fix(codex): proactive context-pressure retire + handoff on app-server runtime (#36801)

[ahmadalzaro1]

Fixes #36801.

Problem

The Codex app-server runtime (run_codex_app_server_turn) hands each turn to a persistent codex app-server subprocess that owns its own conversation thread. Unlike the chat-completions path — which runs a preflight compaction check in build_turn_context() before every model call — the codex path had no proactive compaction. Sessions grow unbounded inside the subprocess until codex hits its hard context ceiling and silently hard-resets, losing all context. The failure is silent: the operator only notices after context is already gone (and on metered Codex seats the retained context is also a per-turn budget burn).

Compressing Hermes' messages list (the naive fix) does not help — codex holds the real history internally; run_turn() only forwards the new user message. The only lever is to retire the thread and reseed a fresh one.

Fix

A proactive, post-turn guard on the codex path:

1. After _record_codex_app_server_usage() — where the real prompt-token count and codex-reported model_context_window are both known — compute prompt_tokens / context_window.
2. When it reaches compression.codex_retire_threshold (new config, default 0.85, 0 disables), build a structured handoff summary from Hermes' projected messages via the auxiliary summary model (not codex — so the summarization itself can't overflow the already-full window), then retire the codex thread (close() + drop) and stash the summary.
3. On the next turn, the respawned (blank) codex thread is reseeded by prepending the summary to the first user input — so Goal/Discoveries/Files survive instead of a cold history=0 reset.

This implements the flow proposed in the issue (extract usage → proactive threshold → handoff-on-retire) plus the structured-summary handoff suggested in the thread.

Notes

• Opt-out / tunable via compression.codex_retire_threshold (0 disables).
• Reuses existing summary machinery via a small public ContextCompressor.build_handoff_summary() seam.
• Best-effort: if summarization is unavailable it still retires (announced via logger.warning) rather than ride an over-full window into a hard reset.
• No behavior change on the chat-completions path, or when the codex window/usage is unknown.

Tests

New tests/agent/transports/test_codex_proactive_compaction.py (10 cases): over/under threshold, summary-less retire, disabled (flag + threshold=0), unknown window, already-retired session, and the build_handoff_summary seam (empty / delegate / error-swallow). All pass; existing test_context_compressor.py + test_codex_app_server_session.py (154 tests) remain green.

[ahmadalzaro1]

Ready for review. This is rebased onto current main (5ff11a689) and conflict-free.

The codex app-server path still has no proactive compaction, so long sessions keep growing until codex hits its hard ceiling and silently resets (#36801). This adds a post-turn guard: when prompt tokens reach compression.codex_retire_threshold (new config, default 0.85, 0 disables), it summarizes via the auxiliary model, retires the codex thread, and reseeds the next turn with the summary instead of a cold history=0 reset.

10 new unit tests; existing test_context_compressor.py + test_codex_app_server_session.py remain green. Happy to adjust the default threshold or split this down if that's easier to review.

[teknium1]

Thanks for the focused investigation and tests. This is now superseded by the Codex-native compaction implementation on current main.

• agent/conversation_compression.py:1018-1126 routes codex_app_server compression to the app server's own thread/compact/start, targeting the actual Codex-owned thread rather than Hermes' local transcript mirror.
• hermes_cli/config.py:1455-1462 provides compression.codex_app_server_auto, including hermes mode for Hermes-threshold-initiated thread compaction and native compaction tracking.
• The Codex app-server runtime: long sessions grow unbounded → hard context reset (no proactive compaction) #36801 maintainer update identifies PR feat(compression): route Codex app-server compaction through thread/compact/start (#36801) #60114 / main commit 87b65e24a as the fix for this exact failure mode; the implementation was introduced by d1c8c03416d7c6780d31ce2371c5b0777f10afaf and is included in v2026.7.7.

This is an automated hermes-sweeper review.