fix(error_classifier): read FastAPI-style {"detail": ...} error bodies; recognise session/weekly usage limits

[i-Hun]

Problem 1: FastAPI-style error bodies read as blank

Proxies built on FastAPI / Starlette (Open WebUI, LiteLLM, most self-hosted OpenAI-compatible gateways) report errors as {"detail": "..."}.

_classify_400() only looked at error.message, message, errorMessage and errorArgs.reason. With none of those present, a long and perfectly descriptive rejection looked like an empty message, the "generic 400 on a large session means context overflow" heuristic fired, and a healthy session was pushed into the compression loop until it was auto-reset with "max compression attempts reached". The real cause (an upstream router error) never reached the user.

This only triggers when context_length <= 256_000 and the session has more than 80 messages, so the same error is harmless on one model and kills the session on another. That made it hard to reproduce and easy to blame on the wrong thing.

Problem 2: subscription usage walls classified as unknown or billing

Subscription-backed providers name the wall by its window: "You've hit your session limit · resets 2:20pm", "You've hit your weekly limit ...". Neither matched the usage-limit patterns, so the error classified as unknown (no backoff, no Retry-After) or, when it did match, as a permanent billing failure because "resets 2:20pm" matched neither "resets at" nor "resets in".

Fix

• New _detail_message(body) reads detail as a string, a dict (proxies nesting an OpenAI-ish object) or a list (FastAPI's own validation errors, keeping loc and type so the message stays actionable).
• It is consulted after the existing keys in classify_api_error(), _classify_400() and _extract_message(), so the standard OpenAI shape keeps precedence.
• "session limit" and "weekly limit" added to the usage-limit patterns; bare "resets " added to the transient signals.

Tests

New tests/agent/test_error_classifier_detail_body.py (11 tests). The fix: a descriptive detail is a format_error, not a compression trigger, even when str(error) does not carry the body; the operator sees the router text; dict and list shapes do not crash. Controls: a genuinely bare 400 on a large session still compresses; a real overflow reported through detail is still an overflow; error.message still wins over detail. Session and weekly limits classify as rate_limit.

Existing error-classifier / failover tests pass (191 tests).

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• Correct fix: FastAPI/Starlette proxy errors ({"detail": "..."}) previously read as blank and tripped the "generic 400 on large session = context overflow" heuristic, forcing healthy sessions into the compression loop; _detail_message now surfaces the real rejection text (agent/error_classifier.py:2197).
• The three detail shapes (str/dict/list) are handled defensively; list items get loc: msg (type) formatting that keeps FastAPI validation errors actionable, and no shape crashes.
• Precedence is preserved: error.message still wins over detail, and a genuinely bare 400 on a large session still compresses — the control tests pin both behaviors.
• New "session limit" / "weekly limit" / "resets " patterns correctly reclassify subscription-backed usage walls as transient rate_limit; the bare "resets " match is broad but only applies within the usage-limit context, low risk.
• Tests are thorough, including SDK body-not-in-str(error) and detail-carried-overflow cases.

[teknium1]

Thanks @i-Hun — the work in this PR has landed on main via:

• 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) #115881 — fix(errors): Codex entitlement/retired-model 400 with a top-level detail body is reported as a model rejection, not context overflow ([Bug]: openai-codex / gpt-5.6-sol returns HTTP 400 "model not supported" despite valid OAuth, valid version (v0.19.1), and working Codex CLI on the same account #81558, [Bug]: unentitled primary + unentitled fallback oscillate forever on openai-codex; "Primary model restored" is emitted without any recovery #106475, salvage fix(error_classifier): read FastAPI-style {"detail": ...} error bodies; recognise session/weekly usage limits #100783) (landed in 05c95d216eb7)

Your contribution is credited there (cherry-picked authorship / co-author trailer or credit in the PR body; see the linked PR for what was kept and what was trimmed). Closing this one as landed / superseded so the backlog reflects reality. If something in your original diff is still missing on current main, please comment and we'll reopen or follow up.