fix(compression): break the protected-tail dead-end on oversized messages

[teknium1]

Summary

Breaks the #61932 "Cannot compress further" dead-end: when the protected tail (token budget + the 8-message floor) alone holds oversized recent messages, compress_start >= compress_end made every compression pass a pure no-op — the transcript stayed over the context window and the retry loop exhausted. Root cause: _prune_old_tool_results passed the uncapped protect_last_n (default 20) into the prune boundary while tail-cut already bounds it at _MAX_TAIL_MESSAGE_FLOOR, and there was no path at all that could shrink oversized tool bodies inside the protected region.

Changes

• agent/context_compressor.py:
  • Cap the prune path's message-count floor at _MAX_TAIL_MESSAGE_FLOOR (same bound tail-cut already uses at _find_tail_cut_by_tokens), so a default protect_last_n=20 cannot freeze a bulky recent tool run against pruning.
  • New Pass 4 pressure demotion: when the protected region still exceeds the soft tail budget (tail_token_budget * 1.5), demote large completed tool/file bodies inside that region — oldest first, keeping a _PRESSURE_KEEP_RECENT_MESSAGES=3 recent floor and the active user ask verbatim. Escalation demotes everything except the single newest tool body; absolute last resort summarizes even that (one 200KB read can exceed the soft budget alone). Demotion rewrites content in place and never touches tool_call_ids, so tool_call/tool_result pairing is preserved.
  • Pass 2/3 refactored into _demote_tool_result_at / _truncate_tool_call_args_at helpers reused by the pressure pass (behavior unchanged outside pressure).
• tests/agent/test_protected_tail_pressure_61932.py: contributor's 5 tests (pressure demotion, last-resort branch incl. the sweeper-requested tool_call_id assertion, full-compress escape, light-tail no-op) + a new regression test pinning the exact [Regression]: auto-compression still reaches Cannot compress further when protected recent tail is large (Desktop/openai-codex) #61932 dead-end shape: head + an 8-message all-oversized tail that previously returned unchanged now compresses below the context window in one pass, with no orphaned tool calls/results.

Validation

	Before (origin/main)	After
All-oversized protected tail, 128K ctx (200,248 tok)	150,330 tok after compress — still over window, dead-end	27K tok (~87% reclaimed), under window in one pass
bash scripts/run_tests.sh tests/agent/ -q -k 'tail or oversized'	—	137 passed, 0 failed
tests/agent/test_protected_tail_pressure_61932.py	new-test counterfactual vs main: dead-end reproduced	6 passed, 0 failed
tests/agent/test_context_compressor.py + tool-pruning siblings	—	207 passed, 0 failed

Credit

Salvaged from #61952 by @giggling-ginger (first submitted Jul 10) — cherry-picked with authorship preserved. Independent approaches to the same bug class, credited: #68377 by @trac3r00 (hard head+tail char cap on tail tool results), #67108 by @Kenmege (tail token budget hard-ceiling override of the message-count floor), #43311 by @josenaicipa (pre-compaction tool-output truncation subsystem for huge tool logs). #61952 was chosen as the base for its correct scoping (targets the exact protected-tail no-op path, data-preserving summaries via _summarize_tool_result, keeps the active ask verbatim) and its earliest submission.

Fixes #61932.

Infographic

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 7cf254d}

all good!