Sessions Are Event Streams — But We Can't Query Them

Every OpenClaw session produces a rich stream of events: messages, tool calls, compactions, spawns, errors. These are stored in .jsonl files — append-only, ordered, timestamped. This is event sourcing by accident.

But there's no API to query this stream. The only option is to parse raw .jsonl files, which means:

• No way to ask "what tools did this agent call in the last hour?"
• No way to aggregate "how many errors across all sessions today?"
• No way to build dashboards, alerts, or analytics without custom file parsing
• The richest data source in OpenClaw is effectively invisible

Proposed: Session Event Query API

Tool-Level (for agents)

// Query own session events
session_events({
  types: ["tool_call", "error"],   // filter by event type
  since: "2026-02-15T10:00:00Z",  // time range
  limit: 50                        // pagination
})

// Query across sessions (for orchestrators)
sessions_events({
  sessionKey: "agent:main:*",      // wildcard
  types: ["spawn", "complete"],
  since: "1h"                      // relative time
})

HTTP API (for external systems)

GET /api/sessions/:key/events?types=error,tool_call&since=1h&limit=100
GET /api/events?agentId=main&types=spawn,complete&since=24h

Event Types to Expose

Why This Enables Everything Else

This is foundational infrastructure. Without queryable events:

• Agent health (perf(hooks): skip cache-busting for immutable bundled hooks #17053-equivalent) has no data source for condition codes
• Cost tracking requires parsing files instead of querying an API
• Orchestration can't react to events in real-time
• Dashboards need custom ETL instead of a simple API call
• Audit logging ([Security] Extension Runtime: No Security Event Logging or Audit Trail #12553, Feature: Gateway-level audit logging for all tool calls #13131) needs structured event access

Event sourcing is the #1 pattern in distributed systems (Kafka, EventStore, Axon). OpenClaw already HAS the event log — it just needs an API on top.

Minimal First Step

Expose the existing .jsonl content via a read-only session_events tool with type filtering and time range. No new storage, no new indexing — just structured access to what's already there.

Happy to help design the schema and test on Windows.