fix(kiro): trigger Claude compact for overlength requests

[acking-you]

Summary

• Normalize Kiro content-length/context-window failures into Anthropic-compatible 413 errors whose message starts with Prompt is too long.
• Handle both upstream HTTP failures and Kiro eventstream ContentLengthExceededException before returning normal assistant/SSE output.
• Preserve usage recording for these failures with client-visible 413 status.

Root cause

Claude Code 2.1.156 only enters its reactive compact path when its API-error assistant text is normalized to Prompt is too long. The gateway was exposing Kiro overlength failures as generic 400 Input is too long / Input content length exceeds threshold, and eventstream exceptions could still be treated like normal assistant responses or max_tokens.

What changed

• Added Kiro overlength detection for HTTP error bodies and eventstream error/exception frames.
• Return invalid_request_error JSON with HTTP 413 and parseable <actual> tokens > <limit> tokens details.
• Map ContentLengthExceededException stop reasons to model_context_window_exceeded.

Test Plan

• cargo test -p llm-access-kiro --jobs 4
• cargo test -p llm-access --jobs 4 -- --test-threads=1
• cargo clippy -p llm-access-kiro -p llm-access --jobs 4 -- -D warnings

[gemini-code-assist]

Code Review

This pull request maps upstream content length exceeded errors and exceptions to a specific "prompt too long" error with a 413 Payload Too Large status code, and updates the stream stop reason to "model_context_window_exceeded". Feedback suggests expanding the detection logic in kiro_text_is_content_length_exceeded to also match the generic "input is too long" string to ensure robust error handling.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6733c094b6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".