Problem

When Clawdbot runs in embedded mode, extensionRunner.initialize() is never called. This means ExtensionRunner.getModel stays at its default () => undefined.

When the compaction safeguard extension runs, it reads ctx.model which calls getModel() → returns undefined. The safeguard then returns a fallback summary with no actual content:

Summary unavailable due to context limits. Older messages were truncated.

This effectively discards all conversation history on every compaction. It also causes cascading re-compactions: the empty summary is only ~50 tokens, so the next tool call immediately fills the context again, triggering another compaction, which again produces an empty summary.

In a single session I observed 22 consecutive compactions, all with empty summaries.

Root Cause

ExtensionRunner.getModel defaults to () => undefined and is only set to the correct getter during initialize(). The three built-in pi-coding-agent modes (interactive, print, rpc) all call initialize(), but Clawdbot's embedded runner (compact.js, run/attempt.js) does not.

In AgentSession._runAutoCompaction(), the model and API key are correctly validated via this.model before emitting the session_before_compact event. So the model is available -- just not through the extension context's ctx.model getter.

Fix

When ctx.model or apiKey is unavailable, return undefined instead of the fallback summary. This lets the built-in compaction code in AgentSession handle summarization, which correctly accesses the model via this.model (agent.state.model) and produces a proper summary.

Alternative

The deeper fix would be to ensure extensionRunner.initialize() is called in the embedded runner flow, but that's a larger change. This fix is safe and minimal -- the built-in compaction path is well-tested and produces correct summaries.

Greptile Overview

Greptile Summary

This PR adjusts the session_before_compact compaction-safeguard extension so it no longer emits an “empty fallback” compaction result when ctx.model or its API key is unavailable (notably in embedded runner mode where extensionRunner.initialize() is not invoked). Instead it returns undefined, allowing AgentSession’s built-in compaction path to run with a valid model sourced from the session state, preventing repeated compactions caused by near-empty summaries.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk.
• Change is narrowly scoped to early-return behavior in the compaction safeguard; it removes a harmful fallback output in cases where the extension lacks model/apiKey context and defers to the core, already-validated compaction logic. No API surface changes, and failure path for real summarization errors is unchanged.
• No files require special attention

_{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!}

Context used:

• Context from dashboard - CLAUDE.md (source)
• Context from dashboard - AGENTS.md (source)