OpenClaw Digest: Resolving Runtime Regressions and Enhancing Gateway Stability

Merged PRs

• Fix Claude CLI runtime migration for gateway turns (Original PR)
• Fix Kimi Coding tool-call replay (Original PR)
• fix(channels): install externalized same-id channel adds (Original PR)
• fix(maintainer): gate body notifications after redaction (Original PR)
• fix: limit Codex attribution to local transcripts (Original PR)
• fix: hide decorative emoji on unsupported terminals (Original PR)
• fix(agents): remove transient session-repair backups (Original PR)
• fix(agents): yield during model stream bursts (Original PR)
• fix: gate dashboard on gateway readiness (Original PR)
• fix(codex): release raw assistant app-server completions (Original PR)
• fix(backup): hide manifest parser internals (Original PR)
• fix(agents): fall back to PI when Codex harness is unavailable (Original PR)
• fix(cli): finalize context engine turns (Original PR)
• fix(agents): preserve Pi tool result error flags (Original PR)
• fix(gateway): abort gmail watcher reload starts (Original PR)
• Fix Telegram context loss after restart replay (Original PR)
• Fix doctor repair for disabled Codex runtime plugin (Original PR)
• fix(codex): enforce native tool policy (Original PR)
• fix(gateway): isolate gmail watcher restart and abort handling (Original PR)
• fix(telegram): drop expired approval callbacks (Original PR)
• Fix session reset files and ACPX orphan reaping (Original PR)

Key Changes

Runtime & Model Integration Fixes

Several critical regressions affecting model runtimes were addressed to ensure consistent behavior across different interfaces (CLI vs. Gateway).

• Claude CLI Migration: A regression was fixed where gateway-mediated turns (via Telegram or Dashboard) ignored legacy runtime fields, causing them to fall back to Anthropic API billing instead of using the local Claude CLI runtime. The fix ensures model-scoped runtime policies are correctly written during migration.
• Codex Harness Stability: To prevent HTTP 500 errors when the Codex harness is unregistered, OpenClaw now falls back to the embedded PI harness. Additionally, openclaw doctor --fix can now automatically repair configurations where the Codex plugin is disabled but required by configured routes.
• Kimi Coding Replay: Fixed a failure in Kimi thinking mode where tool-call messages lacked the required reasoning_content field during replay, which previously caused the Kimi API to reject requests with a 400 error.
• Pi Tool Error Handling: The embedded Pi tool result bridge now correctly preserves error or timeout status flags, preventing agents from silently abandoning tasks when a tool fails (e.g., ambiguity errors in file edits).

Gateway & Channel Reliability

Significant effort was put into eliminating "deadlock" scenarios in messaging channels and improving the gateway's lifecycle management.

• Telegram Ingress Recovery: Resolved a critical issue where process restarts could leave .json.processing claims on disk without active handlers. This effectively blocked subsequent updates in the same lane. The system now implements durable claims and immediate recovery for restart-era markers.
• Telegram Approval Loops: Fixed an infinite retry loop triggered by expired approval callbacks. Previously, these were treated as retryable, blocking all other incoming DMs; they are now treated as terminal and discarded.
• Gmail Watcher Lifecycle: Improved the isolation of Gmail watcher restarts during config hot-reloads. The gateway now ensures that queued post-ready sidecars are stopped and abort signals are propagated to Tailscale setup, preventing stale endpoints from being applied.
• Dashboard Readiness: The openclaw dashboard command is now gated by a gateway readiness helper, ensuring the URL is only provided if the gateway is actually usable or recoverable.

Core Agent & Storage Improvements

• Context Engine Finalization: CLI-routed turns now correctly trigger the context-engine post-turn lifecycle, ensuring that non-legacy engines (like LCM) receive turn data and maintenance for claude-cli runs.
• Session Repair Cleanup: Fixed a bug where successful session repairs left behind unbounded .bak files, which in some field cases grew to over 2 GB of debris for a single session.
• Codex App-Server Delivery: Fixed a deadlock where raw assistant completions without a turn/completed event left OpenClaw waiting for a timeout. The system now interrupts the native Codex turn to release the session lane early.

Impact

These changes directly address several high-severity pain points reported by users:

• Billing Protection: Users of the Claude CLI will no longer see unexpected Anthropic API charges when interacting via the Gateway.
• Messaging Continuity: Telegram users will experience fewer "stuck" bots and a reduction in the frequency of manual spool directory cleanups.
• Resource Management: The removal of transient session backups prevents disk pressure on long-running installations.
• Security Hardening: The enforcement of native tool policies for the Codex harness ensures that before_tool_call hooks are no longer bypassed for native shell/file operations, bringing Codex-native tools under the same security umbrella as OpenClaw dynamic tools.

"The error was delivered as text in toolResult.content with isError=false, so OpenClaw treated it as a successful delivery — it doesn't recognize this as a failure that should trigger retry."

By fixing the Pi tool result flags and the Codex delivery pipeline, OpenClaw significantly improves the reliability of autonomous agent loops, ensuring that failures are signaled and handled rather than resulting in silent stalls.