OpenClaw Update 4.14 is HUGE Value for GPT 5.4 Users!

OpenClaw 4.14 fixes the GPT-5 minimal validation error and adds a retry gate for empty turns, but leaves the Telegram plugin broken.

GPT-5 Model Provider Fixes

Two failure modes are addressed. First, OpenClaw was sending reasoning_effort: minimal to the OpenAI API, which it rejects outright — every request failed with a validation error. The 4.14 release auto-maps minimal to low, so existing configs work without edits. Second, when GPT-5.4 returned reasoning-only or empty turns, sessions would hang or fail silently. A new bounded continuation system with a retry gate now attempts a follow-up and falls back gracefully if nothing visible comes back. The creator is clear this is "not a 100% fixed" — but the direction is now OpenAI-native, reinforced by creator Peter Steinberger joining OpenAI and the 4.12 change that routes GPT/Codex through a dedicated Codex-managed path.

Core Refactor — Performance

Background housekeeping now runs as idle-aware work instead of inline, so prompts stop waiting on it. Mismatched plugin IDs in the context engine now fail fast instead of producing silent bugs. The gateway entry point is unified, killing stale-path startup failures on reinstalls. The recalled-memory path moved out of the trusted system-prompt slot into a hidden untrusted prefix — important because recalled memory could carry instructions the agent would otherwise treat as authoritative. The resolved recall provider and model now show up in /verbose gateway debug logs and on the dashboard.

Core Refactor — Reliability

Heartbeat, cron, and exec event turns no longer overwrite shared session routing metadata, ending context bleeding into the next user session. Auto-reply queues are now split by per-message authorization context — previously, two team members messaging the same agent simultaneously could cause a follow-up to run under the wrong human's permissions. The QMD memory backend was double-loading a legacy lowercase memory.md as a second root collection, which the creator pins as the root cause of recent memory hallucinations.

Dreaming Loop and Telegram Caveat

The dreaming feature still has a known bug: consumed scheduled runs can re-trigger on subsequent heartbeats, causing duplicate memory consolidation sweeps. Separately, the Telegram plugin currently throws a delete my commands failed error that puts OpenClaw into a restart loop — do not update if you run Telegram until the GitHub issue is closed.

Minor Integrations

Telegram now surfaces forum topic names in agent context for multi-topic setups. Four security patches ship alongside the platform fixes.