[imtiyazakiwat]

Summary

runtime.kiro.dev has no /ListAvailableModels, so FALLBACK_MODELS is the entire /v1/models catalog for every account. It had fallen behind Kiro: 9 current models were missing and one end-of-life model was still listed.

Fixes #208.

Changes

Catalog (kiro/config.py): synced with kiro-cli chat --list-models --format json (checked 2026-10-20) and the Kiro models changelog.

• Added claude-opus-5.5, claude-sonnet-5.5, claude-opus-5, claude-sonnet-5, claude-opus-4.8, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, kimi-k3.
• Removed claude-sonnet-4. It reached end of life in Kiro on 2026-10-14 and Kiro now rejects it with Invalid model ID. It is still passed through if a client asks for it, so Kiro returns its own error.

Context windows: every entry now carries tokenLimits.maxInputTokens (from kiro-cli context_window_tokens).

• Prompt tokens are derived from contextUsagePercentage × maxInputTokens.
• Without per-model limits, every model used the 200K default, so usage for 1M-context models (sonnet-4.6, opus-4.6+, all 5.x, kimi-k3) was reported about 5x too low.

Context-window lookup (kiro/streaming_core.py): calculate_tokens_from_context_usage() now normalizes the client's model name before the cache lookup.

• Routes pass the raw request name. Without this, Claude Code style names (claude-sonnet-4-6, claude-sonnet-4-6[1m]) missed the cache and fell back to 200K.
• One place covers OpenAI and Anthropic, streaming and non-streaming.

Normalization (kiro/model_resolver.py): the no-minor-version pattern now strips -latest the same way the minor-version pattern already does (claude-opus-5-latest → claude-opus-5).

• Bare numbers are still read as minor versions (claude-opus-5-5 → claude-opus-5.5).

Docs: added a paid-tier model note and a Sonnet 4 end-of-life note, and removed the Sonnet 4 entry, in all 8 READMEs.

Live verification (Builder ID account, runtime.us-east-1.kiro.dev)

• GET /v1/models: 21 entries (was 13).
• POST /v1/chat/completions (non-streaming): HTTP 200 for every listed model. claude-sonnet-4 returns 400 Invalid model ID, which confirms the end of life.
• New models on all four paths (OpenAI and Anthropic, streaming and non-streaming): HTTP 200 for all 9 × 4 combinations.
• Client name formats via /v1/messages: claude-opus-5-5, claude-sonnet-5-latest, claude-sonnet-5-5[1m] and claude-opus-4-8 all return 200.

Reported prompt tokens for the same ~10.5K-token prompt:

model	before	after
claude-haiku-4.5 (200K, control)	16,258	16,259
claude-sonnet-4.6 (1M)	3,062	16,271
claude-sonnet-4-6 via /v1/messages	n/a	16,267
claude-sonnet-4-6[1m] via /v1/messages	n/a	16,270

Tests

• Full suite: 1773 passed. Base a5292ca has 1691.
• New tests cover:
  • The exact catalog snapshot and the end-of-life removal.
  • Unique IDs, and IDs that normalization leaves unchanged.
  • tokenLimits structure (rejects bool) and the per-model context window values.
  • New models resolving from cache in every client format.
  • Pass-through for the removed model.
  • The /v1/models list with default aliases.
  • Cache limits, and the normalized lookup in token calculation (mock cache and real cache).
  • -latest positive cases, plus negative cases (-latest-20260724, -latestx, 7- and 9-digit suffixes, trailing dash, non-Claude names).
• With the three source files reverted and the tests kept, 56 of the new tests fail.

Known issues (not changed here)

1. Claude Opus 5 (always) and Claude Sonnet 5.5 (sometimes) refuse fake-reasoning requests.
   • Kiro answers with a metadataEvent containing "stopReason":"CONTENT_FILTERED" and refusal category REASONING_EXTRACTION.
   • The gateway does not parse that event, so the client gets HTTP 200 with empty content.
   • reasoning_effort: "none" does not help, because the thinking system-prompt addition is still injected. FAKE_REASONING=false fixes it (Opus 5 then answers normally).
   • Surfacing the refusal and adjusting fake reasoning for these models seems better as a separate PR.
2. The auto-kiro alias is listed but not resolved when the Kiro payload is built, so Kiro rejects it (auto works). This predates this PR and is covered by fix(model_resolver): resolve MODEL_ALIASES in get_model_id_for_kiro #184 and fix: aliases not resolved when building Kiro API payload #198.

Related

• Covers the model additions in fix(config): add claude-opus-4.8 to FALLBACK_MODELS #242, fix(models): accept opus-4.8 and hoist inline system messages #246, feat(models): add claude-opus-5 to fallback model list #279 and part of feat(models): add Claude Sonnet 5 support #260.
• Complements feat(models): auto-discover working-but-unadvertised models #285: listed models show up at startup, and discovery would cover models released later.

[cla-bot]

Thanks for the PR! 🎉

Before merge, we need a one-time CLA confirmation.
It confirms that you have the right to contribute this code and allow the project to use it.

Full CLA text:
https://github.com/jwadow/kiro-gateway/blob/main/CLA.md

Please reply once with:

I have read the CLA and I accept its terms

You need to write once, all further messages from me can be ignored.

[imtiyazakiwat]

I have read the CLA and I accept its terms