@moondeck/station 0.0.1 → 1.0.0

package/lib/agents/kimi.ts

@@ -0,0 +1,329 @@
+/**
+ * KimiSession — a real AgentSession backed by Moonshot's **Kimi Code** CLI,
+ * run inside the task environment:
+ *
+ *   ~/.kimi-code/bin/kimi \
+ *     -p "<prompt>" -m kimi-code/k3 --output-format stream-json [-r <sessionId>]
+ *
+ * Unlike Claude (persistent stdin session) and Codex (`app-server` JSON-RPC),
+ * Kimi Code is **one-shot per turn**: `-p` runs a single prompt to completion,
+ * streams newline-delimited JSON to stdout, and exits. Multi-turn continuity is
+ * `-r <sessionId>` (the session lives in Kimi's cloud + the container's
+ * ~/.kimi-code). So each `send()` spawns a fresh provider process; sends are
+ * serialized so a second message can't race a running turn.
+ *
+ * Stream-json vocabulary (reverse-engineered — OpenAI-chat-style role lines):
+ *   {"role":"assistant","content":"…"}                       → message_complete
+ *   {"role":"assistant","tool_calls":[{function:{name,arguments}}]} → tool_call
+ *   {"role":"tool","tool_call_id":…,"content":…}             → (no event; folded)
+ *   {"role":"meta","type":"session.resume_hint","session_id"} → captured for -r
+ *
+ * AUTH is the operator's Kimi Code **subscription**: `~/.kimi-code` (config +
+ * credentials/kimi-code.json) is docker-cp'd into the container at task-up
+ * (task-up.sh), same pattern as Codex's ~/.codex — never the cloud (ADR-015).
+ * `-p` runs non-interactively and auto-runs tools (the container is the
+ * isolation boundary), so there are no permission prompts to resolve.
+ */
+import { existsSync } from "node:fs";
+import { homedir } from "node:os";
+import { join } from "node:path";
+
+import { requireContainerRuntimeOperational } from "../runtime-guard";
+import type {
+  TaskEnvironmentAgentSessionSurface,
+  TaskEnvironmentSessionRequest,
+} from "../task-environment/types";
+import { newId } from "../ulid";
+import {
+  AGENT_SESSION_OUTPUT_BUFFER_BYTES,
+  spawnEnvironmentProcess,
+  type AgentEnvironmentProcess,
+} from "./proc";
+import { register } from "./registry";
+import type {
+  AgentEvent,
+  AgentEventHandler,
+  AgentKind,
+  AgentSession,
+  RosterAgent,
+} from "./types";
+
+/** Installed by the image's install.sh (code.kimi.com); absolute so a
+ *  non-login provider process doesn't depend on PATH. */
+const KIMI_BIN = "/home/node/.kimi-code/bin/kimi";
+
+// Model ids from `~/.kimi-code/config.toml` (`kimi -m <id>`). k3 is the
+// frontier model; the K2.7 coding pair are cheaper/faster. UPDATE WHEN KIMI
+// CHANGES its lineup. Order = display order in the cloud picker.
+const KIMI_MODELS = [
+  "kimi-code/k3",
+  "kimi-code/kimi-for-coding",
+  "kimi-code/kimi-for-coding-highspeed",
+];
+const KIMI_DEFAULT_MODEL = "kimi-code/k3";
+
+// Kimi Code has no per-invocation reasoning-effort flag (effort is a
+// config.toml model default), so nothing to offer in the picker.
+const KIMI_EFFORTS: string[] = [];
+
+/** Host path to the subscription credential — its presence gates the engine
+ *  (mirrors Codex's ~/.codex/auth.json check). */
+function kimiCredPath(): string {
+  return join(
+    process.env.UAI_OWNER_HOME?.trim() || homedir(),
+    ".kimi-code",
+    "credentials",
+    "kimi-code.json",
+  );
+}
+
+// ---------------------------------------------------------------------------
+// Pure protocol mapping — one stream-json line → AgentEvent[] (+ session id).
+// ---------------------------------------------------------------------------
+
+/** Best-effort one-line detail for a tool card from the tool's JSON args. */
+function toolDetail(rawArgs: unknown): string {
+  if (typeof rawArgs !== "string") return "";
+  try {
+    const a = JSON.parse(rawArgs) as Record<string, unknown>;
+    const pick = a.command ?? a.path ?? a.file_path ?? a.pattern ?? a.query;
+    return typeof pick === "string" ? pick : rawArgs;
+  } catch {
+    return rawArgs;
+  }
+}
+
+export interface MappedLine {
+  events: AgentEvent[];
+  /** A session.resume_hint id, when this line carried one. */
+  sessionId?: string;
+}
+
+/**
+ * Map a single Kimi stream-json line. Non-JSON lines (a tool's raw stdout that
+ * Kimi echoes) and lines with no user-facing meaning (`role:"tool"` results,
+ * `role:"user"` echoes) map to no events.
+ */
+export function mapKimiLine(line: string): MappedLine {
+  const trimmed = line.trim();
+  if (!trimmed.startsWith("{")) return { events: [] };
+  let msg: Record<string, unknown>;
+  try {
+    msg = JSON.parse(trimmed) as Record<string, unknown>;
+  } catch {
+    return { events: [] };
+  }
+
+  if (msg.role === "assistant") {
+    if (typeof msg.content === "string" && msg.content.length > 0) {
+      return { events: [{ type: "message_complete", text: msg.content }] };
+    }
+    if (Array.isArray(msg.tool_calls)) {
+      const events: AgentEvent[] = [];
+      for (const raw of msg.tool_calls) {
+        const tc = (raw ?? {}) as Record<string, unknown>;
+        const fn = (tc.function ?? {}) as Record<string, unknown>;
+        const name = typeof fn.name === "string" ? fn.name : "tool";
+        events.push({
+          type: "tool_call",
+          id: typeof tc.id === "string" ? tc.id : newId(),
+          title: name,
+          detail: toolDetail(fn.arguments),
+        });
+      }
+      return { events };
+    }
+    return { events: [] };
+  }
+
+  if (
+    msg.role === "meta" &&
+    msg.type === "session.resume_hint" &&
+    typeof msg.session_id === "string"
+  ) {
+    return { events: [], sessionId: msg.session_id };
+  }
+
+  return { events: [] };
+}
+
+// ---------------------------------------------------------------------------
+// Session
+// ---------------------------------------------------------------------------
+
+/** Injectable spawn seam (tests provide a fake). */
+export type Spawner = (
+  request: TaskEnvironmentSessionRequest,
+) => AgentEnvironmentProcess;
+
+export class KimiSession implements AgentSession {
+  readonly agentId: string;
+  readonly kind: AgentKind = "kimi";
+
+  private readonly environment: TaskEnvironmentAgentSessionSurface;
+  private readonly model?: string;
+  private readonly systemPreamble: string;
+  private readonly agentEnv: Record<string, string>;
+  private readonly spawner: Spawner;
+
+  private readonly handlers = new Set<AgentEventHandler>();
+  private sessionId: string | null = null;
+  private current: AgentEnvironmentProcess | null = null;
+  private queue: Promise<void> = Promise.resolve();
+  private sentPreamble = false;
+  private closed = false;
+
+  constructor(args: {
+    taskId: string;
+    agent: RosterAgent;
+    environment: TaskEnvironmentAgentSessionSurface;
+    systemPreamble: string;
+    agentEnv?: Record<string, string>;
+    spawner?: Spawner;
+  }) {
+    this.agentId = args.agent.id;
+    this.environment = args.environment;
+    this.model = args.agent.model;
+    this.systemPreamble = args.systemPreamble;
+    this.agentEnv = args.agentEnv ?? {};
+    this.spawner =
+      args.spawner ?? ((request) => spawnEnvironmentProcess(this.environment, request));
+  }
+
+  onEvent(handler: AgentEventHandler): () => void {
+    this.handlers.add(handler);
+    return () => this.handlers.delete(handler);
+  }
+
+  private emit(event: AgentEvent): void {
+    if (this.closed && event.type !== "exit") return;
+    for (const h of this.handlers) h(event);
+  }
+
+  /** Kimi is one-shot; serialize turns so a second message queues behind the
+   *  running one instead of spawning a racing `kimi -p`. */
+  async send(text: string): Promise<void> {
+    if (this.closed) return;
+    const turn = this.queue.catch(() => {}).then(() => this.runTurn(text));
+    this.queue = turn;
+    return turn;
+  }
+
+  private buildRequest(prompt: string): TaskEnvironmentSessionRequest {
+    const argv = [
+      KIMI_BIN,
+      "-p",
+      prompt,
+      "--output-format",
+      "stream-json",
+    ];
+    if (this.model) argv.push("-m", this.model);
+    // Continue the same Kimi session across turns (captured from the first
+    // turn's resume_hint). First turn has none → a fresh session.
+    if (this.sessionId) argv.push("-r", this.sessionId);
+    return {
+      argv: argv as [string, ...string[]],
+      projectEnvProfile: "task",
+      user: "node",
+      env: this.agentEnv,
+      maxOutputBytes: AGENT_SESSION_OUTPUT_BUFFER_BYTES,
+    };
+  }
+
+  private async runTurn(text: string): Promise<void> {
+    if (this.closed) return;
+    // A turn may have spent minutes queued behind its predecessor. Admission
+    // at send() time is stale; this synchronous check immediately precedes the
+    // provider process launch and leaves the first-turn preamble unconsumed on fail.
+    requireContainerRuntimeOperational();
+    // Kimi's `-p` has no system-prompt flag, so fold the channel briefing into
+    // the FIRST turn's prompt (later turns carry it via the resumed session).
+    const prompt =
+      !this.sentPreamble && this.systemPreamble.trim().length > 0
+        ? `${this.systemPreamble}\n\n---\n\n${text}`
+        : text;
+    this.sentPreamble = true;
+
+    const child = this.spawner(this.buildRequest(prompt));
+    this.current = child;
+
+    let buf = "";
+    const consume = (chunk: string): void => {
+      buf += chunk;
+      let nl: number;
+      while ((nl = buf.indexOf("\n")) >= 0) {
+        const line = buf.slice(0, nl);
+        buf = buf.slice(nl + 1);
+        const { events, sessionId } = mapKimiLine(line);
+        if (sessionId) this.sessionId = sessionId;
+        for (const e of events) this.emit(e);
+      }
+    };
+    child.stdout?.on("data", (b: Buffer) => consume(b.toString("utf8")));
+    let stderr = "";
+    child.stderr?.on("data", (b: Buffer) => {
+      stderr += b.toString("utf8");
+    });
+
+    await new Promise<void>((resolve) => {
+      child.on("exit", (code) => {
+        this.current = null;
+        if (buf.trim().length > 0) consume("\n"); // flush a trailing line
+        if (!this.closed) {
+          if (code !== 0) {
+            const tail = stderr.trim().slice(-500);
+            this.emit({
+              type: "error",
+              message: `kimi exited ${code ?? "null"}${tail ? `: ${tail}` : ""}`,
+            });
+          }
+          this.emit({ type: "turn_complete" });
+        }
+        resolve();
+      });
+      child.on("error", (err) => {
+        this.current = null;
+        if (!this.closed) {
+          this.emit({ type: "error", message: `kimi spawn failed: ${err.message}` });
+          this.emit({ type: "turn_complete" });
+        }
+        resolve();
+      });
+    });
+  }
+
+  async interrupt(): Promise<void> {
+    // Kill the running turn's exec. Best-effort — the in-container `kimi` may
+    // finish the current tool, but no further output is emitted.
+    this.current?.kill("SIGKILL");
+  }
+
+  // `-p` mode auto-approves tools (the container is the sandbox), so no
+  // permission requests are ever emitted — nothing to resolve.
+  async resolvePermission(): Promise<void> {
+    /* no-op */
+  }
+
+  async close(): Promise<void> {
+    if (this.closed) return;
+    this.closed = true;
+    this.current?.kill("SIGKILL");
+    this.current = null;
+    this.emit({ type: "exit", code: 0 });
+    this.handlers.clear();
+  }
+}
+
+// Register the Kimi Code adapter at module load (ADR-021).
+register({
+  kind: "kimi",
+  label: "Kimi Code",
+  supportedModels: () => [...KIMI_MODELS],
+  defaultModel: KIMI_DEFAULT_MODEL,
+  supportedEfforts: () => [...KIMI_EFFORTS],
+  // Gated on the operator's Kimi Code subscription credential (copied into
+  // containers at task-up). No login → the engine isn't advertised.
+  available: () => existsSync(kimiCredPath()),
+  create: async ({ taskId, agent, environment, systemPreamble, agentEnv }) =>
+    new KimiSession({ taskId, agent, environment, systemPreamble, agentEnv }),
+});