeDeps = {
   getSnapshot: () => getAutoDashboardData(),
+  resetTransientRetryState,
+  resetSessionTimeoutState,
   startAuto,
 };
 
@@ -47,8 +51,8 @@ export async function resumeAutoAfterProviderDelay(
   // Reset retry counters before restarting — without this, counters
   // accumulate across pause/resume cycles and permanently lock out
   // auto-resume after their respective MAX thresholds.
-  resetTransientRetryState();
-  resetSessionTimeoutState();
+  deps.resetTransientRetryState();
+  deps.resetSessionTimeoutState();
 
   await deps.startAuto(
     ctx as ExtensionCommandContext,

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

@@ -8,6 +8,7 @@ import type { GSDEcosystemBeforeAgentStartHandler } from "../ecosystem/gsd-exten
 import { loadEcosystemExtensions } from "../ecosystem/loader.js";
 import { registerDbTools } from "./db-tools.js";
 import { registerDynamicTools } from "./dynamic-tools.js";
+import { registerExecTools } from "./exec-tools.js";
 import { registerJournalTools } from "./journal-tools.js";
 import { registerMemoryTools } from "./memory-tools.js";
 import { registerQueryTools } from "./query-tools.js";
@@ -15,6 +16,13 @@ import { registerHooks } from "./register-hooks.js";
 import { registerShortcuts } from "./register-shortcuts.js";
 import { writeCrashLog } from "./crash-log.js";
 import { logWarning } from "../workflow-logger.js";
+// Static import so cmux event listeners are registered synchronously during
+// extension bootstrap. Prior implementation used `void import().then()` which
+// queued listener registration as a microtask — any CMUX_CHANNELS emit fired
+// in the same event loop turn as registration (e.g. from a provider-error
+// session hook calling startAuto) would be silently dropped because Node's
+// EventEmitter does not buffer events for late subscribers.
+import { initCmuxEventListeners } from "../../cmux/index.js";
 
 export { writeCrashLog } from "./crash-log.js";
 
@@ -66,6 +74,19 @@ export function registerGsdExtension(pi: ExtensionAPI): void {
   registerWorktreeCommand(pi);
   registerExitCommand(pi);
 
+  // Wire the Layer 2 event emitter bridge so deeply-nested GSD code can emit
+  // extension events (git lifecycle, verify, budget, milestone, unit) without
+  // threading `pi` through every call site.
+  import("../hook-emitter.js")
+    .then(({ setHookEmitter }) => setHookEmitter(pi))
+    .catch((err) => {
+      // Non-fatal — emitters simply become no-ops if this import fails, but
+      // surface the failure so silent bootstrap breakage is debuggable.
+      process.stderr.write(
+        `[gsd] Failed to bootstrap hook-emitter bridge: ${err instanceof Error ? err.stack ?? err.message : String(err)}\n`,
+      );
+    });
+
   installEpipeGuard();
 
   // Ecosystem handlers captured by the GSDExtensionAPI wrapper for the
@@ -87,7 +108,12 @@ export function registerGsdExtension(pi: ExtensionAPI): void {
     ["journal-tools", () => registerJournalTools(pi)],
     ["query-tools", () => registerQueryTools(pi)],
     ["memory-tools", () => registerMemoryTools(pi)],
+    ["exec-tools", () => registerExecTools(pi)],
     ["shortcuts", () => registerShortcuts(pi)],
+    // cmux is a library (no pi), so gsd sets up the event listeners on its
+    // behalf using the shared event channel contract. Registration is
+    // synchronous — see the import comment above for the rationale.
+    ["cmux-events", () => initCmuxEventListeners(pi.events)],
     ["hooks", () => registerHooks(pi, ecosystemHandlers)],
     ["ecosystem", () => {
       void loadEcosystemExtensions(pi, ecosystemHandlers).catch((err) => {

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

@@ -18,12 +18,13 @@ import { loadToolApiKeys } from "../commands-config.js";
 import { loadFile, saveFile, formatContinue } from "../files.js";
 import { deriveState } from "../state.js";
 import { getAutoDashboardData, isAutoActive, isAutoPaused, markToolEnd, markToolStart, recordToolInvocationError } from "../auto.js";
-import { hideFooter } from "../auto-dashboard.js";
+
 import { isParallelActive, shutdownParallel } from "../parallel-orchestrator.js";
 import { checkToolCallLoop, resetToolCallLoopGuard } from "./tool-call-loop-guard.js";
 import { saveActivityLog } from "../activity-log.js";
 import { resetAskUserQuestionsCache } from "../../ask-user-questions.js";
-import { recordToolCall as safetyRecordToolCall, recordToolResult as safetyRecordToolResult } from "../safety/evidence-collector.js";
+import { recordToolCall as safetyRecordToolCall, recordToolResult as safetyRecordToolResult, saveEvidenceToDisk } from "../safety/evidence-collector.js";
+import { parseUnitId } from "../unit-id.js";
 import { classifyCommand } from "../safety/destructive-guard.js";
 import { logWarning as safetyLogWarning } from "../workflow-logger.js";
 import { installNotifyInterceptor } from "./notify-interceptor.js";
@@ -48,7 +49,9 @@ export function registerHooks(
     initNotificationStore(process.cwd());
     installNotifyInterceptor(ctx);
     initNotificationWidget(ctx);
-    initHealthWidget(ctx);
+    if (!isAutoActive()) {
+      initHealthWidget(ctx);
+    }
     resetWriteGateState();
     resetToolCallLoopGuard();
     resetAskUserQuestionsCache();
@@ -90,7 +93,7 @@ export function registerHooks(
     }
     loadToolApiKeys();
     if (isAutoActive()) {
-      ctx.ui.setFooter(hideFooter);
+      ctx.ui.setWidget("gsd-health", undefined);
     }
   });
 
@@ -113,7 +116,7 @@ export function registerHooks(
     }
     loadToolApiKeys();
     if (isAutoActive()) {
-      ctx.ui.setFooter(hideFooter);
+      ctx.ui.setWidget("gsd-health", undefined);
     }
   });
 
@@ -225,6 +228,42 @@ export function registerHooks(
     }));
   });
 
+  // Context-mode snapshot: write .gsd/last-snapshot.md before compaction so
+  // agents can call gsd_resume (or Read the file) to re-orient. Opt-in via
+  // preferences.context_mode.enabled. Runs after the auto-cancel handler
+  // above — if that one returned cancel:true, pi still fires us but the
+  // compaction won't actually happen; the snapshot is still useful then,
+  // since auto may pause and resume later.
+  pi.on("session_before_compact", async () => {
+    try {
+      const { loadEffectiveGSDPreferences } = await import("../preferences.js");
+      const { isContextModeEnabled } = await import("../preferences-types.js");
+      const prefs = loadEffectiveGSDPreferences();
+      if (!isContextModeEnabled(prefs?.preferences)) return;
+      const { writeCompactionSnapshot } = await import("../compaction-snapshot.js");
+      const { ensureDbOpen } = await import("./dynamic-tools.js");
+      await ensureDbOpen();
+      const basePath = process.cwd();
+      let activeContext: string | null = null;
+      try {
+        const state = await deriveState(basePath);
+        if (state.activeMilestone && state.activeSlice && state.activeTask) {
+          activeContext =
+            `Active: ${state.activeMilestone.id} / ${state.activeSlice.id} / ${state.activeTask.id}` +
+            (state.activeTask.title ? ` — ${state.activeTask.title}` : "");
+        }
+      } catch {
+        /* non-fatal */
+      }
+      writeCompactionSnapshot(basePath, { activeContext });
+    } catch (err) {
+      safetyLogWarning(
+        "context-mode",
+        `failed to write compaction snapshot: ${err instanceof Error ? err.message : String(err)}`,
+      );
+    }
+  });
+
   pi.on("session_shutdown", async (_event, ctx: ExtensionContext) => {
     if (isParallelActive()) {
       try {
@@ -445,7 +484,7 @@ export function registerHooks(
 
   pi.on("tool_execution_start", async (event) => {
     if (!isAutoActive()) return;
-    markToolStart(event.toolCallId);
+    markToolStart(event.toolCallId, event.toolName);
   });
 
   pi.on("tool_execution_end", async (event) => {
@@ -461,6 +500,15 @@ export function registerHooks(
     // Safety harness: record tool execution results for evidence cross-referencing
     if (isAutoActive()) {
       safetyRecordToolResult(event.toolCallId, event.toolName, event.result, event.isError);
+      // Persist evidence to disk after each tool result so it survives a session
+      // restart mid-unit (Bug #4385 — non-persisted evidence false positives).
+      const dash = getAutoDashboardData();
+      if (dash.basePath && dash.currentUnit?.type === "execute-task") {
+        const { milestone: pMid, slice: pSid, task: pTid } = parseUnitId(dash.currentUnit.id);
+        if (pMid && pSid && pTid) {
+          saveEvidenceToDisk(dash.basePath, pMid, pSid, pTid);
+        }
+      }
     }
   });