feat(compaction): preserve recent user messages during compression (Codex-style)

[hermes-agent-dhabibi]

Enhanced Compaction (Codex-style)

Background

We studied OpenAI Codex's compact.rs to understand how they handle context compaction. Their key insight: don't just summarize — preserve the user's actual messages alongside the summary.

The Codex Approach

Codex's compaction keeps real user messages from the compressed region (up to 20K tokens, newest-first) and places them before the summary. This ensures the model sees user intent as the primary signal, with the summary as supplementary context.

Before:

[head] → [summary] → [tail]

After:

[head] → [preserved user messages] → [summary] → [tail]

Changes

• Add user_message_max_tokens constructor param (default 20K, matching Codex)
  • Configurable: set to 0 to disable user message preservation entirely
• Add _collect_user_messages() — extracts user messages from the compressed region, filtering out previous compaction summaries
• Modify compress() to insert preserved messages between head and summary
• Add tests for new functionality

[teknium1]

Thanks for exploring a real compaction-quality gap. Current main still summarizes the middle window without retaining its user turns verbatim (agent/context_compressor.py:2885-2935), so the idea remains relevant.

Problems

• agent/context_compressor.py:835 emits preserved messages as raw user turns before the summary-role logic. Multiple preserved messages can become consecutive user roles, violating the invariant enforced by current main (agent/context_compressor.py:3063-3080; tests/agent/test_context_compressor.py:2002-2007).
• The 20K preserved-message budget is inserted after tail selection, outside the token budget calculated at agent/context_compressor.py:2860-2865; it can eliminate compaction savings before the post-assembly metric runs (agent/context_compressor.py:3141-3149).
• _collect_user_messages() calls .strip() on content (agent/context_compressor.py:692 in the PR), which is unsafe for multimodal list content supported by the current assembly path.

Suggested changes

• Preserve selected text inside a single reference-only handoff or pass it to the summarizer, rather than adding raw user-role messages. The linked #499 discussion records the same concern and suggests summarizer-prompt injection.
• Rework against current compression assembly, then cover alternation, multimodal content, and bounded post-compression size.

Automated hermes-sweeper review.

[teknium1]

content is not guaranteed to be a string: the current compressor supports multimodal list content during assembly. Calling .strip() here will raise for such a user message; normalize through the existing content helpers before filtering or truncating.

[teknium1]

Appending each preserved item as a raw user message bypasses the summary collision logic below. Two user turns that were originally separated by a summarized assistant turn become adjacent here, violating the strict role-alternation invariant. Preserve them in one reference-only summary/handoff payload, or include them in the role-selection logic.

[teknium1]

Thanks @hermes-agent-dhabibi for surfacing Codex's compact.rs approach — the underlying insight (user intent as the primary preservation signal) is real and influenced where main ended up.

The acute intent-loss case is now covered: the latest real user and assistant turns are guaranteed to anchor the protected tail (_ensure_last_user_message_in_tail / _ensure_last_assistant_message_in_tail in agent/context_compressor.py), and every summary carries a deterministic, grounded ## Historical Task Snapshot of the latest user ask.

The verbatim multi-user-row mechanism as implemented can't land against current invariants: raw consecutive user rows violate the role-alternation contract, the 20K insert bypasses the tail token budget, .strip() crashes on multimodal list content, and the provenance machinery that has since landed (#69292/#69860 start-of-content markers, zero-user validation) treats raw user-row injection into compressed history as a hazard class.

If you want to pursue the surviving half (preserving N>1 user messages), the viable direction is feeding preserved user text to the summarizer input or a single reference-only handoff block — happy to review a reworked version against current assembly. Closing with credit for the design reference.