Isolation resolution stuck in stale_cleaned loop for convers

Error message

Isolation resolution stuck in stale_cleaned loop for conversation ${conversation.id}

What it means

validateAndResolveIsolation detects that retrying isolation resolution after a stale_cleaned environment still produced a stale-cleaned state, meaning the loop guard fired. The engine retried once with a cleared isolation_env_id and still could not obtain a valid isolation environment, so it throws rather than recursing forever.

Source

Thrown at packages/core/src/orchestrator/orchestrator.ts:245

      // Clear stale reference
      await db.updateConversation(conversation.id, { isolation_env_id: null }).catch(e => {
        if (!(toError(e) instanceof ConversationNotFoundError)) {
          getLog().error(
            { err: toError(e), conversationId: conversation.id },
            'stale_isolation_clear_failed'
          );
        }
      });
      const staleMsg = codebase
        ? 'Detected a stale isolated workspace reference and cleared it. Creating a new isolated workspace now.'
        : 'Detected a stale isolated workspace reference and cleared it. Continuing without an isolated workspace.';
      await platform.sendMessage(conversationId, staleMsg).catch(e => {
        getLog().error({ err: toError(e), conversationId }, 'stale_isolation_notice_failed');
      });
      // Retry without existing env (guard against infinite recursion)
      if (!codebase) return { status: 'none', cwd: conversation.cwd ?? '/workspace', env: null };
      if (_isRetry) {
        throw new Error(
          `Isolation resolution stuck in stale_cleaned loop for conversation ${conversation.id}`
        );
      }
      return validateAndResolveIsolation(
        { ...conversation, isolation_env_id: null },
        codebase,
        platform,
        conversationId,
        hints,
        true,
        userId
      );
    }

    case 'none':
      return { status: 'none', cwd: result.cwd, env: null };

    case 'blocked':

View on GitHub (pinned to 0773b97458)

Solutions

Inspect the isolation environment store/records for conversation ${conversation.id} and purge stale entries manually.
Check for concurrent cleanup jobs racing with run dispatch and pause them.
Clear the conversation's isolation_env_id and retry the dispatch fresh.
Verify the isolation provider is healthy and returns non-stale environments.

Example fix

// before (persistence)
conversation.isolation_env_id = "env-stale-123"; // repeatedly resolves to stale_cleaned
// after (repair)
await db.updateConversation(conversation.id, { isolation_env_id: null });
// then re-dispatch

Defensive patterns

Strategy: retry

Validate before calling

const env = await getIsolationEnv(conversation.isolation_env_id);
if (env?.status === 'stale_cleaned') await clearIsolationEnvId(conversation.id);

Try / catch

try {
  await dispatch(conversationId);
} catch (e) {
  if (e.message.includes('stale_cleaned loop')) {
    await purgeStaleIsolationRecords(conversationId);
    await clearIsolationEnvId(conversationId);
    await dispatch(conversationId); // one manual retry after repair
  } else throw e;
}

Prevention

Purge stale isolation env records instead of leaving them attached to conversations.
Avoid running cleanup jobs concurrently with run dispatch.
Monitor the isolation provider health so stale statuses are caught early.

When it happens

Trigger: A conversation carries an isolation_env_id whose environment is reported stale/cleaned; the retry path clears the id and re-resolves, but validation still fails with stale_cleaned and _isRetry is true — e.g. the environment store keeps returning a stale status for the freshly resolved env.

Common situations: A stale environment record that the store fails to purge or replace; a misbehaving isolation provider marking new envs as cleaned; concurrent cleanup jobs racing with resolution; corrupted isolation env state in the database.

Related errors

AI-assisted analysis of coleam00/Archon@0773b97458 (2026-09-01). Data as JSON: /api/errors/e15672ff62fa21ba. Report an issue: GitHub.