@phren/agent 0.3.19 → 0.3.20

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,73 @@
1
+ /**
2
+ * `phren-agent models --json`: the models this computer can run, for the
3
+ * phren Hook's model picker. Only providers with credentials here are
4
+ * listed, in the order auto-detection picks them, so the first is the
5
+ * default. Each id is `<provider>/<model>`; the Hook passes it back as
6
+ * `--provider` and `--model`. Loads no agent runtime and makes no request.
7
+ */
8
+ import { resolveApiKey } from "@phren/cli/auth/profiles";
9
+ import { codexConfiguredModel, hasCodexToken } from "./providers/codex-auth.js";
10
+ import { getBuiltinModels, getDefaultModel } from "./models.js";
11
+ const PROVIDER_NAMES = {
12
+ "openai-codex": "ChatGPT subscription",
13
+ openai: "OpenAI API",
14
+ openrouter: "OpenRouter",
15
+ anthropic: "Anthropic API",
16
+ deepseek: "DeepSeek API",
17
+ ollama: "Ollama",
18
+ };
19
+ /** Providers with credentials on this computer, in resolveProvider's order.
20
+ * openai-compat needs an endpoint and model of its own, so it is never listed. */
21
+ export function credentialedProviders(env = process.env) {
22
+ const providers = [];
23
+ if (hasCodexToken())
24
+ providers.push("openai-codex");
25
+ if (resolveApiKey("openai", "OPENAI_API_KEY"))
26
+ providers.push("openai");
27
+ if (resolveApiKey("openrouter", "OPENROUTER_API_KEY"))
28
+ providers.push("openrouter");
29
+ if (resolveApiKey("anthropic", "ANTHROPIC_API_KEY"))
30
+ providers.push("anthropic");
31
+ if (env.DEEPSEEK_API_KEY)
32
+ providers.push("deepseek");
33
+ if (env.PHREN_OLLAMA_URL && env.PHREN_OLLAMA_URL !== "off")
34
+ providers.push("ollama");
35
+ return providers;
36
+ }
37
+ export function listModels(providers = credentialedProviders()) {
38
+ const models = [];
39
+ for (const provider of providers) {
40
+ const entries = getBuiltinModels(provider);
41
+ // Codex's own config can name a model the catalog doesn't have yet.
42
+ const configured = provider === "openai-codex" ? codexConfiguredModel() : undefined;
43
+ if (configured && !entries.some(entry => entry.id === configured)) {
44
+ entries.unshift({ ...entries[0], id: configured, label: configured });
45
+ }
46
+ const preferred = configured ?? getDefaultModel(provider);
47
+ for (const entry of entries) {
48
+ models.push({
49
+ id: `${provider}/${entry.id}`,
50
+ name: entry.label,
51
+ provider,
52
+ description: PROVIDER_NAMES[provider] ?? provider,
53
+ ...(!models.some(model => model.isDefault) && entry.id === preferred ? { isDefault: true } : {}),
54
+ ...(entry.reasoningDefault && entry.reasoningDefault !== "none" ? { defaultReasoningEffort: entry.reasoningDefault } : {}),
55
+ supportedReasoningEfforts: entry.reasoningRange.filter(level => level !== "none"),
56
+ });
57
+ }
58
+ }
59
+ return models;
60
+ }
61
+ export function printModels(args) {
62
+ const models = listModels();
63
+ if (args.includes("--json")) {
64
+ console.log(JSON.stringify({ models }));
65
+ return;
66
+ }
67
+ if (!models.length) {
68
+ console.log("No provider credentials found. See phren-agent --help.");
69
+ return;
70
+ }
71
+ for (const model of models)
72
+ console.log(`${model.isDefault ? "*" : " "} ${model.id.padEnd(48)} ${model.name} (${model.description})`);
73
+ }
package/dist/models.js CHANGED
@@ -1,4 +1,21 @@
1
+ /** Levels the pickers step through ("none" is CLI/env only). */
1
2
  export const REASONING_LEVELS = ["low", "medium", "high", "xhigh"];
3
+ /**
4
+ * DeepSeek V4.1 Flash, direct or through an OpenAI-compatible relay: 1M
5
+ * context, 393,216 max output, effort low/high/max (picker "xhigh" = max).
6
+ * Off-peak rates; DeepSeek bills 2x at peak (weekdays 01-04 and 06-10 UTC).
7
+ * Checked against https://api-docs.deepseek.com/quick_start/pricing on
8
+ * 2026-10-01: cache hit $0.003 / miss $0.15 / output $0.60 off-peak
9
+ * ($0.006 / $0.30 / $1.20 peak).
10
+ */
11
+ const DEEPSEEK_FLASH = {
12
+ vision: true,
13
+ contextWindow: 1_000_000,
14
+ maxOutputTokens: 393_216,
15
+ reasoningDefault: null,
16
+ reasoningRange: ["low", "high", "xhigh"],
17
+ pricing: { inputPer1M: 0.15, outputPer1M: 0.6, cacheReadPer1M: 0.003 },
18
+ };
2
19
  const BUILTIN_MODELS = {
3
20
  anthropic: [
4
21
  {
@@ -279,16 +296,8 @@ const BUILTIN_MODELS = {
279
296
  },
280
297
  ],
281
298
  deepseek: [
282
- {
283
- id: "deepseek-flash",
284
- provider: "deepseek",
285
- label: "DeepSeek V4.1 Flash (direct)",
286
- contextWindow: 1_000_000,
287
- maxOutputTokens: 65_536,
288
- reasoningDefault: null,
289
- reasoningRange: ["low", "medium", "high"],
290
- pricing: { inputPer1M: 0.3, outputPer1M: 1.2 },
291
- },
299
+ { ...DEEPSEEK_FLASH, id: "deepseek-flash", provider: "deepseek", label: "DeepSeek V4.1 Flash (direct)" },
300
+ { ...DEEPSEEK_FLASH, id: "deepseek-v4-flash", provider: "deepseek", label: "DeepSeek V4.1 Flash (legacy id)" },
292
301
  {
293
302
  id: "deepseek-v4-pro",
294
303
  provider: "deepseek",
@@ -296,10 +305,15 @@ const BUILTIN_MODELS = {
296
305
  contextWindow: 1_000_000,
297
306
  maxOutputTokens: 65_536,
298
307
  reasoningDefault: null,
299
- reasoningRange: ["low", "medium", "high"],
300
- pricing: { inputPer1M: 1.32, outputPer1M: 3.96 },
308
+ reasoningRange: ["low", "high", "xhigh"],
309
+ pricing: { inputPer1M: 0.66, outputPer1M: 1.98, cacheReadPer1M: 0.022 },
301
310
  },
302
311
  ],
312
+ // Model ids for a generic OpenAI-compatible endpoint (OpenCode Go/Zen…).
313
+ // Pricing applies to metered relays; see isUnmeteredEndpoint for Go.
314
+ "openai-compat": [
315
+ { ...DEEPSEEK_FLASH, id: "deepseek-v4.1-flash", provider: "openai-compat", label: "DeepSeek V4.1 Flash" },
316
+ ],
303
317
  ollama: [
304
318
  {
305
319
  id: "qwen2.5-coder:14b",
@@ -373,6 +387,7 @@ const LEGACY_CONTEXT_LIMITS = [
373
387
  ["o3", 200_000],
374
388
  ["o4-mini", 200_000],
375
389
  ["gemini", 1_000_000],
390
+ ["deepseek-v4", 1_000_000],
376
391
  ["deepseek", 128_000],
377
392
  ["llama", 128_000],
378
393
  ["qwen", 128_000],
@@ -426,11 +441,14 @@ export function normalizeProviderId(provider) {
426
441
  return undefined;
427
442
  if (provider === "codex")
428
443
  return "openai-codex";
444
+ if (provider === "compat")
445
+ return "openai-compat";
429
446
  if (provider === "openrouter" ||
430
447
  provider === "anthropic" ||
431
448
  provider === "openai" ||
432
449
  provider === "openai-codex" ||
433
450
  provider === "deepseek" ||
451
+ provider === "openai-compat" ||
434
452
  provider === "ollama") {
435
453
  return provider;
436
454
  }
@@ -442,10 +460,19 @@ export function normalizeReasoningEffort(raw) {
442
460
  const value = raw.toLowerCase();
443
461
  if (value === "max")
444
462
  return "xhigh";
445
- if (value === "low" || value === "medium" || value === "high" || value === "xhigh")
463
+ if (value === "off")
464
+ return "none";
465
+ if (value === "none" || value === "low" || value === "medium" || value === "high" || value === "xhigh")
446
466
  return value;
447
467
  return undefined;
448
468
  }
469
+ /**
470
+ * OpenCode Go is a flat subscription: requests through it are not billed per
471
+ * token, so cost shows as included and --budget does not trip.
472
+ */
473
+ export function isUnmeteredEndpoint(baseUrl) {
474
+ return !!baseUrl && /^https?:\/\/opencode\.ai\/zen\/go(?:\/|$)/i.test(baseUrl);
475
+ }
449
476
  export function getDefaultModel(provider) {
450
477
  return BUILTIN_MODELS[provider][0].id;
451
478
  }
@@ -502,9 +529,9 @@ export function lookupContextWindow(model, provider) {
502
529
  }
503
530
  return 200_000;
504
531
  }
505
- export function lookupPricing(model, provider) {
532
+ export function lookupPricing(model, provider, baseUrl) {
506
533
  const normalizedProvider = normalizeProviderId(provider);
507
- if (normalizedProvider === "openai-codex") {
534
+ if (normalizedProvider === "openai-codex" || isUnmeteredEndpoint(baseUrl)) {
508
535
  return { pricing: { inputPer1M: 0, outputPer1M: 0 }, metered: false };
509
536
  }
510
537
  const metadata = getModelMetadata(provider, model);
@@ -9,6 +9,7 @@
9
9
  * 6. Shuts down gracefully on ShutdownRequest
10
10
  */
11
11
  import { MAX_SPAWN_DEPTH } from "./types.js";
12
+ import { createIpcHooks } from "./ipc-hooks.js";
12
13
  import { resolveProvider } from "../providers/resolve.js";
13
14
  import { ToolRegistry } from "../tools/registry.js";
14
15
  import { readFileTool } from "../tools/read-file.js";
@@ -34,6 +35,7 @@ import { buildPhrenContext, } from "../memory/context.js";
34
35
  import { startSession, endSession, } from "../memory/session.js";
35
36
  import { runAgent, } from "../agent-loop.js";
36
37
  import { createCostTracker } from "../cost.js";
38
+ import { scopeModelOverrides } from "../model-overrides.js";
37
39
  import { getAgentType, applyAgentType } from "./agent-types.js";
38
40
  import { AgentSpawner } from "./spawner.js";
39
41
  import { createSpawnAgentTool, createSendMessageTool, createListAgentsTool } from "../tools/spawn-agent.js";
@@ -43,29 +45,6 @@ function send(msg) {
43
45
  process.send(msg);
44
46
  }
45
47
  }
46
- /** Build TurnHooks that relay all events to the parent via IPC. */
47
- function createIpcHooks(agentId) {
48
- return {
49
- onTextDelta(text) {
50
- send({ type: "text_delta", agentId, text });
51
- },
52
- onTextDone() {
53
- // No-op — parent reconstructs from deltas
54
- },
55
- onTextBlock(text) {
56
- send({ type: "text_block", agentId, text });
57
- },
58
- onToolStart(name, input, count) {
59
- send({ type: "tool_start", agentId, toolName: name, input, count });
60
- },
61
- onToolEnd(name, input, output, isError, durationMs) {
62
- send({ type: "tool_end", agentId, toolName: name, input, output, isError, durationMs });
63
- },
64
- onStatus(msg) {
65
- send({ type: "status", agentId, message: msg });
66
- },
67
- };
68
- }
69
48
  /** Build a lightweight system prompt for child agents. No "search memory first" forcing. */
70
49
  function buildChildPrompt(task) {
71
50
  return [
@@ -82,8 +61,13 @@ async function initAgentState(payload) {
82
61
  const { agentId, task: _task, cwd, provider: providerName, model, project, permissions, maxTurns, budget, plan, verbose } = payload;
83
62
  // Set cwd (use worktree path if provided)
84
63
  process.chdir(payload.worktreePath ?? cwd);
85
- // Resolve LLM provider
86
- const provider = resolveProvider(providerName, model);
64
+ // Resolve LLM provider; the parent's --context-window / --price-* apply
65
+ // only if this child runs the model they were given for.
66
+ if (payload.modelOverrides) {
67
+ const { model: overridden, ...overrides } = payload.modelOverrides;
68
+ scopeModelOverrides(overridden, overrides);
69
+ }
70
+ const provider = resolveProvider(providerName, model, undefined, undefined, { baseUrl: payload.baseUrl });
87
71
  // Child agents get a lightweight prompt — no "search memory first" forcing
88
72
  const systemPrompt = buildChildPrompt(_task);
89
73
  // Quick phren init — just resolve context, skip slow FTS5 search
@@ -142,11 +126,11 @@ async function initAgentState(payload) {
142
126
  registry.register(gitCommitTool);
143
127
  // Cost tracker
144
128
  const modelName = provider.model ?? model ?? provider.name;
145
- const costTracker = createCostTracker(modelName, budget, provider.name);
129
+ const costTracker = createCostTracker(modelName, budget, provider.name, provider.baseUrl);
146
130
  let spawner = null;
147
131
  const depth = payload.depth ?? 0;
148
132
  if (depth < MAX_SPAWN_DEPTH) {
149
- spawner = new AgentSpawner({ costTracker, depth, getPermissionDefaults: () => registry.permissionConfig });
133
+ spawner = new AgentSpawner({ costTracker, depth, getPermissionDefaults: () => registry.permissionConfig, getParentProvider: () => provider });
150
134
  registry.register(createSpawnAgentTool(spawner, () => registry.permissionConfig));
151
135
  registry.register(createSendMessageTool(spawner));
152
136
  registry.register(createListAgentsTool(spawner));
@@ -169,7 +153,7 @@ async function initAgentState(payload) {
169
153
  maxTurns,
170
154
  verbose,
171
155
  plan,
172
- hooks: createIpcHooks(agentId),
156
+ hooks: createIpcHooks(agentId, send),
173
157
  spawner,
174
158
  pendingDms: [],
175
159
  taskCount: 0,
@@ -179,6 +163,8 @@ async function initAgentState(payload) {
179
163
  async function runTask(state, task) {
180
164
  const beforeInput = state.costTracker.totalInputTokens;
181
165
  const beforeOutput = state.costTracker.totalOutputTokens;
166
+ const beforeCacheRead = state.costTracker.totalCacheReadTokens;
167
+ const beforeCacheWrite = state.costTracker.totalCacheWriteTokens;
182
168
  const beforeCost = state.costTracker.totalCost;
183
169
  const config = {
184
170
  provider: state.provider,
@@ -200,6 +186,8 @@ async function runTask(state, task) {
200
186
  totalCost: result.totalCost,
201
187
  inputTokens: state.costTracker.totalInputTokens - beforeInput,
202
188
  outputTokens: state.costTracker.totalOutputTokens - beforeOutput,
189
+ cacheReadTokens: state.costTracker.totalCacheReadTokens - beforeCacheRead,
190
+ cacheWriteTokens: state.costTracker.totalCacheWriteTokens - beforeCacheWrite,
203
191
  costUsd: state.costTracker.totalCost - beforeCost,
204
192
  };
205
193
  }
@@ -0,0 +1,34 @@
1
+ /** Build TurnHooks that relay all events to the parent via IPC. */
2
+ export function createIpcHooks(agentId, sendMessage) {
3
+ // Characters streamed by the current model call, so a retry can tell the
4
+ // parent how much of its text to drop.
5
+ let attemptChars = 0;
6
+ return {
7
+ onTextDelta(text) {
8
+ attemptChars += text.length;
9
+ sendMessage({ type: "text_delta", agentId, text });
10
+ },
11
+ onStreamRetry() {
12
+ sendMessage({ type: "stream_retry", agentId, discard: attemptChars });
13
+ attemptChars = 0;
14
+ },
15
+ onAssistantMessage() {
16
+ attemptChars = 0;
17
+ },
18
+ onTextDone() {
19
+ // No-op — parent reconstructs from deltas
20
+ },
21
+ onTextBlock(text) {
22
+ sendMessage({ type: "text_block", agentId, text });
23
+ },
24
+ onToolStart(name, input, count) {
25
+ sendMessage({ type: "tool_start", agentId, toolName: name, input, count });
26
+ },
27
+ onToolEnd(name, input, output, isError, durationMs) {
28
+ sendMessage({ type: "tool_end", agentId, toolName: name, input, output, isError, durationMs });
29
+ },
30
+ onStatus(msg) {
31
+ sendMessage({ type: "status", agentId, message: msg });
32
+ },
33
+ };
34
+ }
@@ -13,16 +13,37 @@ import { fileURLToPath } from "node:url";
13
13
  import { dirname, join } from "node:path";
14
14
  import { EventEmitter } from "node:events";
15
15
  import { MAX_SPAWN_DEPTH } from "./types.js";
16
+ import { scopedModelOverrides } from "../model-overrides.js";
16
17
  import { createWorktree, hasWorktreeChanges, removeWorktree, } from "./worktree.js";
17
18
  const __filename = fileURLToPath(import.meta.url);
18
19
  const __dirname = dirname(__filename);
19
20
  /** Resolved path to the child-entry module. */
20
21
  const CHILD_ENTRY = join(__dirname, "child-entry.js");
22
+ /** Providers that reach a configurable endpoint (--base-url). */
23
+ const ENDPOINT_PROVIDERS = new Set(["openai-compat", "deepseek"]);
24
+ /**
25
+ * Where a child on the parent's endpoint provider should connect. A child
26
+ * that names no provider, or the parent's, gets the parent's endpoint, and
27
+ * its model unless it asked for another; openai-compat cannot be found by
28
+ * auto-detection and needs both. Any other provider is the child's own.
29
+ */
30
+ export function childEndpoint(parent, provider, model) {
31
+ if (!parent?.baseUrl || !ENDPOINT_PROVIDERS.has(parent.name))
32
+ return undefined;
33
+ if (provider !== undefined && provider !== parent.name)
34
+ return undefined;
35
+ const childModel = model ?? parent.model;
36
+ return { provider: parent.name, ...(childModel !== undefined ? { model: childModel } : {}), baseUrl: parent.baseUrl };
37
+ }
21
38
  /** Keys forwarded from the parent env into child processes. */
22
39
  const ENV_FORWARD_KEYS = [
23
40
  "OPENROUTER_API_KEY",
24
41
  "ANTHROPIC_API_KEY",
25
42
  "OPENAI_API_KEY",
43
+ "DEEPSEEK_API_KEY",
44
+ // Key for an openai-compat endpoint; the endpoint itself goes in the
45
+ // spawn payload (childEndpoint). Env only: never argv or logs.
46
+ "PHREN_AGENT_API_KEY",
26
47
  "PHREN_AGENT_PROVIDER",
27
48
  "PHREN_AGENT_MODEL",
28
49
  "PHREN_OLLAMA_URL",
@@ -43,12 +64,14 @@ export class AgentSpawner extends EventEmitter {
43
64
  depth;
44
65
  maxDepth;
45
66
  getPermissionDefaults;
67
+ getParentProvider;
46
68
  constructor(opts = {}) {
47
69
  super();
48
70
  this.costTracker = opts.costTracker ?? null;
49
71
  this.depth = opts.depth ?? 0;
50
72
  this.maxDepth = opts.maxDepth ?? MAX_SPAWN_DEPTH;
51
73
  this.getPermissionDefaults = opts.getPermissionDefaults;
74
+ this.getParentProvider = opts.getParentProvider;
52
75
  }
53
76
  canSpawn() {
54
77
  return this.depth < this.maxDepth;
@@ -63,6 +86,7 @@ export class AgentSpawner extends EventEmitter {
63
86
  }
64
87
  const agentId = `agent-${this.nextId++}`;
65
88
  const defaults = this.getPermissionDefaults?.();
89
+ const endpoint = childEndpoint(this.getParentProvider?.(), opts.provider, opts.model);
66
90
  // Build forwarded env
67
91
  const childEnv = {};
68
92
  for (const key of ENV_FORWARD_KEYS) {
@@ -76,9 +100,11 @@ export class AgentSpawner extends EventEmitter {
76
100
  agentId,
77
101
  task: opts.task,
78
102
  cwd: opts.cwd ?? defaults?.projectRoot ?? process.cwd(),
79
- provider: opts.provider,
80
- model: opts.model,
103
+ provider: endpoint?.provider ?? opts.provider,
104
+ model: endpoint?.model ?? opts.model,
105
+ ...(endpoint ? { baseUrl: endpoint.baseUrl } : {}),
81
106
  project: opts.project,
107
+ modelOverrides: scopedModelOverrides(),
82
108
  permissions: opts.permissions ?? defaults?.mode ?? "auto-confirm",
83
109
  maxTurns: opts.maxTurns ?? 50,
84
110
  budget: opts.budget ?? null,
@@ -173,6 +199,9 @@ export class AgentSpawner extends EventEmitter {
173
199
  case "text_block":
174
200
  this.emit("text_block", msg.agentId, msg.text);
175
201
  break;
202
+ case "stream_retry":
203
+ this.emit("stream_retry", msg.agentId, msg.discard);
204
+ break;
176
205
  case "tool_start":
177
206
  this.emit("tool_start", msg.agentId, msg.toolName, msg.input, msg.count);
178
207
  break;
@@ -211,17 +240,22 @@ export class AgentSpawner extends EventEmitter {
211
240
  break;
212
241
  }
213
242
  }
243
+ /** Fold a child's usage into the parent's tracker (its cost already priced by the child). */
214
244
  aggregateChildCost(result) {
215
245
  const tracker = this.costTracker;
216
246
  if (!tracker)
217
247
  return;
218
248
  const inputTokens = result.inputTokens ?? 0;
219
249
  const outputTokens = result.outputTokens ?? 0;
250
+ const cacheReadTokens = result.cacheReadTokens ?? 0;
251
+ const cacheWriteTokens = result.cacheWriteTokens ?? 0;
220
252
  const costUsd = result.costUsd ?? 0;
221
- if (inputTokens === 0 && outputTokens === 0 && costUsd === 0)
253
+ if (inputTokens === 0 && outputTokens === 0 && cacheReadTokens === 0 && cacheWriteTokens === 0 && costUsd === 0)
222
254
  return;
223
255
  tracker.totalInputTokens += inputTokens;
224
256
  tracker.totalOutputTokens += outputTokens;
257
+ tracker.totalCacheReadTokens += cacheReadTokens;
258
+ tracker.totalCacheWriteTokens += cacheWriteTokens;
225
259
  tracker.totalCost += costUsd;
226
260
  }
227
261
  /** Route a direct message from one agent to another. */