@evident-ai/cli 3.0.1-dev.bf495ea → 3.0.1-dev.c152b49

package/dist/index.js

@@ -1059,6 +1059,14 @@ function finishOf(m) {
   const infoFinish = m.info?.finish;
   return typeof infoFinish === "string" ? infoFinish : void 0;
 }
+function errorOf(m) {
+  if (!m || typeof m !== "object") return void 0;
+  return m.info?.error ?? m.error;
+}
+function isAssistantInFlight(m) {
+  if (completedOf(m) == null) return true;
+  return finishOf(m) === "tool-calls";
+}
 async function createOpenCodeSession(port, directory) {
   const url = new URL(`${opencodeBase(port)}/session`);
   if (directory && directory.trim()) {
@@ -1140,9 +1148,27 @@ function messageRunState(messages, userMessageId) {
     if (!reply) return "unknown";
   }
   if (!reply) return "queued";
-  if (completedOf(reply) == null) return "running";
-  if (finishOf(reply) === "tool-calls") return "running";
-  return "done";
+  if (isAssistantInFlight(reply)) return "running";
+  return errorOf(reply) != null ? "failed" : "done";
+}
+function messageError(messages, userMessageId) {
+  const reply = findLastAssistantReplyFor(messages, userMessageId);
+  const error2 = errorOf(reply);
+  if (error2 == null) return null;
+  if (typeof error2 === "string") return error2;
+  if (typeof error2 === "object") {
+    const e = error2;
+    const dataMessage = e.data?.message;
+    if (typeof dataMessage === "string") return dataMessage;
+    if (typeof e.message === "string") return e.message;
+  }
+  return "The agent run failed.";
+}
+function hasRunningAssistantExcept(messages, exceptUserMessageId) {
+  if (!messages || messages.length === 0) return false;
+  return messages.some(
+    (m) => roleOf(m) === "assistant" && parentIdOf(m) !== exceptUserMessageId && isAssistantInFlight(m)
+  );
 }
 function opencodeMessageIdFor2(queuedMessageId) {
   return opencodeMessageIdFor(queuedMessageId);
@@ -1552,6 +1578,7 @@ var DEFAULT_RETRY_POLICY = {
 var DEFAULT_PAUSED_POLL_INTERVAL_MS = 2e3;
 var DEFAULT_PAUSED_MAX_WAIT_MS = 10 * 60 * 1e3;
 var DEFAULT_DISPATCH_CONFIRM_MS = 6e3;
+var DEFAULT_STUCK_QUEUED_MS = 6e4;
 var ChannelAuthError = class extends Error {
   constructor(message) {
     super(message);
@@ -1587,6 +1614,7 @@ var ChannelDriver = class {
   pausedPollIntervalMs;
   pausedMaxWaitMs;
   dispatchConfirmMs;
+  stuckQueuedMs;
   now;
   /** Cache of conversationId → opencode sessionId. */
   sessions = /* @__PURE__ */ new Map();
@@ -1606,6 +1634,40 @@ var ChannelDriver = class {
    * a steady-state-poll re-dispatch will not double-run the message.
    */
   dispatched = /* @__PURE__ */ new Set();
+  /**
+   * Re-adopted (ADR-0046) Evident message ids currently tracked by a watcher.
+   * Used only to distinguish a RE-ADOPTED give-up from a normal-dispatch give-up
+   * so the former can be parked in `dontRedispatch` (Bug 2). A row is added when
+   * it is re-adopted and removed when its watcher settles or it is observed off
+   * the processing list.
+   */
+  readopted = /* @__PURE__ */ new Set();
+  /**
+   * "Don't re-DISPATCH / re-attach this orphan again" (Bug 2/5). Set when a
+   * re-adopted running/orphan row's watcher hit its `processed_at`-anchored
+   * deadline (or an orphan whose window already elapsed): the still-`processing`
+   * server row would otherwise be re-adopted (and re-dispatched) on EVERY ~2s
+   * drain until the 15-min cron resets it — spamming new turns.
+   *
+   * CRITICAL (Bugbot #202): this suppresses ONLY the dispatch/re-attach paths, it
+   * does NOT suppress DONE delivery. A row parked here whose reply later COMPLETES
+   * in opencode must still be delivered via `markDone` on the next drain — so
+   * `readoptOne` computes `state` FIRST and this set is checked only on the
+   * non-done path. It is cleared once the row leaves the processing list (cron
+   * reset → it drains normally as `pending`), so it can never leak.
+   */
+  dontRedispatch = /* @__PURE__ */ new Set();
+  /**
+   * "markDone for this row is TERMINALLY undeliverable" (Bug 4). Set ONLY when a
+   * re-adopted DONE row's `markDone` returned a terminal 4xx (a status that will
+   * never succeed). Checked at the TOP of the `done` branch so we do NOT re-attempt
+   * that markDone every ~2s drain while the row stays `processing`. A TRANSIENT
+   * markDone failure must NOT land here (it must still retry next drain). Separate
+   * from `dontRedispatch` because the two concerns are independent: a row can need
+   * "stop re-dispatching" without "stop delivering", and vice versa. Cleared once
+   * the row leaves the processing list, exactly like `dontRedispatch`.
+   */
+  doneUndeliverable = /* @__PURE__ */ new Set();
   /**
    * Cache of the opencode root directory (from `GET /path`). Resolved lazily on
    * first session creation so drain-created sessions are rooted at the project
@@ -1629,6 +1691,7 @@ var ChannelDriver = class {
     this.pausedPollIntervalMs = config2.pausedPollIntervalMs ?? DEFAULT_PAUSED_POLL_INTERVAL_MS;
     this.pausedMaxWaitMs = config2.pausedMaxWaitMs ?? DEFAULT_PAUSED_MAX_WAIT_MS;
     this.dispatchConfirmMs = config2.dispatchConfirmMs ?? DEFAULT_DISPATCH_CONFIRM_MS;
+    this.stuckQueuedMs = config2.stuckQueuedMs ?? DEFAULT_STUCK_QUEUED_MS;
     this.now = config2.now ?? (() => Date.now());
   }
   /** The IPv4-loopback base URL for the local `opencode serve`. */
@@ -1661,6 +1724,7 @@ var ChannelDriver = class {
       for (const conv of conversations) {
         dispatched += await this.processConversation(conv);
       }
+      await this.readoptProcessing();
     } finally {
       this.draining = false;
     }
@@ -1747,6 +1811,7 @@ var ChannelDriver = class {
       this.dispatched.add(message.id);
       this.registerInFlight(conv, sessionId, message, opencodeMessageId);
       dispatched += 1;
+      void this.postSignal(conv.id, message.id, "dispatched");
     }
     if (messages.length > 0 && dispatched === 0 && skippedAlreadyDispatched === messages.length) {
       this.log({
@@ -1812,7 +1877,54 @@ var ChannelDriver = class {
       dispatchedAt: now,
       deadline: now + this.pausedMaxWaitMs,
       started: false,
-      done: false
+      done: false,
+      stuckReported: false
+    });
+  }
+  /**
+   * Register a RE-ADOPTED `processing` message with its session watcher
+   * (ADR-0046, WI-4). Mirrors `registerInFlight` but anchors the give-up
+   * `deadline` to the row's SERVER-SIDE `processed_at` (Invariant 1), NEVER to
+   * `now`: a row already `processing` for e.g. 5 min must give up ~5 min from now
+   * (10 min after `processed_at`), not 10 min from now — otherwise its deadline
+   * lands ~15 min after `processed_at`, coinciding with the cron reset →
+   * double-drive race. `dispatchedAt` stays `now` (only the appear-guard uses it).
+   *
+   * `evidentMessageId` addresses the SERVER row (for markProcessing/markDone);
+   * `opencodeMessageId` is the id the watcher polls for a reply — for the orphan
+   * fresh-run path these differ (a fresh opencode id under the same server row).
+   *
+   * `started` is set true so the watcher does NOT re-`markProcessing` a row the
+   * server already flipped to `processing`; the running/done transitions still
+   * fire from the watcher's normal branches.
+   */
+  registerReadopted(conv, sessionId, message, opencodeMessageId, processedAtMs) {
+    let watcher = this.watchers.get(sessionId);
+    if (!watcher) {
+      watcher = {
+        conv,
+        inFlight: /* @__PURE__ */ new Map(),
+        loop: null,
+        reportedQuestions: /* @__PURE__ */ new Set(),
+        reportedPermissions: /* @__PURE__ */ new Set()
+      };
+      this.watchers.set(sessionId, watcher);
+    }
+    watcher.inFlight.set(message.id, {
+      evidentMessageId: message.id,
+      opencodeMessageId,
+      message,
+      dispatchedAt: this.now(),
+      deadline: processedAtMs + this.pausedMaxWaitMs,
+      // The server row is ALREADY `processing`; do not re-fire markProcessing.
+      started: true,
+      done: false,
+      // Not yet reported stuck-queued. The once-guard (`stuckReported`) applies,
+      // AND the stuck-queued observer now INCLUDES re-adopted queued wedges: it
+      // gates on `state === 'queued'` (turn produced no reply), not on `started`,
+      // so a re-adopted row left wedged in `queued` still emits the signal once
+      // (queued-followup-redrive, #210).
+      stuckReported: false
     });
   }
   /**
@@ -1878,6 +1990,7 @@ var ChannelDriver = class {
           conversation_id: watcher.conv.id
         });
         for (const evidentMessageId of [...watcher.inFlight.keys()]) {
+          this.readopted.delete(evidentMessageId);
           this.removeInFlight(watcher, evidentMessageId);
         }
         return;
@@ -1898,7 +2011,7 @@ var ChannelDriver = class {
   async serviceInFlightMessage(sessionId, watcher, inFlight, messages) {
     const conv = watcher.conv;
     const state = messageRunState(messages, inFlight.opencodeMessageId);
-    if ((state === "running" || state === "done") && !inFlight.started) {
+    if ((state === "running" || state === "done" || state === "failed") && !inFlight.started) {
       let claimed;
       try {
         claimed = await this.markProcessing(conv.id, inFlight.evidentMessageId, sessionId);
@@ -1967,11 +2080,63 @@ var ChannelDriver = class {
       this.removeInFlight(watcher, inFlight.evidentMessageId);
       return;
     }
+    if (state === "failed") {
+      if (!inFlight.done) {
+        const error2 = messageError(messages, inFlight.opencodeMessageId) ?? void 0;
+        this.log({
+          level: "error",
+          message: `Message ${inFlight.evidentMessageId.slice(0, 8)} errored \u2014 marking failed: ${error2 ?? "(no error text)"}`,
+          conversation_id: conv.id,
+          message_id: inFlight.evidentMessageId
+        });
+        try {
+          await this.markFailed(conv.id, inFlight.evidentMessageId, sessionId, error2);
+        } catch (err) {
+          if (err instanceof ChannelAuthError) throw err;
+          if (err instanceof ChannelTerminalError) {
+            this.log({
+              level: "error",
+              message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} failed (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: "error",
+              message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} failed 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: "error",
+            message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} failed (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);
+      return;
+    }
     if (state === "unknown") {
       if (this.now() - inFlight.dispatchedAt >= this.dispatchConfirmMs) {
         await this.redispatchInFlight(sessionId, inFlight);
       }
     }
+    if (state === "queued" && !inFlight.stuckReported && this.now() - inFlight.dispatchedAt >= this.stuckQueuedMs && !hasRunningAssistantExcept(messages, inFlight.opencodeMessageId)) {
+      inFlight.stuckReported = true;
+      void this.postSignal(conv.id, inFlight.evidentMessageId, "stuck_queued", {
+        stuck_for_ms: this.now() - inFlight.dispatchedAt
+      });
+    }
     if (this.now() >= inFlight.deadline) {
       this.log({
         level: "info",
@@ -2015,12 +2180,345 @@ var ChannelDriver = class {
     }
     inFlight.dispatchedAt = this.now();
   }
+  // -------------------------------------------------------------------------
+  // 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).
+   *
+   * The pending drain only re-drives `pending` rows; a message already flipped to
+   * `processing` before the runner died is watched by nobody until the 15-min
+   * cron resets it. Here we fetch those rows, and per row resolve its correlated
+   * reply against opencode's OWN session store — completing, re-attaching, or
+   * (for an orphan) forcing a genuine fresh run. Runs on EVERY drain tick, so it
+   * is idempotent per message (Invariant 2): a row a watcher already tracks is
+   * skipped in `readoptOne` — one driver, no double-drive.
+   *
+   * Only `ChannelAuthError` propagates (to `drainPending`, like the pending
+   * path); every other early return LOGS a reason with context — no silent drop.
+   */
+  async readoptProcessing() {
+    const rows = await this.getProcessingMessages();
+    if (this.dontRedispatch.size > 0 || this.doneUndeliverable.size > 0) {
+      const stillProcessing = new Set(rows.map((r) => r.id));
+      for (const id of [...this.dontRedispatch, ...this.doneUndeliverable]) {
+        if (!stillProcessing.has(id)) {
+          const cleared = this.dontRedispatch.delete(id);
+          const clearedUndeliverable = this.doneUndeliverable.delete(id);
+          if (cleared || clearedUndeliverable) {
+            this.log({
+              level: "info",
+              message: `Re-adopt: message ${id.slice(0, 8)} left the processing list (cron reset) \u2014 cleared gave-up marker`,
+              message_id: id
+            });
+          }
+        }
+      }
+    }
+    if (rows.length === 0) return;
+    const bySession = /* @__PURE__ */ new Map();
+    for (const row of rows) {
+      if (!row.opencode_session_id) {
+        this.log({
+          level: "error",
+          message: `Cannot re-adopt processing message ${row.id.slice(0, 8)} \u2014 no opencode session id; leaving for the cron safety net`,
+          conversation_id: row.conversation_id,
+          message_id: row.id
+        });
+        continue;
+      }
+      const list = bySession.get(row.opencode_session_id) ?? [];
+      list.push(row);
+      bySession.set(row.opencode_session_id, list);
+    }
+    for (const [sessionId, sessionRows] of bySession) {
+      let messages;
+      try {
+        const res = await this.fetchImpl(`${this.opencodeBase}/session/${sessionId}/message`);
+        if (!res.ok) {
+          this.log({
+            level: "error",
+            message: `Re-adopt: polling session ${sessionId.slice(0, 8)} returned HTTP ${res.status} \u2014 skipping this session this tick`
+          });
+          continue;
+        }
+        const body = await res.json();
+        if (!Array.isArray(body)) {
+          this.log({
+            level: "error",
+            message: `Re-adopt: polling session ${sessionId.slice(0, 8)} returned a non-array message body \u2014 skipping this session this tick`
+          });
+          continue;
+        }
+        messages = body;
+      } catch (err) {
+        this.log({
+          level: "error",
+          message: `Re-adopt: failed to poll session ${sessionId.slice(0, 8)} (will retry next drain): ${err instanceof Error ? err.message : String(err)}`
+        });
+        continue;
+      }
+      for (const row of sessionRows) {
+        await this.readoptOne(sessionId, row, messages);
+      }
+    }
+  }
+  /**
+   * Re-adopt ONE `processing` row against the tick's session message snapshot
+   * (ADR-0046 Decision §1/§2). Idempotent: skips a row already being driven.
+   *
+   * Branches on `messageRunState(messages, opencodeMessageIdFor(row.id))`:
+   *   - `done`             → `markDone` now (guarded like the watcher's done branch);
+   *   - `failed`           → `markFailed` with the surfaced error (issue #182), so an
+   *                          errored turn is reported failed on restart, NOT re-dispatched;
+   *   - `running`/`queued` → re-attach a watcher via `registerReadopted` (no re-dispatch);
+   *   - `unknown`          → re-dispatch the STABLE id + attach a watcher (orphan).
+   *
+   * Only `ChannelAuthError` propagates.
+   */
+  async readoptOne(sessionId, row, messages) {
+    if (this.isTracked(sessionId, row.id)) {
+      this.log({
+        level: "info",
+        message: `Re-adopt: message ${row.id.slice(0, 8)} already tracked in-flight \u2014 skipping (owned by the watcher loop)`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      return;
+    }
+    const ocId = opencodeMessageIdFor2(row.id);
+    const state = messageRunState(messages, ocId);
+    if (state === "done") {
+      if (this.doneUndeliverable.has(row.id)) {
+        this.log({
+          level: "info",
+          message: `Re-adopt: message ${row.id.slice(0, 8)} markDone is terminally undeliverable \u2014 left to the cron; skipping until it leaves processing`,
+          conversation_id: row.conversation_id,
+          message_id: row.id
+        });
+        return;
+      }
+      this.log({
+        level: "info",
+        message: `Re-adopt: message ${row.id.slice(0, 8)} completed while unwatched \u2014 marking done`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      try {
+        await this.markDone(row.conversation_id, row.id, sessionId);
+      } catch (err) {
+        if (err instanceof ChannelAuthError) throw err;
+        if (err instanceof ChannelTerminalError) {
+          this.doneUndeliverable.add(row.id);
+          this.log({
+            level: "error",
+            message: `Re-adopt: failed to mark message ${row.id.slice(0, 8)} done (terminal HTTP ${err.status}) \u2014 parking until it leaves processing; leaving for the cron safety net: ${err.message}`,
+            conversation_id: row.conversation_id,
+            message_id: row.id
+          });
+          return;
+        }
+        this.log({
+          level: "error",
+          message: `Re-adopt: failed to mark message ${row.id.slice(0, 8)} done (will retry next drain): ${err instanceof Error ? err.message : String(err)}`,
+          conversation_id: row.conversation_id,
+          message_id: row.id
+        });
+        return;
+      }
+      this.dontRedispatch.delete(row.id);
+      return;
+    }
+    if (state === "failed") {
+      const error2 = messageError(messages, ocId) ?? void 0;
+      this.log({
+        level: "error",
+        message: `Re-adopt: message ${row.id.slice(0, 8)} errored while unwatched \u2014 marking failed: ${error2 ?? "(no error text)"}`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      try {
+        await this.markFailed(row.conversation_id, row.id, sessionId, error2);
+      } catch (err) {
+        if (err instanceof ChannelAuthError) throw err;
+        if (err instanceof ChannelTerminalError) {
+          this.doneUndeliverable.add(row.id);
+          this.log({
+            level: "error",
+            message: `Re-adopt: failed to mark message ${row.id.slice(0, 8)} failed (terminal HTTP ${err.status}) \u2014 parking until it leaves processing; leaving for the cron safety net: ${err.message}`,
+            conversation_id: row.conversation_id,
+            message_id: row.id
+          });
+          return;
+        }
+        this.log({
+          level: "error",
+          message: `Re-adopt: failed to mark message ${row.id.slice(0, 8)} failed (will retry next drain): ${err instanceof Error ? err.message : String(err)}`,
+          conversation_id: row.conversation_id,
+          message_id: row.id
+        });
+        return;
+      }
+      this.dontRedispatch.delete(row.id);
+      return;
+    }
+    if (this.dontRedispatch.has(row.id)) {
+      this.log({
+        level: "info",
+        message: `Re-adopt: message ${row.id.slice(0, 8)} already gave up \u2014 left to the cron; skipping until it leaves processing`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      return;
+    }
+    if (state === "running" || state === "queued") {
+      const conv = this.convForRow(sessionId, row);
+      const message = this.queuedMessageForRow(row);
+      this.registerReadopted(conv, sessionId, message, ocId, this.processedAtMs(row));
+      this.dispatched.add(row.id);
+      this.readopted.add(row.id);
+      this.ensureWatcherRunning(sessionId);
+      this.log({
+        level: "info",
+        message: `Re-adopt: message ${row.id.slice(0, 8)} ${state} \u2014 re-attached watcher (stable id, no re-dispatch)`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      return;
+    }
+    await this.forceReadoptRun(sessionId, row);
+  }
+  /**
+   * Re-dispatch an orphaned (`unknown`) `processing` row (ADR-0046 Decision §2).
+   *
+   * The stable-id user message is absent from the session, so we (re-)dispatch with
+   * the STABLE id (`opencodeMessageIdFor(row.id)`) — NOT a divergent per-attempt id.
+   * This is what keeps the reply correlatable: the server's completion
+   * notification looks for the reply under the stable id, so the fresh turn's reply
+   * (which hangs off the stable id) is found and delivered. The residual
+   * duplicate-incomplete-turn semantics (ADR §2, `.harness/restart-recovery-orphan-finding.md`)
+   * are unchanged and, for an ABSENT id, cannot bite — there is no existing turn
+   * to swallow the duplicate.
+   *
+   * `evidentMessageId = row.id` addresses the SERVER row; the stable
+   * `opencodeMessageId` is what the watcher polls. Deadline anchored to
+   * `processed_at` (Invariant 1).
+   */
+  async forceReadoptRun(sessionId, row) {
+    const ocId = opencodeMessageIdFor2(row.id);
+    if (this.processedAtMs(row) + this.pausedMaxWaitMs <= this.now()) {
+      this.dontRedispatch.add(row.id);
+      this.log({
+        level: "info",
+        message: `Re-adopt: message ${row.id.slice(0, 8)} orphaned but its re-adopt window has already elapsed \u2014 not dispatching an unwatchable turn; parking until it leaves processing (cron will reset it)`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      return;
+    }
+    const options = {
+      agent: row.opencode_agent ?? void 0,
+      model: row.opencode_model ?? void 0
+    };
+    this.log({
+      level: "info",
+      message: `Re-adopt: message ${row.id.slice(0, 8)} orphaned (user message absent) \u2014 re-dispatching with the stable id`,
+      conversation_id: row.conversation_id,
+      message_id: row.id
+    });
+    try {
+      await sendPromptAsync(this.port, sessionId, row.content, options, ocId);
+    } catch (err) {
+      if (err instanceof ChannelAuthError) throw err;
+      this.log({
+        level: "error",
+        message: `Re-adopt: re-dispatch failed for message ${row.id.slice(0, 8)} (will retry next drain): ${err instanceof Error ? err.message : String(err)}`,
+        conversation_id: row.conversation_id,
+        message_id: row.id
+      });
+      return;
+    }
+    const conv = this.convForRow(sessionId, row);
+    const message = this.queuedMessageForRow(row);
+    this.registerReadopted(conv, sessionId, message, ocId, this.processedAtMs(row));
+    this.dispatched.add(row.id);
+    this.readopted.add(row.id);
+    this.ensureWatcherRunning(sessionId);
+  }
+  /**
+   * True if `evidentMessageId` is already being driven — either in the
+   * authoritative `dispatched` set or a live watcher's in-flight set for this
+   * session (Invariant 2, WI-5). Either signal means a watcher owns the row.
+   */
+  isTracked(sessionId, evidentMessageId) {
+    if (this.dispatched.has(evidentMessageId)) return true;
+    const watcher = this.watchers.get(sessionId);
+    return watcher?.inFlight.has(evidentMessageId) ?? false;
+  }
+  /**
+   * Parse a re-adopt row's `processed_at` (ISO string) to epoch ms for the
+   * deadline anchor (Invariant 1). The endpoint guarantees `processed_at` is set
+   * for `processing` rows, but if it is somehow null/unparseable fall back to
+   * `now` (defensive) AND log — a fallback means the anchor is weaker than
+   * intended, which is worth surfacing.
+   */
+  processedAtMs(row) {
+    const parsed = row.processed_at ? Date.parse(row.processed_at) : NaN;
+    if (!Number.isNaN(parsed)) return parsed;
+    this.log({
+      level: "error",
+      message: `Re-adopt: message ${row.id.slice(0, 8)} has null/unparseable processed_at (${String(row.processed_at)}) \u2014 anchoring deadline to now (defensive)`,
+      conversation_id: row.conversation_id,
+      message_id: row.id
+    });
+    return this.now();
+  }
+  /** Build the `PendingConversation` shape the watcher needs from a re-adopt row. */
+  convForRow(sessionId, row) {
+    return {
+      id: row.conversation_id,
+      agent_id: this.agentId,
+      opencode_session_id: sessionId,
+      pending_message_count: 0,
+      oldest_pending_at: row.processed_at
+    };
+  }
+  /** Build the `QueuedMessage` shape the watcher/re-dispatch needs from a re-adopt row. */
+  queuedMessageForRow(row) {
+    return {
+      id: row.id,
+      content: row.content,
+      status: "processing",
+      opencode_agent: row.opencode_agent,
+      opencode_model: row.opencode_model,
+      source_message_id: row.source_message_id,
+      slack_user_id: row.slack_user_id
+    };
+  }
   /**
    * Remove a message from the in-flight set AND the authoritative dispatched
    * set. Once the in-flight set empties, the watcher loop's `while` guard exits
    * and its `.finally` removes the session entry from `this.watchers`.
+   *
+   * Bug 2: if a RE-ADOPTED message is removed WITHOUT having completed
+   * (`!inFlight.done` — i.e. a give-up: deadline reached, or markDone left to the
+   * cron), park it in `dontRedispatch` so the next drain does NOT re-adopt (and
+   * re-dispatch) the still-`processing` row every ~2s until the 15-min cron. A
+   * re-adopted message that completed (`done`) needs no marker — it's leaving
+   * `processing`. This suppresses only re-dispatch: if its reply later completes,
+   * the done branch still delivers it (Bugbot #202).
    */
   removeInFlight(watcher, evidentMessageId) {
+    const inFlight = watcher.inFlight.get(evidentMessageId);
+    if (this.readopted.delete(evidentMessageId) && inFlight && !inFlight.done) {
+      this.dontRedispatch.add(evidentMessageId);
+      this.log({
+        level: "info",
+        message: `Re-adopt: message ${evidentMessageId.slice(0, 8)} gave up \u2014 parking until it leaves the processing list (cron reset)`,
+        conversation_id: watcher.conv.id,
+        message_id: evidentMessageId
+      });
+    }
     watcher.inFlight.delete(evidentMessageId);
     this.dispatched.delete(evidentMessageId);
   }
@@ -2164,6 +2662,35 @@ var ChannelDriver = class {
     }
     return await res.json();
   }
+  /**
+   * Fetch this agent's `processing` messages for re-adoption (ADR-0046, WI-1).
+   * The pending path (`getPendingConversations`/`getPendingMessages`) only
+   * surfaces `pending` rows, so a message already `processing` when the runner
+   * died is invisible to it — this dedicated endpoint returns exactly those rows
+   * with the fields the re-adopt path needs (`processed_at`,
+   * `opencode_session_id`, routing).
+   *
+   * Response is an OBJECT WRAPPER `{ messages: [...] }` (snake_case) — NOT a bare
+   * array. Propagates `ChannelAuthError` on 401/403; throws a plain `Error` on
+   * other non-ok so `drainPending`'s try/finally leaves `draining` false and the
+   * next tick retries.
+   */
+  async getProcessingMessages() {
+    const res = await this.fetchImpl(
+      `${this.apiUrl}/agents/${this.agentId}/conversations/processing`,
+      { headers: { Authorization: this.getAuthHeader() } }
+    );
+    this.assertAuth(res, "fetching processing messages");
+    if (!res.ok) {
+      throw new Error(`Failed to get processing messages: HTTP ${res.status}`);
+    }
+    const data = await res.json();
+    let messages = data.messages ?? [];
+    if (this.conversationFilter) {
+      messages = messages.filter((m) => m.conversation_id === this.conversationFilter);
+    }
+    return messages;
+  }
   /**
    * EXISTING combinedAuth route — now fired by the watcher on queued→running
    * (Task 3.3), NOT at dispatch/claim time. `{status:'processing',
@@ -2245,7 +2772,17 @@ var ChannelDriver = class {
     }
     throw new ChannelTerminalError(`marking message as done: HTTP ${res.status}`, res.status);
   }
-  async markFailed(conversationId, messageId) {
+  /**
+   * Mark a message `failed`. `sessionId` / `error` are threaded to the API ONLY
+   * when provided (issue #182): a bare `markFailed(conv, msg)` sends
+   * `{status:'failed'}` unchanged (the dispatch-failure path), while an errored
+   * OpenCode turn sends `{status:'failed', opencode_session_id, error}` so the
+   * failure reason reaches the channel.
+   */
+  async markFailed(conversationId, messageId, sessionId, error2) {
+    const body = { status: "failed" };
+    if (sessionId !== void 0) body.opencode_session_id = sessionId;
+    if (error2 !== void 0) body.error = error2;
     await this.callWithRetry(
       "marking message as failed",
       () => this.fetchImpl(
@@ -2253,11 +2790,47 @@ var ChannelDriver = class {
         {
           method: "PATCH",
           headers: { Authorization: this.getAuthHeader(), "Content-Type": "application/json" },
-          body: JSON.stringify({ status: "failed" })
+          body: JSON.stringify(body)
         }
       )
     );
   }
+  /**
+   * Best-effort, SINGLE-ATTEMPT runner→server telemetry ping
+   * (queued-followup-redrive). `POST .../messages/:id/signal {signal, ...extra}`
+   * — the server records it via `log()` (no DB write, no notification). This is
+   * fire-and-forget: it MUST NEVER throw into the drain or the watcher tick, and
+   * MUST NOT use `callWithRetry` (a telemetry ping must not block the sequential
+   * watcher tick — one attempt is enough). A failure is SWALLOWED but LOGGED with
+   * context (no silent catch, per development-workflow).
+   */
+  async postSignal(conversationId, messageId, signal, extra) {
+    try {
+      const res = await this.fetchImpl(
+        `${this.apiUrl}/agents/${this.agentId}/threads/${conversationId}/messages/${messageId}/signal`,
+        {
+          method: "POST",
+          headers: { Authorization: this.getAuthHeader(), "Content-Type": "application/json" },
+          body: JSON.stringify({ signal, ...extra })
+        }
+      );
+      if (!res.ok) {
+        this.log({
+          level: "error",
+          message: `Signal '${signal}' for message ${messageId.slice(0, 8)} returned HTTP ${res.status} (telemetry-only, ignored)`,
+          conversation_id: conversationId,
+          message_id: messageId
+        });
+      }
+    } catch (err) {
+      this.log({
+        level: "error",
+        message: `Signal '${signal}' for message ${messageId.slice(0, 8)} failed (telemetry-only, ignored): ${err instanceof Error ? err.message : String(err)}`,
+        conversation_id: conversationId,
+        message_id: messageId
+      });
+    }
+  }
   async persistSession(conversationId, sessionId) {
     const res = await this.fetchImpl(
       `${this.apiUrl}/agents/${this.agentId}/threads/${conversationId}`,