fix: Model flag parsing and model args (#42)

2025-10-30 20:23:21 -07:00
parent 77d78c22da
commit 32a3f7c7ab
7 changed files with 106 additions and 60 deletions
--- a/src/agent/create.ts
+++ b/src/agent/create.ts
@@ -7,7 +7,6 @@ import type {
  Block,
  CreateBlock,
 } from "@letta-ai/letta-client/resources/agents/agents";
-import { formatAvailableModels, resolveModel } from "../model";
 import {
  loadProjectSettings,
  updateProjectSettings,
@@ -16,12 +15,15 @@ import { loadSettings, updateSettings } from "../settings";
 import { getToolNames } from "../tools/manager";
 import { getClient } from "./client";
 import { getDefaultMemoryBlocks } from "./memory";
+import { formatAvailableModels, resolveModel } from "./model";
+import { updateAgentLLMConfig } from "./modify";
 import { SYSTEM_PROMPT } from "./promptAssets";

 export async function createAgent(
  name = "letta-cli-agent",
  model?: string,
  embeddingModel = "openai/text-embedding-3-small",
+  updateArgs?: Record<string, unknown>,
 ) {
  // Resolve model identifier to handle
  let modelHandle: string;
@@ -168,5 +170,13 @@ export async function createAgent(
    include_base_tool_rules: false,
    initial_message_sequence: [],
  });
+
+  // Apply updateArgs if provided (e.g., reasoningEffort, contextWindow, etc.)
+  if (updateArgs && Object.keys(updateArgs).length > 0) {
+    await updateAgentLLMConfig(agent.id, modelHandle, updateArgs);
+    // Refresh agent state to get updated config
+    return await client.agents.retrieve(agent.id);
+  }
+
  return agent; // { id, ... }
 }
--- a/src/agent/model.ts
+++ b/src/agent/model.ts
@@ -0,0 +1,64 @@
+/**
+ * Model resolution and handling utilities
+ */
+import modelsData from "../models.json";
+
+export const models = modelsData;
+
+/**
+ * Resolve a model by ID or handle
+ * @param modelIdentifier - Can be either a model ID (e.g., "opus") or a full handle (e.g., "anthropic/claude-opus-4-1-20250805")
+ * @returns The model handle if found, null otherwise
+ */
+export function resolveModel(modelIdentifier: string): string | null {
+  const byId = models.find((m) => m.id === modelIdentifier);
+  if (byId) return byId.handle;
+
+  const byHandle = models.find((m) => m.handle === modelIdentifier);
+  if (byHandle) return byHandle.handle;
+
+  return null;
+}
+
+/**
+ * Get the default model handle
+ */
+export function getDefaultModel(): string {
+  const defaultModel = models.find((m) => m.isDefault);
+  return defaultModel?.handle || models[0].handle;
+}
+
+/**
+ * Format available models for error messages
+ */
+export function formatAvailableModels(): string {
+  return models.map((m) => `  ${m.id.padEnd(20)} ${m.handle}`).join("\n");
+}
+
+/**
+ * Get model info by ID or handle
+ * @param modelIdentifier - Can be either a model ID (e.g., "opus") or a full handle (e.g., "anthropic/claude-opus-4-1-20250805")
+ * @returns The model info if found, null otherwise
+ */
+export function getModelInfo(modelIdentifier: string) {
+  const byId = models.find((m) => m.id === modelIdentifier);
+  if (byId) return byId;
+
+  const byHandle = models.find((m) => m.handle === modelIdentifier);
+  if (byHandle) return byHandle;
+
+  return null;
+}
+
+/**
+ * Get updateArgs for a model by ID or handle
+ * @param modelIdentifier - Can be either a model ID (e.g., "opus") or a full handle (e.g., "anthropic/claude-opus-4-1-20250805")
+ * @returns The updateArgs if found, undefined otherwise
+ */
+export function getModelUpdateArgs(
+  modelIdentifier?: string,
+): Record<string, unknown> | undefined {
+  if (!modelIdentifier) return undefined;
+  const modelInfo = getModelInfo(modelIdentifier);
+  return modelInfo?.updateArgs;
+}
--- a/src/cli/components/ModelSelector.tsx
+++ b/src/cli/components/ModelSelector.tsx
@@ -1,7 +1,7 @@
 // Import useInput from vendored Ink for bracketed paste support
 import { Box, Text, useInput } from "ink";
 import { useState } from "react";
-import { models } from "../../model";
+import { models } from "../../agent/model";
 import { colors } from "./colors";

 interface ModelSelectorProps {
--- a/src/headless.ts
+++ b/src/headless.ts
@@ -8,6 +8,7 @@ import type { StopReasonType } from "@letta-ai/letta-client/resources/runs/runs"
 import { getClient } from "./agent/client";
 import { createAgent } from "./agent/create";
 import { sendMessageStream } from "./agent/message";
+import { getModelUpdateArgs } from "./agent/model";
 import { SessionStats } from "./agent/stats";
 import { createBuffers, toLines } from "./cli/helpers/accumulator";
 import { safeJsonParseOr } from "./cli/helpers/safeJsonParse";
@@ -25,6 +26,8 @@ export async function handleHeadlessCommand(argv: string[], model?: string) {
      continue: { type: "boolean", short: "c" },
      new: { type: "boolean" },
      agent: { type: "string", short: "a" },
+      model: { type: "string", short: "m" },
+      prompt: { type: "boolean", short: "p" },
      "output-format": { type: "string" },
    },
    strict: false,
@@ -70,7 +73,8 @@ export async function handleHeadlessCommand(argv: string[], model?: string) {

  // Priority 2: Check if --new flag was passed (skip all resume logic)
  if (!agent && forceNew) {
-    agent = await createAgent(undefined, model);
+    const updateArgs = getModelUpdateArgs(model);
+    agent = await createAgent(undefined, model, undefined, updateArgs);
  }

  // Priority 3: Try to resume from project settings (.letta/settings.local.json)
@@ -101,7 +105,8 @@ export async function handleHeadlessCommand(argv: string[], model?: string) {

  // Priority 5: Create a new agent
  if (!agent) {
-    agent = await createAgent(undefined, model);
+    const updateArgs = getModelUpdateArgs(model);
+    agent = await createAgent(undefined, model, undefined, updateArgs);
  }

  // Save agent ID to both project and global settings
--- a/src/index.ts
+++ b/src/index.ts
@@ -218,6 +218,7 @@ async function main() {

        setLoadingState("initializing");
        const { createAgent } = await import("./agent/create");
+        const { getModelUpdateArgs } = await import("./agent/model");
        const { updateSettings, loadProjectSettings, updateProjectSettings } =
          await import("./settings");

@@ -238,7 +239,8 @@ async function main() {
        // Priority 2: Check if --new flag was passed (skip all resume logic)
        if (!agent && forceNew) {
          // Create new agent, don't check any lastAgent fields
-          agent = await createAgent(undefined, model);
+          const updateArgs = getModelUpdateArgs(model);
+          agent = await createAgent(undefined, model, undefined, updateArgs);
        }

        // Priority 3: Try to resume from project settings (.letta/settings.local.json)
@@ -270,7 +272,8 @@ async function main() {

        // Priority 5: Create a new agent
        if (!agent) {
-          agent = await createAgent(undefined, model);
+          const updateArgs = getModelUpdateArgs(model);
+          agent = await createAgent(undefined, model, undefined, updateArgs);
        }

        // Save agent ID to both project and global settings
--- a/src/model.ts
+++ b/src/model.ts
@@ -1,36 +0,0 @@
-/**
- * Model resolution and handling utilities
- */
-import modelsData from "./models.json";
-
-export const models = modelsData;
-
-/**
- * Resolve a model by ID or handle
- * @param modelIdentifier - Can be either a model ID (e.g., "opus") or a full handle (e.g., "anthropic/claude-opus-4-1-20250805")
- * @returns The model handle if found, null otherwise
- */
-export function resolveModel(modelIdentifier: string): string | null {
-  const byId = models.find((m) => m.id === modelIdentifier);
-  if (byId) return byId.handle;
-
-  const byHandle = models.find((m) => m.handle === modelIdentifier);
-  if (byHandle) return byHandle.handle;
-
-  return null;
-}
-
-/**
- * Get the default model handle
- */
-export function getDefaultModel(): string {
-  const defaultModel = models.find((m) => m.isDefault);
-  return defaultModel?.handle || models[0].handle;
-}
-
-/**
- * Format available models for error messages
- */
-export function formatAvailableModels(): string {
-  return models.map((m) => `  ${m.id.padEnd(20)} ${m.handle}`).join("\n");
-}
--- a/src/models.json
+++ b/src/models.json
@@ -5,21 +5,21 @@
    "label": "Claude Sonnet 4.5 (default)",
    "description": "The recommended default model (currently Sonnet 4.5)",
    "isDefault": true,
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  },
  {
    "id": "opus",
    "handle": "anthropic/claude-opus-4-1-20250805",
    "label": "Claude Opus 4.1",
    "description": "Anthropic's smartest (and slowest) model",
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  },
  {
    "id": "haiku",
    "handle": "anthropic/claude-haiku-4-5-20251001",
    "label": "Claude Haiku 4.5",
    "description": "Anthropic's fastest model",
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  },
  {
    "id": "gpt-5-codex",
@@ -27,9 +27,9 @@
    "label": "GPT-5-Codex",
    "description": "A variant of GPT-5 optimized for agentic coding",
    "updateArgs": {
-      "reasoningEffort": "medium",
+      "reasoning_effort": "medium",
      "verbosity": "medium",
-      "contextWindow": 272000
+      "context_window": 272000
    }
  },
  {
@@ -38,7 +38,7 @@
    "label": "GLM-4.6",
    "description": "The best open weights coding model",
    "updateArgs": {
-      "contextWindow": 200000
+      "context_window": 200000
    }
  },
  {
@@ -47,9 +47,9 @@
    "label": "GPT-5 (minimal)",
    "description": "OpenAI's latest model (limited reasoning, fastest GPT-5 option)",
    "updateArgs": {
-      "reasoningEffort": "minimal",
+      "reasoning_effort": "minimal",
      "verbosity": "medium",
-      "contextWindow": 272000
+      "context_window": 272000
    }
  },
  {
@@ -58,9 +58,9 @@
    "label": "GPT-5 (low)",
    "description": "OpenAI's latest model (some reasoning enabled)",
    "updateArgs": {
-      "reasoningEffort": "low",
+      "reasoning_effort": "low",
      "verbosity": "medium",
-      "contextWindow": 272000
+      "context_window": 272000
    }
  },
  {
@@ -69,9 +69,9 @@
    "label": "GPT-5 (medium)",
    "description": "OpenAI's latest model (using their recommended reasoning level)",
    "updateArgs": {
-      "reasoningEffort": "medium",
+      "reasoning_effort": "medium",
      "verbosity": "medium",
-      "contextWindow": 272000
+      "context_window": 272000
    }
  },
  {
@@ -80,9 +80,9 @@
    "label": "GPT-5 (high)",
    "description": "OpenAI's latest model (maximum reasoning depth)",
    "updateArgs": {
-      "reasoningEffort": "high",
+      "reasoning_effort": "high",
      "verbosity": "medium",
-      "contextWindow": 272000
+      "context_window": 272000
    }
  },
  {
@@ -90,27 +90,27 @@
    "handle": "google_ai/gemini-2.5-flash",
    "label": "Gemini 2.5 Flash",
    "description": "Google's fastest model",
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  },
  {
    "id": "gemini-pro",
    "handle": "google_ai/gemini-2.5-pro",
    "label": "Gemini 2.5 Pro",
    "description": "Google's smartest model",
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  },
  {
    "id": "gpt-4.1",
    "handle": "openai/gpt-4.1",
    "label": "GPT-4.1",
    "description": "OpenAI's most recent non-reasoner model",
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  },
  {
    "id": "o4-mini",
    "handle": "openai/o4-mini",
    "label": "o4-mini",
    "description": "OpenAI's latest o-series reasoning model",
-    "updateArgs": { "contextWindow": 180000 }
+    "updateArgs": { "context_window": 180000 }
  }
 ]