(MOT-4080) feat(provider-kimi): add Moonshot provider

provider-kimi/Cargo.toml

@@ -0,0 +1,35 @@
+[workspace]
+
+[package]
+name = "provider-kimi"
+version = "1.0.0"
+edition = "2021"
+publish = false
+license = "Apache-2.0"
+description = "Moonshot (Kimi) Chat Completions provider worker behind llm-router."
+
+[[bin]]
+name = "provider-kimi"
+path = "src/main.rs"
+
+[lib]
+path = "src/lib.rs"
+
+[dependencies]
+llm-router = { path = "../llm-router" }
+# Must match llm-router's pin so the provider and router share one IIIClient type.
+iii-sdk = "=0.21.6"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+# Must stay on the same schemars major as iii-sdk so the derived
+# request/response schemas match what the SDK emits at registration.
+schemars = "0.8"
+tokio = { version = "1", features = ["macros", "rt-multi-thread", "sync", "time", "signal"] }
+futures = "0.3"
+reqwest = { version = "0.12", default-features = false, features = ["json", "stream", "rustls-tls"] }
+clap = { version = "4", features = ["derive", "env"] }
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["fmt", "env-filter"] }
+
+[dev-dependencies]
+uuid = { version = "1", features = ["v4"] }

provider-kimi/README.md

@@ -0,0 +1,77 @@
+# provider-kimi
+
+Moonshot (Kimi) Chat Completions provider worker behind
+[llm-router](../llm-router/). Moonshot's API is OpenAI Chat
+Completions–compatible, so this worker forks the shared provider scaffolding
+and adds Kimi's `reasoning_content` thinking stream.
+
+Implements the provider protocol from
+`tech-specs/2026-06-agentic/llm-router.md`: `provider::kimi::stream`
+(SSE chunks → `AssistantMessageEvent` frames into a router-owned channel) and
+`provider::kimi::refresh_models` (live `GET /v1/models` filtered to Kimi /
+Moonshot chat families, enriched with a curated capability snapshot →
+`router::models::reconcile`).
+
+## Behavior
+
+- **Registration:** self-declares via `router::provider::register` with
+  backoff until acked, and re-declares on the `router::ready` trigger type.
+  The declaration carries no static `models` slice — `GET /v1/models` is the
+  source of truth and a refresh fires right after registration. Defaults:
+  `api_url: https://api.moonshot.ai/v1/chat/completions`,
+  `credential_env_var: MOONSHOT_API_KEY`.
+- **Identity binding:** the router returns a `registration_token` on first
+  registration; it is persisted in iii-state (scope `provider-kimi`,
+  key `registration_token`) and presented on every later
+  `register`/`resolve`/`reconcile`. If that state is lost the router rejects
+  re-registration — the operator must clear the binding on the router side.
+- **Credentials:** resolved per request via `router::provider::resolve`
+  (config slice → `MOONSHOT_API_KEY` env on the router → none). Both
+  `api_key` and `oauth` credential shapes are sent as `Authorization:
+  Bearer`; v1 performs no OAuth refresh.
+- **Liveness:** `ping` at least every 30s of upstream silence; a failed
+  channel write (caller gone / `router::abort`) drops the SSE receiver and
+  aborts the in-flight HTTP request.
+- **Errors:** 401/403 → `auth_expired`, 429 → `rate_limited` (except a quota
+  wall, `exceeded_current_quota_error` → `permanent`),
+  context-length errors → `context_overflow`, 5xx/`engine_overloaded_error` /
+  network → `transient`, other 4xx → `permanent`. No transport retries here —
+  the router owns retry policy.
+- **Request shape vs OpenAI:** Moonshot accepts the classic `max_tokens` param
+  (not `max_completion_tokens`) and has no `reasoning_effort` knob.
+- **Thinking:** Kimi thinking models (Kimi K2 Thinking, `kimi-thinking-preview`)
+  stream their reasoning in `delta.reasoning_content` before the answer text.
+  This worker surfaces it as a `Thinking` content block via
+  `ThinkingStart`/`ThinkingDelta`/`ThinkingEnd` (no replay signature). The
+  model itself decides whether it thinks; `thinking_level` is advisory.
+- **Structured output:** Moonshot supports JSON mode only
+  (`response_format: {"type":"json_object"}`), not OpenAI's strict
+  `json_schema` mode. A requested schema is mapped to `json_object` and a
+  report-and-continue warning rides the final message (the caller must mention
+  "JSON" in the prompt). Curated records declare
+  `supports_structured_output: true`.
+- **Prompt caching:** Moonshot context caching is automatic — no request
+  markers. `prompt_tokens_details.cached_tokens` lands on `usage.cache_read`.
+- **Curated snapshot:** `src/curated.rs` carries display names / context
+  windows / output ceilings / capability flags / pricing for known Kimi and
+  Moonshot families; conservative defaults for unknown ones. Context windows
+  and pricing are best-effort placeholders — **verify against
+  platform.moonshot.ai before release**. Discovery supplies only bare ids.
+
+## Tests
+
+```bash
+cargo test                                            # unit (pure modules + TCP stubs)
+III_ENGINE_BIN=$(which iii) cargo test --test integration -- --test-threads=1
+```
+
+The integration suite spawns a real engine, the real router (path dep), this
+provider, and a local stub upstream — no external API calls anywhere.
+
+## Running
+
+The binary takes the standard worker CLI flags: `--url` (engine WebSocket,
+default `ws://127.0.0.1:49134`, falls back to the `III_WS_URL` environment
+variable), `--manifest` (print the registry manifest and exit), and
+`--config` (accepted but ignored with a warning — provider config comes
+from the `llm-router` configuration entry).

provider-kimi/build.rs

@@ -0,0 +1,6 @@
+fn main() {
+    println!(
+        "cargo:rustc-env=TARGET={}",
+        std::env::var("TARGET").expect("TARGET must be set by Cargo build scripts")
+    );
+}

provider-kimi/config.yaml

@@ -0,0 +1,8 @@
+# provider-kimi has no file-based configuration.
+#
+# Credentials, `api_url`, and `max_tokens` arrive per request from
+# llm-router's resolve step; the provider block lives in the engine's
+# `llm-router` configuration entry (README § Configuration).
+#
+# This file exists to satisfy the standard worker layout
+# (docs/sops/new-worker.md §2). Keys placed here are ignored with a warning.

provider-kimi/iii-permissions.yaml

@@ -0,0 +1,10 @@
+# Agent permissions for the provider-kimi worker.
+# Spec: tech-specs/2026-06-agentic/llm-router.md § Security.
+version: 1
+
+rules:
+  # Direct provider calls bypass the router's accounting, budgets, and retry
+  # policy — never agent-callable. The router invokes these worker-to-worker.
+  - '!provider::kimi::stream'
+  - '!provider::kimi::refresh_models'
+  - '!provider::kimi::on_router_ready'

provider-kimi/iii.worker.yaml

@@ -0,0 +1,11 @@
+iii: v1
+name: provider-kimi
+language: rust
+deploy: binary
+manifest: Cargo.toml
+bin: provider-kimi
+description: Moonshot (Kimi) Chat Completions provider worker; implements provider::kimi::stream and provider::kimi::refresh_models behind llm-router.
+
+dependencies:
+  iii-state: "^0.21.6"
+  llm-router: "^1.0.0"