feat: make context compaction preserve recent turns

[BestLemoon]

Summary

• add a separate protect_recent_n floor so the newest turns stay verbatim during compaction
• add request-time micro-compaction to prune old tool outputs before triggering full compression
• add a heavier repeated-compaction handoff mode so later summaries compress more aggressively under sustained pressure

Testing

• source venv/bin/activate && python -m pytest tests/agent/test_context_compressor.py -q

[teknium1]

Thanks for the detailed compaction work. The title-level recent-turn problem has since been addressed on current main by aec38855b5792e4a912d527b64df1a43cbf09c90 (fix(agent): preserve recent turns during compression), which added a bounded preserved tail and regression coverage in tests/agent/test_context_compressor.py:2517.

Problems

• The proposed per-request _prepare_messages_for_request() call (run_agent.py, diff RIGHT line 7793) prunes retained history before every API request. That mutates the cached prefix outside the current full-compaction path (agent/conversation_loop.py:1007-1058), conflicting with the prompt-cache invariant in AGENTS.md:19-23.
• compression.protect_recent_n is added (hermes_cli/config.py, diff RIGHT line 357) without corresponding tests or documentation, despite compression settings being documented in website/docs/user-guide/configuration.md:737-745.

Suggested changes

• Use the merged bounded-tail behavior as the baseline and scope any remaining micro/heavy-compaction idea as a cache-safe compression design on current main.
• Add end-to-end coverage for config propagation and user/assistant/tool-group preservation before introducing a separate floor.

Automated hermes-sweeper review.

[teknium1]

This runs a history-mutating tool-result prune before every model request, not only at an explicit compression boundary. That invalidates the cached conversation prefix during the tool loop; please keep retained-message mutation within the full compaction path, with its session/cache bookkeeping.

[teknium1]

Thanks @BestLemoon — this PR correctly identified the preserve-recent-turns gap before it was fixed on main. Where each piece stands now:

• Recent-turn floor: landed via aec38855b5 — _MAX_TAIL_MESSAGE_FLOOR = 8 plus the token-budget tail with 1.5x soft ceiling in _find_tail_cut_by_tokens (agent/context_compressor.py). This is your protect_recent_n floor, generalized and capped.
• Heavy repeated-compaction mode: the sustained-pressure case is now handled by protected-tail pressure demotion (fix(compression): break the protected-tail dead-end on oversized messages #69830) plus the per-turn compaction cap compression.max_attempts (fix(compression): make the attempt cap config-driven and enforce it at every compression site #69315), and head protection decays to system-prompt-only after the first pass.
• Per-request micro-compaction: we can't take this mechanism — mutating the message prefix before every API call breaks the per-conversation prompt-cache invariant (see AGENTS.md), which is a hard design constraint here.

Closing as superseded — credit for early identification of the recent-turn loss problem. Thanks for the contribution!