Summary

Compaction can truncate a large amount of session history even when summary generation fails, leaving only the fallback summary text:

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

This causes catastrophic context loss (“amnesia”) within the same sessionKey .

I believe the issue is in the compaction safeguard extension: when summarization fails / model is missing / apiKey is missing, it still returns a compaction result with firstKeptEntryId , allowing truncation to proceed. The fallback summary is hard-coded and not a usable summary.
Code pointers:

• Fallback summary constant: src/agents/pi-extensions/compaction-safeguard.ts (FALLBACK_SUMMARY)
• Error path explicitly says “truncating history” on summarization failure: src/agents/pi-extensions/compaction-safeguard.ts (catch block logs: Compaction summarization failed; truncating history )

Steps to reproduce

1. Start a long-running conversation until compaction triggers (very large token history).
2. Make summary generation fail during compaction (e.g. invalid/missing API key, provider outage, forced timeout, or otherwise making session_before_compact summarization throw).
3. Continue the conversation and observe that history is truncated while the summary becomes the fallback “Summary unavailable…” text.
   (Alternative deterministic path)
4. Run with a configuration/environment where ctx.model or apiKey cannot be resolved during compaction.
5. Trigger compaction.
6. Observe compaction still proceeds with fallback summary and truncation.

Expected behavior

If summary generation fails (or model/apiKey is unavailable), compaction should NOT truncate history.

Possible acceptable fallbacks:

• Abort compaction and keep original history intact.
• Retry summarization with a more conservative strategy.
• Degrade by keeping more raw history (instead of truncating to a tiny tail) until a valid summary is produced.

Actual behavior

• Summary becomes: "Summary unavailable due to context limits. Older messages were truncated."
• A massive drop in retained context occurs within the same sessionKey (example evidence below shows tokensBefore dropping from ~181k to ~378).
• The assistant appears to “forget everything” because old messages are truncated and no meaningful summary is preserved.

Environment

• Version: 2026.1.24-3
• OS: ubuntu
• Install method: npm

Logs or screenshots

Compaction event 1

sessionKey: agent:main:feishu:dm:ou_2f460687acb188f14f3c8d6af3fd30e7
tokensBefore: 181134
summary: "Summary unavailable due to context limits. Older messages were truncated."

Compaction event 2

tokensBefore: 378
summary: "Summary unavailable due to context limits. Older messages were truncated."