openclaw - 💡(How to fix) Fix [Bug]: Tool call replay loop: runtime repeats agent's previous tool call after failure, causing infinite loops [1 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Bug Description

After upgrading to OpenClaw 2026.4.25, the agent enters an infinite tool loop where the runtime replays the agent's previous tool call after a tool failure, rather than allowing the agent to reason and take a new action. The agent does not "decide" to repeat — the runtime automatically re-emits the same tool call.

Symptoms

1. Agent issues a tool call (e.g., exec)
2. Tool returns failure/empty output
3. Runtime feeds result back to agent
4. Agent attempts to fix/reason
5. Runtime replays the original tool call (not issued by the agent in this turn)
6. Loop repeats indefinitely (observed up to 9+ iterations)

This also manifests as:

• Successful workflows completing but then repeating anyway (e.g., file sent to Feishu, but the entire workflow replays)
• Context window filling up with duplicate tool calls (ctx 100% before raising token limit to 500k)
• Files growing with duplicate data appended each iteration

Environment

• OpenClaw version: 2026.4.25 (aa36ee6)
• Model: bailian/qwen3.5-plus
• OS: macOS (Darwin 25.3.0, arm64)
• Node.js: v22.22.0
• Channel: Feishu (DM)

Timeline

• Pre-2026.4.25: No tool loop issues
• After 2026.4.25 upgrade: First symptom was ctx 100% (context window exhaustion)
• Raised context limit to 200k, then 500k — ctx 100% stopped, but tool loops persist
• Agent behavior rules (AGENTS.md) with anti-loop constraints are ineffective — the loop is at the runtime level, not agent decision level

Relevant Changelog Items (v2026.4.25)

The following changes in 2026.4.25 may be related:

• Agents/runtime: submit heartbeat, cron, and exec wakeups as transient runtime context instead of visible user prompts
• Auto-reply: poison inbound message dedupe after replay-unsafe provider/runtime failures
• New openclaw.tool.loop OpenTelemetry spans suggest the team is aware of tool loop diagnostics

Expected Behavior

After a tool call fails and the agent receives the output, the agent should be able to reason and issue a new tool call (or no tool call). The runtime should not replay the previous tool call.

Actual Behavior

The runtime replays the agent's previous tool call automatically, creating an infinite loop that can only be stopped by manually killing the agent process.

Steps to Reproduce

1. Run a multi-step workflow involving exec tool calls
2. Trigger a tool failure (e.g., command returns non-zero exit code)
3. Observe: agent attempts to reason/fix
4. Observe: runtime replays the same tool call
5. Loop continues until manually stopped

Workarounds Attempted

• Adding anti-loop rules to AGENTS.md (agent behavior config) — ineffective because the loop is at the runtime level
• Increasing context window (100k → 200k → 500k) — delays ctx exhaustion but doesn't stop the loop
• Manual stop command — only way to interrupt

Logs / Evidence

In the WebUI, the agent shows repeated tool call → tool output → tool call sequences with identical tool inputs. The agent's own text output shows it attempting to reason, but the runtime keeps replaying the same tool call.

Example pattern (repeated 9 times): Agent: "让我查一下..." → exec: cat /path/to/config | python3 -c "..." → Tool output: (empty) Agent: "你说到了最关键的问题。让我查一下..." → exec: cat /path/to/config | python3 -c "..." (SAME call, replayed by runtime) → Tool output: (empty) ... repeats ...

Suggested Investigation

1. Check the tool loop / replay logic introduced or modified in 2026.4.25
2. Review poison inbound message dedupe after tool failures
3. Verify that runtime context injection after tool failure doesn't trigger tool replay
4. Consider adding a hard tool call loop limit at the runtime level (not agent level)

Steps to reproduce

Agent: "让我查一下..." → exec: cat /path/to/config | python3 -c "..." → Tool output: (empty) Agent: "你说到了最关键的问题。让我查一下..." → exec: cat /path/to/config | python3 -c "..." (SAME call, replayed by runtime) → Tool output: (empty) ... repeats ...

Expected behavior

Suggested Investigation

1. Check the tool loop / replay logic introduced or modified in 2026.4.25
2. Review poison inbound message dedupe after tool failures
3. Verify that runtime context injection after tool failure doesn't trigger tool replay
4. Consider adding a hard tool call loop limit at the runtime level (not agent level)

Actual behavior

No replay

OpenClaw version

2026.04.25

Operating system

macOS

Install method

No response

Model

qwen3.5-plus

Provider / routing chain

openclaw ->gateway ->qwen3.5

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response