@link-assistant/hive-mind 2.12.2 → 2.12.3

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # @link-assistant/hive-mind
 
+## 2.12.3
+
+### Patch Changes
+
+- 97b034f: Make a refused work session explain itself. A Formal AI sidecar (or any isolation) launch that never produced a container now logs the session UUID, backend, tool and reason to stderr instead of failing silently, keeps the UUID in the Telegram failure reply together with a sentence saying the session has no log and is not listed by `--list`, records the reason on the `session_untracked` event and in the durable session store, and is reported by `/queue` as a failed item rather than as `Finished: … (started)`. Registry pull refusals are classified, so a permanent `unauthorized`/`denied`/`not-found` escalates with remediation instead of repeating the same bland warning, and a task image that cannot be pulled falls back to a locally present one. Telegram replies also show start-command's execution UUID — the identifier `$ --list` prints — next to the session UUID, so a running or finished task can finally be found in the session list.
+
 ## 2.12.2
 
 ### Patch Changes

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@link-assistant/hive-mind",
-  "version": "2.12.2",
+  "version": "2.12.3",
   "description": "AI-powered issue solver and hive mind for collaborative problem solving",
   "main": "src/hive.mjs",
   "type": "module",

package/src/formal-ai-image.lib.mjs

@@ -0,0 +1,222 @@
+/**
+ * Resolve — and prove available — the image the Formal AI sidecar boots from
+ * (issue #2154).
+ *
+ * ## Why this module exists
+ *
+ * Before this module, `acquireFormalAiSidecar` handed
+ * `ghcr.io/link-assistant/formal-ai:<version>` straight to `docker run` and
+ * relied on Docker's implicit pull. The published package is private, so every
+ * Formal AI task died with the raw daemon dump:
+ *
+ * ```text
+ * Command failed: docker run --detach --name hive-mind-formal-ai …
+ * Unable to find image 'ghcr.io/link-assistant/formal-ai:0.339.1' locally
+ * docker: Error response from daemon: error from registry: unauthorized
+ * ```
+ *
+ * Three separate defects made that fatal:
+ *
+ *  1. **No preflight.** The registry problem surfaced as a failed *container
+ *     launch* rather than as a failed *image resolution*, so nothing could tell
+ *     the operator what to do about it.
+ *  2. **No diagnosis.** `unauthorized` from a registry has exactly three causes
+ *     (private package, missing/insufficient credentials, wrong reference) and
+ *     each has a different fix. The raw dump named none of them.
+ *  3. **No alternative.** The Hive Mind images themselves bake
+ *     `/usr/local/bin/formal-ai` at the same pinned version (`Dockerfile`,
+ *     `Dockerfile.dind`: `cargo install formal-ai --version ${FORMAL_AI_VERSION}
+ *     --locked`), and that image is already present on the host because every
+ *     isolated task runs from it. A registry outage therefore never had to stop
+ *     Formal AI work at all.
+ *
+ * ## Contract
+ *
+ * - `HIVE_MIND_FORMAL_AI_IMAGE` is an **exact** operator pin: it is the only
+ *   candidate, and if it cannot be resolved the acquire fails. An operator who
+ *   names an image means that image.
+ * - Otherwise the published image is preferred, and the local Hive Mind image is
+ *   the fallback. The fallback is only *used*; it is never pulled, because its
+ *   whole point is that it is already on the host.
+ * - Nothing here downgrades a Formal AI task to another model. Issue #2146's
+ *   fail-closed rule still holds: when no candidate resolves, the task is
+ *   refused with an actionable message.
+ *
+ * @see https://github.com/link-assistant/hive-mind/issues/2154
+ * @see https://github.com/link-assistant/hive-mind/issues/2146
+ */
+
+import { execFile } from 'node:child_process';
+import { promisify } from 'node:util';
+
+import { FORMAL_AI_BOOTSTRAP_VERSION } from './formal-ai-version.lib.mjs';
+import { getDockerIsolationImage } from './hive-mind-image.lib.mjs';
+
+const execFileAsync = promisify(execFile);
+
+/** Image published by every Formal AI release (`:latest` plus the bare version). */
+export const FORMAL_AI_IMAGE_REPOSITORY = 'ghcr.io/link-assistant/formal-ai';
+
+const DEFAULT_DOCKER_TIMEOUT_MS = 600_000;
+
+/** Where each candidate came from, so logs and errors can explain themselves. */
+export const FORMAL_AI_IMAGE_SOURCES = Object.freeze({
+  PINNED: 'operator-pin',
+  PUBLISHED: 'published-image',
+  HIVE_MIND: 'hive-mind-image',
+});
+
+const errorText = error => error?.stderr?.toString?.().trim() || error?.stdout?.toString?.().trim() || error?.message || String(error);
+
+/**
+ * Classify a `docker pull` failure into the operator action that fixes it.
+ *
+ * Registry errors are famously interchangeable-looking; the distinction that
+ * matters in practice is that GHCR answers `unauthorized` for a package that
+ * *exists but is private* and `denied` for one that does not exist or that the
+ * presented token may not see.
+ *
+ * @param {string} message - Raw stderr from the failed pull.
+ * @returns {{kind: string, reason: string, remediation: string[]}}
+ */
+export const classifyDockerRegistryError = message => {
+  const text = String(message || '');
+  if (/unauthorized|authentication required|requires authentication/i.test(text)) {
+    return {
+      kind: 'unauthorized',
+      reason: 'the registry refused an anonymous or under-scoped pull',
+      remediation: ['make the package public (GHCR packages published by GITHUB_TOKEN are private by default), or', 'authenticate the daemon with a token carrying the `read:packages` scope: `echo $TOKEN | docker login ghcr.io -u <user> --password-stdin`, or', 'point HIVE_MIND_FORMAL_AI_IMAGE at an image this host can already pull'],
+    };
+  }
+  if (/denied|forbidden|insufficient_scope|permission_denied/i.test(text)) {
+    return {
+      kind: 'denied',
+      reason: 'the registry denied access to that reference',
+      remediation: ['check the repository name and tag exist', 'check the credentials in use carry the `read:packages` scope'],
+    };
+  }
+  if (/manifest unknown|not found|no such (?:image|manifest)/i.test(text)) {
+    return { kind: 'not-found', reason: 'the registry has no such tag', remediation: ['check the tag exists in the registry', 'pin a published tag with HIVE_MIND_FORMAL_AI_IMAGE'] };
+  }
+  if (/no space left on device/i.test(text)) {
+    return { kind: 'disk-full', reason: 'the daemon ran out of disk while unpacking the image', remediation: ['free disk space on the Docker data root, then retry'] };
+  }
+  if (/timeout|timed out|temporary failure|dial tcp|i\/o timeout|connection refused|network is unreachable|EOF/i.test(text)) {
+    return { kind: 'network', reason: 'the registry was unreachable', remediation: ['check network/proxy access to the registry, then retry'] };
+  }
+  return { kind: 'unknown', reason: 'the pull failed', remediation: ['inspect the daemon error above'] };
+};
+
+/**
+ * Resolve the local Hive Mind image, which bakes `formal-ai` at the same pinned
+ * version as the published sidecar image and is already present on any host
+ * that runs Docker-isolated tasks.
+ */
+export const resolveFormalAiFallbackImage = (env = process.env) => String(env.HIVE_MIND_FORMAL_AI_FALLBACK_IMAGE || '').trim() || getDockerIsolationImage({ env });
+
+/**
+ * Ordered list of images to try, most preferred first.
+ *
+ * @param {object} [env]
+ * @returns {Array<{image: string, source: string, pullable: boolean}>}
+ */
+export const resolveFormalAiSidecarImageCandidates = (env = process.env) => {
+  const pinned = String(env.HIVE_MIND_FORMAL_AI_IMAGE || '').trim();
+  if (pinned) return [{ image: pinned, source: FORMAL_AI_IMAGE_SOURCES.PINNED, pullable: true }];
+  const candidates = [{ image: `${FORMAL_AI_IMAGE_REPOSITORY}:${FORMAL_AI_BOOTSTRAP_VERSION}`, source: FORMAL_AI_IMAGE_SOURCES.PUBLISHED, pullable: true }];
+  const fallback = resolveFormalAiFallbackImage(env);
+  // Never pulled: the fallback's value is that it is already on the host. A
+  // deployment that has to pull it would be pulling the *bigger* image.
+  if (fallback && fallback !== candidates[0].image) candidates.push({ image: fallback, source: FORMAL_AI_IMAGE_SOURCES.HIVE_MIND, pullable: false });
+  return candidates;
+};
+
+/** The preferred image, kept for callers and logs that only need the headline reference. */
+export const resolveFormalAiSidecarImage = (env = process.env) => resolveFormalAiSidecarImageCandidates(env)[0].image;
+
+const inspectLocalImage = async (image, { run, timeoutMs }) => {
+  try {
+    const result = await run('docker', ['image', 'inspect', image, '--format', '{{.Id}}'], { encoding: 'utf8', timeout: timeoutMs });
+    return String(result?.stdout ?? '').trim() || null;
+  } catch {
+    return null;
+  }
+};
+
+/**
+ * Render the aggregated failure so an operator reading the Telegram reply or the
+ * bot log knows the cause and the fix without opening a shell.
+ */
+const describeFailure = attempts => {
+  const lines = ['No Formal AI image could be resolved, so the sidecar was not started.'];
+  for (const attempt of attempts) {
+    if (attempt.source === FORMAL_AI_IMAGE_SOURCES.HIVE_MIND && attempt.kind === 'absent') {
+      lines.push(`• ${attempt.image} (local Hive Mind image, fallback): not present on this host — Docker-isolated tasks would have to pull it too.`);
+      continue;
+    }
+    lines.push(`• ${attempt.image} (${attempt.source}): ${attempt.reason}${attempt.error ? ` — ${attempt.error}` : ''}`);
+    for (const step of attempt.remediation ?? []) lines.push(`    → ${step}`);
+  }
+  lines.push('Formal AI tasks fail closed by design (issue #2146): Hive Mind will not silently run them on another model.');
+  return lines.join('\n');
+};
+
+/**
+ * Return the first candidate image that is usable on this host, pulling only the
+ * pullable ones and never throwing for a candidate that simply is not there.
+ *
+ * @param {object} params
+ * @param {object} [params.env]
+ * @param {Function} [params.run] - `execFile`-shaped seam, so tests can drive a fake daemon.
+ * @param {number} [params.timeoutMs]
+ * @param {Function|null} [params.log]
+ * @param {boolean} [params.verbose]
+ * @param {boolean} [params.pull] - Set false to accept only images already on the host.
+ * @returns {Promise<{image: string, source: string, digest: string|null, pulled: boolean, attempts: object[]}>}
+ * @throws {Error} When no candidate resolves; the message names every attempt and its fix.
+ */
+export const ensureFormalAiSidecarImage = async ({ env = process.env, run = execFileAsync, timeoutMs = DEFAULT_DOCKER_TIMEOUT_MS, log = null, verbose = false, pull = true, candidates = resolveFormalAiSidecarImageCandidates(env) } = {}) => {
+  const attempts = [];
+
+  for (const candidate of candidates) {
+    const localDigest = await inspectLocalImage(candidate.image, { run, timeoutMs });
+    if (localDigest) {
+      if (verbose && log) await log(`[VERBOSE] formal-ai-image: using '${candidate.image}' (${candidate.source}) already present locally, digest=${localDigest}`);
+      return { image: candidate.image, source: candidate.source, digest: localDigest, pulled: false, attempts };
+    }
+
+    if (!candidate.pullable || !pull) {
+      attempts.push({ ...candidate, kind: 'absent', reason: 'not present locally and not pulled', error: null, remediation: [] });
+      continue;
+    }
+
+    if (log) await log(`⬇️ Pulling the Formal AI sidecar image ${candidate.image}`);
+    try {
+      await run('docker', ['pull', candidate.image], { encoding: 'utf8', timeout: timeoutMs });
+    } catch (error) {
+      const message = errorText(error);
+      const diagnosis = classifyDockerRegistryError(message);
+      attempts.push({ ...candidate, kind: diagnosis.kind, reason: diagnosis.reason, error: message, remediation: diagnosis.remediation });
+      if (log) await log(`⚠️ Could not pull ${candidate.image}: ${diagnosis.reason} (${diagnosis.kind}). ${message}`);
+      continue;
+    }
+
+    const digest = await inspectLocalImage(candidate.image, { run, timeoutMs });
+    if (verbose && log) await log(`[VERBOSE] formal-ai-image: pulled '${candidate.image}' (${candidate.source}), digest=${digest ?? 'unknown'}`);
+    return { image: candidate.image, source: candidate.source, digest, pulled: true, attempts };
+  }
+
+  const error = new Error(describeFailure(attempts));
+  error.formalAiImageAttempts = attempts;
+  throw error;
+};
+
+export default {
+  FORMAL_AI_IMAGE_REPOSITORY,
+  FORMAL_AI_IMAGE_SOURCES,
+  classifyDockerRegistryError,
+  ensureFormalAiSidecarImage,
+  resolveFormalAiFallbackImage,
+  resolveFormalAiSidecarImage,
+  resolveFormalAiSidecarImageCandidates,
+};

package/src/formal-ai-isolation.lib.mjs

@@ -32,7 +32,14 @@ export const acquireFormalAiSidecarForTask = async ({ backend, args = [], model
   try {
     return { sidecar: await acquire({ sessionId, tool, model, env, verbose, log }), error: null };
   } catch (error) {
-    return { sidecar: null, error: `Formal AI sidecar could not be started, so the task was not launched (issue #2146): ${error?.message || error}` };
+    const message = `Formal AI sidecar could not be started, so the task was not launched (issue #2146): ${error?.message || error}`;
+    // Issue #2154: this used to be returned and nothing else. The reply went to
+    // Telegram, the session was untracked, and the bot log showed only the
+    // untracking — so the operator could see that a task had vanished but never
+    // why. Every refusal is now on the record, with the session UUID that names
+    // the task in `$ --list` and in the session store.
+    console.error(`[formal-ai-isolation] Session ${sessionId}: ${message}`);
+    return { sidecar: null, error: message };
   }
 };
 

package/src/formal-ai-sidecar.lib.mjs

@@ -46,7 +46,8 @@ import fs from 'node:fs';
 import path from 'node:path';
 import { promisify } from 'node:util';
 
-import { FORMAL_AI_BOOTSTRAP_VERSION } from './formal-ai-version.lib.mjs';
+import { FORMAL_AI_MINIMUM_VERSION, isFormalAiVersionAtLeast } from './formal-ai-version.lib.mjs';
+import { ensureFormalAiSidecarImage, resolveFormalAiSidecarImage } from './formal-ai-image.lib.mjs';
 import { isFormalAiModel } from './formal-ai-model.lib.mjs';
 import { getModelFromArgs } from './model-args.lib.mjs';
 import { resolveBotStateDir } from './session-store.lib.mjs';
@@ -77,8 +78,14 @@ export const FORMAL_AI_SIDECAR_PORT = 8080;
 export const FORMAL_AI_MEMORY_MOUNT = '/home/box/.formal-ai';
 export const FORMAL_AI_MEMORY_PATH = `${FORMAL_AI_MEMORY_MOUNT}/memory.lino`;
 
-/** Image published by every Formal AI release (`:latest` plus the bare version). */
-export const FORMAL_AI_IMAGE_REPOSITORY = 'ghcr.io/link-assistant/formal-ai';
+/**
+ * Image published by every Formal AI release (`:latest` plus the bare version).
+ *
+ * Re-exported from `formal-ai-image.lib.mjs`, which owns image resolution since
+ * issue #2154 taught us that "which image" and "is it actually pullable" are the
+ * same question.
+ */
+export { FORMAL_AI_IMAGE_REPOSITORY, resolveFormalAiSidecarImage, resolveFormalAiSidecarImageCandidates } from './formal-ai-image.lib.mjs';
 
 /** Applied to the sidecar, its network and its volume so reconciliation can find them. */
 export const FORMAL_AI_SIDECAR_LABEL = 'com.link-assistant.hive-mind.formal-ai';
@@ -86,6 +93,9 @@ export const FORMAL_AI_SIDECAR_LABEL = 'com.link-assistant.hive-mind.formal-ai';
 const STATE_FILE_NAME = 'formal-ai-sidecar.json';
 const SIDECAR_LOCK_NAME = 'formal-ai-sidecar';
 const DEFAULT_DOCKER_TIMEOUT_MS = 120_000;
+// Pulling a sidecar image is the one Docker call that legitimately takes many
+// minutes, so it gets its own budget instead of the general command timeout.
+const DEFAULT_IMAGE_TIMEOUT_MS = 600_000;
 const DEFAULT_HEALTH_ATTEMPTS = 60;
 const DEFAULT_HEALTH_DELAY_MS = 1000;
 
@@ -107,9 +117,6 @@ const sleep = ms => new Promise(resolve => setTimeout(resolve, ms));
  */
 export const isFormalAiTask = ({ args = [], model = null } = {}) => isFormalAiModel(model || getModelFromArgs(args));
 
-/** Resolve the image the sidecar boots with. Operators may pin their own build. */
-export const resolveFormalAiSidecarImage = (env = process.env) => String(env.HIVE_MIND_FORMAL_AI_IMAGE || '').trim() || `${FORMAL_AI_IMAGE_REPOSITORY}:${FORMAL_AI_BOOTSTRAP_VERSION}`;
-
 /** Build the endpoint origin for a host name or address. */
 export const buildFormalAiSidecarBaseUrl = (host = FORMAL_AI_SIDECAR_NETWORK_ALIAS) => `http://${host}:${FORMAL_AI_SIDECAR_PORT}`;
 
@@ -433,18 +440,14 @@ export const stopFormalAiSidecar = async ({ env = process.env, fsImpl = fs, run
  * Must be called *before* the task container's command is allowed to run.
  * Returns the endpoint the task should be pointed at.
  */
-export const acquireFormalAiSidecar = async ({ sessionId, tool = null, model = null, env = process.env, fsImpl = fs, run = execFileAsync, timeoutMs, log = null, verbose = false, now = () => new Date(), healthAttempts, healthDelayMs, sleepImpl = sleep, lockOptions = {} } = {}) => {
+export const acquireFormalAiSidecar = async ({ sessionId, tool = null, model = null, env = process.env, fsImpl = fs, run = execFileAsync, timeoutMs, imageTimeoutMs = DEFAULT_IMAGE_TIMEOUT_MS, log = null, verbose = false, now = () => new Date(), healthAttempts, healthDelayMs, sleepImpl = sleep, lockOptions = {} } = {}) => {
   if (!sessionId) throw new Error('acquireFormalAiSidecar requires a sessionId');
 
   return withFormalAiSidecarLock(
     async () => {
-      const image = resolveFormalAiSidecarImage(env);
       const state = readFormalAiSidecarState({ env, fsImpl });
       const leases = await reconcileLeases(state.leases, { run, timeoutMs, log, verbose });
 
-      await ensureFormalAiNetwork({ run, timeoutMs, log, verbose });
-      await ensureFormalAiMemoryVolume({ image, run, timeoutMs, log, verbose });
-
       let container = await inspectDockerContainer(FORMAL_AI_SIDECAR_CONTAINER_NAME, { run, timeoutMs });
       if (container.exists && !container.running) {
         // A stopped container may predate an image change; recreate instead of
@@ -453,8 +456,18 @@ export const acquireFormalAiSidecar = async ({ sessionId, tool = null, model = n
         container = { exists: false, running: false, image: null, imageDigest: null };
       }
 
+      // Resolve the image *before* anything shells out with it. Until issue
+      // #2154 the reference went straight into `docker run`, so a registry that
+      // refused the pull surfaced as an unreadable `Command failed: docker run …`
+      // dump and the task died even though a usable image sat on the host.
+      const resolved = container.exists && container.image ? { image: container.image, source: 'running-sidecar', pulled: false } : await ensureFormalAiSidecarImage({ env, run, timeoutMs: imageTimeoutMs, log, verbose });
+      const image = resolved.image;
+
+      await ensureFormalAiNetwork({ run, timeoutMs, log, verbose });
+      await ensureFormalAiMemoryVolume({ image, run, timeoutMs, log, verbose });
+
       if (!container.exists) {
-        if (log) await log(`🧠 Starting the Formal AI sidecar (${image}) on the internal network '${FORMAL_AI_SIDECAR_NETWORK_NAME}'`);
+        if (log) await log(`🧠 Starting the Formal AI sidecar (${image}, ${resolved.source}) on the internal network '${FORMAL_AI_SIDECAR_NETWORK_NAME}'`);
         await dockerText(run, buildFormalAiSidecarRunArgs({ image, env }), { timeoutMs });
         container = await inspectDockerContainer(FORMAL_AI_SIDECAR_CONTAINER_NAME, { run, timeoutMs });
       }
@@ -466,12 +479,21 @@ export const acquireFormalAiSidecar = async ({ sessionId, tool = null, model = n
         throw new Error(`Formal AI sidecar '${FORMAL_AI_SIDECAR_CONTAINER_NAME}' did not become healthy: ${health.error}`);
       }
 
+      // The sidecar may now boot from any of several images (issue #2154), so
+      // the version floor is enforced against what the process actually reports
+      // rather than assumed from the tag. A too-old binary would answer /health
+      // and then fail on the agent-mode API the tasks depend on.
+      const reportedVersion = health.health?.version ?? null;
+      if (reportedVersion && !isFormalAiVersionAtLeast(reportedVersion, FORMAL_AI_MINIMUM_VERSION)) {
+        throw new Error(`Formal AI sidecar image ${image} (${resolved.source}) runs formal-ai ${reportedVersion}, but Hive Mind requires >= ${FORMAL_AI_MINIMUM_VERSION}. Rebuild or repin the image (HIVE_MIND_FORMAL_AI_IMAGE) before running Formal AI tasks.`);
+      }
+
       const address = await readFormalAiSidecarAddress({ run, timeoutMs });
       const acquiredAt = now().toISOString();
       const nextLeases = [...leases.filter(lease => lease.sessionId !== sessionId), { sessionId, tool, model, acquiredAt }];
       writeFormalAiSidecarState({ ...state, image: container.image || image, imageDigest: container.imageDigest, startedAt: state.startedAt || acquiredAt, leases: nextLeases }, { env, fsImpl });
 
-      if (verbose && log) await log(`[VERBOSE] formal-ai-sidecar: lease '${sessionId}' acquired (${nextLeases.length} active), image=${container.image || image}, digest=${container.imageDigest ?? 'unknown'}, address=${address ?? 'unknown'}, formal-ai=${health.health?.version ?? 'unknown'}, memory schema=${health.health?.memory?.schema_version ?? 'unknown'}`);
+      if (verbose && log) await log(`[VERBOSE] formal-ai-sidecar: lease '${sessionId}' acquired (${nextLeases.length} active), image=${container.image || image} (${resolved.source}), digest=${container.imageDigest ?? 'unknown'}, address=${address ?? 'unknown'}, formal-ai=${health.health?.version ?? 'unknown'}, memory schema=${health.health?.memory?.schema_version ?? 'unknown'}`);
 
       return {
         address,
@@ -482,6 +504,7 @@ export const acquireFormalAiSidecar = async ({ sessionId, tool = null, model = n
         containerName: FORMAL_AI_SIDECAR_CONTAINER_NAME,
         memoryVolume: FORMAL_AI_MEMORY_VOLUME_NAME,
         image: container.image || image,
+        imageSource: resolved.source,
         imageDigest: container.imageDigest,
         health: health.health,
         leaseCount: nextLeases.length,

package/src/formal-ai-updater.lib.mjs

@@ -33,6 +33,7 @@ import { execFile } from 'node:child_process';
 import fs from 'node:fs';
 import { promisify } from 'node:util';
 
+import { classifyDockerRegistryError } from './formal-ai-image.lib.mjs';
 import { FORMAL_AI_IMAGE_REPOSITORY, FORMAL_AI_MEMORY_MOUNT, FORMAL_AI_MEMORY_PATH, FORMAL_AI_MEMORY_VOLUME_NAME, FORMAL_AI_SIDECAR_CONTAINER_NAME, buildFormalAiSidecarRunArgs, ensureFormalAiMemoryVolume, ensureFormalAiNetwork, inspectDockerContainer, readDockerImageDigest, readFormalAiSidecarState, reconcileFormalAiSidecar, stopFormalAiSidecar, waitForFormalAiSidecarHealth, withFormalAiSidecarLock, writeFormalAiSidecarState } from './formal-ai-sidecar.lib.mjs';
 
 const execFileAsync = promisify(execFile);
@@ -207,8 +208,19 @@ export const updateFormalAiSidecarWhenIdle = async ({ env = process.env, fsImpl
         await dockerText(run, ['pull', '--quiet', image], { timeoutMs: pullTimeoutMs });
       } catch (error) {
         const message = error?.stderr?.toString?.().trim() || error?.message || String(error);
-        if (log) await log(`⚠️ Could not pull ${image}; keeping the current Formal AI image: ${message}`);
-        return { status: 'failed', stage: 'pull', image, error: message };
+        // Issue #2154: this warning fired every ~5 minutes for hours ("Could not
+        // pull … unauthorized") without ever saying that the registry was
+        // refusing us, and the operator only learned of it when a Formal AI task
+        // failed to start. A permanent refusal (unauthorized/denied/not-found)
+        // is a configuration fault, not a transient blip: name it, say how to
+        // fix it, and report the classification to the caller.
+        const classification = classifyDockerRegistryError(message);
+        const permanent = ['unauthorized', 'denied', 'not-found'].includes(classification.kind);
+        if (log) {
+          const remediation = classification.remediation.length ? ` Fix it by one of: ${classification.remediation.join('; ')}.` : '';
+          await log(`${permanent ? '🚨' : '⚠️'} Could not pull ${image} — ${classification.reason} (${classification.kind}); keeping the current Formal AI image: ${message}${permanent ? remediation : ''}`);
+        }
+        return { status: 'failed', stage: 'pull', image, error: message, classification: classification.kind, permanent, remediation: classification.remediation };
       }
 
       const pulledDigest = await readDockerImageDigest(image, { run, timeoutMs });

package/src/hive-mind-image.lib.mjs

@@ -0,0 +1,56 @@
+/**
+ * Resolve the Hive Mind container image references shared by the Docker
+ * isolation runner and the Formal AI sidecar.
+ *
+ * Extracted from `isolation-runner.lib.mjs` for issue #2154: the Formal AI
+ * sidecar needs the very same reference to fall back to the locally present
+ * Hive Mind image (which bakes `/usr/local/bin/formal-ai`) when the published
+ * Formal AI image cannot be pulled, and importing the isolation runner from the
+ * sidecar would create an import cycle.
+ *
+ * @see https://github.com/link-assistant/hive-mind/issues/2154
+ */
+
+export const HIVE_MIND_IMAGE_REPO = 'konard/hive-mind';
+export const HIVE_MIND_DIND_IMAGE_REPO = 'konard/hive-mind-dind';
+export const DEFAULT_HIVE_MIND_IMAGE_TAG = 'latest';
+
+/**
+ * Resolve the tag used for the Docker isolation image.
+ *
+ * Release Docker images bake this env var from `HIVE_MIND_VERSION`, so a parent
+ * container started via `:latest` still launches child isolation containers from
+ * the same immutable release tag. Local/PR builds fall back to `latest`, and
+ * operators can override the tag explicitly when using custom images. Pinning
+ * matters for Docker-in-Docker deployments: the nested daemon starts with an
+ * empty image store, so a `:latest` digest drift from the host copy forces a
+ * fresh multi-gigabyte pull. See issue #1879.
+ */
+export function resolveDockerIsolationImageTag({ env = process.env } = {}) {
+  const explicit = String(env.HIVE_MIND_DOCKER_ISOLATION_IMAGE_TAG || '').trim();
+  return explicit || DEFAULT_HIVE_MIND_IMAGE_TAG;
+}
+
+/**
+ * Pick the Docker image used for `--isolation docker`.
+ *
+ * start-command defaults its Docker backend to a base OS image. Hive Mind needs
+ * an image with the same CLI/tooling baseline as the parent process instead.
+ *
+ * `HIVE_MIND_DOCKER_ISOLATION_IMAGE` is a full override (repo:tag). Otherwise
+ * the repo is chosen by image variant and the tag by
+ * `resolveDockerIsolationImageTag()`.
+ */
+export function getDockerIsolationImage({ env = process.env } = {}) {
+  if (env.HIVE_MIND_DOCKER_ISOLATION_IMAGE) return env.HIVE_MIND_DOCKER_ISOLATION_IMAGE;
+  const repo = String(env.HIVE_MIND_IMAGE_VARIANT || '').toLowerCase() === 'dind' ? HIVE_MIND_DIND_IMAGE_REPO : HIVE_MIND_IMAGE_REPO;
+  return `${repo}:${resolveDockerIsolationImageTag({ env })}`;
+}
+
+export default {
+  DEFAULT_HIVE_MIND_IMAGE_TAG,
+  HIVE_MIND_DIND_IMAGE_REPO,
+  HIVE_MIND_IMAGE_REPO,
+  getDockerIsolationImage,
+  resolveDockerIsolationImageTag,
+};

package/src/isolation-runner.lib.mjs

@@ -21,6 +21,12 @@ import os from 'node:os';
 import path from 'node:path';
 import { isExecutingSessionStatus, isTerminalSessionStatus } from './session-status.lib.mjs';
 import { acquireFormalAiSidecarForTask, attachFormalAiTaskContainer, releaseFormalAiSidecarForTask } from './formal-ai-isolation.lib.mjs';
+// The image references live in their own module so the Formal AI sidecar can
+// resolve the locally present Hive Mind image (which bakes `formal-ai`) without
+// importing this runner and creating a cycle. Re-exported here because callers
+// and tests have always reached them through the isolation runner. See #2154.
+import { getDockerIsolationImage } from './hive-mind-image.lib.mjs';
+export { getDockerIsolationImage, resolveDockerIsolationImageTag } from './hive-mind-image.lib.mjs';
 let commandStreamDollarPromise = null;
 async function getCommandStreamDollar() {
   if (!commandStreamDollarPromise) {
@@ -43,9 +49,6 @@ async function getCommandStreamDollar() {
 export { isExecutingSessionStatus, isTerminalSessionStatus, isKilledSessionStatus } from './session-status.lib.mjs';
 // Valid isolation backends
 const VALID_ISOLATION_BACKENDS = ['screen', 'tmux', 'docker'];
-const HIVE_MIND_IMAGE_REPO = 'konard/hive-mind';
-const HIVE_MIND_DIND_IMAGE_REPO = 'konard/hive-mind-dind';
-const DEFAULT_HIVE_MIND_IMAGE_TAG = 'latest';
 const DOCKER_CONTAINER_HOME = '/home/box';
 const FORMAL_AI_COMPOSE_HOSTNAME = 'link-assistant-formal-ai';
 // Default path where the host Docker socket is bind-mounted inside a DinD container so box's host-image passthrough can copy host images into the nested daemon. Matches box's own DIND_HOST_DOCKER_SOCK default. The deploy must mount it (`-v /var/run/docker.sock:/var/run/host-docker.sock:ro`) or the nested daemon starts empty and the first isolated task pulls the full, multi-gigabyte image. See issue #1914.
@@ -96,36 +99,6 @@ function maybeAddMount(mounts, source, target, existsSync) {
   if (!existsSync(source)) return;
   mounts.push({ source, target });
 }
-/**
- * Resolve the tag used for the Docker isolation image.
- *
- * Release Docker images bake this env var from `HIVE_MIND_VERSION`, so a parent
- * container started via `:latest` still launches child isolation containers from
- * the same immutable release tag. Local/PR builds fall back to `latest`, and
- * operators can override the tag explicitly when using custom images. Pinning
- * matters for Docker-in-Docker deployments: the nested daemon starts with an
- * empty image store, so a `:latest` digest drift from the host copy forces a
- * fresh multi-gigabyte pull. See issue #1879.
- */
-export function resolveDockerIsolationImageTag({ env = process.env } = {}) {
-  const explicit = String(env.HIVE_MIND_DOCKER_ISOLATION_IMAGE_TAG || '').trim();
-  return explicit || DEFAULT_HIVE_MIND_IMAGE_TAG;
-}
-/**
- * Pick the Docker image used for `--isolation docker`.
- *
- * start-command defaults its Docker backend to a base OS image. Hive Mind needs
- * an image with the same CLI/tooling baseline as the parent process instead.
- *
- * `HIVE_MIND_DOCKER_ISOLATION_IMAGE` is a full override (repo:tag). Otherwise
- * the repo is chosen by image variant and the tag by
- * `resolveDockerIsolationImageTag()`.
- */
-export function getDockerIsolationImage({ env = process.env } = {}) {
-  if (env.HIVE_MIND_DOCKER_ISOLATION_IMAGE) return env.HIVE_MIND_DOCKER_ISOLATION_IMAGE;
-  const repo = String(env.HIVE_MIND_IMAGE_VARIANT || '').toLowerCase() === 'dind' ? HIVE_MIND_DIND_IMAGE_REPO : HIVE_MIND_IMAGE_REPO;
-  return `${repo}:${resolveDockerIsolationImageTag({ env })}`;
-}
 /**
  * Resolve the path where the host Docker socket is expected to be mounted inside
  * a DinD container. box's entrypoint reads this socket to copy host images into
@@ -296,6 +269,48 @@ async function runStartCommand(binPath, startCommandArgs) {
 export function generateSessionId() {
   return crypto.randomUUID();
 }
+const UUID_PATTERN = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;
+/**
+ * Extract start-command's own execution UUID from a launch banner.
+ *
+ * Issue #2154: an isolated task has two UUIDs. Hive Mind generates the session
+ * name and passes it as `--session` (it also becomes the container name);
+ * start-command mints a separate execution UUID and prints it as the `session`
+ * field of its launch banner:
+ *
+ * ```
+ * │ session   edc7b051-e12f-4f7b-b677-c885f3208407
+ * │ container 0a3627ef-f1f1-4801-a073-3678b9453db7
+ * ```
+ *
+ * `$ --list` shows the execution UUID, while Telegram and the logs showed the
+ * session UUID, so the two views could not be joined — which is why three
+ * refused tasks and two healthy ones looked equally unaccounted for. Returning
+ * it lets the caller record both.
+ *
+ * Only a well-formed UUID is returned; a banner we do not recognise yields
+ * null rather than a guess, because a wrong correlation is worse than none.
+ *
+ * @param {string} output - Raw stdout from the detached `$` launch
+ * @returns {string|null}
+ */
+export function parseStartCommandExecutionUuid(output) {
+  const raw = (output || '').trim();
+  if (!raw) return null;
+  try {
+    const parsed = JSON.parse(raw);
+    const data = Array.isArray(parsed) ? parsed[0] : parsed;
+    const uuid = data?.uuid || data?.session || null;
+    if (typeof uuid === 'string' && UUID_PATTERN.test(uuid.trim())) return uuid.trim();
+  } catch {
+    // Human-readable banner — fall through.
+  }
+  // The banner is box-drawn (`│ session   <uuid>`); tolerate the prefix, an
+  // ASCII `|`, or no prefix at all.
+  const match = raw.match(/^[\s│|]*session\s+([^\s]+)\s*$/im);
+  const candidate = match?.[1]?.trim();
+  return candidate && UUID_PATTERN.test(candidate) ? candidate : null;
+}
 /**
  * Parse output from `$ --status <session>`.
  *
@@ -544,23 +559,22 @@ async function logDockerIsolationPostLaunchDiagnostics(sessionId, env = process.
 export async function executeWithIsolation(command, args, options = {}) {
   const { backend, verbose = false } = options;
   const sessionId = options.sessionId || generateSessionId();
+  // Issue #2154: a launch that never produced a container left no trace in the
+  // bot log — the reply went to Telegram and the log jumped straight to
+  // "session untracked", so an operator could see that tasks were disappearing
+  // but not why, and could not even name them. Every unsuccessful return from
+  // this function now records the session UUID (the same one `$ --list` and the
+  // session store use) together with the reason.
+  const failLaunch = (error, extra = {}) => {
+    console.error(`[isolation-runner] Session ${sessionId} was not launched (backend=${backend}, tool=${options.tool ?? 'claude'}, model=${options.model ?? 'default'}): ${error}`);
+    return { success: false, sessionId, output: '', error, ...extra };
+  };
   if (!VALID_ISOLATION_BACKENDS.includes(backend)) {
-    return {
-      success: false,
-      sessionId,
-      output: '',
-      error: `Invalid isolation backend: '${backend}'. Must be one of: ${VALID_ISOLATION_BACKENDS.join(', ')}`,
-    };
+    return failLaunch(`Invalid isolation backend: '${backend}'. Must be one of: ${VALID_ISOLATION_BACKENDS.join(', ')}`);
   }
   const binPath = await findStartCommandBinary();
   if (!binPath) {
-    return {
-      success: false,
-      sessionId,
-      output: '',
-      warning: '⚠️ WARNING: start-command ($) not found in PATH\nPlease install: npm install -g start-command',
-      error: 'start-command ($) not found',
-    };
+    return failLaunch('start-command ($) not found', { warning: '⚠️ WARNING: start-command ($) not found in PATH\nPlease install: npm install -g start-command' });
   }
   if (verbose) {
     console.log(`[VERBOSE] isolation-runner: Using $ binary at: ${binPath}`);
@@ -573,7 +587,7 @@ export async function executeWithIsolation(command, args, options = {}) {
   // fails. Fail closed — a Formal AI task must never start without Formal AI.
   const hostEnv = options.env || process.env;
   const { sidecar, error: sidecarError } = await acquireFormalAiSidecarForTask({ backend, args, model: options.model ?? null, tool: options.tool ?? null, sessionId, env: hostEnv, verbose });
-  if (sidecarError) return { success: false, sessionId, output: '', error: sidecarError };
+  if (sidecarError) return failLaunch(sidecarError);
   const taskEnv = sidecar ? { ...hostEnv, HIVE_MIND_FORMAL_AI_BASE_URL: sidecar.baseUrl } : hostEnv;
   const effectiveOptions =
     backend === 'docker'
@@ -628,7 +642,7 @@ export async function executeWithIsolation(command, args, options = {}) {
     if (formalAiAttachError) await removeDockerContainer(sessionId, verbose);
     await releaseFormalAiSidecarForTask({ sidecar, sessionId, env: hostEnv, verbose });
     if (formalAiAttachError) {
-      return { success: false, sessionId, output: result.output, error: `Formal AI task container could not be attached to the internal Formal AI network, so the task was stopped instead of falling back to another model (issue #2146): ${formalAiAttachError}` };
+      return failLaunch(`Formal AI task container could not be attached to the internal Formal AI network, so the task was stopped instead of falling back to another model (issue #2146): ${formalAiAttachError}`, { output: result.output });
     }
   }
   // Issue #1939: capture the freshly-launched docker session's reported status
@@ -639,19 +653,22 @@ export async function executeWithIsolation(command, args, options = {}) {
     await logDockerIsolationPostLaunchDiagnostics(sessionId, options.env || process.env);
   }
   if (result.success) {
+    // Issue #2154: hand the caller start-command's own execution UUID as well.
+    // It is the identifier `$ --list` prints, so without it the bot and the
+    // session list cannot be joined by an operator.
+    const executionUuid = parseStartCommandExecutionUuid(result.output);
+    if (verbose) {
+      console.log(executionUuid ? `[VERBOSE] isolation-runner: start-command execution UUID for session ${sessionId}: ${executionUuid} (this is what '$ --list' shows)` : `[VERBOSE] isolation-runner: start-command reported no execution UUID for session ${sessionId}; '$ --list' cannot be correlated for this session`);
+    }
     return {
       success: true,
       sessionId,
+      executionUuid,
       output: result.output,
       containerFilesystemStartBytes,
     };
   }
-  return {
-    success: false,
-    sessionId,
-    output: result.output,
-    error: result.error,
-  };
+  return failLaunch(result.error, { output: result.output });
 }
 /**
  * Query the status of an isolated session via `$ --status <uuid>`

package/src/locales/en.lino

@@ -654,6 +654,7 @@ en
         runner_also_failed "The runner also failed; its exit code is preserved for investigation."
         killed "Work session {{reason}}{{exitSuffix}}"
         stopped "Work session stopped by user{{requestedBy}}{{exitSuffix}}"
+        not_launched "The work session was not launched, so it has no log and is not listed by `--list`."
     duration
       label "Duration"
     session
@@ -670,6 +671,8 @@ en
         resumed_attempt "🔄 A new working session was started to recover from this kill (attempt {{attempt}}): {{sessionId}}"
     isolation
       label "Isolation"
+    execution
+      label "Execution"
     error
       executing
         command "❌ Error executing {{commandName}} command"

package/src/locales/hi.lino

@@ -654,6 +654,7 @@ hi
         runner_also_failed "रनर भी विफल हुआ; जाँच के लिए उसका exit code सुरक्षित रखा गया है।"
         killed "कार्य सत्र रोका गया: {{reason}}{{exitSuffix}}"
         stopped "कार्य सत्र उपयोगकर्ता द्वारा रोका गया{{requestedBy}}{{exitSuffix}}"
+        not_launched "कार्य सत्र शुरू ही नहीं हुआ, इसलिए उसका कोई log नहीं है और वह `--list` में नहीं दिखता।"
     duration
       label "अवधि"
     session
@@ -670,6 +671,8 @@ hi
         resumed_attempt "🔄 इस समाप्ति से पुनर्प्राप्ति हेतु नया कार्य सत्र शुरू किया गया (प्रयास {{attempt}}): {{sessionId}}"
     isolation
       label "Isolation"
+    execution
+      label "निष्पादन"
     error
       executing
         command "❌ {{commandName}} command चलाने में त्रुटि"

package/src/locales/ru.lino

@@ -654,6 +654,7 @@ ru
         runner_also_failed "Средство запуска также завершилось с ошибкой; код выхода сохранён для расследования."
         killed "Рабочий сеанс остановлен: {{reason}}{{exitSuffix}}"
         stopped "Рабочий сеанс остановлен пользователем{{requestedBy}}{{exitSuffix}}"
+        not_launched "Рабочий сеанс не был запущен, поэтому у него нет журнала и он не отображается в `--list`."
     duration
       label "Длительность"
     session
@@ -670,6 +671,8 @@ ru
         resumed_attempt "🔄 Запущена новая рабочая сессия для восстановления после этого завершения (попытка {{attempt}}): {{sessionId}}"
     isolation
       label "Изоляция"
+    execution
+      label "Запуск"
     error
       executing
         command "❌ Ошибка выполнения команды {{commandName}}"

package/src/locales/zh.lino

@@ -654,6 +654,7 @@ zh
         runner_also_failed "运行器也失败了；其退出代码已保留以供调查。"
         killed "工作会话已终止：{{reason}}{{exitSuffix}}"
         stopped "工作会话已由用户停止{{requestedBy}}{{exitSuffix}}"
+        not_launched "工作会话未启动，因此没有日志，也不会出现在 `--list` 中。"
     duration
       label "耗时"
     session
@@ -670,6 +671,8 @@ zh
         resumed_attempt "🔄 已启动新的工作会话以从此次终止中恢复（第 {{attempt}} 次尝试）：{{sessionId}}"
     isolation
       label "隔离"
+    execution
+      label "执行"
     error
       executing
         command "❌ 执行 {{commandName}} 命令时出错"

package/src/session-monitor.lib.mjs

@@ -164,6 +164,9 @@ export function trackSession(sessionName, sessionInfo, verbose = false) {
     url: sessionInfo.url || null,
     command: sessionInfo.command || null,
     sessionId: sessionInfo.sessionId || null,
+    // Issue #2154: `$ --list` prints start-command's execution UUID, not the
+    // session name. Logging both is what makes the two views joinable.
+    executionUuid: sessionInfo.executionUuid || null,
     startTime: sessionInfo.startTime instanceof Date ? sessionInfo.startTime.toISOString() : sessionInfo.startTime || null,
   });
 }
@@ -230,24 +233,40 @@ export function markSessionStopRequested(sessionId, { requestedBy = null, verbos
  * map and the durable store without emitting a `session_completed` audit event —
  * the session never ran, so it has no exit code to record (issue #1946).
  *
+ * Issue #2154: the durable structured log recorded only
+ * `session_untracked {"sessionName":…}`. Why the session disappeared a second
+ * after it was announced lived on untimestamped console lines, so an incident
+ * could not be reconstructed from the timestamped log alone. Callers now pass
+ * the reason and it is recorded with the event.
+ *
  * @param {string} sessionName - Name/UUID of the session to drop
  * @param {boolean} verbose - Whether to log verbose output
+ * @param {object} [details] - Extra context recorded with the `session_untracked` event
+ * @param {string} [details.reason] - Why the session was dropped (e.g. the launch error)
  */
-export function untrackSession(sessionName, verbose = false) {
+export function untrackSession(sessionName, verbose = false, details = {}) {
   if (!sessionName) return;
   const sessionInfo = activeSessions.get(sessionName) || null;
   const existed = activeSessions.delete(sessionName);
+  const reason = typeof details?.reason === 'string' && details.reason.trim() ? details.reason.trim() : null;
   if (verbose && existed) {
-    console.log(`[VERBOSE] Session ${sessionName} untracked (launch failed before it started)`);
+    console.log(`[VERBOSE] Session ${sessionName} untracked (launch failed before it started)${reason ? `: ${reason}` : ''}`);
   }
   if (sessionStore && isPersistableSession(sessionInfo)) {
     try {
-      sessionStore.remove(sessionName, { status: 'launch-failed', exitCode: null });
+      sessionStore.remove(sessionName, { status: 'launch-failed', exitCode: null, reason });
     } catch (error) {
       console.error(`[session-monitor] Could not remove untracked session ${sessionName}: ${error.message}`);
     }
   }
-  logEvent('session_untracked', { sessionName });
+  logEvent('session_untracked', {
+    sessionName,
+    reason,
+    url: sessionInfo?.url || null,
+    command: sessionInfo?.command || null,
+    tool: sessionInfo?.tool || null,
+    isolationBackend: sessionInfo?.isolationBackend || null,
+  });
 }
 /**
  * Get the number of active sessions being tracked
@@ -646,6 +665,18 @@ export async function monitorSessions(bot, verbose = false, options = {}) {
         sessionInfo.logPath = statusResult.logPath;
         persistSessionSnapshot(sessionName, sessionInfo);
       }
+      // Issue #2154: the same status record carries start-command's *execution*
+      // UUID — the only identifier `$ --list` prints. A session launched before
+      // this fix (or by a start-command whose banner we could not parse) has
+      // none, so backfill it here; otherwise that session stays impossible to
+      // find in the session list for its whole lifetime.
+      if (statusResult?.uuid && sessionInfo.executionUuid !== statusResult.uuid) {
+        if (verbose) {
+          console.log(`[VERBOSE] Session ${sessionName}: recorded start-command execution UUID ${statusResult.uuid} (this is what '$ --list' shows)`);
+        }
+        sessionInfo.executionUuid = statusResult.uuid;
+        persistSessionSnapshot(sessionName, sessionInfo);
+      }
     } else {
       // Issue #1586: Non-isolation screen sessions cannot reliably detect
       // completion because start-screen keeps the screen alive via `exec bash`.

package/src/session-store.lib.mjs

@@ -35,7 +35,10 @@ import path from 'node:path';
 // with its exact original invocation plus `--resume <lastSessionId>`.
 // `commandAlias` (#2109) preserves the Telegram spelling (`solve`, `codex`,
 // `claude`, etc.) so a bot notification never suggests a terminal-only command.
-const PERSISTABLE_FIELDS = ['chatId', 'messageId', 'startTime', 'url', 'command', 'commandAlias', 'isolationBackend', 'sessionId', 'containerFilesystemStartBytes', 'containerFilesystemLastBytes', 'containerFilesystemLastObservedAt', 'tool', 'infoBlock', 'urlContext', 'requesterUserId', 'showLimits', 'locale', 'logPath', 'args'];
+// `executionUuid` (#2154) is start-command's own identifier for the execution —
+// the one `$ --list` prints. It differs from `sessionId`, so persisting it is
+// what lets a restarted bot still correlate its sessions with the session list.
+const PERSISTABLE_FIELDS = ['chatId', 'messageId', 'startTime', 'url', 'command', 'commandAlias', 'isolationBackend', 'sessionId', 'executionUuid', 'containerFilesystemStartBytes', 'containerFilesystemLastBytes', 'containerFilesystemLastObservedAt', 'tool', 'infoBlock', 'urlContext', 'requesterUserId', 'showLimits', 'locale', 'logPath', 'args'];
 
 /**
  * Resolve the directory durable bot state is written to. Honors
@@ -213,7 +216,9 @@ export function createSessionStore(options = {}) {
         delete sessions[sessionName];
         writeSnapshotMap(sessions);
       }
-      appendEvent('complete', sessionName, { status: meta.status ?? null, exitCode: meta.exitCode ?? null });
+      // Issue #2154: carry the reason (when the caller knows it) so the durable
+      // event log says why a session ended, not just that it did.
+      appendEvent('complete', sessionName, { status: meta.status ?? null, exitCode: meta.exitCode ?? null, reason: meta.reason ?? null });
       log('debug', `Removed session ${sessionName} from snapshot`, meta);
     },
 

package/src/telegram-command-execution.lib.mjs

@@ -2,7 +2,7 @@ import { spawn } from 'child_process';
 import { describeChildExit } from './child-exit.lib.mjs';
 import { promisify } from 'util';
 import { exec as execCallback } from 'child_process';
-import { t } from './i18n.lib.mjs';
+import { formatFailedLaunchMessage as defaultFormatFailedLaunchMessage } from './work-session-formatting.lib.mjs';
 
 const exec = promisify(execCallback);
 
@@ -101,7 +101,7 @@ function executeWithCommand(startScreenCmd, command, args, verbose = false) {
  * @returns {Function} executeAndUpdateMessage(ctx, startingMessage, commandName, args, infoBlock, perCommandIsolation, tool, urlContext, sessionExtras)
  */
 export function buildExecuteAndUpdateMessage(deps) {
-  const { resolveIsolation, ISOLATION_BACKEND, isolationRunner, VERBOSE, executeStartScreen, trackSession, untrackSession, AUTO_WATCH_MESSAGE, startAutoTerminalWatchForSession, bot, formatExecutingWorkSessionMessage, formatStartingWorkSessionMessage } = deps;
+  const { resolveIsolation, ISOLATION_BACKEND, isolationRunner, VERBOSE, executeStartScreen, trackSession, untrackSession, AUTO_WATCH_MESSAGE, startAutoTerminalWatchForSession, bot, formatExecutingWorkSessionMessage, formatStartingWorkSessionMessage, formatFailedLaunchMessage = defaultFormatFailedLaunchMessage } = deps;
   return async function executeAndUpdateMessage(ctx, startingMessage, commandName, args, infoBlock, perCommandIsolation = null, tool = 'claude', urlContext = null, { showLimits = false, limitsAtStart = null, locale = null, commandAlias = null } = {}) {
     const { chat, message_id: msgId } = startingMessage;
     const safeEdit = async text => {
@@ -130,14 +130,27 @@ export function buildExecuteAndUpdateMessage(deps) {
       trackSession(session, sessionInfo, VERBOSE);
       await safeEdit(formatStartingWorkSessionMessage({ sessionName: session, isolationBackend: iso.backend, infoBlock, locale }));
       result = await iso.runner.executeWithIsolation(commandName, args, { backend: iso.backend, sessionId: session, tool, verbose: VERBOSE });
-      if (result.success && sessionInfo && Number.isFinite(result.containerFilesystemStartBytes)) {
-        sessionInfo.containerFilesystemStartBytes = result.containerFilesystemStartBytes;
+      if (result.success && sessionInfo && (Number.isFinite(result.containerFilesystemStartBytes) || result.executionUuid)) {
+        if (Number.isFinite(result.containerFilesystemStartBytes)) sessionInfo.containerFilesystemStartBytes = result.containerFilesystemStartBytes;
+        // Issue #2154: `$ --list` identifies executions by start-command's own
+        // UUID, not by the session name the bot shows. Keep both on the session
+        // so the two views can be joined — in the reply, in the structured log
+        // and in the durable snapshot after a restart.
+        if (result.executionUuid) sessionInfo.executionUuid = result.executionUuid;
         trackSession(session, sessionInfo, VERBOSE);
       }
       if (!result.success) {
         // The launch never produced a live container — drop the optimistic
         // tracking so a phantom session is not monitored or resumed.
-        if (typeof untrackSession === 'function') untrackSession(session, VERBOSE);
+        // Issue #2154: the untracking used to be the *only* trace of the
+        // failure anywhere outside Telegram. Record the UUID, the backend and
+        // the reason first, so the bot log explains why a session that was
+        // announced a second ago is gone and absent from `--list`.
+        const launchError = result.error || result.output || 'unknown error';
+        console.error(`[telegram-bot] ${commandName} session ${session} was not launched (isolation=${iso.backend}, tool=${tool}): ${launchError}`);
+        // The reason also goes into the structured `session_untracked` event, so
+        // the timestamped log explains the disappearance on its own.
+        if (typeof untrackSession === 'function') untrackSession(session, VERBOSE, { reason: launchError });
         sessionInfo = undefined;
       }
     } else {
@@ -152,9 +165,15 @@ export function buildExecuteAndUpdateMessage(deps) {
     }
     if (result.warning) return safeEdit(`⚠️  ${result.warning}`);
     if (result.success) {
-      await safeEdit(formatExecutingWorkSessionMessage({ sessionName: session, isolationBackend: iso?.backend || null, infoBlock, locale }));
+      await safeEdit(formatExecutingWorkSessionMessage({ sessionName: session, executionUuid: result.executionUuid || null, isolationBackend: iso?.backend || null, infoBlock, locale }));
       if (AUTO_WATCH_MESSAGE && commandName === 'solve' && sessionInfo?.isolationBackend) await startAutoTerminalWatchForSession({ bot, ctx, sessionId: session, sessionInfo, verbose: VERBOSE });
-    } else await safeEdit(`${t('telegram.error_executing_command', { commandName }, { locale })}:\n\n\`\`\`\n${result.error || result.output}\n\`\`\`\n\n${infoBlock}`);
+    } else {
+      // Issue #2154: keep the session UUID in the failure reply. It is the only
+      // handle the operator has on the attempt, and this edit replaces the
+      // "🔄 Starting..." message that used to carry it.
+      if (!iso) console.error(`[telegram-bot] ${commandName} command failed to start (no isolation, tool=${tool}): ${result.error || result.output || 'unknown error'}`);
+      await safeEdit(formatFailedLaunchMessage({ commandName, sessionName: iso ? session : null, isolationBackend: iso?.backend || null, infoBlock, error: result.error || result.output, locale }));
+    }
   };
 }
 

package/src/telegram-solve-queue.lib.mjs

@@ -15,7 +15,7 @@ import { collectExecutingItems, formatDuration, formatQueueToolSection, formatWa
 export { QUEUE_CONFIG, THRESHOLD_STRATEGIES } from './queue-config.lib.mjs';
 import { QUEUE_CONFIG } from './queue-config.lib.mjs';
 import { reserveStartSlotForQueue } from './queue-start-reservation.lib.mjs';
-import { formatExecutingWorkSessionMessage, formatStartingWorkSessionMessage } from './work-session-formatting.lib.mjs';
+import { formatExecutingWorkSessionMessage, formatFailedLaunchMessage, formatStartingWorkSessionMessage } from './work-session-formatting.lib.mjs';
 import { t } from './i18n.lib.mjs';
 import { lt } from './limits-i18n.lib.mjs';
 export const QueueItemStatus = {
@@ -1070,9 +1070,24 @@ export class SolveQueue {
         // This was a bug where the final message update never happened because messageInfo was null
         // See: https://github.com/link-assistant/hive-mind/issues/1062
         const savedMessageInfo = item.messageInfo;
-        // Update to Started status (terminal - forgets message tracking)
-        item.setStarted(sessionName);
-        this.stats.totalCompleted++;
+        // Issue #2154: a launch that never produced a container was still marked
+        // STARTED, counted in `totalCompleted`, pushed onto `completed` and
+        // logged as `Finished: […] (started)`. Three Formal AI tasks that never
+        // ran therefore appear in the bot log as successful starts — a false
+        // positive that hid the incident and contradicted `$ --list`, which had
+        // no such sessions. A refused launch is a failure of the queue item.
+        const launchFailed = Boolean(result) && result.success === false;
+        if (launchFailed) {
+          item.setFailed(result.error || result.output || 'the launch was refused before the container started');
+          item.sessionName = sessionName;
+          item.messageInfo = null; // terminal status — stop tracking the message
+          this.stats.totalFailed++;
+          console.error(`[solve_queue] Item ${item.id} was not launched (session ${sessionName}): ${item.error}`);
+        } else {
+          // Update to Started status (terminal - forgets message tracking)
+          item.setStarted(sessionName);
+          this.stats.totalCompleted++;
+        }
         // Final message update using saved messageInfo
         if (item.ctx && result && savedMessageInfo) {
           const { chatId, messageId } = savedMessageInfo;
@@ -1089,7 +1104,17 @@ export class SolveQueue {
                 });
                 await item.ctx.telegram.editMessageText(chatId, messageId, undefined, response, { parse_mode: 'Markdown' });
               } else {
-                const response = `${t('telegram.error_executing_command', { commandName: 'solve' }, { locale: item.locale })}:\n\n\`\`\`\n${result.error || result.output}\n\`\`\`\n\n${item.infoBlock}`;
+                // Issue #2154: a queued /solve that fails to launch reports the
+                // same way as a direct one — with its session UUID and a note
+                // that nothing was started, instead of a bare error dump.
+                const response = formatFailedLaunchMessage({
+                  commandName: 'solve',
+                  sessionName: sessionName === 'unknown' ? null : sessionName,
+                  isolationBackend: result.isolationBackend || null,
+                  infoBlock: item.infoBlock,
+                  error: result.error || result.output,
+                  locale: item.locale,
+                });
                 await item.ctx.telegram.editMessageText(chatId, messageId, undefined, response, { parse_mode: 'Markdown' });
               }
             } catch (error) {

package/src/work-session-formatting.lib.mjs

@@ -80,12 +80,58 @@ export function formatStartingWorkSessionMessage({ sessionName = null, isolation
   return `${header}\n\n📊 ${sessionLabel}: \`${sessionName}\`${isolationInfo}${details}`;
 }
 
-export function formatExecutingWorkSessionMessage({ sessionName = 'unknown', isolationBackend = null, infoBlock = '', locale = null } = {}) {
+export function formatExecutingWorkSessionMessage({ sessionName = 'unknown', executionUuid = null, isolationBackend = null, infoBlock = '', locale = null } = {}) {
   const sessionLabel = text(locale, 'telegram.session_label', 'Session');
   const isolationLabel = text(locale, 'telegram.isolation_label', 'Isolation');
   const isolationInfo = isolationBackend ? `\n🔒 ${isolationLabel}: \`${isolationBackend}\`` : '';
   const details = infoBlock ? `\n\n${infoBlock}` : '';
-  return `${text(locale, 'telegram.work_session_executing', '⏳ Executing...')}\n\n📊 ${sessionLabel}: \`${sessionName}\`${isolationInfo}${details}`;
+  // Issue #2154: `$ --list` identifies an execution by start-command's own UUID,
+  // which is *not* the session name shown above it. Printing only one of the two
+  // left the operator unable to match a running task to the session list. The
+  // line is omitted entirely when start-command reported no UUID, so a session
+  // never carries an empty label.
+  const executionLabel = text(locale, 'telegram.execution_label', 'Execution');
+  const executionInfo = executionUuid ? `\n🆔 ${executionLabel}: \`${executionUuid}\`` : '';
+  return `${text(locale, 'telegram.work_session_executing', '⏳ Executing...')}\n\n📊 ${sessionLabel}: \`${sessionName}\`${executionInfo}${isolationInfo}${details}`;
+}
+
+/**
+ * Render the reply for a work session that never started (issue #2154).
+ *
+ * The previous reply was the raw runner error in a code fence and nothing else.
+ * Two things were missing, and both were reported as separate symptoms of the
+ * same incident:
+ *
+ *   - **The session UUID.** It was generated before the launch and shown in the
+ *     "🔄 Starting..." message, but the failure reply *overwrote* that message,
+ *     so the only identifier the task ever had was destroyed by its own error
+ *     report. Nothing then connected the Telegram thread to the bot log lines,
+ *     to the session store, or to a `--log <uuid>` lookup.
+ *   - **Why the task is missing from `--list`.** A failed launch produces no
+ *     container, so the session is untracked and never appears in the listing.
+ *     Without saying so, the reply reads as if the task were running somewhere.
+ *
+ * @param {Object} params
+ * @param {string} [params.commandName] - Command the user invoked (`solve`, `hive`, …)
+ * @param {string|null} [params.sessionName] - Session UUID, when one was generated
+ * @param {string|null} [params.isolationBackend]
+ * @param {string} [params.infoBlock]
+ * @param {string} [params.error] - Runner error text
+ * @param {string|null} [params.locale]
+ * @returns {string} Markdown reply
+ *
+ * @see https://github.com/link-assistant/hive-mind/issues/2154
+ */
+export function formatFailedLaunchMessage({ commandName = 'command', sessionName = null, isolationBackend = null, infoBlock = '', error = '', locale = null } = {}) {
+  const header = text(locale, 'telegram.error_executing_command', `❌ Error executing ${commandName} command`, { commandName });
+  const sessionLabel = text(locale, 'telegram.session_label', 'Session');
+  const isolationLabel = text(locale, 'telegram.isolation_label', 'Isolation');
+  const sessionLine = sessionName ? `\n📊 ${sessionLabel}: \`${sessionName}\`` : '';
+  const isolationLine = isolationBackend ? `\n🔒 ${isolationLabel}: \`${isolationBackend}\`` : '';
+  const body = String(error ?? '').trim() || 'unknown error';
+  const notLaunched = sessionName ? `\n\n${text(locale, 'telegram.work_session_not_launched', 'The work session was not launched, so it has no log and is not listed by `--list`.')}` : '';
+  const details = infoBlock ? `\n\n${infoBlock}` : '';
+  return `${header}:${sessionLine}${isolationLine}\n\n\`\`\`\n${body}\n\`\`\`${notLaunched}${details}`;
 }
 
 /**
@@ -176,6 +222,13 @@ export function formatSessionCompletionMessage({ sessionName, sessionInfo, statu
   const sessionLabel = text(messageLocale, 'telegram.session_label', 'Session');
   const isolationLabel = text(messageLocale, 'telegram.isolation_label', 'Isolation');
   const isolationInfo = sessionInfo?.isolationBackend ? `\n🔒 ${isolationLabel}: \`${sessionInfo.isolationBackend}\`` : '';
+  // Issue #2154: the completion reply is the last word the Telegram thread has
+  // on a task, and the handle an operator uses afterwards to fetch its log. Keep
+  // start-command's execution UUID (the one `$ --list` prints) next to the
+  // session UUID, so the finished task can still be found in the session list.
+  const executionLabel = text(messageLocale, 'telegram.execution_label', 'Execution');
+  const executionUuid = sessionInfo?.executionUuid || statusResult?.uuid || null;
+  const executionInfo = executionUuid ? `\n🆔 ${executionLabel}: \`${executionUuid}\`` : '';
   const startTime = parseDateValue(statusResult?.startTime) || parseDateValue(sessionInfo?.startTime) || observedEndTime;
   const endTime = parseDateValue(statusResult?.endTime) || observedEndTime;
   const durationSeconds = Math.max(0, (endTime.getTime() - startTime.getTime()) / 1000);
@@ -188,7 +241,7 @@ export function formatSessionCompletionMessage({ sessionName, sessionInfo, statu
   const statusEmoji = statusEmojiOverride || (failed ? '❌' : '✅');
   let message = `${statusEmoji} *${statusText}*\n\n`;
   message += `⏱️ ${durationLabel}: ${formatSessionDurationSeconds(durationSeconds)}\n`;
-  message += `📊 ${sessionLabel}: \`${sessionName || 'unknown'}\`${isolationInfo}${details}`;
+  message += `📊 ${sessionLabel}: \`${sessionName || 'unknown'}\`${executionInfo}${isolationInfo}${details}`;
 
   // Issue #594: --show-limits virtual option appends snapshot/delta sections
   // (Markdown code blocks) below the standard completion details.