By devasher · Edited by Nominiclaw
This update focuses on improving subagent delivery reliability, refining Telegram forum topic concurrency, and hardening Codex sandbox and configuration policies.
Significant effort was directed toward the reliability of subagent completion announces. A critical fix was implemented to recover stale subagent completion announces by retrying requester wake and forcing a message-tool handoff when direct-agent replies are invisible. Additionally, the system now skips dormant completion wake probes, routing late announcements directly through the requester-agent handoff to reduce no_active_run log noise.
In the Codex extension, a fix ensures that Codex-native subagent tasks are marked terminal when a spawnAgent call fails or is blocked, preventing tasks from remaining stuck in a running state in the registry.
Telegram integration received substantial upgrades to support forum topics more effectively. The update fixes parallel flow issues by replacing global flush chains with per-topic queues and introducing a GroupFairQueue to prevent a single topic from monopolizing the shared group send pipe.
Other channel improvements include:
runtimeOnly prompt path, ensuring the model sees the replied-to message body.Security and policy enforcement in the Codex app-server were strengthened. The system now honors Docker bind write policies, ensuring that :rw bind mounts are included in writableRoots even when workspaceAccess is set to ro.
Further Codex improvements include:
Configuration validation was refined to allow bundled provider timeout overlays (e.g., models.providers.openai.timeoutSeconds) without requiring a full provider redeclaration. CLI usability was improved by making root help config-sensitive, ensuring that the correct memory plugin commands (like ltm for LanceDB) are advertised based on the active configuration.
These changes collectively reduce operational noise and increase the robustness of complex agent workflows. The improvements to subagent delivery and Codex task reconciliation directly address high-severity issues where completed work failed to surface or tasks remained ghost-running in the registry.
For Telegram users, the transition to per-topic concurrency and fair-share queuing removes serialization bottlenecks, allowing multiple forum topics to operate in parallel without blocking one another. The WhatsApp periodic drain eliminates significant delivery delays, ensuring a more responsive user experience. Finally, the hardening of the Codex sandbox and configuration validator provides a more flexible yet secure environment for deploying custom model providers and bind-mounted data sources.