Use Fireworks Responses endpoint for Kimi K3

[fontanierh]

Description

• Route Kimi K3 through Fireworks Responses so reasoning item IDs, function calls, and tool results are replayed without flattening them into assistant text.
• Preserve Responses reasoning IDs based on the persisted wire metadata rather than the model provider, including interleaved reasoning and tool-call items.
• Preserve ID-bearing reasoning items even when the provider returns an empty visible summary, matching Kimi Code multi-step replay behavior.
• Send service_tier: "priority", disable provider storage, and adapt named forced tools to the tool-choice shape Fireworks accepts.

Tests

• Ran the live Kimi K3 endpoint suite against Fireworks with service_tier: "priority": all 46 cases passed, including multi-turn following, forced tools, reasoning efforts, and structured output.
• Ran a live stateless tool continuation with store: false: Fireworks accepted the replayed reasoning item ID and tool result, then returned fresh reasoning and the expected final answer.
• Added focused coverage for raw interleaved item ordering, Fireworks reasoning-ID persistence and replay, empty-summary replay state, forced tools, and hosted tool-search filtering.

Risk

The provider endpoint change is scoped to Kimi K3. Reasoning-ID detection now applies to any Responses provider that persists an id, matching the Responses wire contract. The endpoint ID and pricing remain unchanged, so rollback is a normal code revert.

Fireworks' Responses frontend parses service_tier, but the current Responses reference does not document it and responses do not expose the selected tier. Confirm priority routing and billing with Fireworks before production rollout.

Deploy Plan

Standard deploy after Fireworks confirms Responses routes and bills the priority tier.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

2 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
playground	Ignored	Preview	Jul 28, 2026 12:56pm
storybook	Ignored	Preview	Jul 28, 2026 12:56pm

[dust-agent]

Coding Rules LGTM \o/