@staix/agent-hub 0.12.39 → 0.12.41

package/templates/routing.toml

@@ -1,6 +1,5 @@
 # agent-hub routing: who takes which kind of task (L1), and which model route the local worker asks for (L2).
-# [targets.*] and [routes.*] use Switchyard's own table shapes (v0.2.0 toml_schema): the hub copies them into the
-# sidecar config and adds the gateway client, so anything Switchyard accepts can be written here.
+# Hub routes run in-process and name gateway aliases directly.
 
 # Top-level keys, before any table. Pi's hub/auto and the hub stage routes keep a session's model through a tool loop and
 # switch only where it pays (#197): "off", "shadow" (record in events.jsonl what the planner would do, routing unchanged)
@@ -9,41 +8,8 @@ stay_switch = "shadow"
 max_switch_prefill_tokens = 32000
 
 [local]
-route = "sy/coding"                                        # Switchyard route the local worker asks for
-fixed_model = "vllm/deepseek-ai/DeepSeek-V4-Flash-0731"    # straight to OmniRoute when the sidecar is absent or unhealthy
-
-[targets.dsv4f]
-id = "vllm/deepseek-ai/DeepSeek-V4-Flash-0731"
-
-[routes."sy/coding"]
-type = "passthrough"
-target = "dsv4f"
-
-[routes."sy/fast"]
-type = "passthrough"
-target = "dsv4f"
-
-# With a second model served (here GLM-5.3-Flash), replace sy/coding with the two-tier route and add the review escalation:
-#
-# [targets.glm53f]
-# id = "vllm/zai-org/GLM-5.3-Flash"
-#
-# [routes."sy/coding"]
-# type = "stage_router"
-# efficient_target = "glm53f"
-# capable_target = "dsv4f"
-# picker = "efficient_first"
-# confidence_threshold = 0.5
-#
-# [routes."sy/review"]
-# type = "llm_classifier"
-# mode = "escalation"
-# classifier_target = "glm53f"
-# weak_target = "glm53f"
-# strong_target = "dsv4f"
-#
-# [routes."sy/review".escalation]      # required by the 0.2.0 binary for mode = "escalation"
-# confirmations = 2
+route = "hub/stage"             # hub-owned model routing
+fixed_model = "coding"          # direct gateway fallback when the route is unavailable
 
 # L1: task assignment. `peers` is the preference order among attached, unpaused peers (idle before busy).
 # `route` / `fixed_model` apply when the task lands on the local worker. `tasks.escalate_after` consecutive
@@ -70,19 +36,19 @@ local_allowed = false
 
 [classes.implement]
 peers = ["pi", "local", "codex", "kimi"]
-route = "sy/coding"
+route = "hub/stage"
 pi_backend = "dgx"
 escalate_to = ["codex", "kimi", "claude"]
 
 [classes.bulk_edit]
 peers = ["pi", "local", "kimi"]
-route = "sy/fast"
+fixed_model = "fast"
 pi_backend = "dgx"
 escalate_to = ["codex", "kimi"]
 
 [classes.test]
 peers = ["pi", "local", "kimi", "codex"]
-route = "sy/fast"
+fixed_model = "fast"
 pi_backend = "dgx"
 escalate_to = ["codex", "kimi"]
 
@@ -92,18 +58,18 @@ local_allowed = false         # judgment classes never go to local
 
 [classes.summarize]
 peers = ["pi", "local", "kimi"]
-route = "sy/fast"
+fixed_model = "fast"
 pi_backend = "mlx"
 escalate_to = ["codex", "kimi"]
 
 [classes.triage]
 peers = ["pi", "local", "kimi"]
-route = "sy/fast"
+fixed_model = "fast"
 pi_backend = "mlx"
 escalate_to = ["codex", "kimi"]
 
-# In-process L2 routes are opt-in: set local.route or a class route to hub/<id>.
-# They never enter the optional Switchyard sidecar TOML. Defaults are efficient fast, capable coding.
+# In-process L2 routes: set local.route or a class route to hub/<id>.
+# A class with fixed_model and no route calls that alias directly. Defaults are efficient fast, capable coding.
 [hub_routes."hub/stage"]
 type = "stage"
 efficient = "fast"

package/src/switchyard/config.ts

@@ -1,45 +0,0 @@
-import type { Routing } from "../hub/routing.ts";
-
-export const KEY_ENV = "OMNIROUTE_API_KEY";
-
-const scalar = (v: unknown): string => {
-  if (typeof v === "string") return JSON.stringify(v);
-  if (typeof v === "number" || typeof v === "boolean") return String(v);
-  if (Array.isArray(v)) return `[${v.map(scalar).join(", ")}]`;
-  throw new Error(`unsupported TOML value: ${JSON.stringify(v)}`);
-};
-
-/** Bare where TOML allows it, quoted otherwise: `glm-5.3` as a bare header would parse as the nested table "glm-5"."3". */
-const key = (k: string) => (/^[A-Za-z0-9_-]+$/.test(k) ? k : JSON.stringify(k));
-
-function table(header: string, entries: Record<string, unknown>): string {
-  const flat = Object.entries(entries).filter(([, v]) => typeof v !== "object" || Array.isArray(v));
-  const nested = Object.entries(entries).filter(([, v]) => v && typeof v === "object" && !Array.isArray(v));
-  return [
-    `[${header}]`,
-    ...flat.map(([k, v]) => `${key(k)} = ${scalar(v)}`),
-    "",
-    ...nested.map(([k, v]) => table(`${header}.${key(k)}`, v as Record<string, unknown>)),
-  ].join("\n");
-}
-
-/**
- * Switchyard v0.2.0 server config. routing.toml already uses Switchyard's table shapes, so this only adds what the
- * hub knows: the gateway client, `llm_client` on targets and the public `id` on routes. The API key is referenced
- * by env var name and never written; Access headers, when the live URL needs them, are the one secret in the file.
- */
-export function switchyardToml(routing: Routing, gateway: { baseUrl: string; extraHeaders: Record<string, string> }): string {
-  const headers = Object.entries(gateway.extraHeaders);
-  const client = [
-    "[llm_clients.gateway]",
-    'format = "openai_chat"',
-    `base_url = ${scalar(gateway.baseUrl)}`,
-    `api_key_env = ${scalar(KEY_ENV)}`,
-    ...(headers.length ? [`extra_headers = { ${headers.map(([k, v]) => `${scalar(k)} = ${scalar(v)}`).join(", ")} }`] : []),
-    // no timeout_ms: the 0.2.0 binary rejects it (its llm_clients fields are format, base_url, api_key_env, extra_headers, max_retries)
-    "",
-  ].join("\n");
-  const targets = Object.entries(routing.targets).map(([name, t]) => table(`targets.${key(name)}`, { ...t, llm_client: "gateway" }));
-  const routes = Object.entries(routing.routes).map(([id, r], i) => table(`routes.r${i}`, { id, ...r }));
-  return ["schema_version = 1", "", client, ...targets, ...routes].join("\n");
-}

package/src/switchyard/sidecar.ts

@@ -1,136 +0,0 @@
-import { spawn, spawnSync, type ChildProcess } from "node:child_process";
-import { chmodSync, rmSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import type { Routing } from "../hub/routing.ts";
-import type { OmniRoute } from "../omniroute/client.ts";
-import { KEY_ENV, switchyardToml } from "./config.ts";
-import { stopOwnedProcess } from "../hub/child-process.ts";
-
-class GatewayDown extends Error {}
-
-
-export interface SidecarOptions {
-  routing: Routing;
-  omni: OmniRoute;
-  stateDir: string;
-  port: number;
-  /** `AGENTHUB_SWITCHYARD_BIN`, else `switchyard-server` on PATH. */
-  bin?: string;
-  log: (line: string) => void;
-}
-
-/**
- * L2 as a session-scoped sidecar: started on the first hub-owned model call, stopped with the hub, never resident.
- * Every failure (no binary, bad config, unhealthy, a failed call) turns it off for the rest of the hub run and the
- * caller falls back to `fixed_model` on OmniRoute directly.
- */
-export class Sidecar {
-  private proc: ChildProcess | undefined;
-  private starting: Promise<string | undefined> | undefined;
-  private generation = 0;
-  private stopped = false;
-  private stopping: Promise<void> | undefined;
-  private off = "";
-  /** The gateway URL the running sidecar was generated with; calls through it go there whatever the client probes later. */
-  upstream: string | undefined;
-
-  constructor(private readonly opts: SidecarOptions) {}
-
-  get status(): string {
-    return this.off ? `off (${this.off})` : this.proc ? `127.0.0.1:${this.opts.port}` : "not started";
-  }
-
-  /** Base URL to send chat calls to, or undefined when the hub should call OmniRoute directly. */
-  endpoint(): Promise<string | undefined> {
-    if (this.off || this.stopped) return Promise.resolve(undefined);
-    return (this.starting ??= this.start().catch((e: Error) => {
-      // An unreachable gateway says nothing about Switchyard: this call goes direct, the next one tries the sidecar again.
-      if (e instanceof GatewayDown) this.starting = undefined;
-      else this.disable(e.message);
-      return undefined;
-    }));
-  }
-
-  /** A call through the sidecar failed: stop trusting it. Logged once. */
-  disable(reason: string): void {
-    if (this.off) return;
-    this.off = reason;
-    this.opts.log(`switchyard: off, falling back to fixed_model on OmniRoute (${reason})`);
-    void this.stop().catch((e: Error) => this.opts.log(`switchyard: shutdown incomplete (${e.message})`));
-  }
-
-  async stop(): Promise<void> {
-    if (this.stopping) return this.stopping;
-    this.stopped = true;
-    ++this.generation;
-    const proc = this.proc;
-    // Preserve the historical synchronous file cleanup; the returned promise
-    // still represents the child process shutdown and must be awaited by the hub.
-    rmSync(join(this.opts.stateDir, "switchyard.toml"), { force: true });
-    this.stopping = (async () => {
-      if (proc) await stopOwnedProcess(proc);
-      if (this.proc === proc) this.proc = undefined;
-    })();
-    return this.stopping;
-  }
-
-  private async start(): Promise<string | undefined> {
-    const generation = this.generation;
-    if (this.stopped) return undefined;
-    const { routing, omni, stateDir, port, log } = this.opts;
-    if (!Object.keys(routing.routes).length) throw new Error("routing.toml defines no routes");
-    const bin = this.opts.bin ?? process.env.AGENTHUB_SWITCHYARD_BIN ?? "switchyard-server";
-    const baseUrl = await omni.base();
-    const key = omni.apiKey();
-    if (!baseUrl || !key) throw new GatewayDown("no OmniRoute endpoint or key");
-    if (this.stopped || generation !== this.generation) return undefined;
-    this.upstream = baseUrl;
-
-    const file = join(stateDir, "switchyard.toml");
-    writeFileSync(file, switchyardToml(routing, { baseUrl, extraHeaders: omni.accessHeaders(baseUrl) }), { mode: 0o600 });
-    chmodSync(file, 0o600);
-    const env = { PATH: process.env.PATH ?? "", HOME: process.env.HOME ?? "", [KEY_ENV]: key };
-    const check = spawnSync(bin, ["--config", file, "--dry-run"], { env, encoding: "utf8" });
-    if (check.error) throw new Error(`${bin} not runnable: ${check.error.message}`);
-    if (check.status !== 0) throw new Error(`config rejected: ${(check.stderr || check.stdout).trim().slice(0, 300)}`);
-
-    // Switchyard's default host is 0.0.0.0; the hub binds loopback only.
-    const proc = spawn(bin, ["--config", file, "--host", "127.0.0.1", "--port", String(port)], { env, stdio: ["ignore", "ignore", "pipe"] });
-    if (this.stopped || generation !== this.generation) {
-      proc.kill("SIGTERM");
-      await stopOwnedProcess(proc);
-      return undefined;
-    }
-    this.proc = proc;
-    let exited = false;
-    // Switchyard logs every request at INFO on stderr; only problems belong in hub.log.
-    proc.stderr?.on("data", (d) => {
-      for (const line of String(d).split("\n")) if (/\b(WARN|ERROR)\b/.test(line)) log(`[switchyard] ${line.trim().slice(0, 500)}`);
-    });
-    proc.on("error", () => (exited = true));
-    proc.on("exit", (code) => {
-      exited = true;
-      if (this.proc === proc) this.disable(`exited with code ${code}`);
-    });
-    for (let i = 0; i < 100 && !exited; i++) {
-      if (this.stopped || generation !== this.generation) {
-        await stopOwnedProcess(proc);
-        if (this.proc === proc) this.proc = undefined;
-        return undefined;
-      }
-      if (await fetch(`http://127.0.0.1:${port}/health`).then((r) => r.ok, () => false)) {
-        if (this.stopped || generation !== this.generation) {
-          await stopOwnedProcess(proc);
-          if (this.proc === proc) this.proc = undefined;
-          return undefined;
-        }
-        log(`switchyard: up on 127.0.0.1:${port}, routes ${Object.keys(routing.routes).join(", ")}`);
-        return `http://127.0.0.1:${port}/v1`;
-      }
-      await Bun.sleep(100);
-    }
-    await stopOwnedProcess(proc);
-    if (this.proc === proc) this.proc = undefined;
-    throw new Error(exited ? "exited during startup" : "not healthy within 10 s");
-  }
-}