[Bug]: Infinite compaction / continuation loop prevents coding from starting

[chelebyy]

Prerequisites

• I will write this issue in English (see our Language Policy)
• I have searched existing issues to avoid duplicates
• I am using the latest version of oh-my-opencode
• I have read the documentation or asked an AI coding agent with this project's GitHub URL loaded and couldn't find the answer

Bug Description

When using oh-my-openagent, the agent sometimes gets stuck in a compaction / continuation loop and never starts actual coding.

I tested this with multiple models and also with a single agent setup (Hephaestus only). The issue still happens, so it does not seem to be specific to a particular model or multi-agent orchestration.

Observed behavior:

• The agent keeps generating continuation / compaction style content repeatedly
• It does not move on to implementation
• It appears to re-enter the same summarization / continuation pattern over and over
• Reinstalling the tool did not fix it

Important notes:

• I also tested different agents previously (including Sisyphus and Hermes), but I can reproduce the same issue even with only Hephaestus enabled
• I tested with multiple models and saw the same behavior
• I removed and reinstalled multiple times, but the problem remains
• I run wsl and windows

This looks like a recursive continuation / compaction loop rather than a model-specific issue.

I searched existing issues before posting this...

loop.txt

Steps to Reproduce

1. Install and run oh-my-openagent on the latest version
2. Use a normal coding task that includes enough context/history for compaction or continuation behavior to become relevant
3. Run the task with a single agent setup (for example, Hephaestus only)
4. Wait until the agent starts generating compaction / continuation style outputs
5. Observe that it keeps repeating that pattern instead of starting code changes

I was able to reproduce this across multiple models and also after reinstalling.

Expected Behavior

The agent should compact context if needed, then continue the original coding task and begin implementation instead of repeatedly generating continuation / summary-style outputs.

Actual Behavior

The agent repeatedly produces continuation / compaction style output and never proceeds to actual coding or file modification.

Doctor Output

oMoMoMoMo Doctor

 ✓ System OK (opencode 1.2.5 · oh-my-opencode 3.6.0)

Error Logs

Agent output showing repeated compaction / continuation loop:

[search-mode]
MAXIMIZE SEARCH EFFORT...

[analyze-mode]
ANALYSIS MODE...

SYSTEM DIRECTIVE: OH-MY-OPENCODE - COMPACTION CONTEXT
...
Provide a detailed prompt for continuing our conversation above.

(This pattern repeats continuously and never proceeds to coding)

Configuration

{
  "$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-openagent/dev/assets/oh-my-opencode.schema.json",
  "agents": {
    "sisyphus": {
      "model": "openai/gpt-5.4",
      "variant": "medium"
    },
    "hephaestus": {
      "model": "openai/gpt-5.3-codex",
      "variant": "medium"
    },
    "oracle": {
      "model": "openai/gpt-5.4",
      "variant": "high"
    },
    "librarian": {
      "model": "google/gemini-3-flash-preview"
    },
    "explore": {
      "model": "opencode/gpt-5-nano"
    },
    "multimodal-looker": {
      "model": "openai/gpt-5.3-codex",
      "variant": "medium"
    },
    "prometheus": {
      "model": "openai/gpt-5.4",
      "variant": "high"
    },
    "metis": {
      "model": "openai/gpt-5.4",
      "variant": "high"
    },
    "momus": {
      "model": "openai/gpt-5.4",
      "variant": "xhigh"
    },
    "atlas": {
      "model": "openai/gpt-5.4",
      "variant": "medium"
    },
    "sisyphus-junior": {
      "model": "openai/gpt-5.4",
      "variant": "medium"
    }
  },
  "categories": {
    "visual-engineering": {
      "model": "google/gemini-3.1-pro-preview",
      "variant": "high"
    },
    "ultrabrain": {
      "model": "openai/gpt-5.4",
      "variant": "xhigh"
    },
    "deep": {
      "model": "openai/gpt-5.3-codex",
      "variant": "medium"
    },
    "artistry": {
      "model": "google/gemini-3.1-pro-preview",
      "variant": "high"
    },
    "quick": {
      "model": "openai/gpt-5.4-mini"
    },
    "unspecified-low": {
      "model": "openai/gpt-5.3-codex",
      "variant": "medium"
    },
    "unspecified-high": {
      "model": "openai/gpt-5.4",
      "variant": "high"
    },
    "writing": {
      "model": "google/gemini-3-flash-preview"
    }
  }
}

Additional Context

This issue blocks actual development because the agent never reaches the coding phase.

It keeps looping in compaction/continuation instead of executing tasks.

I also tried disabling multi-agent setups and testing with only Hephaestus, but the issue still occurs.

Operating System

Windows

OpenCode Version

1.2.27

[code-yeongyu]

[sisyphus-bot]

We're sorry you ran into this issue. We've identified the root cause of the infinite compaction/continuation loop.

Root Cause

The preemptive-compaction hook lacks a cooldown mechanism between successive compactions. Here's the problematic flow:

1. Agent generates output with high context usage (>78%)
2. preemptive-compaction triggers and calls session.summarize()
3. The compactionContextInjector injects a detailed prompt asking for context sections
4. Agent generates a summary response
5. message.updated event fires, which clears the compactedSessions set (line 170)
6. If context is still high, preemptive-compaction triggers again immediately
7. Loop repeats without any delay between compactions

The issue is in /src/hooks/preemptive-compaction.ts:

• Line 90-91: Checks if session is in compactedSessions to prevent duplicate compactions
• Line 131: Adds session to compactedSessions after compaction
• Line 170: Clears compactedSessions on every message.updated event

Without a cooldown, the hook can trigger multiple times in rapid succession, causing the agent to get stuck in a summarization loop instead of proceeding to actual work.

Fix Approach

Add a cooldown mechanism to preemptive-compaction.ts:

1. Track lastCompactionTime per session (similar to how todo-continuation-enforcer does it)
2. Add a constant PREEMPTIVE_COMPACTION_COOLDOWN_MS (suggest 30-60 seconds)
3. In toolExecuteAfter, check if enough time has passed since last compaction before triggering
4. This prevents rapid successive compactions while still allowing compaction when truly needed

The fix should be minimal and follow the existing pattern used in todo-continuation-enforcer/idle-event.ts (lines 132-141).

Severity

HIGH - This blocks users from using the agent for any substantial work when context window gets high.

[chelebyy]

Thanks for identifying the root cause! The cooldown mechanism sounds like the perfect fix. I'll keep this issue open until the fix is merged and I can verify it on my end. Looking forward to the update!

[code-yeongyu]

Fixed on dev branch. See commit history.