Skip to content

Add Cocoon Stack Gateway to the AI gateway benchmarks - #409

Open
CMGS wants to merge 1 commit into
computesdk:masterfrom
CMGS:add-cocoonstack-gateway
Open

Add Cocoon Stack Gateway to the AI gateway benchmarks#409
CMGS wants to merge 1 commit into
computesdk:masterfrom
CMGS:add-cocoonstack-gateway

Conversation

@CMGS

@CMGS CMGS commented Sep 10, 2026

Copy link
Copy Markdown

Adds Cocoon Stack Gateway (cocoonstack-gateway, github.com/cocoonstack/gateway) to all four AI-gateway families. The participant is served from gw-bench.simular.cloud (us-east4).

Family Wire Path Model
Anthropic anthropic /v1/messages claude-haiku-4-5-20251001
OpenAI responses /v1/responses gpt-5.4-mini
Gemini openai /v1/chat/completions gemini-3.6-flash
Kimi openai /v1/chat/completions kimi-k3

Notes on the entries:

  • Gemini goes through the OpenAI-compatible chat wire: the gateway translates to Gemini's native streamGenerateContent; there is no native Gemini inbound, same as the other translation-layer participants in that family. Thinking tokens are reported under completion_tokens_details.reasoning_tokens and folded into completion_tokens.
  • Kimi omits temperature (kimi-k3 accepts only 1, as documented at the top of providers-kimi.ts) and sets reasoningCountsAsFirstToken: true; the stream carries reasoning_content in Moonshot's own shape.
  • Every wire ends with a usage frame (stream_options.include_usage is honored on the chat wire; Responses and Messages report usage natively), so tokens/sec resolves for all four families.

Credential: a single env var, COCOONSTACK_GATEWAY_API_KEY (a bearer / x-api-key value issued by us). It is wired through .env.example, provider-vars.json and the workflow's vault list; a missing key reports the participant as SKIPPED as usual. We will send the key to the maintainers out of band.

Verification: pnpm typecheck clean; this repo's runner from a us-east4 client with --iterations 50 (50 cold + 50 warm) on all four families, interleaved with the family's direct baseline, 800/800 probes succeeded (--no-ingest, results not included here).


Devin Review

Adds `cocoonstack-gateway` (github.com/cocoonstack/gateway, served from
gw-bench.simular.cloud in us-east4) to all four families:

- Anthropic: native `/v1/messages`, claude-haiku-4-5-20251001
- OpenAI: `/v1/responses`, gpt-5.4-mini
- Gemini: OpenAI-compatible `/v1/chat/completions`, gemini-3.6-flash (the
  gateway translates to Gemini's native API; there is no native Gemini inbound)
- Kimi: `/v1/chat/completions`, kimi-k3 on Moonshot's international endpoint;
  `temperature` omitted (kimi-k3 accepts only 1) and
  `reasoningCountsAsFirstToken: true` (the stream carries `reasoning_content`)

One credential, COCOONSTACK_GATEWAY_API_KEY, wired through .env.example,
provider-vars.json and the workflow's vault list; a missing key reports
SKIPPED as usual. Confirmed live with this repo's runner from us-east4:
`--iterations 50` on all four families, 800/800 probes succeeded, and every
wire ends with a usage frame so tokens/sec resolves.
@open-cla

open-cla Bot commented Sep 10, 2026

Copy link
Copy Markdown

Contributor License Agreement

All contributors are covered by a CLA.

@devin-ai-integration devin-ai-integration Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

Devin Review

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant