@evident-ai/cli 3.1.1-dev.5a3c6a5 → 3.1.1-dev.63072ee

package/dist/index.js

@@ -1532,6 +1532,9 @@ function isPreamblePinnedRunning(messages, userMessageId) {
   const reply = findLastAssistantReplyFor(messages, userMessageId);
   return completedOf(reply) != null && finishOf(reply) === "tool-calls";
 }
+function isB2AbandonmentConfirmed(params) {
+  return params.pinnedForMs >= params.minPinnedMs && params.descendantOngoing === false;
+}
 function messageError(messages, userMessageId) {
   const reply = findLastAssistantReplyFor(messages, userMessageId);
   const error2 = errorOf(reply);
@@ -1545,6 +1548,42 @@ function messageError(messages, userMessageId) {
   }
   return "The agent run failed.";
 }
+function messageFailure(messages, userMessageId) {
+  const reply = findLastAssistantReplyFor(messages, userMessageId);
+  const error2 = errorOf(reply);
+  if (error2 == null || typeof error2 !== "object") return null;
+  const e = error2;
+  const replyProviderId = reply?.info?.providerID ?? null;
+  const replyModelId = reply?.info?.modelID ?? null;
+  if (e.name === "ProviderAuthError") {
+    const data = e.data;
+    const providerId = typeof data?.providerID === "string" && data.providerID || replyProviderId;
+    return { kind: "model_auth", providerId, modelId: replyModelId, reason: "missing" };
+  }
+  if (e.name === "APIError") {
+    const data = e.data;
+    const statusCode = data?.statusCode;
+    if (statusCode === 401 || statusCode === 403) {
+      return {
+        kind: "model_auth",
+        providerId: replyProviderId,
+        modelId: replyModelId,
+        reason: "rejected"
+      };
+    }
+  }
+  return null;
+}
+function applyZeroProviderFallback(classified, hasConfiguredProvider, replyProviderId, replyModelId = null) {
+  if (classified != null) return classified;
+  if (hasConfiguredProvider !== false) return null;
+  return {
+    kind: "model_auth",
+    providerId: replyProviderId,
+    modelId: replyModelId,
+    reason: "missing"
+  };
+}
 function hasRunningAssistantExcept(messages, exceptUserMessageId) {
   if (!messages || messages.length === 0) return false;
   return messages.some(
@@ -2073,6 +2112,18 @@ var RunnerConnection = class {
   }
 };
 
+// src/lib/tunnel/ready-marker.ts
+import { writeFileSync } from "fs";
+function writeTunnelReadyMarker(path, agentId) {
+  try {
+    writeFileSync(path, `${agentId}
+`);
+    return { ok: true };
+  } catch (error2) {
+    return { ok: false, error: error2 instanceof Error ? error2.message : String(error2) };
+  }
+}
+
 // src/lib/channels/driver.ts
 import { homedir } from "os";
 
@@ -2499,6 +2550,8 @@ var DEFAULT_PAUSED_MAX_WAIT_MS = 10 * 60 * 1e3;
 var DEFAULT_STUCK_QUEUED_MS = 6e4;
 var HEARTBEAT_MS = 6e4;
 var ABSOLUTE_MAX_PROCESSING_MS = 6 * 60 * 60 * 1e3;
+var B2_ABANDONMENT_MIN_PINNED_MS = 3 * 6e4;
+var B2_ABANDONMENT_RECHECK_MS = HEARTBEAT_MS;
 var POLL_MISS_GRACE_MS = HEARTBEAT_MS;
 var MAX_SUPERSEDED_CONVERSATIONS = 256;
 var ChannelAuthError = class extends Error {
@@ -3317,12 +3370,17 @@ var ChannelDriver = class _ChannelDriver {
       stuckReported: false,
       lastAliveAt: 0,
       aliveInFlight: false,
+      titleSynced: false,
+      titleSyncInFlight: false,
       awaitingHumanLatched: false,
       pausedOnQuestion: false,
       pausedOnPermission: false,
       pausedClearConfirmed: false,
       pausedInFlight: false,
-      deliveryDeadlineAnchored: false
+      deliveryDeadlineAnchored: false,
+      b2PinnedSinceMs: 0,
+      b2LastDescendantCheckMs: 0,
+      b2AbandonedSignalled: false
     });
   }
   /**
@@ -3390,12 +3448,17 @@ var ChannelDriver = class _ChannelDriver {
       // with no extra `re_adopted` signal needed (folds old WI-6).
       lastAliveAt: 0,
       aliveInFlight: false,
+      titleSynced: false,
+      titleSyncInFlight: false,
       awaitingHumanLatched: false,
       pausedOnQuestion: false,
       pausedOnPermission: false,
       pausedClearConfirmed: false,
       pausedInFlight: false,
-      deliveryDeadlineAnchored: false
+      deliveryDeadlineAnchored: false,
+      b2PinnedSinceMs: 0,
+      b2LastDescendantCheckMs: 0,
+      b2AbandonedSignalled: false
     });
   }
   /**
@@ -3557,58 +3620,7 @@ var ChannelDriver = class _ChannelDriver {
       }
     }
     if (state === "done") {
-      this.anchorDeliveryDeadline(inFlight);
-      if (!inFlight.done) {
-        this.log({
-          level: "info",
-          message: `Message ${inFlight.evidentMessageId.slice(0, 8)} completed \u2014 marking done`,
-          conversation_id: conv.id,
-          message_id: inFlight.evidentMessageId
-        });
-        const title = await this.resolveSessionTitle(sessionId, watcher.conv.id);
-        const usage = messageUsage(messages, inFlight.opencodeMessageId);
-        try {
-          await this.markDone(
-            conv.id,
-            inFlight.evidentMessageId,
-            sessionId,
-            inFlight.opencodeMessageId,
-            title,
-            usage
-          );
-        } catch (err) {
-          if (err instanceof ChannelAuthError) throw err;
-          if (err instanceof ChannelTerminalError) {
-            this.log({
-              level: "warn",
-              message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (terminal HTTP ${err.status}) \u2014 leaving for the cron safety net: ${err.message}`,
-              conversation_id: conv.id,
-              message_id: inFlight.evidentMessageId
-            });
-            this.removeInFlight(watcher, inFlight.evidentMessageId);
-            return;
-          }
-          if (this.now() >= inFlight.deadline) {
-            this.log({
-              level: "warn",
-              message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done within the watch window \u2014 leaving for the cron safety net: ${err instanceof Error ? err.message : String(err)}`,
-              conversation_id: conv.id,
-              message_id: inFlight.evidentMessageId
-            });
-            this.removeInFlight(watcher, inFlight.evidentMessageId);
-            return;
-          }
-          this.log({
-            level: "warn",
-            message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (will retry next tick): ${err instanceof Error ? err.message : String(err)}`,
-            conversation_id: conv.id,
-            message_id: inFlight.evidentMessageId
-          });
-          return;
-        }
-        inFlight.done = true;
-      }
-      this.removeInFlight(watcher, inFlight.evidentMessageId);
+      await this.settleMessageDone(sessionId, watcher, inFlight, messages);
       return;
     }
     if (state === "failed") {
@@ -3622,8 +3634,16 @@ var ChannelDriver = class _ChannelDriver {
           message_id: inFlight.evidentMessageId
         });
         const usage = messageUsage(messages, inFlight.opencodeMessageId);
+        const failure = await this.classifyModelAuthFailure(messages, inFlight.opencodeMessageId);
         try {
-          await this.markFailed(conv.id, inFlight.evidentMessageId, sessionId, error2, usage);
+          await this.markFailed(
+            conv.id,
+            inFlight.evidentMessageId,
+            sessionId,
+            error2,
+            usage,
+            failure
+          );
         } catch (err) {
           if (err instanceof ChannelAuthError) throw err;
           if (err instanceof ChannelTerminalError) {
@@ -3668,6 +3688,44 @@ var ChannelDriver = class _ChannelDriver {
       });
     }
     const activelyRunning = state === "running" && !awaitingHuman;
+    const pinnedNow = activelyRunning && isPreamblePinnedRunning(messages, inFlight.opencodeMessageId);
+    const snapshotReadable = messages != null && messages.length > 0;
+    if (!pinnedNow) {
+      if (snapshotReadable) {
+        inFlight.b2PinnedSinceMs = 0;
+        inFlight.b2LastDescendantCheckMs = 0;
+        inFlight.b2AbandonedSignalled = false;
+      }
+    } else {
+      if (inFlight.b2AbandonedSignalled) {
+        await this.settleMessageDone(sessionId, watcher, inFlight, messages);
+        return;
+      }
+      if (inFlight.b2PinnedSinceMs === 0) inFlight.b2PinnedSinceMs = this.now();
+      const pinnedForMs = this.now() - inFlight.b2PinnedSinceMs;
+      if (pinnedForMs >= B2_ABANDONMENT_MIN_PINNED_MS && this.now() - inFlight.b2LastDescendantCheckMs >= B2_ABANDONMENT_RECHECK_MS) {
+        inFlight.b2LastDescendantCheckMs = this.now();
+        const descendantOngoing = await this.isAnyDescendantSessionOngoing(sessionId);
+        if (isB2AbandonmentConfirmed({
+          pinnedForMs,
+          minPinnedMs: B2_ABANDONMENT_MIN_PINNED_MS,
+          descendantOngoing
+        })) {
+          inFlight.b2AbandonedSignalled = true;
+          this.log({
+            level: "warn",
+            message: `Message ${id.slice(0, 8)} b2-pinned for ${Math.round(pinnedForMs / 1e3)}s with no ongoing descendant sub-agent session (status-map confirmed) \u2014 treating the delegated/tool turn as abandoned, resolving done`,
+            conversation_id: conv.id,
+            message_id: id
+          });
+          void this.postSignal(conv.id, id, "b2_abandoned_resolved", {
+            watched_for_ms: pinnedForMs
+          });
+          await this.settleMessageDone(sessionId, watcher, inFlight, messages);
+          return;
+        }
+      }
+    }
     if (activelyRunning && this.now() - inFlight.processingAnchorMs >= ABSOLUTE_MAX_PROCESSING_MS) {
       this.log({
         level: "warn",
@@ -3687,6 +3745,18 @@ var ChannelDriver = class _ChannelDriver {
         inFlight.aliveInFlight = false;
         if (ok) inFlight.lastAliveAt = this.now();
       });
+      if (!inFlight.titleSynced && !inFlight.titleSyncInFlight) {
+        inFlight.titleSyncInFlight = true;
+        void this.resolveSessionTitle(sessionId, conv.id).then(async (title) => {
+          if (!title) {
+            inFlight.titleSyncInFlight = false;
+            return;
+          }
+          const ok = await this.patchConversationTitle(conv.id, title);
+          inFlight.titleSyncInFlight = false;
+          if (ok) inFlight.titleSynced = true;
+        });
+      }
     }
     if (awaitingHuman) {
       if (!inFlight.awaitingHumanLatched) {
@@ -3724,6 +3794,70 @@ var ChannelDriver = class _ChannelDriver {
       this.removeInFlight(watcher, inFlight.evidentMessageId);
     }
   }
+  /**
+   * Settle a message whose run-state has resolved `'done'` — extracted verbatim
+   * (pure refactor, no behavior change) from `serviceInFlightMessage`'s former
+   * inline `state === 'done'` branch body, so a SECOND caller (the #721
+   * b2-abandonment resolution) can reach the exact same completion behavior
+   * (delivery-deadline anchoring, title resolution, usage extraction, and
+   * `markDone`'s auth/terminal/transient-retry discipline) without duplicating it
+   * and risking the two copies silently drifting apart.
+   */
+  async settleMessageDone(sessionId, watcher, inFlight, messages) {
+    const conv = watcher.conv;
+    this.anchorDeliveryDeadline(inFlight);
+    if (!inFlight.done) {
+      this.log({
+        level: "info",
+        message: `Message ${inFlight.evidentMessageId.slice(0, 8)} completed \u2014 marking done`,
+        conversation_id: conv.id,
+        message_id: inFlight.evidentMessageId
+      });
+      const title = await this.resolveSessionTitle(sessionId, watcher.conv.id);
+      const usage = messageUsage(messages, inFlight.opencodeMessageId);
+      try {
+        await this.markDone(
+          conv.id,
+          inFlight.evidentMessageId,
+          sessionId,
+          inFlight.opencodeMessageId,
+          title,
+          usage
+        );
+      } catch (err) {
+        if (err instanceof ChannelAuthError) throw err;
+        if (err instanceof ChannelTerminalError) {
+          this.log({
+            level: "warn",
+            message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (terminal HTTP ${err.status}) \u2014 leaving for the cron safety net: ${err.message}`,
+            conversation_id: conv.id,
+            message_id: inFlight.evidentMessageId
+          });
+          this.removeInFlight(watcher, inFlight.evidentMessageId);
+          return;
+        }
+        if (this.now() >= inFlight.deadline) {
+          this.log({
+            level: "warn",
+            message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done within the watch window \u2014 leaving for the cron safety net: ${err instanceof Error ? err.message : String(err)}`,
+            conversation_id: conv.id,
+            message_id: inFlight.evidentMessageId
+          });
+          this.removeInFlight(watcher, inFlight.evidentMessageId);
+          return;
+        }
+        this.log({
+          level: "warn",
+          message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (will retry next tick): ${err instanceof Error ? err.message : String(err)}`,
+          conversation_id: conv.id,
+          message_id: inFlight.evidentMessageId
+        });
+        return;
+      }
+      inFlight.done = true;
+    }
+    this.removeInFlight(watcher, inFlight.evidentMessageId);
+  }
   // Restart recovery: re-adopt `processing` messages (ADR-0046, WI-3/4/5)
   /**
    * Re-adopt this agent's `processing` messages on drain (ADR-0046 Decision §1).
@@ -3890,6 +4024,7 @@ var ChannelDriver = class _ChannelDriver {
     if (state === "failed") {
       const error2 = messageError(messages, ocId ?? "") ?? void 0;
       const usage = messageUsage(messages, ocId ?? "");
+      const failure = await this.classifyModelAuthFailure(messages, ocId ?? "");
       this.log({
         level: "error",
         message: `Re-adopt: message ${row.id.slice(0, 8)} errored while unwatched \u2014 marking failed: ${error2 ?? "(no error text)"}`,
@@ -3897,7 +4032,7 @@ var ChannelDriver = class _ChannelDriver {
         message_id: row.id
       });
       try {
-        await this.markFailed(row.conversation_id, row.id, sessionId, error2, usage);
+        await this.markFailed(row.conversation_id, row.id, sessionId, error2, usage, failure);
       } catch (err) {
         if (err instanceof ChannelAuthError) throw err;
         if (err instanceof ChannelTerminalError) {
@@ -4303,6 +4438,47 @@ var ChannelDriver = class _ChannelDriver {
     }
     return false;
   }
+  /**
+   * Tri-state variant of the upward parentID membership walk (#721), used ONLY
+   * by `isAnyDescendantSessionOngoing`. Walks the SAME cached
+   * `resolveSessionParent` chain `sessionBelongsTo` uses above, but — unlike
+   * `sessionBelongsTo`, which deliberately collapses "confirmed not a
+   * descendant" and "the walk's fetch failed" into the same `false` (safe for
+   * its OTHER callers: interaction attribution and the recovery-path
+   * `isAnyDescendantSessionAlive`, both of which just retry next tick with no
+   * safety consequence either way) — this variant keeps those two outcomes
+   * SEPARATE, because `isAnyDescendantSessionOngoing`'s caller
+   * (`isB2AbandonmentConfirmed`) must never treat "couldn't tell" as "confirmed
+   * not ongoing".
+   *
+   * Return contract:
+   *  - `true`  → the walk reached `rootSessionId` — `sessionId` IS a descendant.
+   *  - `false` → the walk reached a definitive, parent-less root session
+   *              WITHOUT ever matching `rootSessionId` — `sessionId` is
+   *              CONFIRMED NOT a descendant of it.
+   *  - `null`  → INDETERMINATE: a `GET /session/:id` fetch failed partway
+   *              through the walk (`resolveSessionParent` returned `undefined`),
+   *              or the depth cap (32) was hit without a definitive answer (a
+   *              pathological/cyclic chain proves nothing either way). NEVER
+   *              treat this the same as `false` — see `sessionBelongsTo`'s own
+   *              doc comment above for why that collapse is safe THERE but not
+   *              here.
+   *
+   * `sessionBelongsTo` itself is UNCHANGED — this is an additive helper scoped
+   * to the live-path descendant check, not a modification of shared code used
+   * by interaction attribution or the recovery path.
+   */
+  async resolveSessionMembership(sessionId, rootSessionId) {
+    let current = sessionId;
+    for (let depth = 0; current && depth < 32; depth++) {
+      if (current === rootSessionId) return true;
+      const parent = await this.resolveSessionParent(current);
+      if (parent === void 0) return null;
+      if (parent === null) return false;
+      current = parent;
+    }
+    return null;
+  }
   /**
    * Resolve (and cache) a session's `parentID` via `GET /session/:id`. Returns
    * `null` for a root session (no parent) and `undefined` when opencode is
@@ -4387,6 +4563,54 @@ var ChannelDriver = class _ChannelDriver {
     }
     return null;
   }
+  /**
+   * Best-effort mid-turn title sync (#711 follow-up): PATCH a resolved OpenCode
+   * session title onto the conversation via the PLAIN conversation-update
+   * endpoint (`PATCH /runners/:agentId/conversations/:conversationId`) — NOT the
+   * message-status endpoint `markProcessing`/`markDone` use. Deliberately a
+   * separate, lighter call: it carries no `status`, so it cannot re-trigger the
+   * `processing`/`done` transition side effects (Slack notices, activity-log
+   * rows, delivery jobs) those PATCHes gate on `transitioned` — this call only
+   * ever touches `conversations.title`. That route (`routes/conversations.ts`)
+   * skips a title write matching the stored value, so a redundant call with the
+   * same title is a real no-op — it does not bump `updated_at`, which the
+   * conversation list sorts and paginates on. (Note this is a DIFFERENT guard
+   * from `threads.ts`'s "non-empty AND changed" one, which only covers the
+   * message-status PATCH; the non-empty half is enforced here instead, by
+   * `resolveSessionTitle` never returning an empty/placeholder title.)
+   *
+   * Telemetry-only / never blocks the caller, mirroring `postSignal`: a failure
+   * is logged and the title is simply retried on the next heartbeat tick (the
+   * caller only latches `titleSynced` on `true`).
+   */
+  async patchConversationTitle(conversationId, title) {
+    try {
+      const res = await this.fetchImpl(
+        `${this.apiUrl}/runners/${this.agentId}/conversations/${conversationId}`,
+        {
+          method: "PATCH",
+          headers: { Authorization: this.getAuthHeader(), "Content-Type": "application/json" },
+          body: JSON.stringify({ title })
+        }
+      );
+      if (!res.ok) {
+        this.log({
+          level: "debug",
+          message: `Mid-turn title sync PATCH for conversation ${conversationId.slice(0, 8)} returned HTTP ${res.status} (best-effort, will retry next heartbeat)`,
+          conversation_id: conversationId
+        });
+        return false;
+      }
+      return true;
+    } catch (err) {
+      this.log({
+        level: "debug",
+        message: `Best-effort mid-turn title sync PATCH failed for conversation ${conversationId.slice(0, 8)} (will retry next heartbeat): ${err instanceof Error ? err.message : String(err)}`,
+        conversation_id: conversationId
+      });
+      return false;
+    }
+  }
   /**
    * DEFENSIVE cross-check for the restart-recovery path (WI-2): is any descendant
    * (`task` sub-agent) session under `rootSessionId` still genuinely doing work?
@@ -4445,6 +4669,84 @@ var ChannelDriver = class _ChannelDriver {
     }
     return false;
   }
+  /**
+   * LIVE-PATH descendant-liveness check (#721): is any descendant (`task`
+   * sub-agent) session under `rootSessionId` currently ONGOING per OpenCode's own
+   * in-memory status map (`isSessionOngoing` — `busy`/`retry`)?
+   *
+   * Deliberately NOT `isAnyDescendantSessionAlive` (the RECOVERY-path
+   * cross-check above): that method judges liveness from the child's OWN
+   * TRANSCRIPT (`isSessionActivelyGenerating`), which is the right (only) option
+   * on the recovery path because a restart WIPES `SessionStatus`. On the LIVE
+   * path the local opencode server IS running, so its in-memory status map is
+   * live and authoritative — and per ADR-0047 §4a ("the child has its own entry
+   * [in the map]"), a `task` descendant's OWN busy/retry entry reflects its
+   * ENTIRE turn (including any tool call it is itself executing), not a
+   * per-message transcript snapshot. This sidesteps the "child's own tool is
+   * executing, between its step's completion and the next generation step"
+   * transcript gap that a transcript-based check would need a second,
+   * sustained-window bound to guard against — it is simply not derived from
+   * message timestamps at all.
+   *
+   * Why not just check `isSessionOngoing(port, rootSessionId)` (the ROOT's own
+   * status, as the recovery path does per §4a)? Because on the LIVE path the
+   * root session can be shared: a SECOND, unrelated user message can land on the
+   * SAME session (issue #721's own root cause) and keep the root `busy` for a
+   * reason that has nothing to do with THIS message's delegation. A `task`
+   * descendant session is spawned for exactly one delegated turn and never
+   * reused, so its OWN status-map entry is unambiguous evidence about that one
+   * delegation — which the root's status is not.
+   *
+   * Why membership is checked via `resolveSessionMembership`, NOT
+   * `sessionBelongsTo`: `sessionBelongsTo` collapses a transient
+   * `GET /session/:id` fetch failure into "not a descendant", which would
+   * silently drop a genuinely-live candidate from consideration on the one
+   * unlucky tick its membership-walk fetch hiccups (#721).
+   * `resolveSessionMembership` keeps that failure mode as a distinct `null`
+   * (indeterminate) so it is folded into THIS method's own `indeterminate` flag
+   * instead.
+   *
+   * Return contract (note the DIFFERENT judge vs. `isAnyDescendantSessionAlive`):
+   *  - `true`  → some descendant session is `busy`/`retry` (genuinely ongoing).
+   *  - `false` → enumeration succeeded, EVERY candidate's MEMBERSHIP was
+   *              confirmed either way (`resolveSessionMembership` never
+   *              returned `null`), and every CONFIRMED descendant's status read
+   *              succeeded and is not ongoing (includes "no descendant session
+   *              exists at all" — e.g. a plain, non-`task` tool call).
+   *  - `null`  → INDETERMINATE: `listSessions` failed, OR at least one
+   *              candidate's MEMBERSHIP could not be confirmed
+   *              (`resolveSessionMembership` returned `null` — a fetch failure
+   *              or pathological chain partway through the parent walk), OR at
+   *              least one CONFIRMED descendant's `isSessionOngoing` read
+   *              failed — and no OTHER candidate was already confirmed `true`.
+   *              The caller MUST NOT treat `null` the same as `false` here
+   *              (unlike the recovery cross-check's contract) — see
+   *              `isB2AbandonmentConfirmed`.
+   */
+  async isAnyDescendantSessionOngoing(rootSessionId) {
+    const sessions = await listSessions(this.port);
+    if (!sessions) {
+      this.log({
+        level: "warn",
+        message: `Could not enumerate sessions to check descendant liveness for root ${rootSessionId} (listSessions failed) \u2014 treating descendant liveness as indeterminate`
+      });
+      return null;
+    }
+    let indeterminate = false;
+    for (const candidate of sessions) {
+      if (!candidate?.id || candidate.id === rootSessionId) continue;
+      const membership = await this.resolveSessionMembership(candidate.id, rootSessionId);
+      if (membership === null) {
+        indeterminate = true;
+        continue;
+      }
+      if (membership === false) continue;
+      const ongoing = await isSessionOngoing(this.port, candidate.id);
+      if (ongoing === true) return true;
+      if (ongoing === null) indeterminate = true;
+    }
+    return indeterminate ? null : false;
+  }
   /**
    * Cheap decision-telemetry label for a running row's LAST correlated reply
    * (WI-2 Task 2.2): which running SHAPE it is, for the status-gated recovery log.
@@ -4708,7 +5010,7 @@ var ChannelDriver = class _ChannelDriver {
    *     exists but is wedged, so the next attempt must get a fresh one
    *     instead of reusing it — see WI-1's server-side null-clearing PATCH).
    */
-  async markFailed(conversationId, messageId, sessionId, error2, usage) {
+  async markFailed(conversationId, messageId, sessionId, error2, usage, failure) {
     const body = { status: "failed" };
     if (sessionId === null) {
       body.opencode_session_id = null;
@@ -4717,6 +5019,12 @@ var ChannelDriver = class _ChannelDriver {
     }
     if (error2 !== void 0) body.error = error2;
     if (usage) Object.assign(body, usage);
+    if (failure) {
+      body.failure_kind = failure.kind;
+      body.failure_provider_id = failure.providerId;
+      body.failure_model_id = failure.modelId;
+      body.failure_reason = failure.reason;
+    }
     await this.callWithRetry(
       "marking message as failed",
       () => this.fetchImpl(
@@ -4729,6 +5037,29 @@ var ChannelDriver = class _ChannelDriver {
       )
     );
   }
+  /**
+   * Classify an errored turn as a model-auth failure (#736 P1-4), for `markFailed`.
+   *
+   * `messageFailure` alone (structured OpenCode error → `model_auth`) covers
+   * most cases; when it returns `null` on this ALREADY-FAILED turn, fall back
+   * to the P1-2b zero-provider check — one extra loopback call to
+   * `hasAnyConfiguredProvider`, only reached when the structured classifier
+   * couldn't place it. Fails open (never throws): a fallback probe failure
+   * (`null`/indeterminate) leaves the classification `null`, which produces
+   * today's byte-identical PATCH body via `markFailed`'s `if (failure)` guard.
+   */
+  async classifyModelAuthFailure(messages, userMessageId) {
+    const classified = messageFailure(messages, userMessageId);
+    if (classified != null) return classified;
+    const reply = findLastAssistantReplyFor(messages, userMessageId);
+    const hasProvider = await hasAnyConfiguredProvider(this.port);
+    return applyZeroProviderFallback(
+      classified,
+      hasProvider,
+      reply?.info?.providerID ?? null,
+      reply?.info?.modelID ?? null
+    );
+  }
   /**
    * Best-effort, SINGLE-ATTEMPT runner→server telemetry ping
    * (queued-followup-redrive). `POST .../messages/:id/signal {signal, ...extra}`
@@ -4905,7 +5236,7 @@ async function ensureOpenCodeRunning(ctx) {
       console.log(chalk5.yellow("Tip: Run with the correct port:"));
       console.log(
         chalk5.dim(
-          `  ${getCliName()} run --agent ${ctx.agentId} --port ${runningInstances[0].port}`
+          `  ${getCliName()} run --runner ${ctx.agentId} --port ${runningInstances[0].port}`
         )
       );
     }
@@ -5035,19 +5366,21 @@ async function resolveAgentIdFromKey(authHeader) {
       return { agent_id: data.agent_id };
     }
     return {
-      error: "Cannot resolve runner ID: auth type is not agent_key. Please provide --agent explicitly."
+      error: "Cannot resolve runner ID: auth type is not agent_key. Please provide --runner explicitly."
     };
   } catch (error2) {
     const message = error2 instanceof Error ? error2.message : "Unknown error";
     return { error: `Failed to resolve runner from key: ${message}` };
   }
 }
+var BEST_EFFORT_NOTIFY_TIMEOUT_MS = 2e3;
 async function notifyAgentDisconnected(agentId, authHeader) {
   const apiUrl = getApiUrlConfig();
   try {
     const response = await fetch(`${apiUrl}/runners/${agentId}/disconnect`, {
       method: "POST",
-      headers: { Authorization: authHeader }
+      headers: { Authorization: authHeader },
+      signal: AbortSignal.timeout(BEST_EFFORT_NOTIFY_TIMEOUT_MS)
     });
     if (!response.ok) {
       const serverMessage = await readErrorMessage(response);
@@ -5058,7 +5391,35 @@ async function notifyAgentDisconnected(agentId, authHeader) {
     }
     return { ok: true };
   } catch (error2) {
-    return { ok: false, error: error2 instanceof Error ? error2.message : String(error2) };
+    return { ok: false, error: describeBestEffortError(error2) };
+  }
+}
+function describeBestEffortError(error2) {
+  const name = error2?.name;
+  if (name === "TimeoutError" || name === "AbortError") {
+    return `timed out after ${BEST_EFFORT_NOTIFY_TIMEOUT_MS}ms`;
+  }
+  return error2 instanceof Error ? error2.message : String(error2);
+}
+async function reportMicrovmId(agentId, authHeader, microvmId) {
+  try {
+    const apiUrl = getApiUrlConfig();
+    const response = await fetch(`${apiUrl}/runners/${agentId}/microvm`, {
+      method: "POST",
+      headers: { Authorization: authHeader, "Content-Type": "application/json" },
+      body: JSON.stringify({ microvm_id: microvmId }),
+      signal: AbortSignal.timeout(BEST_EFFORT_NOTIFY_TIMEOUT_MS)
+    });
+    if (!response.ok) {
+      const serverMessage = await readErrorMessage(response);
+      return {
+        ok: false,
+        error: `HTTP ${response.status}${serverMessage ? `: ${serverMessage}` : ""}`
+      };
+    }
+    return { ok: true };
+  } catch (error2) {
+    return { ok: false, error: describeBestEffortError(error2) };
   }
 }
 async function getAgentInfo(agentId, authHeader) {
@@ -5108,6 +5469,7 @@ var MAX_ACTIVITY_LOG_ENTRIES = 10;
 var CHANNEL_POLL_INTERVAL_MS = Number(process.env.EVIDENT_CHANNEL_POLL_INTERVAL_MS) || 2e3;
 var CHANNEL_STUCK_QUEUED_MS = Number(process.env.EVIDENT_STUCK_QUEUED_MS) || void 0;
 var SHUTDOWN_DRAIN_TIMEOUT_MS = Number(process.env.EVIDENT_SHUTDOWN_DRAIN_MS) || 25e3;
+var TELEMETRY_SHUTDOWN_TIMEOUT_MS = 5e3;
 function resolveLogLevel(options) {
   const accepted = Object.keys(LOG_LEVELS);
   const validate = (value, source) => {
@@ -5434,7 +5796,18 @@ async function notifyOffline(state) {
     if (state.interactive) displayStatus(state);
   }
 }
+async function timeShutdownPhase(state, durations, name, run2) {
+  const startedAt = Date.now();
+  try {
+    return await run2();
+  } finally {
+    const elapsedMs = Date.now() - startedAt;
+    durations[name] = elapsedMs;
+    log2(state, `Shutdown phase ${name}: ${elapsedMs}ms`);
+  }
+}
 async function cleanup(state, opts = {}) {
+  const durations = {};
   state.running = false;
   for (const timer of state.sessionCleanupTimers) {
     clearInterval(timer);
@@ -5448,7 +5821,13 @@ async function cleanup(state, opts = {}) {
       logActivity(state, { type: "info", message: "Draining in-flight work before shutdown..." });
       displayStatus(state);
     }
-    const settled = await state.channelDriver.waitForInFlight(SHUTDOWN_DRAIN_TIMEOUT_MS);
+    const driver = state.channelDriver;
+    const settled = await timeShutdownPhase(
+      state,
+      durations,
+      "drain",
+      () => driver.waitForInFlight(SHUTDOWN_DRAIN_TIMEOUT_MS)
+    );
     if (!settled) {
       logActivity(state, {
         type: "info",
@@ -5457,13 +5836,15 @@ async function cleanup(state, opts = {}) {
       if (state.interactive) displayStatus(state);
     }
   }
-  await notifyOffline(state);
+  await timeShutdownPhase(state, durations, "offline_notify", () => notifyOffline(state));
   if (state.connection) {
-    state.connection.close();
+    const connection = state.connection;
+    await timeShutdownPhase(state, durations, "tunnel_close", () => connection.close());
     state.connection = null;
   }
   if (state.opencodeProcess) {
-    stopOpenCode(state.opencodeProcess);
+    const opencodeProcess = state.opencodeProcess;
+    await timeShutdownPhase(state, durations, "opencode_stop", () => stopOpenCode(opencodeProcess));
     if (state.interactive) {
       logActivity(state, { type: "info", message: "Stopped OpenCode process" });
       displayStatus(state);
@@ -5472,6 +5853,7 @@ async function cleanup(state, opts = {}) {
     }
     state.opencodeProcess = null;
   }
+  return durations;
 }
 async function run(options) {
   const interactive = isInteractive(options.json);
@@ -5542,14 +5924,38 @@ async function run(options) {
   const handleSignal = async () => {
     if (state.shuttingDown) return;
     state.shuttingDown = true;
+    const shutdownStartedAt = Date.now();
     if (state.interactive) {
       logActivity(state, { type: "info", message: "Shutting down..." });
       displayStatus(state);
     } else {
       log2(state, "Shutting down...");
     }
-    await cleanup(state, { graceful: true });
-    await shutdownTelemetry();
+    const durations = await cleanup(state, { graceful: true });
+    const telemetryBudgetMs = Number(process.env.EVIDENT_TELEMETRY_SHUTDOWN_MS) || TELEMETRY_SHUTDOWN_TIMEOUT_MS;
+    await timeShutdownPhase(state, durations, "telemetry_flush", async () => {
+      let timer;
+      const flushed = shutdownTelemetry().then(
+        () => true,
+        (error2) => {
+          log2(
+            state,
+            `Telemetry flush failed during shutdown: ${error2 instanceof Error ? error2.message : String(error2)}`,
+            "warn"
+          );
+          return true;
+        }
+      );
+      const timedOut = new Promise((resolve3) => {
+        timer = setTimeout(() => resolve3(false), telemetryBudgetMs);
+      });
+      if (!await Promise.race([flushed, timedOut])) {
+        log2(state, `Telemetry flush exceeded ${telemetryBudgetMs}ms \u2014 exiting anyway`, "warn");
+      }
+      clearTimeout(timer);
+    });
+    const breakdown = Object.entries(durations).map(([phase, ms]) => `${phase}=${ms}ms`).join(" ");
+    log2(state, `Shutdown complete in ${Date.now() - shutdownStartedAt}ms (${breakdown})`);
     process.exit(0);
   };
   process.on("SIGINT", handleSignal);
@@ -5663,6 +6069,21 @@ async function run(options) {
     }
     spinner?.succeed(`Runner: ${validation.agent.name || state.agentId}`);
     state.agentName = validation.agent.name;
+    const microvmId = process.env.MICROVM_ID?.trim();
+    if (microvmId) {
+      const reported = await reportMicrovmId(state.agentId, state.authHeader, microvmId);
+      if (reported.ok) {
+        log2(state, "Reported MicroVM identity so this runner can be resumed rather than restarted");
+      } else {
+        const message = `Could not report MicroVM identity (future wakes will cold-start): ${reported.error}`;
+        log2(state, message, "warn");
+        if (state.interactive && !state.json) {
+          logActivity(state, { type: "info", level: "warn", message });
+        }
+      }
+    } else {
+      log2(state, "Not running in a MicroVM (MICROVM_ID unset) \u2014 nothing to report", "debug");
+    }
     const ocSpinner = interactive && !state.json ? ora3("Checking OpenCode...").start() : null;
     try {
       const oc = await ensureOpenCodeRunning({
@@ -5741,6 +6162,18 @@ async function run(options) {
             type: "info",
             message: `Tunnel ${isReconnect ? "reconnected" : "connected"} (runner: ${agentId})`
           });
+          if (options.tunnelReadyFile) {
+            const marker = writeTunnelReadyMarker(options.tunnelReadyFile, agentId);
+            if (marker.ok) {
+              log2(state, `Wrote tunnel readiness marker to ${options.tunnelReadyFile}`, "debug");
+            } else {
+              log2(
+                state,
+                `Failed to write tunnel readiness marker to ${options.tunnelReadyFile}: ${marker.error}`,
+                "error"
+              );
+            }
+          }
           emitAgentConnected(state.agentId, {
             port: state.port,
             cli_version: getCliVersion(),
@@ -5905,6 +6338,9 @@ program.command("run").description("Connect to Evident and process messages").op
   "Allow Evident to write files into this directory (repeatable). Omit to disable file sync entirely.",
   (value, previous) => previous.concat([value]),
   []
+).option(
+  "--tunnel-ready-file <path>",
+  "Path to write once the tunnel is connected (set by the MicroVM hooks; unused on a developer machine)"
 ).action(
   (options) => {
     run({
@@ -5924,7 +6360,8 @@ program.command("run").description("Connect to Evident and process messages").op
       sessionCleanupInterval: options.sessionCleanupInterval,
       // Raw values — expansion/validation is single-sourced in run.ts's
       // resolveFileSyncDirectories.
-      enableFileSyncTo: options.enableFileSyncTo
+      enableFileSyncTo: options.enableFileSyncTo,
+      tunnelReadyFile: options.tunnelReadyFile
     });
   }
 );