fix(errors): Codex entitlement/retired-model 400 with a top-level detail body is reported as a model rejection, not context overflow (#81558, #106475, salvage #100783)

[teknium1]

A Codex entitlement/retired-model 400 {"detail": "The '<model>' model is not supported when using Codex with a ChatGPT account."} on a large session is now reported as a descriptive request rejection with the provider's own text, instead of being mistaken for context overflow and burning the compression loop.

• agent/error_classifier.py::_body_message_candidates now yields a top-level detail (string, or nested {"message": ...}) after the existing OpenAI / flat / litellm-Bedrock keys, so every consumer of the body text sees it: the generic-400 "bare body + large session ⇒ overflow" heuristic in _classify_400, _extract_message (the user-visible message), and the Codex masked-replay check.
• The pydantic list shape of detail is untouched — it is still read only by _oversized_message_content_rejection.
• Because is_client_error excludes overflow, the existing entitlement marker from fix(agent): unentitled Codex primary + fallback no longer oscillate or announce a false restore (#106475, salvage #106482) #106549 (agent/turn_api_error.py → fallback_cooldown._mark_entitlement_rejected_model) never ran on large sessions; with the body read correctly it now does.
• One invariant test in tests/agent/test_error_classifier.py (string + nested shapes not overflow, no compression, fallback, message is the provider text; control: a bare Error body on the same session still compresses). Red on base, green on this head.

Live probe (real openai SDK _make_status_error_from_response on an httpx.Response(400) carrying the exact #81558 body → public classify_api_error, provider=openai-codex, 223 messages / ~110k tokens, temp HERMES_HOME):

case	base 8df0a037	this PR
entitlement {"detail": …} 400, large session	context_overflow, should_compress=True, message Error code: 400 - {'detail': …}	format_error, should_compress=False, should_fallback=True, message The 'gpt-5.5-codex' model is not supported when using Codex with a ChatGPT account.
same body, small session (control)	format_error	format_error (unchanged)
{"detail": {"message": …}}, large session	context_overflow	format_error
bare {"error": {"message": "Error"}}, large session (heuristic control)	context_overflow	context_overflow (unchanged)

Root cause: _body_message_candidates never read the FastAPI-style top-level detail key, so a long, descriptive rejection counted as an empty body and tripped the large-session overflow heuristic.

Tests: tests/agent/test_error_classifier.py tests/agent/test_auxiliary_error_classification.py tests/agent/test_entitlement_fail_closed.py → 177 passed.

Refs #81558 (the commenter's "400 + large session ⇒ Context length exceeded … Cannot compress further" loop; the OP's wire-format hypothesis is contradicted in-thread and stays unreproduced)
Refs #106475 (already closed by #106549; this makes that fix reachable on large sessions)
Salvages #100783 (@i-Hun) — slim re-port of its detail-body half onto the current rule-table classifier; co-authored.

Dropped hunks

• fix(error_classifier): read FastAPI-style {"detail": ...} error bodies; recognise session/weekly usage limits #100783's "session limit" / "weekly limit" usage-limit patterns and the bare "resets " transient signal: a different error class (subscription usage walls), not part of this fix. Re-probed on this head: a 429 carrying You've hit your session limit · resets 2:20pm already classifies rate_limit (retryable) via the status path; the same text as a 400 is format_error. Left for its own PR if a real 400-shaped report appears.
• fix(error_classifier): read FastAPI-style {"detail": ...} error bodies; recognise session/weekly usage limits #100783's _detail_message list flattening (loc: msg (type)): the list shape is already consumed by _oversized_message_content_rejection and needs no message text.

Infographic

[whyyagswhy]

Independent verification on the PR head (7c40b50): error-classifier suite 169/169 green on Linux. Reading a top-level detail body turns a bare 400 into a descriptive error. No findings.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 7c40b50 — fix: read a top-level detail error body so a descriptive 4}

all good!