Bug type

Behavior bug

Summary

In a multi-agent setup, native /new / before_reset hook handling derives agent identity incorrectly and can cause the bundled session-memory hook to write reset/session memory into workspace-main even when the live bound agent is a different agent (for example, Navi with workspace-navi).

Steps to reproduce

1. Configure a non-main agent bound to a channel/DM, for example:
   • agentId: navi
   • agents.list[navi].workspace = ~/.openclaw/workspace-navi
2. Ensure bundled internal hooks include:
   • session-memory
3. Use the bound non-main agent in a real chat surface (Discord DM in my case).
4. Trigger /new or /reset.
5. Inspect the resulting session-memory artifacts under ~/.openclaw/.

Expected behavior

Reset/session-memory flow should resolve to the bound active agent and write memory/session bridge files into that agent's workspace.

Example:

• bound agent navi → writes to ~/.openclaw/workspace-navi/memory/

Actual behavior

Session-start/reset memory artifacts are being written into ~/.openclaw/workspace-main/memory/ instead of the bound agent workspace.

Observed files:

• ~/.openclaw/workspace-main/memory/2026-03-07-2106.md
• ~/.openclaw/workspace-main/memory/2026-03-07-session-greeting.md
• ~/.openclaw/workspace-main/memory/2026-03-08-session-start.md

These are session-start style summaries and reference:

• Session Key: agent:main:main

Meanwhile the actual bound agent continuity files live in:

• ~/.openclaw/workspace-navi/memory/

So the system ends up in a workspace split-brain state.

OpenClaw version

2026.3.7

Also checked on current main source as of 2026-03-08; the relevant code path still appears unchanged there.

Operating system

Linux (WSL2 / Ubuntu-style environment)

Install method

Git/source install (main) and stable release both inspected.

Logs, screenshots, and evidence

Current stable tag inspected:
- v2026.3.7
- commit: 42a1394c5c0fb86706f61598e68e0db30e8c99c1

Current upstream main still contains this in src/auto-reply/reply/commands-core.ts:

  await hookRunner.runBeforeReset(
    { sessionFile, messages, reason: params.action },
    {
      agentId: params.sessionKey?.split(":")[0] ?? "main",
      sessionKey: params.sessionKey,
      sessionId: prevEntry?.sessionId,
      workspaceDir: params.workspaceDir,
    },
  );

For a session key like:
- agent:navi:main

split(":")[0] returns:
- "agent"

not:
- "navi"

Bundled session-memory handler resolves workspace from session identity:

src/hooks/bundled/session-memory/handler.ts

  const agentId = resolveAgentIdFromSessionKey(event.sessionKey);
  const workspaceDir = cfg
    ? resolveAgentWorkspaceDir(cfg, agentId)
    : path.join(resolveStateDir(process.env, os.homedir), "workspace");

Observed split on disk:
- Navi configured workspace: ~/.openclaw/workspace-navi
- stray session-memory files created in: ~/.openclaw/workspace-main/memory/

Example files created there:
- 2026-03-07-2106.md
- 2026-03-07-session-greeting.md
- 2026-03-08-session-start.md

Impact and severity

Affected:

• multi-agent setups using non-main agents
• bundled session-memory on /new / /reset
• likely any workflow that assumes reset/session-memory writes land in the active bound agent workspace

Severity:

• High behavior bug

Frequency:

• Reproducible in my setup

Consequence:

• session/reset memory written into the wrong workspace
• split-brain continuity (workspace-navi for real agent files, workspace-main for reset/session-memory artifacts)
• likely contributes to stale continuity, wrong memory reads, and reset weirdness in multi-agent setups

Additional information

Regression timing

This appears to have been introduced when native reset hooks were backfilled in:

• commit aec41a588b
• message: fix(hooks): backfill reset command hooks for native /new path
• date: 2026-02-24

That commit is contained in releases including:

• v2026.3.2
• v2026.3.7

Related issues

Potentially related, but not the same exact bug:

• Bug: session-memory hook (and command-logger) do not fire on /new or /reset commands #23027 — session-memory / command-logger do not fire on /new or /reset
• Internal hooks not triggered in auto-reply path due to missing sessionKey parameter #29203 — internal hooks not triggered in auto-reply path due to missing sessionKey
• agent_end hook passes wrong agentId (always 'agent' instead of actual agent ID) #39521 — agent_end hook passes wrong agentId using the same split(":")[0] pattern

Suggested fix direction

At minimum:

• stop using params.sessionKey?.split(":")[0]
• use proper session-key parsing such as resolveAgentIdFromSessionKey(params.sessionKey)

But ideally:

• reset hooks should prefer the bound/active agent identity (or explicit hook context agent/workspace) rather than relying on a legacy-looking session key that can be agent:main:main

Why I think this is runtime, not config

The relevant config values remained stable across backups and live config:

• agents.defaults.workspace = ~/.openclaw/workspace
• agents.list[navi].workspace = ~/.openclaw/workspace-navi

So this does not look like a plain config-path mistake. It looks like a hook/runtime identity-resolution bug.