OpenClaw Update: Optimizing Session Storage, Realtime Voice, and Channel Reliability

Merged PRs

• Fix compaction checkpoints to avoid transcript copies Original PR
• feat(signal): support reaction approvals Original PR
• fix: use context-aware reserveTokensFloor in overflow recovery hint Original PR
• feat(gateway): forward frequency_penalty, presence_penalty, and seed via OpenAI-compatible HTTP gateway Original PR
• perf: cache model cost indexes Original PR
• Refactor Google Meet realtime output activity tracking Original PR
• fix: prefer source public artifacts in source checkouts Original PR
• fix(gateway): dampen repeated device-required probes Original PR
• fix(ui): keep local file markdown links inert Original PR
• fix(update): avoid duplicate plugin smoke failures Original PR
• fix(gateway): cap retained compaction checkpoint bytes Original PR
• perf: reduce runtime metadata hotpath churn Original PR
• Refactor realtime output activity tracking Original PR
• fix(memory-wiki): bound compile page reads Original PR
• fix(cli): error on unknown command root when --help/--version is appended Original PR
• test: improve test profiling helpers Original PR
• Refactor Discord realtime forced consult coordination Original PR
• fix(scripts): use process.exitCode instead of process.exit in codex-cli branch Original PR
• fix(codex): disable Codex built-in personality for native threads Original PR
• Refactor realtime voice turn context tracking Original PR
• fix(discord): surface silent reply-delivery skips and remove runtime.error optional-chain Original PR
• fix(hooks): clear Gmail watcher renewal interval on re-entry Original PR
• fix(codex): honor yolo app-server approval policy Original PR
• fix(usage-cost): surface unpriced-model spend as missingCostEntries instead of a confident $0 Original PR
• fix(irc): store inbound channel routes as channel:#name and join before send Original PR
• fix(agents): strip markdown code spans from IDENTITY.md values and labels Original PR
• fix(memory-core): use CJK-aware tokenizer for dreaming dedupe Original PR
• perf: reduce runtime cache churn Original PR
• fix(pi-runner): flush block replies after compaction retry Original PR
• Fix Telegram DM reply delivery for overlapping turns Original PR

Key Changes

Session Storage & Performance Optimizations

Significant efforts were made to reduce the disk and CPU footprint of session management. A major shift in how compaction checkpoints are handled now avoids duplicating full transcripts, instead using metadata-backed markers to record stable pre- and post-compaction identities. To prevent disk bloat in long-running sessions, a fixed 128 MiB byte budget has been introduced for retained checkpoint snapshots per session.

On the performance front, several "hotpath" optimizations were implemented. These include caching model cost indexes to avoid repeated rebuilds and reducing runtime metadata churn by making plugin metadata snapshots immutable. Additionally, the session-store cache now clones objects rather than reparsing serialized JSON on hits, further reducing CPU overhead.

Realtime Voice & Coordination Refactoring

OpenClaw is moving toward a provider-neutral architecture for realtime voice. New shared trackers have been introduced to the SDK for:

• Output Activity: Tracking audio counters, playback state, and barge-in gating, now reused across Discord and Google Meet.
• Turn Context: Managing pending audio contexts and speaker attribution, migrated from Discord-specific logic to a shared helper.
• Forced Consults: Coordinating pending fallback consults and deduplicating native tool calls, now shared across realtime surfaces.

Channel Reliability & Feature Enhancements

Several critical fixes were merged to ensure messages actually reach users:

• Telegram: Fixed a bug where overlapping direct-message turns would supersede active reply delivery, causing replies to stay in webchat but never reach the user.
• Discord: Added verbose logging to surface "silent" reply skips that occurred during gateway restarts or when reasoning payloads were handled, making these events debuggable for operators.
• Signal: Introduced native reaction approvals, allowing users to approve or deny execution requests using 👍/👎 reactions.
• IRC: Normalized group route identities to channel:#name and ensured transient sends perform a JOIN before PRIVMSG.

Codex & Agent Logic

To improve agent consistency, Codex's built-in personality layer is now disabled for native threads. This ensures that OpenClaw's workspace-defined identity and instructions remain the authoritative source of behavior, preventing prompt conflicts. Additionally, the Codex app-server now correctly honors "yolo" approval policies, auto-accepting requests when the runtime is configured for full access.

Impact

These changes directly address several high-impact user pain points. The fix for Telegram DM delivery resolves a scenario where users perceived the bot as ignoring them despite the session transcript showing successful processing. Similarly, the new Discord logging transforms invisible abort-skips into traceable events, reducing the time required to diagnose delivery failures during system instability.

For operators, the reduction in checkpoint disk usage and the optimization of metadata lookups will lead to lower I/O pressure and improved gateway responsiveness, especially in high-concurrency environments. The shift toward a shared SDK for realtime voice lays the groundwork for more consistent behavior across different voice-enabled channels, while the CJK-aware tokenizer in the memory dreaming pipeline ensures that deduplication is effective for non-English workspaces, preventing redundant entries in MEMORY.md.