fix(compressor): preserve memory across repeated compression

[KeyArgo]

Summary

This PR implements the compression fixes from the May 30 context-compressor huddle:

• add calibrated rough-token accounting so compression decisions can learn from real token usage
• reserve a fixed output/summarizer buffer by capping the effective compression trigger at context_length - 7120
• soften the summary prefix so prior context remains usable for anaphora without reviving stale active-task hijacks
• add boundary tests for tool-call/tool-result tail splitting behavior
• replace repeated summary-of-summary rewriting with append-only summary segments, so each compressed turn range is summarized once and older facts are rendered forward rather than rewritten by the LLM

Why

Repeated prose re-summarization can degrade operational memory over long sessions. The append-only segment design keeps prior compressed ranges stable while still giving the summarizer read-only prior context for continuity.

Validation

• venv/bin/python -m pytest tests/agent/test_context_compressor.py tests/agent/test_context_compressor_huddle_fixes.py tests/agent/test_context_compressor_summary_continuity.py
  • 129 passed, 1 warning
• Deterministic session-compression eval against a real 22 MB Codex JSONL transcript:
  • typed ledger retention after 5 passes: 100.0%
  • typed ledger retention after 10 passes: 100.0%
  • typed ledger retention after 25 passes: 100.0%
  • typed ledger retention after 50 passes: 100.0%
  • recursive summary baseline after 5 passes: 53.4%

Notes

The fork branch is currently behind upstream main; this is opened as a draft so maintainers can review the compression design before merge-readiness cleanup.

[mxnstrexgl]

LGTM — automated review passed. No security, quality, or test coverage issues detected.

[teknium1]

Thanks for tackling a real long-session failure mode. Current main still recursively feeds PREVIOUS SUMMARY into the compressor prompt and replaces _previous_summary with the rewrite (agent/context_compressor.py:1956-1971, :2063-2071).

Problems

• agent/context_compressor.py:1399-1401 concatenates old segment text, while :1360-1385 renders every segment. The segment-count threshold does not bound rendered tokens, so repeated compaction can recreate the context-pressure problem this change is intended to avoid.
• The added state is reset in on_session_reset (:582-583) but the PR's on_session_end only clears _previous_summary (:648-662). _context_for_new_turns() reads _summary_segments first (:1421-1423), allowing cross-session prior context to leak on a reused compressor.

Suggested changes

• Define and test a hard rendered/prior-context token budget across many compactions; concatenating segment bodies is not a bound.
• Clear the new state at every real session boundary and add a reused-compressor regression test.
• Reconcile the reserve portion with current main's effective-input-budget logic in agent/context_compressor.py:995-1035.

This is an automated hermes-sweeper review.

[teknium1]

This concatenates every old segment body rather than reducing it. Because _render_summary_from_segments() emits all segment text, keeping five segment objects does not bound either the handoff or the next PRIOR CONTEXT prompt. Please enforce a real token bound and cover repeated passes.

[teknium1]

Please clear _summary_segments and _turns_seen in on_session_end too. That method is a separate lifecycle path and currently only clears _previous_summary; _context_for_new_turns() will prefer surviving segments and leak them into the next reused session.

[teknium1]

Thanks @KeyArgo — the quantified demonstration of recursive-summary degradation here (53.4% retention vs 100% ledger over 5+ passes on a real transcript) is the best evidence anyone has produced for the _previous_summary compounding problem, and it remains a documented known weakness.

Where each half stands on current main:

• Absolute output reserve: superseded — _compute_threshold_tokens already subtracts the real per-model max_tokens reservation (Context compaction trigger ignores the output-token reservation (custom-provider default max_tokens=65536 halves the usable input budget) #43547) with the degenerate-window guard, and feat(compression): absolute token threshold (compression.threshold_tokens) #69335 added a user-facing absolute compression.threshold_tokens cap.
• SUMMARY_PREFIX softening: contrary to a deliberate, documented decision — after the stale-task-resumption regressions, the contract is additive-clause-only (fix(compression): affirm tool use stays active in the compaction handoff prefix #66291); the tool-use-suppression concern was fixed additively.
• Tail-boundary pairing: covered by _sanitize_tool_pairs + boundary alignment helpers.
• Rough-token calibration: main reconciles rough-vs-real differently (update_from_response + should_defer_preflight_to_real_usage), and the CJK estimator fix (fix(compression): CJK-aware token estimates and stop re-copying stale head summaries #69302) removed the biggest skew.
• Append-only summary segments: STILL the live design direction — main does the recursive rewrite today and nothing merged addresses compounding.

Closing this vehicle (stale base, mixed scope), but the segments half is explicitly invited back as a focused rebase: segments-only, hard rendered-token budget, session-boundary clearing. Credit for the retention analysis — it will be the benchmark for whatever lands.