OpenClaw Issue Digest: Session Stability and Context Management

Recent activity in the OpenClaw repository highlights a critical focus on session stability, particularly regarding how the system handles context overflows and external runtime integrations. While several feature requests aim to improve the user experience for multi-agent setups, a cluster of high-severity bugs related to 'silent' failures in embedded sessions is currently a priority for maintainers.

Open Issues

Session Stability and Critical Failures

Several reports indicate that agent sessions are entering terminal states without notifying the user, leading to a frozen experience:

• Preemptive Context Overflow (#84536): Embedded sessions may silently die during tool loops when a hardcoded overflow ratio is hit, leaving the session in a processing state for hours without delivering an error to the channel.
• Codex App-Server Wedges (#85251): Sessions using the Codex backend may hang indefinitely after a turn/started notification, resulting in a silent failure that is only resolved by the system's 360-second stuck-session recovery.
• Subagent Result Loss (#44925): Orchestration failures are causing subagent completions to be lost silently, with no retries or notifications when the announce-back to the parent session fails.

Context and Memory Management

Issues regarding how OpenClaw manages long-term and short-term memory are surfacing as regression bugs and feature gaps:

• Compaction Preflight Errors (#84857): A regression in 2026.5.20-beta.1 causes compaction preflights to throw a MissingAgentHarnessError for claude-cli runtimes, blocking session compaction and eventually leading to context overflow.
• BOOT.md Persistence (#84970): The boot-md startup hook is incorrectly persisting synthetic prompts into active main sessions, causing hidden context bloat across gateway restarts.
• Memory Search Precision (#44395, #44540): Users are reporting retrieval inaccuracies due to fixed-size chunking that blends semantically distinct sections. Additionally, the BM25 component of hybrid search is penalizing multimodal (image/audio) results, effectively hiding them from users.

Tooling and UX Enhancements

Contributors are pushing for more robust identity and interface tools:

• Identity Resolution (#45427): A request to include trusted sender_name in inbound metadata to prevent agents from misrouting DMs based on untrusted conversational context.
• Control UI Improvements (#45323): A proposal for Slack-style @mention autocomplete in the Control UI to facilitate multi-agent coordination.
• Tool Invocation Enforcement (#45049): Reports that agents frequently simulate tool usage in text rather than triggering real tool calls, suggesting a need for stronger enforcement in the agent loop.

Key Themes

The "Silent Failure" Pattern

Across multiple issues (#84536, #85251, #44925), a recurring theme is the lack of user-facing error propagation. When a backend process (like a tool loop or a Codex socket) hangs or crashes, the system often logs the error internally but fails to notify the user, leaving them in a state of perpetual waiting.

Context Bloat and Regression

There is a clear tension between the introduction of new context management features (like LCM and Lossless) and the stability of existing runtimes. The claude-cli harness error (#84857) and the boot-md persistence bug (#84970) demonstrate how internal maintenance tasks can inadvertently pollute the very context they are meant to optimize.

Multimodal and Structural Memory

There is an emerging demand for