From f0b085ebcaa37396e4164e049b254cabce846fef Mon Sep 17 00:00:00 2001 From: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com> Date: Mon, 6 Jul 2026 19:22:02 -0300 Subject: [PATCH] fix(executors): inject reasoning_content for native Kimi provider (port from 9router#1480) (#6419) inject reasoning_content for native Kimi (port #1480) (net +1/-0, test OK). Integrated into release/v3.8.46. --- CHANGELOG.md | 1 + open-sse/executors/default.ts | 21 ++++++ ...imi-native-reasoning-injected-1480.test.ts | 65 +++++++++++++++++++ 3 files changed, 87 insertions(+) create mode 100644 tests/unit/kimi-native-reasoning-injected-1480.test.ts diff --git a/CHANGELOG.md b/CHANGELOG.md index a08b4493dd..1faa1baa46 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -21,6 +21,7 @@ ### 🐛 Bug Fixes +- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider. Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345) - **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400. **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot) - **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher. When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk) - **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one. With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a *different* model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr) diff --git a/open-sse/executors/default.ts b/open-sse/executors/default.ts index afabd75645..4b844f4c76 100644 --- a/open-sse/executors/default.ts +++ b/open-sse/executors/default.ts @@ -17,6 +17,10 @@ import { import { isOfficialAnthropicBaseUrl } from "../utils/anthropicHost.ts"; import { applyProviderRequestDefaults } from "../services/providerRequestDefaults.ts"; import { stripUnsupportedParams } from "../translator/paramSupport.ts"; +import { + injectReasoningContentForThinkingModel, + isThinkingMessageModel, +} from "../utils/reasoningContentInjector.ts"; import { detectFormat, getOpenAICompatibleType, @@ -695,6 +699,23 @@ export class DefaultExecutor extends BaseExecutor { this.ensureThinkingBudget(withDefaults as Record, model); } + // 9router#1480: the native Moonshot `kimi` provider (executor "default") + // is a thinking-mode upstream that 400s with "reasoning_content must be + // passed back" when a prior assistant turn lacks it. OpencodeExecutor + // already injects a placeholder for OpenCode-routed thinking models; the + // direct kimi connection hit neither injection path. Scope to `kimi` so + // gateway-served models that merely match the thinking-model name pattern + // (and may reject an extra field) are unaffected. + if (this.provider === "kimi") { + const outboundModel = + typeof (withDefaults as Record)?.model === "string" + ? ((withDefaults as Record).model as string) + : model; + if (isThinkingMessageModel(outboundModel)) { + withDefaults = injectReasoningContentForThinkingModel(withDefaults); + } + } + return withDefaults; } diff --git a/tests/unit/kimi-native-reasoning-injected-1480.test.ts b/tests/unit/kimi-native-reasoning-injected-1480.test.ts new file mode 100644 index 0000000000..1654a8185f --- /dev/null +++ b/tests/unit/kimi-native-reasoning-injected-1480.test.ts @@ -0,0 +1,65 @@ +import test from "node:test"; +import assert from "node:assert/strict"; + +import { DefaultExecutor } from "../../open-sse/executors/default.ts"; + +/** + * Regression guard for upstream 9router#1480. + * + * The native Moonshot `kimi` provider (executor "default") is a thinking-mode + * upstream that returns 400 "reasoning_content must be passed back" when a prior + * assistant turn in the history lacks `reasoning_content`. OpencodeExecutor + * already injects a placeholder for OpenCode-routed thinking models, but the + * direct kimi connection went through DefaultExecutor, which did not — so + * multi-turn kimi conversations 400'd. The injection must fire for `kimi`, and + * must NOT fire for unrelated providers that merely serve a matching model name. + */ + +const STREAM = true; +const CREDENTIALS = { apiKey: "k" } as Record; + +function multiTurnBody(model: string) { + return { + model, + stream: STREAM, + messages: [ + { role: "user", content: "hi" }, + { role: "assistant", content: "previous answer" }, // no reasoning_content + { role: "user", content: "follow up" }, + ], + }; +} + +test("DefaultExecutor(kimi) injects reasoning_content on assistant turns that lack it", () => { + const out = new DefaultExecutor("kimi").transformRequest( + "kimi-k2.6", + multiTurnBody("kimi-k2.6"), + STREAM, + CREDENTIALS + ) as Record; + const messages = out.messages as Array>; + const assistant = messages.find((m) => m.role === "assistant") as Record; + assert.equal( + typeof assistant.reasoning_content === "string" && + (assistant.reasoning_content as string).length > 0, + true, + "kimi assistant message must carry a non-empty reasoning_content placeholder" + ); +}); + +test("DefaultExecutor(openai) does NOT inject reasoning_content (scoped to kimi)", () => { + // A non-kimi provider must not gain the injection even for a thinking-ish name. + const out = new DefaultExecutor("openai").transformRequest( + "kimi-k2.6", + multiTurnBody("kimi-k2.6"), + STREAM, + CREDENTIALS + ) as Record; + const messages = out.messages as Array>; + const assistant = messages.find((m) => m.role === "assistant") as Record; + assert.equal( + Object.prototype.hasOwnProperty.call(assistant, "reasoning_content"), + false, + "non-kimi providers must not be given a reasoning_content placeholder" + ); +});