openclaw - 💡(How to fix) Fix [Bug]: Sub-agent exec tool calls broken after update to 2026.4.8 [1 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

After updating from 2026.4.5 to 2026.4.8, all sub-agents lost the ability to execute exec tool calls against canonical data files — they reported being unable to find files that were accessible to the main agent in the same session.

Steps to reproduce

1. Install OpenClaw 2026.4.8 (upgrade from 2026.4.5).
2. Configure one or more sub-agents with exec tool access to canonical data files via jq.
3. Delegate a query to any sub-agent that requires an exec/jq call.
4. Observe sub-agent reports it cannot find the canonical data file, despite the file being accessible to the main agent in the same session.
5. Roll back to 2026.4.5 — sub-agent exec calls work correctly immediately.

Expected behavior

Sub-agents should execute exec tool calls against canonical data files as they did in 2026.4.5. The main agent retained full exec access throughout — only sub-agent sessions were affected. Rolling back to 2026.4.5 (3e72c03) restored full sub-agent exec functionality immediately.

Actual behavior

Sub-agents report they cannot find canonical data files that are accessible to the main agent in the same session. Example response: "I can't find the file in the workspace." The main agent can execute identical jq commands against the same files successfully. All sub-agents affected. No error in gateway logs — sub-agents simply fail to resolve file paths via exec.

OpenClaw version

2026.4.8 (9ece252) — broken, 2026.4.5 (3e72c03) — working

Operating system

macOS 15.4

Install method

npm global

Model

openai-codex/gpt-5.4 (main agent), openai/gpt-5.4-nano (sub-agents)

Provider / routing chain

openclaw -> openai-codex (OAuth), [main agent] openclaw -> openai API, [sub-agents] openclaw -> ollama (local, socat relay) [sub-agents fallback]

Additional provider/model setup details

Sub-agents are configured with per-agent model overrides via environment variable substitution in openclaw.json. Main agent uses openai-codex OAuth. Sub-agents use openai API key for primary and local Ollama via socat relay for fallback. exec approval mode is "targets" with all sub-agent IDs explicitly listed in agentFilter. Bug is model-agnostic — occurs regardless of which model is assigned to sub-agents.

Logs, screenshots, and evidence

Impact and severity

Affected: User with sub-agents configured using exec tool access. Severity: High — completely blocks sub-agent functionality. All sub-agents unable to execute exec calls. Only workaround is rollback to 2026.4.5. Frequency: 100% reproducible — occurs on every sub-agent exec call after updating to 2026.4.8. Consequence: Full rollback to previous version required to restore functionality. No partial workaround available.

Additional information

No response