@evident-ai/cli 3.1.1-dev.bedbd30 → 3.1.1-dev.c0fca20

package/dist/index.js

@@ -267,16 +267,28 @@ async function getToken() {
   }
   return null;
 }
+function toError(err) {
+  return err instanceof Error ? err : new Error(String(err));
+}
 async function deleteToken(options = {}) {
   const keytar = await getKeytar();
+  const failures = [];
   if (keytar) {
     if (options.all) {
-      const all = await keytar.findCredentials(SERVICE_NAME).catch(() => []);
+      let accounts = [];
+      try {
+        accounts = await keytar.findCredentials(SERVICE_NAME);
+      } catch (err) {
+        failures.push({ type: "enumerate", error: toError(err) });
+      }
       await Promise.all(
-        all.map(
-          (entry) => keytar.deletePassword(SERVICE_NAME, entry.account).catch(() => {
-          })
-        )
+        accounts.map(async (entry) => {
+          try {
+            await keytar.deletePassword(SERVICE_NAME, entry.account);
+          } catch (err) {
+            failures.push({ type: "delete", account: entry.account, error: toError(err) });
+          }
+        })
       );
     } else {
       await keytar.deletePassword(SERVICE_NAME, keychainAccount());
@@ -287,6 +299,7 @@ async function deleteToken(options = {}) {
   } else {
     clearCredentials();
   }
+  return { failures };
 }
 
 // src/utils/ui.ts
@@ -455,9 +468,22 @@ async function login(options) {
 }
 
 // src/commands/logout.ts
+function describeFailure(failure) {
+  if (failure.type === "enumerate") {
+    return `could not list stored keychain entries (${failure.error.message})`;
+  }
+  return `${failure.account} (${failure.error.message})`;
+}
 async function logout(options = {}) {
   if (options.all) {
-    await deleteToken({ all: true });
+    const result = await deleteToken({ all: true });
+    if (result.failures.length > 0) {
+      printError(
+        `Failed to fully clear your keychain: ${result.failures.map(describeFailure).join("; ")}. Your local credentials file was cleared, but stale keychain entries may remain \u2014 run \`evident logout --all\` again, or remove them manually from your OS keychain / credential manager.`
+      );
+      process.exitCode = 1;
+      return;
+    }
     printSuccess("Logged out of all endpoints.");
     return;
   }
@@ -510,7 +536,10 @@ var TelemetryEventTypes = {
   AGENT_DISCONNECTED: "agent.disconnected",
   AGENT_MESSAGE_PROCESSING: "agent.message_processing",
   AGENT_MESSAGE_DONE: "agent.message_done",
-  AGENT_MESSAGE_FAILED: "agent.message_failed"
+  AGENT_MESSAGE_FAILED: "agent.message_failed",
+  // A `warn`/`error` runner-side log line forwarded server-side for
+  // observability (issue #916) — see `apps/cli/src/lib/runner-activity-telemetry.ts`.
+  RUNNER_ACTIVITY: "runner.activity"
 };
 
 // ../../packages/types/src/tunnel/index.ts
@@ -565,6 +594,13 @@ var isShuttingDown = false;
 var FLUSH_INTERVAL_MS = 5e3;
 var MAX_BUFFER_SIZE = 50;
 var FLUSH_TIMEOUT_MS = 3e3;
+var authProvider = null;
+function setTelemetryAuthProvider(provider) {
+  authProvider = provider;
+}
+var FLUSH_FAILURE_LOG_INTERVAL_MS = 6e4;
+var lastFlushFailureLoggedAt = 0;
+var suppressedFlushFailureCount = 0;
 function logEvent(eventType, options = {}) {
   const event = {
     event_type: eventType,
@@ -599,9 +635,16 @@ async function flushEvents() {
     flushTimeout = null;
   }
   try {
-    const credentials2 = await getToken();
-    if (!credentials2) {
-      return;
+    const providerContext = authProvider?.();
+    let authHeader;
+    if (providerContext?.authHeader) {
+      authHeader = providerContext.authHeader;
+    } else {
+      const credentials2 = await getToken();
+      if (!credentials2) {
+        return;
+      }
+      authHeader = `Bearer ${credentials2.token}`;
     }
     const apiUrl = getApiUrlConfig();
     const controller = new AbortController();
@@ -616,7 +659,7 @@ async function flushEvents() {
         method: "POST",
         headers: {
           "Content-Type": "application/json",
-          Authorization: `Bearer ${credentials2.token}`
+          Authorization: authHeader
         },
         body: JSON.stringify(request),
         signal: controller.signal
@@ -628,8 +671,15 @@ async function flushEvents() {
       clearTimeout(timeout);
     }
   } catch (error2) {
-    if (process.env.DEBUG) {
-      console.error("Telemetry error:", error2);
+    const now = Date.now();
+    if (now - lastFlushFailureLoggedAt >= FLUSH_FAILURE_LOG_INTERVAL_MS) {
+      const message = error2 instanceof Error ? error2.message : String(error2);
+      const suffix = suppressedFlushFailureCount > 0 ? ` (${suppressedFlushFailureCount} more suppressed in the last ${FLUSH_FAILURE_LOG_INTERVAL_MS / 1e3}s)` : "";
+      console.error(`Telemetry flush error: ${message}${suffix}`);
+      lastFlushFailureLoggedAt = now;
+      suppressedFlushFailureCount = 0;
+    } else {
+      suppressedFlushFailureCount++;
     }
   }
 }
@@ -698,6 +748,69 @@ var EventTypes = {
   DEPRECATED_AGENT_KEY_ENV_USED: "cli.deprecated_agent_key_env_used"
 };
 
+// src/lib/runner-activity-telemetry.ts
+var FORWARDED_LEVELS = /* @__PURE__ */ new Set(["warn", "error"]);
+var SEVERITY_BY_LEVEL = {
+  warn: "warning",
+  error: "error"
+};
+var MAX_MESSAGE_LENGTH = 500;
+var TRUNCATION_MARKER = "\u2026";
+function redact(message) {
+  return message.replace(/esk_[A-Za-z0-9_-]+/g, "esk_***").replace(/ct_[A-Za-z0-9_-]+/g, "ct_***").replace(/https?:\/\/\S+/g, "<url>");
+}
+function truncate(message) {
+  if (message.length <= MAX_MESSAGE_LENGTH) return message;
+  return message.slice(0, MAX_MESSAGE_LENGTH - TRUNCATION_MARKER.length) + TRUNCATION_MARKER;
+}
+var RATE_LIMIT_WINDOW_MS = 6e4;
+var RATE_LIMIT_MAX_EVENTS = 30;
+var windowStartedAt = 0;
+var windowCount = 0;
+var windowDroppedCount = 0;
+function admitUnderRateLimit(now) {
+  if (now - windowStartedAt >= RATE_LIMIT_WINDOW_MS) {
+    if (windowDroppedCount > 0) {
+      console.error(
+        `[runner-activity-telemetry] rate cap reached: dropped ${windowDroppedCount} ${windowDroppedCount === 1 ? "entry" : "entries"} in the last ${RATE_LIMIT_WINDOW_MS / 1e3}s (cap ${RATE_LIMIT_MAX_EVENTS}/min)`
+      );
+    }
+    windowStartedAt = now;
+    windowCount = 0;
+    windowDroppedCount = 0;
+  }
+  if (windowCount >= RATE_LIMIT_MAX_EVENTS) {
+    windowDroppedCount++;
+    if (windowDroppedCount === 1) {
+      console.error(
+        `[runner-activity-telemetry] rate cap reached (${RATE_LIMIT_MAX_EVENTS} per ${RATE_LIMIT_WINDOW_MS / 1e3}s) \u2014 dropping further entries this window`
+      );
+    }
+    return false;
+  }
+  windowCount++;
+  return true;
+}
+function forwardRunnerActivity(entry, context) {
+  try {
+    if (!FORWARDED_LEVELS.has(entry.level)) return;
+    if (!context.agentId || !context.authHeader) return;
+    if (!admitUnderRateLimit(Date.now())) return;
+    const rawMessage = entry.error ?? entry.message ?? "";
+    const message = truncate(redact(rawMessage));
+    logEvent(TelemetryEventTypes.RUNNER_ACTIVITY, {
+      severity: SEVERITY_BY_LEVEL[entry.level],
+      message,
+      metadata: { source: "cli.run" },
+      agentId: context.agentId
+    });
+  } catch (err) {
+    console.error(
+      `[runner-activity-telemetry] failed to forward runner activity: ${err instanceof Error ? err.message : String(err)}`
+    );
+  }
+}
+
 // src/lib/auth.ts
 async function getAuthCredentials() {
   const runnerKey = process.env.EVIDENT_RUNNER_KEY;
@@ -1532,6 +1645,9 @@ 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 +1661,42 @@ 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 +2225,18 @@ 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 +2663,8 @@ 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 {
@@ -3324,7 +3490,10 @@ var ChannelDriver = class _ChannelDriver {
       pausedOnPermission: false,
       pausedClearConfirmed: false,
       pausedInFlight: false,
-      deliveryDeadlineAnchored: false
+      deliveryDeadlineAnchored: false,
+      b2PinnedSinceMs: 0,
+      b2LastDescendantCheckMs: 0,
+      b2AbandonedSignalled: false
     });
   }
   /**
@@ -3399,7 +3568,10 @@ var ChannelDriver = class _ChannelDriver {
       pausedOnPermission: false,
       pausedClearConfirmed: false,
       pausedInFlight: false,
-      deliveryDeadlineAnchored: false
+      deliveryDeadlineAnchored: false,
+      b2PinnedSinceMs: 0,
+      b2LastDescendantCheckMs: 0,
+      b2AbandonedSignalled: false
     });
   }
   /**
@@ -3561,58 +3733,7 @@ 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") {
@@ -3626,8 +3747,16 @@ 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) {
@@ -3672,6 +3801,44 @@ 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",
@@ -3740,6 +3907,70 @@ 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).
@@ -3906,6 +4137,7 @@ 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)"}`,
@@ -3913,7 +4145,7 @@ 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) {
@@ -4319,6 +4551,47 @@ 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
@@ -4509,6 +4782,84 @@ 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.
@@ -4772,7 +5123,7 @@ 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;
@@ -4781,6 +5132,12 @@ 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(
@@ -4793,6 +5150,29 @@ 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}`
@@ -4945,10 +5325,16 @@ var ChannelDriver = class _ChannelDriver {
 import chalk5 from "chalk";
 import ora2 from "ora";
 import { select as select2 } from "@inquirer/prompts";
+var INTERACTIVE_START_TIMEOUT_MS = 3e4;
 async function ensureOpenCodeRunning(ctx) {
   const healthCheck = await checkOpenCodeHealth(ctx.port);
   if (healthCheck.healthy) {
-    return { port: ctx.port, process: null, version: healthCheck.version ?? null };
+    return {
+      port: ctx.port,
+      process: null,
+      version: healthCheck.version ?? null,
+      notReadyReason: null
+    };
   }
   const runningInstances = await findHealthyOpenCodeInstances();
   if (runningInstances.length > 0) {
@@ -4969,7 +5355,7 @@ 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}`
         )
       );
     }
@@ -4989,14 +5375,22 @@ async function ensureOpenCodeRunning(ctx) {
   if (!ctx.interactive) {
     ctx.log(`OpenCode is not running on port ${ctx.port}. Starting it automatically...`);
     const proc = await startOpenCode(ctx.port);
-    const health = await waitForOpenCodeHealth(ctx.port, 3e4);
+    const health = await waitForOpenCodeHealth(ctx.port, ctx.startTimeoutMs);
     if (!health.healthy) {
-      throw new Error(
-        `OpenCode failed to start on port ${ctx.port}. Install with: npm install -g opencode-ai`
-      );
+      return {
+        port: ctx.port,
+        process: proc,
+        version: null,
+        notReadyReason: `it did not become healthy within ${Math.round(ctx.startTimeoutMs / 1e3)}s`
+      };
     }
     ctx.log(`OpenCode started on port ${ctx.port}${health.version ? ` (v${health.version})` : ""}`);
-    return { port: ctx.port, process: proc, version: health.version ?? null };
+    return {
+      port: ctx.port,
+      process: proc,
+      version: health.version ?? null,
+      notReadyReason: null
+    };
   }
   let port = ctx.port;
   if (isPortInUse(port)) {
@@ -5049,15 +5443,15 @@ Port ${port} is already in use.`));
   if (action === "start") {
     const spinner = ora2("Starting OpenCode...").start();
     const proc = await startOpenCode(port);
-    const health = await waitForOpenCodeHealth(port, 3e4);
+    const health = await waitForOpenCodeHealth(port, INTERACTIVE_START_TIMEOUT_MS);
     if (!health.healthy) {
       spinner.fail("Failed to start OpenCode");
       throw new Error("OpenCode failed to start");
     }
     spinner.stop();
-    return { port, process: proc, version: health.version ?? null };
+    return { port, process: proc, version: health.version ?? null, notReadyReason: null };
   }
-  return { port, process: null, version: null };
+  return { port, process: null, version: null, notReadyReason: "you chose to continue without it" };
 }
 
 // src/commands/agent-lookup.ts
@@ -5099,7 +5493,7 @@ 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";
@@ -5254,6 +5648,35 @@ function resolveFileSyncDirectories(raw, homeDir) {
   }
   return directories;
 }
+var DEFAULT_OPENCODE_START_TIMEOUT_SECONDS = 180;
+var MAX_OPENCODE_START_TIMEOUT_SECONDS = 3600;
+var OPENCODE_START_TIMEOUT_ENV = "EVIDENT_OPENCODE_START_TIMEOUT";
+function resolveOpenCodeStartTimeoutMs(options, env = process.env) {
+  const defaultMs = DEFAULT_OPENCODE_START_TIMEOUT_SECONDS * 1e3;
+  let raw;
+  let source;
+  if (options.opencodeStartTimeout !== void 0) {
+    raw = options.opencodeStartTimeout;
+    source = "--opencode-start-timeout";
+  } else if (env[OPENCODE_START_TIMEOUT_ENV] !== void 0 && env[OPENCODE_START_TIMEOUT_ENV] !== "") {
+    raw = env[OPENCODE_START_TIMEOUT_ENV];
+    source = OPENCODE_START_TIMEOUT_ENV;
+  } else {
+    return { timeoutMs: defaultMs, warnings: [] };
+  }
+  const trimmed = raw.trim();
+  const seconds = Number(trimmed);
+  const isPositiveInteger = /^\d+$/.test(trimmed) && Number.isInteger(seconds) && seconds > 0;
+  if (!isPositiveInteger || seconds > MAX_OPENCODE_START_TIMEOUT_SECONDS) {
+    return {
+      timeoutMs: defaultMs,
+      warnings: [
+        `Ignoring invalid ${source} "${raw}": expected a positive integer number of seconds (at most ${MAX_OPENCODE_START_TIMEOUT_SECONDS}); using the default ${DEFAULT_OPENCODE_START_TIMEOUT_SECONDS}s`
+      ]
+    };
+  }
+  return { timeoutMs: seconds * 1e3, warnings: [] };
+}
 function meetsThreshold(state, level) {
   return LOG_LEVELS[level] >= LOG_LEVELS[state.logLevel];
 }
@@ -5275,6 +5698,10 @@ function log2(state, message, level = "info") {
 function logActivity(state, entry) {
   const level = entry.level ?? (entry.type === "error" ? "error" : "info");
   if (!meetsThreshold(state, level)) return;
+  forwardRunnerActivity(
+    { level, message: entry.message, error: entry.error },
+    { agentId: state.agentId, authHeader: state.authHeader }
+  );
   const fullEntry = {
     ...entry,
     level,
@@ -5629,6 +6056,7 @@ async function run(options) {
     sessionCleanupTimers: [],
     authHeader: ""
   };
+  setTelemetryAuthProvider(() => ({ authHeader: state.authHeader, agentId: state.agentId }));
   if (fileSyncDirectories.length > 0) {
     log2(state, `File sync enabled for: ${fileSyncDirectories.join(", ")}`);
   } else {
@@ -5817,40 +6245,52 @@ async function run(options) {
     } else {
       log2(state, "Not running in a MicroVM (MICROVM_ID unset) \u2014 nothing to report", "debug");
     }
+    const { timeoutMs: opencodeStartTimeoutMs, warnings: opencodeStartTimeoutWarnings } = resolveOpenCodeStartTimeoutMs(options, process.env);
+    for (const warning2 of opencodeStartTimeoutWarnings) {
+      logActivity(state, { type: "info", level: "warn", message: warning2 });
+    }
     const ocSpinner = interactive && !state.json ? ora3("Checking OpenCode...").start() : null;
     try {
       const oc = await ensureOpenCodeRunning({
         port: state.port,
         interactive: state.interactive,
         agentId: state.agentId,
-        log: (message) => log2(state, message)
+        log: (message) => log2(state, message),
+        startTimeoutMs: opencodeStartTimeoutMs
       });
       state.port = oc.port;
       state.opencodeProcess = oc.process;
       state.opencodeVersion = oc.version;
-      state.opencodeConnected = oc.process !== null || oc.version !== null;
+      state.opencodeConnected = oc.notReadyReason === null;
       const version2 = state.opencodeVersion ? ` (v${state.opencodeVersion})` : "";
       ocSpinner?.succeed(`OpenCode running on port ${state.port}${version2}`);
-      const versionWarning = buildOpenCodeVersionWarning(state.opencodeVersion);
-      if (versionWarning) {
-        log2(state, versionWarning, "warn");
-        if (state.interactive && !state.json) {
-          logActivity(state, { type: "info", level: "warn", message: versionWarning });
+      if (!state.interactive && oc.notReadyReason !== null) {
+        const message = `OpenCode is not ready on port ${state.port}: ${oc.notReadyReason}. The runner will still come online, but messages will fail until opencode answers \u2014 raise the wait with --opencode-start-timeout <seconds> (env ${OPENCODE_START_TIMEOUT_ENV}).`;
+        logActivity(state, { type: "info", level: "warn", message });
+      } else {
+        const versionWarning = buildOpenCodeVersionWarning(state.opencodeVersion);
+        if (versionWarning) {
+          log2(state, versionWarning, "warn");
+          if (state.interactive && !state.json) {
+            logActivity(state, { type: "info", level: "warn", message: versionWarning });
+          }
         }
-      }
-      const noProviderWarning = buildNoProviderWarning(await hasAnyConfiguredProvider(state.port));
-      if (noProviderWarning) {
-        log2(state, noProviderWarning, "warn");
-        if (state.interactive && !state.json) {
-          logActivity(state, { type: "info", level: "warn", message: noProviderWarning });
-          blank();
-          console.log(chalk6.yellow("\u26A0 No OpenCode model provider is configured."));
-          console.log(
-            chalk6.dim(
-              `Run ${chalk6.cyan("opencode auth login")} to set one up \u2014 messages will fail until then.`
-            )
-          );
-          blank();
+        const noProviderWarning = buildNoProviderWarning(
+          await hasAnyConfiguredProvider(state.port)
+        );
+        if (noProviderWarning) {
+          log2(state, noProviderWarning, "warn");
+          if (state.interactive && !state.json) {
+            logActivity(state, { type: "info", level: "warn", message: noProviderWarning });
+            blank();
+            console.log(chalk6.yellow("\u26A0 No OpenCode model provider is configured."));
+            console.log(
+              chalk6.dim(
+                `Run ${chalk6.cyan("opencode auth login")} to set one up \u2014 messages will fail until then.`
+              )
+            );
+            blank();
+          }
         }
       }
     } catch (error2) {
@@ -5895,6 +6335,18 @@ 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(),
@@ -6045,7 +6497,10 @@ program.command("run").description("Connect to Evident and process messages").op
 ).option("-p, --port <port>", "OpenCode port (default: 4096)", "4096").option(
   "--log-level <level>",
   "Log verbosity: debug | info | warn | error (default: info). Env: EVIDENT_LOG_LEVEL"
-).option("-v, --verbose", "Alias for --log-level debug (ignored if --log-level is set)").option("-c, --conversation <id>", "Process only this specific conversation").option("--idle-timeout <seconds>", "Exit after N seconds idle").option("--json", "Output in JSON format").option(
+).option("-v, --verbose", "Alias for --log-level debug (ignored if --log-level is set)").option("-c, --conversation <id>", "Process only this specific conversation").option("--idle-timeout <seconds>", "Exit after N seconds idle").option(
+  "--opencode-start-timeout <seconds>",
+  "Seconds to wait for OpenCode to become healthy when the runner starts it (default: 180). Env: EVIDENT_OPENCODE_START_TIMEOUT"
+).option("--json", "Output in JSON format").option(
   "--session-cleanup-max-age <duration>",
   "Delete OpenCode sessions idle longer than this (e.g. 7d, 24h). Enables cleanup. Env: EVIDENT_SESSION_CLEANUP_MAX_AGE"
 ).option(
@@ -6059,6 +6514,9 @@ 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({
@@ -6071,6 +6529,9 @@ program.command("run").description("Connect to Evident and process messages").op
       verbose: options.verbose,
       conversation: options.conversation,
       idleTimeout: options.idleTimeout ? parseInt(options.idleTimeout, 10) : void 0,
+      // Raw string — validation/precedence is single-sourced in run.ts's
+      // resolveOpenCodeStartTimeoutMs (flag > EVIDENT_OPENCODE_START_TIMEOUT > 180s default).
+      opencodeStartTimeout: options.opencodeStartTimeout,
       json: options.json,
       // Raw strings — the resolver in run.ts single-sources parsing (M1).
       sessionCleanupMaxAge: options.sessionCleanupMaxAge,
@@ -6078,7 +6539,8 @@ 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
     });
   }
 );