AnthropicAdapter.handleError context-length heuristic has no HTTP-status guard

[Anarchid]

Summary

AnthropicAdapter.handleError at src/providers/anthropic.ts:350-351 classifies any error containing "context" or "too long" in its message as a non-retryable context_length error, regardless of HTTP status:

```ts
if (message.includes('context') || message.includes('too long')) {
return contextLengthError(message, error, rawRequest);
}
```

A 5xx response whose body happens to contain those substrings — e.g. \"Internal error: context processing failed\", \"Backend took too long to respond\" — is misclassified as context_length instead of server, which silently suppresses correct retry behavior. The user's request fails permanently on a transient backend error.

The newly-added 400-classification branch this PR (#16) introduced sits structurally adjacent to this heuristic, making it a natural fix-together. Greptile rated this P1 on PR #16's review thread; the PR merged with no author reply and the guard wasn't added.

Suggested fix

Gate the heuristic on status === 400 (and possibly the explicit context_length error code if Anthropic surfaces one):

```ts
if (status === 400 && (message.includes('context') || message.includes('too long'))) {
return contextLengthError(message, error, rawRequest);
}
```

A 4xx-only classifier matches the real failure mode: context-length is a client-side request shape problem, never something a 5xx can signal.

Code

• src/providers/anthropic.ts:335 — handleError signature
• src/providers/anthropic.ts:350-351 — the unguarded heuristic
• src/providers/anthropic.ts:74, 245 — both handleError call sites (request + stream paths)

Severity

High — silently suppresses retries on transient backend failures whose error text happens to contain a common English word. Fresh merge (PR #16 landed 2026-05-21) so impact in production is currently bounded, but every retried request taking this path will fail incorrectly.

Origin

Greptile P1 finding on PR #16 (merged 2026-05-21). Not addressed in merge.

[Anarchid]

Status update 2026-07-02 (verified against main @ edac27d): the heuristic is unchanged, and two changes since filing have made the missing guard more consequential, not less.

• The unguarded check now sits at src/providers/anthropic.ts:410 (was :350), still ahead of any status discrimination other than 429/401.
• PR fix(anthropic): classify 400 as invalid_request instead of unknown #16 (de3e172, merged the day this was filed) added a status === 400 → invalidRequestError branch — but below the heuristic (now :421). So a 400 whose message mentions "context" is intercepted by the heuristic before the explicit 400 branch ever runs, and a non-400 containing "context"/"too long" is still misclassified as context_length.
• PR fix(anthropic): classify mid-stream overloaded_error as retryable #28 (14ec355, 2026-06-15) now recovers real HTTP statuses from SSE error bodies (e.g. overloaded_error → 529). That makes the failure mode concrete rather than hypothetical: a 529 with "took too long"-style body text gets classified as non-retryable context_length before the retryable-server classification can see it — exactly the retry suppression described above.

The one-line fix in the issue body still applies verbatim; it just lives at :410 now.

🤖 Status audit via Claude Code