[italocjs]

Everything this gateway needed to run Claude Code (2.1.304) reliably against Kiro, collected on one branch. Based on upstream main @ a5292ca.

Commits

1. fix(anthropic): accept messages[].role system and merge into system prompt (upstream jwadow#197, by @W0n9)

Cherry-picked unchanged, author kept. Newer Claude Code sends inline role: "system" messages; without this the gateway answers 422.

2. fix(streaming): send keepalive while the client stream is silent (upstream jwadow#288 / issue jwadow#287)

Symptom: a large tool call (e.g. Write of a big file) stalls with "Waiting for API response … check your network", then Claude Code retries the whole turn.

Why: tool calls are buffered until Kiro finishes the response, so the client gets zero bytes for minutes. Claude Code's byte-level idle watchdog (CLAUDE_BYTE_STREAM_IDLE_TIMEOUT_MS, default 180-300 s) then aborts.

Fix: a ping event (Anthropic) or : keepalive SSE comment (OpenAI) every STREAM_KEEPALIVE_INTERVAL seconds (default 10, 0 disables) while the stream is silent.

3. fix(streaming): surface Kiro refusals instead of empty responses (new)

Symptom: replies come back empty. Claude Code shows "No response requested." or sends "Your previous response had no visible output", and sessions feel stuck or very slow.

Why: Kiro now refuses the gateway's fake-reasoning prompt in any conversation with history:

{"stopDetails":{"refusal":{"category":"REASONING_EXTRACTION","explanation":"The selected model cannot continue this conversation. ..."}},"stopReason":"CONTENT_FILTERED"}

The parser had no pattern for stopDetails, so this became an empty end_turn. Reproduced on both claude-sonnet-5.5 and claude-opus-5.5. A one-message request works; user → assistant → user is refused every time; the same with thinking: disabled works.

Fix: parse the refusal and report it the way each API reports a refusal:

• Anthropic: stop_reason: "refusal", with the explanation as a text block.
• OpenAI: finish_reason: "content_filter", with the explanation as content.
• Both streaming and non-streaming.
• When the category is REASONING_EXTRACTION and fake reasoning is on, the message and a WARNING log line say to set FAKE_REASONING=false.

Deliberately not an error event: Claude Code retries api_error, which would turn this into a retry loop.

Required configuration

FAKE_REASONING=false. With it, multi-turn requests answer normally; the refusal fix only makes the failure visible.

Testing

• Full unit suite: 1726 passed. New tests: 15 for keepalive, 17 for refusals (parser, core, Anthropic and OpenAI, streaming and non-streaming). 7 of the refusal tests fail on the unpatched code; the other 3 check that normal behavior is unchanged.
• Keepalive, live: pings every 10 s during a 40-50 s buffered tool call. Claude Code with CLAUDE_BYTE_STREAM_IDLE_TIMEOUT_MS=15000 finished a ~54 s Write (662 lines) without a retry.
• Refusal, live, against the real Kiro API with FAKE_REASONING=true: the client now receives stop_reason=refusal with Kiro's explanation and the fix hint, instead of an empty reply.

Note on prompt caching

Kiro applies prompt caching automatically; no gateway change is needed. Measured from Kiro's meteringEvent on a growing conversation with a ~45k-token prefix:

Model	Turn 1	Turn 2	Turn 3
Sonnet	0.616 credits	0.186	0.201
Opus	0.900 credits	0.351	0.293

Kiro's cachePoint field had no effect. The cache_read_input_tokens: 0 that clients see is only the gateway's own token estimate.