Context Compression Improvements

Tracking issue for seven independent improvements to agent/context_compressor.py that improve information retention, efficiency, and robustness during context compression.

PRs — Information Quality (merged from original batch)

• feat(compressor): smart tool output collapse during context pruning #9661 — Smart tool output collapse: Replace generic placeholder with informative 1-line summaries ([terminal] ran \npm test` -> exit 0, 47 lines`)
• feat(compressor): structured action-log summary format #9663 — Structured action-log format: Redesign summary template with numbered actions, [tool: name] tags, and Active State section
• feat(compressor): preserve user messages verbatim during compression #9665 — Preserve user messages: Append verbatim user messages from compressed turns to the summary

PRs — Robustness & Efficiency (new batch)

• fix(compressor): anti-thrashing protection for ineffective compression loops #9674 — Anti-thrashing protection: Detect ineffective compression loops (<10% savings) and back off after 2 consecutive failures
• feat(compressor): implement preflight compression check #9675 — Preflight compression check: Override should_compress_preflight() to avoid paying for a full API call when already over the limit
• feat(compressor): deduplicate tool results + prune tool_call arguments #9677 — Dedup + argument pruning: Deduplicate identical tool results (same file read 5x) + truncate large tool_call arguments in assistant messages
• fix(compressor): hardening — max_tokens cap, multimodal safety, note fix, adaptive cooldown #9678 — Compressor hardening: Summary max_tokens cap (2x→1.3x), multimodal content safety, compression note fix, adaptive failure cooldown (600s→60s for transient errors)

Benchmark Summary

Smart tool collapse (PR #9661):

User message preservation (PR #9665):

Anti-thrashing (PR #9674): Prevents infinite compression loops that burn LLM summary calls with <10% savings each.

Dedup (PR #9677): 5 reads of the same 10KB file → 50KB in context → deduped to 10KB + 4 stubs.

Hardening (PR #9678): Summary model now gets 1.3x budget instead of 2x (cost saving). Transient errors cool down for 60s instead of 600s.

Independence

All 7 PRs are independent — each can be merged separately. They touch different methods within context_compressor.py:

• feat(compressor): smart tool output collapse during context pruning #9661: _prune_old_tool_results() + new _summarize_tool_result()
• feat(compressor): structured action-log summary format #9663: _generate_summary() prompt template
• feat(compressor): preserve user messages verbatim during compression #9665: New _extract_user_messages() + compress() Phase 3b
• fix(compressor): anti-thrashing protection for ineffective compression loops #9674: should_compress() + compress() savings tracking + on_session_reset()
• feat(compressor): implement preflight compression check #9675: New should_compress_preflight() override
• feat(compressor): deduplicate tool results + prune tool_call arguments #9677: _prune_old_tool_results() dedup + arg pruning passes
• fix(compressor): hardening — max_tokens cap, multimodal safety, note fix, adaptive cooldown #9678: _generate_summary() max_tokens, _prune_old_tool_results() multimodal, compress() note, failure cooldown

Related issues addressed

• Context compaction is too abrupt; propose rolling compaction for smoother continuity #9561 — compression too abrupt / repeated ineffective passes → fix(compressor): anti-thrashing protection for ineffective compression loops #9674
• Feature: Context Compaction Quality Overhaul — Handoff-Oriented Prompts, User Message Preservation, and Configurable Compaction (inspired by Codex CLI) #499 — tool details truncated, quality bottleneck, no dedup → feat(compressor): smart tool output collapse during context pruning #9661, feat(compressor): structured action-log summary format #9663, feat(compressor): deduplicate tool results + prune tool_call arguments #9677
• [Bug]: Context compression drops recent conversation记忆 and injects cross-session fragments #9413, [Bug]: Iterative context compaction summary keeps completed topics alive and overrides the current active topic #9631 — iterative summary quality → feat(compressor): structured action-log summary format #9663
• [Bug]: Context compression causes incoherent responses on small-context models #7133 — incoherent after compression on small models → fix(compressor): anti-thrashing protection for ineffective compression loops #9674 (prevents re-compression loops)