openclaw - 💡(How to fix) Fix [Bug] agents.list[].workspace has no effect on bootstrap injection or file tool CWD (v2026.3.14) [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

agents.list[].workspace has no effect on bootstrap file injection or file tool CWD — all agents use agents.defaults.workspace regardless of per-agent config

Steps to reproduce

1. Configure two agents in agents.list[] with different workspace paths (e.g. li → ~/.openclaw/workspace, kim → ~/.openclaw/workspace/personal)
2. Set agents.defaults.workspace to one agent's path (e.g. li's path)
3. Start gateway and send a message to the other agent (kim via Telegram)
4. Observe that kim's file tool operations and bootstrap files (SOUL.md, AGENTS.md) resolve to agents.defaults.workspace, not her own workspace

Expected behavior

Each agent in agents.list[] loads bootstrap files (SOUL.md, AGENTS.md, IDENTITY.md, USER.md) from its own workspace path, and file tool CWD resolves to that agent's workspace. Per-agent workspace isolation should work as documented at docs.openclaw.ai/concepts/agent-workspace and docs.openclaw.ai/gateway/configuration#configure-multi-agent-routing

Actual behavior

agents.list[].workspace is ignored at runtime. All agents — regardless of per-agent workspace config — load bootstrap files from agents.defaults.workspace and resolve file tool CWD to that same path. Confirmed: agent kim (workspace/personal) reports file paths under workspace/ (Li's path). BOOTSTRAP.md lookup went to defaults workspace. SOUL.md and AGENTS.md loaded are Li's files, not Kim's.

OpenClaw version

2026.3.14 (55e12bd)

Operating system

Windows 11 / WSL2 Ubuntu 24.04

Install method

npm global

Model

ollama/qwen3-kim:8b (local) — agents.defaults.model: openai/gpt-5.1

Provider / routing chain

openclaw -> ollama (local, http://127.0.0.1:11434)

Additional provider/model setup details

Two agents bound to separate channels: li → whatsapp → anthropic/claude-sonnet-4-6 kim → telegram → ollama/qwen3-kim:8b

agents.list[li].workspace = /home/user/.openclaw/workspace agents.list[kim].workspace = /home/user/.openclaw/workspace/personal agents.defaults.workspace = /home/user/.openclaw/workspace

Workarounds attempted — all rejected as schema-invalid (Unrecognized key): agents.list[].systemPrompt channels.telegram.dms.<id>.systemPrompt channels.telegram.dms.<id> (entire block)

Logs, screenshots, and evidence

Gateway log — kim file tool resolving to wrong workspace:
  [agent/kim] file lookup: /home/user/.openclaw/workspace/BOOTSTRAP.md
  (expected: /home/user/.openclaw/workspace/personal/BOOTSTRAP.md)

Bootstrap injection confirmed wrong: kim loads Li's AGENTS.md + SOUL.md
from agents.defaults.workspace instead of her own workspace/personal/.
Kim's IDENTITY.md (injected separately) correctly shows kim persona,
but SOUL.md + AGENTS.md are Li's files.

Impact and severity

Affected: any multi-agent setup using per-agent workspace isolation Severity: High — blocks per-agent persona isolation entirely; all agents share one workspace, making independent agent personalities non-functional without workaround Frequency: always — 100% reproducible Consequence: agents load wrong SOUL.md/AGENTS.md, persona isolation fails silently (no error thrown, schema accepts the workspace key)

Additional information

The workspace key IS schema-valid (no Unrecognized key error at startup), so OpenClaw accepts the config but silently ignores it at runtime. This makes the bug hard to detect — it fails silently with no warning.

Note: prior to disabling model fallback, a more capable commercial model (gpt-5.1) was compensating for this bug by inferring the correct persona from IDENTITY.md alone. The bug was always present but masked by the fallback model's capability. Switching to a local 8B model exposed it.

Last known working state: NOT_ENOUGH_INFO (may never have worked; the masking made it appear functional)