@quintinshaw/pi-dynamic-workflows 1.1.0 → 1.2.0

package/README.md

@@ -86,8 +86,11 @@ return { inventory, summary }
 | `label` | string | Human-readable label for progress display |
 | `phase` | string | Override the current phase for this agent |
 | `schema` | object | JSON Schema for structured output |
+| `model` | string | Run this agent on a specific model — `provider/modelId` or a bare `modelId` |
 | `timeoutMs` | number | Override the default 5-minute agent timeout |
 
+Models can also be set per phase via `meta.phases[].model`. Precedence is `opts.model` > phase model > session default; an unknown model logs a warning and falls back to the default.
+
 ### Structured output
 
 Pass a JSON Schema via `opts.schema` and the subagent returns a validated object:
@@ -117,6 +120,7 @@ Scripts run inside a Node `vm` sandbox. Intentionally unavailable: `Date.now()`,
 - **Core runtime** — `agent` / `parallel` / `pipeline` / `phase` / `log` / `budget` in a sandboxed script
 - **Structured output** — JSON-Schema-validated subagent results
 - **Real token & cost accounting** — read from each subagent's SDK session (input / output / total / cost), with a character estimate only as fallback when a provider reports no usage; `budget` gates on the real total
+- **Real per-agent / per-phase model routing** — `opts.model` and `meta.phases[].model` actually select the model (resolved against your authed model registry), with graceful fallback
 - **Safety limits** — 1000-agent cap (`maxAgents`), per-agent timeout (`agentTimeoutMs`), recoverable-vs-fatal error classification
 - **Live progress + token/cost display**, `Esc` to abort
 - **Log persistence** to `.pi/workflows/runs/`
@@ -125,7 +129,6 @@ Scripts run inside a Node `vm` sandbox. Intentionally unavailable: `Date.now()`,
 
 Tracked toward closer parity with Claude Code dynamic workflows:
 
-- **Real per-agent / per-phase model routing** (`opts.model`, `meta.phases[].model`)
 - **Command surface** — `/workflows` (list / status / stop) and reachable background runs
 - **Resume** — journaled results, replay the unchanged prefix, run the rest live
 - **Worktree isolation** for parallel edits, and **bundled `/deep-research`**

package/dist/agent.d.ts

@@ -30,6 +30,14 @@ export interface AgentRunOptions<TSchemaDef extends TSchema | undefined = undefi
      * usage is never lost. `total === 0` means the provider reported no usage.
      */
     onUsage?: (usage: AgentUsage) => void;
+    /**
+     * Model spec for this subagent: either `provider/modelId` (unambiguous) or a
+     * bare `modelId`. When it can't be resolved, the session default is used and
+     * a warning is logged. When omitted, the session default applies.
+     */
+    model?: string;
+    /** Called with the resolved model id once known (for display/telemetry). */
+    onModelResolved?: (modelId: string) => void;
 }
 export type AgentRunResult<TSchemaDef extends TSchema | undefined> = TSchemaDef extends TSchema ? Static<TSchemaDef> : string;
 export declare class WorkflowAgent {
@@ -37,7 +45,16 @@ export declare class WorkflowAgent {
     private readonly baseTools;
     private readonly sessionOptions;
     private readonly instructions?;
+    /** Lazily built once; shares the SDK's agentDir/auth so resolved models are authed. */
+    private registry?;
     constructor(options?: WorkflowAgentOptions);
+    private getRegistry;
+    /**
+     * Resolve a model spec to a Model. Accepts `provider/modelId` (unambiguous)
+     * or a bare `modelId` (prefers auth-configured models, then any known model).
+     * Returns undefined when nothing matches.
+     */
+    private resolveModel;
     run<TSchemaDef extends TSchema | undefined = undefined>(prompt: string, options?: AgentRunOptions<TSchemaDef>): Promise<AgentRunResult<TSchemaDef>>;
     private buildPrompt;
     private lastAssistantText;

package/dist/agent.js

@@ -1,22 +1,60 @@
-import { createAgentSession, createCodingTools, getAgentDir, SessionManager, SettingsManager, } from "@earendil-works/pi-coding-agent";
+import { join } from "node:path";
+import { AuthStorage, createAgentSession, createCodingTools, getAgentDir, ModelRegistry, SessionManager, SettingsManager, } from "@earendil-works/pi-coding-agent";
 import { createStructuredOutputTool } from "./structured-output.js";
 export class WorkflowAgent {
     cwd;
     baseTools;
     sessionOptions;
     instructions;
+    /** Lazily built once; shares the SDK's agentDir/auth so resolved models are authed. */
+    registry;
     constructor(options = {}) {
         this.cwd = options.cwd ?? process.cwd();
         this.baseTools = options.tools ?? createCodingTools(this.cwd);
         this.sessionOptions = options.session ?? {};
         this.instructions = options.instructions;
     }
+    getRegistry() {
+        if (!this.registry) {
+            const dir = getAgentDir();
+            // Same agentDir/auth files createAgentSession uses by default, so a model
+            // resolved here carries valid credentials.
+            const auth = AuthStorage.create(join(dir, "auth.json"));
+            this.registry = ModelRegistry.create(auth, join(dir, "models.json"));
+        }
+        return this.registry;
+    }
+    /**
+     * Resolve a model spec to a Model. Accepts `provider/modelId` (unambiguous)
+     * or a bare `modelId` (prefers auth-configured models, then any known model).
+     * Returns undefined when nothing matches.
+     */
+    resolveModel(spec) {
+        const registry = this.getRegistry();
+        const slash = spec.indexOf("/");
+        if (slash > 0) {
+            return registry.find(spec.slice(0, slash), spec.slice(slash + 1));
+        }
+        return registry.getAvailable().find((m) => m.id === spec) ?? registry.getAll().find((m) => m.id === spec);
+    }
     async run(prompt, options = {}) {
         const capture = { called: false, value: undefined };
         const customTools = [...this.baseTools, ...(options.tools ?? [])];
         if (options.schema) {
             customTools.push(createStructuredOutputTool({ schema: options.schema, capture }));
         }
+        // Resolve a requested model spec to a Model object. A given-but-unresolved
+        // spec falls back to the session default (with a warning) rather than failing.
+        let resolvedModel;
+        if (options.model) {
+            resolvedModel = this.resolveModel(options.model);
+            if (resolvedModel) {
+                options.onModelResolved?.(`${resolvedModel.provider}/${resolvedModel.id}`);
+            }
+            else {
+                console.warn(`[workflow] model "${options.model}" not found; using session default`);
+            }
+        }
         const agentDir = getAgentDir();
         const { session } = await createAgentSession({
             cwd: this.cwd,
@@ -29,6 +67,8 @@ export class WorkflowAgent {
             settingsManager: SettingsManager.create(this.cwd, agentDir),
             customTools,
             ...this.sessionOptions,
+            // Per-call model wins over any sessionOptions.model.
+            ...(resolvedModel ? { model: resolvedModel } : {}),
         });
         let removeAbortListener;
         try {

package/dist/workflow.d.ts

@@ -31,6 +31,7 @@ export interface WorkflowRunOptions extends WorkflowAgentOptions {
         label: string;
         phase?: string;
         prompt: string;
+        model?: string;
     }) => void;
     onAgentEnd?: (event: {
         label: string;

package/dist/workflow.js

@@ -4,10 +4,13 @@ import { WorkflowAgent } from "./agent.js";
 import { DEFAULT_AGENT_TIMEOUT_MS, MAX_AGENTS_PER_RUN, MAX_CONCURRENCY } from "./config.js";
 import { WorkflowError, WorkflowErrorCode, wrapError } from "./errors.js";
 import { createWorkflowLogger } from "./logger.js";
+import { parseModelRoutingFromMeta, resolveModelForPhase } from "./model-routing.js";
 const DETERMINISM_BLOCKLIST = /\bDate\s*\.\s*now\b|\bMath\s*\.\s*random\b|\bnew\s+Date\s*\(\s*\)/;
 export async function runWorkflow(script, options = {}) {
     const started = Date.now();
     const { meta, body } = parseWorkflowScript(script);
+    // Per-phase model routing from meta.phases[].model (empty when none declared).
+    const routingConfig = parseModelRoutingFromMeta(meta.phases);
     const maxAgents = options.maxAgents ?? MAX_AGENTS_PER_RUN;
     const agentTimeoutMs = options.agentTimeoutMs ?? DEFAULT_AGENT_TIMEOUT_MS;
     const runId = options.runId ?? `run-${started.toString(36)}`;
@@ -62,11 +65,13 @@ export async function runWorkflow(script, options = {}) {
         }
         const assignedPhase = agentOptions.phase ?? state.currentPhase;
         const requestedLabel = agentOptions.label?.trim();
+        // Precedence: explicit agentOptions.model > phase model (meta.phases[].model).
+        const modelSpec = agentOptions.model ?? resolveModelForPhase(assignedPhase, routingConfig);
         return limiter(async () => {
             state.agentCount++;
             const label = requestedLabel || defaultAgentLabel(assignedPhase, state.agentCount);
             const timeout = agentOptions.timeoutMs ?? agentTimeoutMs;
-            options.onAgentStart?.({ label, phase: assignedPhase, prompt });
+            options.onAgentStart?.({ label, phase: assignedPhase, prompt, model: modelSpec });
             // Captured from the subagent's real session usage; falls back to an
             // estimate when the provider reports no usage (total === 0).
             let usage;
@@ -89,6 +94,7 @@ export async function runWorkflow(script, options = {}) {
                     schema: agentOptions.schema,
                     signal: options.signal,
                     instructions: buildAgentInstructions(assignedPhase, agentOptions),
+                    model: modelSpec,
                     onUsage: (u) => {
                         usage = u;
                     },
@@ -350,8 +356,7 @@ function buildAgentInstructions(phase, options) {
         lines.push(`Act as workflow subagent type: ${options.agentType}`);
     if (options.isolation)
         lines.push(`Requested isolation: ${options.isolation}`);
-    if (options.model)
-        lines.push(`Requested model: ${options.model}`);
+    // Note: options.model is applied for real via the session, not injected as prose.
     return lines.length ? lines.join("\n") : undefined;
 }
 function estimateTokens(value) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@quintinshaw/pi-dynamic-workflows",
-  "version": "1.1.0",
+  "version": "1.2.0",
   "description": "Claude-Code-style dynamic workflow orchestration for Pi.",
   "type": "module",
   "main": "./dist/index.js",

package/src/agent.ts

@@ -1,9 +1,12 @@
-import type { AssistantMessage, TextContent } from "@earendil-works/pi-ai";
+import { join } from "node:path";
+import type { AssistantMessage, Model, TextContent } from "@earendil-works/pi-ai";
 import {
+  AuthStorage,
   type CreateAgentSessionOptions,
   createAgentSession,
   createCodingTools,
   getAgentDir,
+  ModelRegistry,
   SessionManager,
   SettingsManager,
   type ToolDefinition,
@@ -43,6 +46,14 @@ export interface AgentRunOptions<TSchemaDef extends TSchema | undefined = undefi
    * usage is never lost. `total === 0` means the provider reported no usage.
    */
   onUsage?: (usage: AgentUsage) => void;
+  /**
+   * Model spec for this subagent: either `provider/modelId` (unambiguous) or a
+   * bare `modelId`. When it can't be resolved, the session default is used and
+   * a warning is logged. When omitted, the session default applies.
+   */
+  model?: string;
+  /** Called with the resolved model id once known (for display/telemetry). */
+  onModelResolved?: (modelId: string) => void;
 }
 
 export type AgentRunResult<TSchemaDef extends TSchema | undefined> = TSchemaDef extends TSchema
@@ -54,6 +65,8 @@ export class WorkflowAgent {
   private readonly baseTools: ToolDefinition[];
   private readonly sessionOptions: Partial<CreateAgentSessionOptions>;
   private readonly instructions?: string;
+  /** Lazily built once; shares the SDK's agentDir/auth so resolved models are authed. */
+  private registry?: ModelRegistry;
 
   constructor(options: WorkflowAgentOptions = {}) {
     this.cwd = options.cwd ?? process.cwd();
@@ -62,6 +75,31 @@ export class WorkflowAgent {
     this.instructions = options.instructions;
   }
 
+  private getRegistry(): ModelRegistry {
+    if (!this.registry) {
+      const dir = getAgentDir();
+      // Same agentDir/auth files createAgentSession uses by default, so a model
+      // resolved here carries valid credentials.
+      const auth = AuthStorage.create(join(dir, "auth.json"));
+      this.registry = ModelRegistry.create(auth, join(dir, "models.json"));
+    }
+    return this.registry;
+  }
+
+  /**
+   * Resolve a model spec to a Model. Accepts `provider/modelId` (unambiguous)
+   * or a bare `modelId` (prefers auth-configured models, then any known model).
+   * Returns undefined when nothing matches.
+   */
+  private resolveModel(spec: string): Model<any> | undefined {
+    const registry = this.getRegistry();
+    const slash = spec.indexOf("/");
+    if (slash > 0) {
+      return registry.find(spec.slice(0, slash), spec.slice(slash + 1));
+    }
+    return registry.getAvailable().find((m) => m.id === spec) ?? registry.getAll().find((m) => m.id === spec);
+  }
+
   async run<TSchemaDef extends TSchema | undefined = undefined>(
     prompt: string,
     options: AgentRunOptions<TSchemaDef> = {},
@@ -73,6 +111,18 @@ export class WorkflowAgent {
       customTools.push(createStructuredOutputTool({ schema: options.schema, capture }) as unknown as ToolDefinition);
     }
 
+    // Resolve a requested model spec to a Model object. A given-but-unresolved
+    // spec falls back to the session default (with a warning) rather than failing.
+    let resolvedModel: Model<any> | undefined;
+    if (options.model) {
+      resolvedModel = this.resolveModel(options.model);
+      if (resolvedModel) {
+        options.onModelResolved?.(`${resolvedModel.provider}/${resolvedModel.id}`);
+      } else {
+        console.warn(`[workflow] model "${options.model}" not found; using session default`);
+      }
+    }
+
     const agentDir = getAgentDir();
     const { session } = await createAgentSession({
       cwd: this.cwd,
@@ -85,6 +135,8 @@ export class WorkflowAgent {
       settingsManager: SettingsManager.create(this.cwd, agentDir),
       customTools,
       ...this.sessionOptions,
+      // Per-call model wins over any sessionOptions.model.
+      ...(resolvedModel ? { model: resolvedModel } : {}),
     });
 
     let removeAbortListener: (() => void) | undefined;

package/src/workflow.ts

@@ -7,6 +7,7 @@ import { WorkflowAgent, type WorkflowAgentOptions } from "./agent.js";
 import { DEFAULT_AGENT_TIMEOUT_MS, MAX_AGENTS_PER_RUN, MAX_CONCURRENCY } from "./config.js";
 import { WorkflowError, WorkflowErrorCode, wrapError } from "./errors.js";
 import { createWorkflowLogger } from "./logger.js";
+import { parseModelRoutingFromMeta, resolveModelForPhase } from "./model-routing.js";
 
 export interface WorkflowMetaPhase {
   title: string;
@@ -37,7 +38,7 @@ export interface WorkflowRunOptions extends WorkflowAgentOptions {
   runId?: string;
   onLog?: (message: string) => void;
   onPhase?: (title: string) => void;
-  onAgentStart?: (event: { label: string; phase?: string; prompt: string }) => void;
+  onAgentStart?: (event: { label: string; phase?: string; prompt: string; model?: string }) => void;
   onAgentEnd?: (event: { label: string; phase?: string; result: unknown; tokens?: number }) => void;
   onTokenUsage?: (usage: { input: number; output: number; total: number; cost: number }) => void;
 }
@@ -93,6 +94,8 @@ export async function runWorkflow<T = unknown>(
 ): Promise<WorkflowRunResult<T>> {
   const started = Date.now();
   const { meta, body } = parseWorkflowScript(script);
+  // Per-phase model routing from meta.phases[].model (empty when none declared).
+  const routingConfig = parseModelRoutingFromMeta(meta.phases);
   const maxAgents = options.maxAgents ?? MAX_AGENTS_PER_RUN;
   const agentTimeoutMs = options.agentTimeoutMs ?? DEFAULT_AGENT_TIMEOUT_MS;
   const runId = options.runId ?? `run-${started.toString(36)}`;
@@ -164,13 +167,15 @@ export async function runWorkflow<T = unknown>(
 
     const assignedPhase = agentOptions.phase ?? state.currentPhase;
     const requestedLabel = agentOptions.label?.trim();
+    // Precedence: explicit agentOptions.model > phase model (meta.phases[].model).
+    const modelSpec = agentOptions.model ?? resolveModelForPhase(assignedPhase, routingConfig);
 
     return limiter(async () => {
       state.agentCount++;
       const label = requestedLabel || defaultAgentLabel(assignedPhase, state.agentCount);
       const timeout = agentOptions.timeoutMs ?? agentTimeoutMs;
 
-      options.onAgentStart?.({ label, phase: assignedPhase, prompt });
+      options.onAgentStart?.({ label, phase: assignedPhase, prompt, model: modelSpec });
 
       // Captured from the subagent's real session usage; falls back to an
       // estimate when the provider reports no usage (total === 0).
@@ -197,6 +202,7 @@ export async function runWorkflow<T = unknown>(
             schema: agentOptions.schema,
             signal: options.signal,
             instructions: buildAgentInstructions(assignedPhase, agentOptions),
+            model: modelSpec,
             onUsage: (u: AgentUsage) => {
               usage = u;
             },
@@ -480,7 +486,7 @@ function buildAgentInstructions(phase: string | undefined, options: AgentOptions
   if (phase) lines.push(`Workflow phase: ${phase}`);
   if (options.agentType) lines.push(`Act as workflow subagent type: ${options.agentType}`);
   if (options.isolation) lines.push(`Requested isolation: ${options.isolation}`);
-  if (options.model) lines.push(`Requested model: ${options.model}`);
+  // Note: options.model is applied for real via the session, not injected as prose.
   return lines.length ? lines.join("\n") : undefined;
 }