[Feature]: Add Anthropic-specific markers to _is_context_overflow for context overflow classification

[bukeeastrey]

Duplicate check

• I searched the existing open and closed issues and this is not a duplicate.

Problem

When using the anthropic provider, context overflow errors are misclassified as BAD_REQUEST instead of CONTEXT_OVERFLOW. This prevents the COMPACT_AND_RETRY recovery path from activating — the error is surfaced directly to the user instead of automatically compacting the conversation and retrying.

The root cause is that _is_context_overflow() in failures.py only contains OpenAI-centric markers ("context length", "context window", "too many tokens", etc.) and doesn't recognize Anthropic's specific error phrasing:

• "exceed context limit" (from "input length and max_tokens exceed context limit: ...")
• "request_too_large" (413 error type)
• "request size exceeds" (413 message variant)
• "prompt_too_long" (error code)

Proposed behavior

Add Anthropic-specific markers to _is_context_overflow() in src/agentos/provider/failures.py:

• "prompt_too_long" — matches the error code
• "exceed context limit" — matches Anthropic's canonical 400 message
• "request_too_large" — matches Anthropic's 413 error type
• "request size exceeds" — matches the 413 message variant

This ensures Anthropic context overflows classify as CONTEXT_OVERFLOW → COMPACT_AND_RETRY, consistent with how OpenAI-compatible providers already behave.

Area

Provider integration

Alternatives considered

• Adding "max_tokens" as a marker was considered but rejected — it appears in nearly every Anthropic error message as a parameter name and would cause widespread false positives.
• Matching on status_code == 413 alone was considered, but a 413 can indicate a payload-size issue unrelated to token count, so text-based markers are more precise.

[s97472091-pixel]

Claiming this issue — I will add the Anthropic-specific markers (prompt_too_long, exceed context limit, request_too_large, request size exceeds) to _is_context_overflow() in src/agentos/provider/failures.py with regression tests. PR coming shortly.

[andreapn]

Accepted, p2 — with a scope trim.

Confirmed that _is_context_overflow() (src/agentos/provider/failures.py:78) carries only the OpenAI-centric marker set, and that anthropic is routed through classify_provider_error (line 160), so the classification really does apply to this provider.

Checking the four proposed markers individually:

• exceed context limit — genuinely missing. Anthropic's canonical 400 is input length and max_tokens exceed context limit: ..., which matches none of the existing markers. This is the case worth fixing.
• request_too_large / request size exceeds — missing, and reasonable to add.
• prompt_too_long — already covered in practice. Anthropic's message text for that code reads prompt is too long: 200000 tokens > 199999 maximum, and "prompt is too long" is already in the list. Adding the underscore form is harmless as a belt-and-braces match on the error code, but the issue overstates it as a gap.

Good judgement on the two alternatives you rejected — max_tokens as a marker would false-positive on nearly every Anthropic error, and bare 413 matching would catch genuine payload-size problems. Keep both out.

Three PRs target this (#640, #627, #615) — we will pick one.