[Bug]: Infinite loop of tool calls

[LeoNardo-LB]

Prerequisites

• I have searched existing issues to avoid duplicates
• I am using the latest version of oh-my-opencode
• I have read the documentation

Bug Description

This situation occurred during tool invocation

After manual intervention, it was prompted to ignore the last callback message, but this situation still occurred later.

Steps to Reproduce

1. chose the zhipuai-coding-plan/glm-4.7 model
2. Make a plan to solve the problem and use Prometheus
3. /start-work to execute the plan
4. This situation occurred after a long period of operation.
5. After actively intervening and interrupting, when using /ulw-loop to continue execution, this situation still occurs.

Expected Behavior

normal tool call

Actual Behavior

Infinite loop tool call, see the picture

Doctor Output

Installation
────────────────────────────────────────
  ✓ OpenCode Installation → 1.1.34
  ✓ Plugin Registration → Registered (pinned: 3.0.0-beta.16)

Configuration
────────────────────────────────────────
  ✓ Configuration Validity → Valid JSON config

Authentication
────────────────────────────────────────
  ✓ Anthropic (Claude) Auth → Auth plugin available
  ○ OpenAI (ChatGPT) Auth → Auth plugin not installed
  ○ Google (Gemini) Auth → Auth plugin not installed

Dependencies
────────────────────────────────────────
  ✓ AST-Grep CLI → installed
  ⚠ AST-Grep NAPI → Not installed (optional)
  ⚠ Comment Checker → Not installed (optional)

Tools & Servers
────────────────────────────────────────
  ⚠ GitHub CLI → Not installed (optional)
  ✓ LSP Servers → 2/4 servers available
  ✓ Built-in MCP Servers → 2 built-in servers enabled
  ○ User MCP Configuration → No user MCP configuration found

Updates
────────────────────────────────────────
  ✓ Version Status → Pinned to version 3.0.0-beta.16

Summary
────────────────────────────────────────

  8 passed, 0 failed, 3 warnings, 3 skipped
  Total: 14 checks in 972ms

⚠ All systems operational with warnings.

Error Logs

Configuration

// oh-my-opencode.json
{
    "$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-opencode/master/assets/oh-my-opencode.schema.json",
    "agents": {
        "sisyphus": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "librarian": {
            "model": "zhipuai-coding-plan/glm-4.5-air"
        },
        "explore": {
            "model": "zhipuai-coding-plan/glm-4.5-air"
        },
        "oracle": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "frontend-ui-ux-engineer": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "document-writer": {
            "model": "zhipuai-coding-plan/glm-4.5-air"
        },
        "multimodal-looker": {
            "model": "zhipuai-coding-plan/glm-4.6V"
        },
        "atlas": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "OpenCode-Builder": {
            "model": "zhipuai-coding-plan/glm-4.7",
            "disable": true
        },
        "build": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "metis": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "momus": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "plan": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "prometheus": {
            "model": "zhipuai-coding-plan/glm-4.7"
        },
        "sisyphus-junior": {
            "model": "zhipuai-coding-plan/glm-4.7"
        }
    }
}

Additional Context

windows11 - wsl2:ubuntu 24.04.1

Operating System

Other

OpenCode Version

oh-my-opencode@3.0.0-beta.16

[justsisyphus]

Investigation Summary: Infinite Loop of Tool Calls

Issue Analysis

User reported infinite loop of tool calls when using /start-work with Prometheus planner on zhipuai models (GLM-4.7). The issue persisted even after manual intervention and using /ulw-loop to continue.

Root Cause Investigation

I've analyzed the codebase for loop prevention mechanisms:

1. Debounce Mechanism (FOUND)

• Location: src/hooks/atlas/index.ts (lines 389, 564-567)
• Implementation:
  • CONTINUATION_COOLDOWN_MS = 5000 (5-second cooldown)
  • Tracks lastContinuationInjectedAt per session
  • Prevents rapid continuation injections on session.idle events
  • Status: ✅ Exists and working

2. Max Iterations Limit (FOUND)

• Location: src/hooks/ralph-loop/index.ts (lines 287-299)
• Implementation:
  • Default: 100 iterations (configurable via default_max_iterations)
  • Stops loop when iteration >= max_iterations
  • Status: ✅ Exists for ralph-loop

3. Tool Call Limits (NOT FOUND)

• No explicit tool call count limit found in:
  • src/hooks/atlas/ - orchestration hook
  • src/features/background-agent/ - background task manager
  • src/tools/delegate-task/ - delegation system
• Status: ⚠️ Missing

Potential Causes

1. Model-Specific Behavior: GLM-4.7 may be generating tool calls in a pattern that:

   • Bypasses the 5-second debounce (rapid tool completions)
   • Doesn't trigger session.idle events properly
   • Generates tool calls faster than the debounce window

2. Tool Call Loop: The infinite loop appears to be:

   • Tool execution → Tool result → New tool call (same tool)
   • Repeats without reaching completion conditions
   • Not caught by ralph-loop's iteration limit (which tracks prompts, not tool calls)

3. Missing Safeguard: No per-session tool call counter exists to:

   • Detect rapid tool call sequences
   • Break loops after N consecutive tool calls
   • Differentiate between legitimate tool use and loops

Recommendations

1. Add Tool Call Counter

   • Track consecutive tool calls per session
   • Break loop if same tool called >5 times consecutively
   • Reset counter on user input or different tool

2. Enhance Debounce

   • Consider tool call frequency, not just idle events
   • Add exponential backoff for rapid tool calls

3. Add Safeguard in Atlas Hook

   • Monitor tool execution patterns
   • Detect and interrupt infinite loops before they escalate

Configuration Note

User is using zhipuai models (GLM-4.7) which may have different behavior patterns than Anthropic/OpenAI models. This could be model-specific.

---

Status: Valid bug - missing tool call loop prevention
Severity: High (can consume tokens indefinitely)
Affected: Users with non-Anthropic models, especially during /start-work with Prometheus

[code-yeongyu]

Hi @LeoNardo-LB,

Thanks for reporting this issue. To help us investigate, could you provide more details?

1. What do the screenshots show? - Could you describe what's happening in the images? (e.g., which tool is being called repeatedly, what messages are appearing)

2. Which specific tool was looping? - Was it delegate_task, Bash, Read, or something else?

3. What was the agent trying to do? - Was it attempting the same action over and over, or different actions?

4. Any error messages or logs? - Even partial logs would help understand the loop pattern.

5. How long did it run? - When you say "long period of operation", approximately how long before the loop started?

This information will help us determine whether this is a model-specific behavior (GLM-4.7) or a general issue in the orchestration logic.

Thanks!

[LeoNardo-LB]

@code-yeongyu

1. As you can see, the infinite loop in the picture occurred because the tool call tags were incorrectly output when invoking the tool (the <tool_call> XML tag). However, this was an issue with the previous version and hasn't occurred since then (but the <tool_call> XML tag still exists; when will this be fixed 😭)
2. It looks like the bash tool
3. It should be generating a work report
4. The logs should no longer be available as it was quite a long time ago
5. After proper planning, Atlas executed the work, which should have been near completion. However, the todo list was never marked as finished, causing the automatic prompt system to keep prompting to advance the work. The problem occurred after 7-8 loops