openclaw - 💡(How to fix) Fix Feature: Cross-session context injection via channel mentions [1 participants]

Problem

OpenClaw sessions are isolated silos. When an agent builds deep context in one channel (e.g. a long-running project discussion in #popt-ai), that knowledge is trapped — inaccessible to other sessions without manual copy-paste or the agent explicitly using sessions_history.

In Discord (and potentially other platforms), users naturally reference other channels with #channel-name mentions. These are already rendered as <#channelId> in the message body, but OpenClaw does not act on them.

Proposed solution

When a message contains a channel mention (<#channelId>), OpenClaw should optionally resolve it to the bound session and inject a compressed summary of that session's recent context into the current run.

Flow:

1. Gateway detects <#channelId> pattern(s) in the inbound message
2. Resolves channel ID → session key via the session store
3. Fetches recent history from the target session (configurable depth)
4. Generates a compressed summary (LLM-generated or last-N messages)
5. Injects the summary into the current run's context (e.g. as a system prompt addition or a synthetic context block)

Example user experience:

User in #frontend: "Use the API design we discussed in #backend — apply it to the new components"

Agent receives the #backend session summary automatically and can act on it without the user re-explaining.

Design considerations

Summary strategy

• Default: LLM-generated summary of last N messages (token-efficient)
• Optional: raw last-N messages for full fidelity (--full or config flag)
• Cache summaries with a TTL to avoid re-summarizing on every mention

Token budget

• Injected cross-session context should respect a configurable token cap (e.g. crossSessionContext.maxTokens: 2000)
• Multiple channel mentions in one message should share the budget

Config surface

{
  session: {
    crossSessionContext: {
      enabled: false,          // opt-in
      maxTokens: 2000,         // per-run budget for injected context
      summaryDepth: 50,        // messages to consider for summary
      cacheTtlMinutes: 30,     // cache summaries
      strategy: "summary",     // "summary" | "raw" | "hybrid"
    }
  }
}

Security / scoping

• Only inject from sessions belonging to the same agent (no cross-agent leaks by default)
• Respect sendPolicy — if a session is deny, don't expose its content
• Option to allowlist/denylist specific sessions for cross-injection

Platform support

• Discord: <#channelId> — natural fit, already parsed
• Slack: <#C123|channel-name> — same pattern
• Other platforms: could support explicit syntax like @session:key or #channel-name

Implementation paths

1. Hook-based (plugin): A message:received or message:preprocessed hook detects mentions, fetches history, and injects context. Could work with existing context engine plugin API (systemPromptAddition).

2. Core feature: Native gateway support in the message pipeline. More efficient (no extra LLM call for caching), better UX, and consistent across channels.

3. Agent-level workaround (available today): The agent can use sessions_list + sessions_history manually, but this is ad-hoc, burns tokens per use, and requires the agent to "know" it should do this.

Prior art / related

• sessions_history tool already supports cross-session reads
• bootstrap-extra-files hook demonstrates dynamic context injection
• Context Engine plugin API (assemble → systemPromptAddition) provides the injection point
• Channel .md files in workspace are a static version of per-channel context

Open questions

• Should the summary be visible to the user (e.g. "📎 Injected context from #backend: ...")?
• Should this work bidirectionally (both sessions get each other's context)?
• How to handle stale sessions (last active 2 weeks ago)?
• Should the agent be able to trigger this programmatically (e.g. sessions_context(sessionKey) tool)?