seb/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(compression): raise compaction trigger to 85% for gpt-5.5 on Codex OAuth (#40957)

Browse Source 
The ChatGPT Codex OAuth backend hard-caps gpt-5.5 at a 272K context window
(verified live: a ~330K-token request to chatgpt.com/backend-api/codex/responses
is rejected with context_length_exceeded while ~250K succeeds; the same slug
exposes 1.05M on the direct OpenAI API / OpenRouter and 400K on Copilot). At the
default 50% trigger, auto-compaction fires at ~136K — half the usable window.

Raise the trigger to 85% (~231K) on this exact route only, gated by a new
compression.codex_gpt55_autoraise config flag (default true). When it fires,
emit a one-time notice (CLI inline print + gateway status_callback replay) with
the exact opt-back-out command. gpt-5.5 on any other provider keeps the user's
global threshold.

- _is_codex_gpt55() matches the 5.5 family only on provider=openai-codex
- _compression_threshold_for_model() now provider-aware + opt-out param
- config key + _config_version bump (27->28) for backfill
- docs + tests (40 cases in test_arcee_trinity_overrides.py)
This commit is contained in:
Teknium
2026-06-07 01:40:50 -07:00
committed by GitHub
parent 2d099fed1e
commit 0524c9b34e
5 changed files with 218 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
agent/agent_init.py
+63 -2
View File
@@ -68,6 +68,24 @@ def _ra():
return run_agent
def _build_codex_gpt55_autoraise_notice(autoraise: Dict[str, float]) -> str:
"""Build the one-time notice shown when Codex gpt-5.5 raises compaction.
``autoraise`` is ``{"from": <old_ratio>, "to": <new_ratio>}``. The same
text is printed inline for CLI users and replayed via ``status_callback``
for gateway users, so it must be self-contained and include the exact
opt-back-out command.
"""
from_pct = int(round(autoraise["from"] * 100))
to_pct = int(round(autoraise["to"] * 100))
return (
f"ℹ Codex gpt-5.5 caps context at 272K, so auto-compaction was raised "
f"to {to_pct}% (from {from_pct}%) to use more of the window before "
f"summarizing.\n"
f" Opt back out: hermes config set compression.codex_gpt55_autoraise false"
)
def _normalized_custom_base_url(value: Any) -> str:
if not isinstance(value, str):
return ""
@@ -1240,11 +1258,41 @@ def init_agent(
if not isinstance(_compression_cfg, dict):
_compression_cfg = {}
compression_threshold = float(_compression_cfg.get("threshold", 0.50))
# Per-model/route compaction-threshold override. Codex gpt-5.5 raises to
# 85% (the Codex backend caps the window at 272K, so the default 50% would
# compact at ~136K — half the usable context). Gated by an opt-out config
# flag so the user can fall back to the global threshold; when the override
# fires we stash a one-time notification (replayed on the first turn) that
# tells the user what changed and how to revert.
_codex_gpt55_autoraise = str(
_compression_cfg.get("codex_gpt55_autoraise", True)
).lower() in {"true", "1", "yes"}
agent._compression_threshold_autoraised = None
try:
from agent.auxiliary_client import _compression_threshold_for_model as _cthresh_fn
_model_cthresh = _cthresh_fn(agent.model)
from agent.auxiliary_client import (
_compression_threshold_for_model as _cthresh_fn,
_is_codex_gpt55 as _is_codex_gpt55_fn,
)
_model_cthresh = _cthresh_fn(
agent.model,
agent.provider,
allow_codex_gpt55_autoraise=_codex_gpt55_autoraise,
)
if _model_cthresh is not None:
_prev_threshold = compression_threshold
compression_threshold = _model_cthresh
# Notify only for the Codex gpt-5.5 autoraise (the Arcee Trinity
# override is a long-standing silent default). Skip the notice when
# the user's global threshold already meets/exceeds the raised
# value, since nothing actually changed for them.
if (
_is_codex_gpt55_fn(agent.model, agent.provider)
and _model_cthresh > _prev_threshold + 1e-9
):
agent._compression_threshold_autoraised = {
"from": _prev_threshold,
"to": _model_cthresh,
}
except Exception:
pass
compression_enabled = str(_compression_cfg.get("enabled", True)).lower() in {"true", "1", "yes"}
@@ -1621,11 +1669,24 @@ def init_agent(
print(f"📊 Context limit: {agent.context_compressor.context_length:,} tokens (compress at {int(compression_threshold*100)}% = {agent.context_compressor.threshold_tokens:,})")
else:
print(f"📊 Context limit: {agent.context_compressor.context_length:,} tokens (auto-compression disabled)")
# One-time notice when the Codex gpt-5.5 autoraise kicked in, with the
# exact opt-back-out command. Printed inline at startup for CLI users;
# gateway users get the same text replayed via _compression_warning on
# turn 1 (set below, after the warning slot is initialized).
_autoraise = getattr(agent, "_compression_threshold_autoraised", None)
if _autoraise and compression_enabled:
print(_build_codex_gpt55_autoraise_notice(_autoraise))
# Check immediately so CLI users see the warning at startup.
# Gateway status_callback is not yet wired, so any warning is stored
# in _compression_warning and replayed in the first run_conversation().
agent._compression_warning = None
# Gateway parity for the Codex gpt-5.5 autoraise notice: the startup print
# above only reaches the CLI, so stash the same text here to be replayed
# through status_callback on the first turn (Telegram/Discord/Slack/etc.).
_autoraise = getattr(agent, "_compression_threshold_autoraised", None)
if _autoraise and compression_enabled:
agent._compression_warning = _build_codex_gpt55_autoraise_notice(_autoraise)
# Lazy feasibility check: deferred to the first turn that approaches the
# compression threshold. Running it eagerly here costs ~400ms cold (network
# probe of the auxiliary provider chain + /models lookup) on every agent