@aaronjmars/openroutercli 0.5.0 → 0.7.0

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.
package/README.md CHANGED
@@ -73,7 +73,7 @@ and run `openrouter login --management`.
73
73
  | **OAuth PKCE login** | `openrouter login` opens the browser, catches the callback locally, exchanges the code, and stores the key - `0600` perms in `~/.config/openrouter/`. No paste-an-API-key dance. |
74
74
  | **Two-slot key management** | User key + management/provisioning key live in separate slots. `keys`, `byok`, `guardrails`, `workspaces`, `observability`, `activity` auto-pick the management key; everything else uses the user key. `--key` always wins. |
75
75
  | **Streaming + JSON modes** | Streams in a TTY, switches to JSON when piped or `--json` is set. No global flag plumbing - works on every command. |
76
- | **Chat with everything** | Tool calls (`--tool @file.json --tool-choice required`), server tools (`--server-tool web_search`), deferred tool loading, structured output (`--schema schema.json`), JSON mode, multimodal images (`--image url\|path`), image output (`--modalities text,image`), reasoning effort, automatic prompt caching (`--cache --cache-ttl 1h`), alignment rules (`--rule`), provider routing, fallback models, service tiers, sessions and traces, full sampling controls. |
76
+ | **Chat with everything** | Tool calls (`--tool @file.json --tool-choice required`), server tools (`--server-tool web_search`), deferred tool loading, structured output (`--schema schema.json`), JSON mode, multimodal images (`--image url\|path`), image output (`--modalities text,image`), reasoning effort, automatic prompt caching (`--cache --cache-ttl 1h`), OpenAI explicit caching (`--cache-ttl 30m|12h`) and free repeat responses (`--response-cache`), alignment rules (`--rule`), provider routing, fallback models, service tiers, sessions and traces, full sampling controls. |
77
77
  | **Interactive REPL** | `openrouter chat -i` - slash commands `/model <id>`, `/reset`, `/exit`. History preserved per session. |
78
78
  | **Multi-format inference** | `chat` (OpenAI), `messages` (Anthropic), `responses` (OpenAI Responses) - same auth, same flags, same JSON contract. |
79
79
  | **Batches** | `batches {create,list,get,wait,results,delete}` - send a JSONL file of chat / responses / messages / embeddings requests as one async job at about half price. |
@@ -81,7 +81,7 @@ and run `openrouter login --management`.
81
81
  | **Model discovery with detail** | `models list` (filter, sort by price/context), `models show <id>` (full pricing breakdown - prompt, completion, cache r/w, reasoning, image, web search), `models endpoints <id> --sort throughput\|latency\|prompt\|completion\|uptime --best` for choosing the right provider variant. |
82
82
  | **Benchmarks + rankings** | `benchmarks`, `classifications`, and `datasets {apps,daily,session-cost}` expose OpenRouter's current public benchmark and usage datasets. |
83
83
  | **Management analytics** | `analytics meta` discovers fields; `analytics query` runs metrics/dimensions/time-range queries with a management key. |
84
- | **Structured decisions** | `decisions` (alias `systemone`) calls System One for typed choice, score, and yes/no outputs, with optional image inputs. |
84
+ | **Structured decisions** | `decisions` (alias `systemone`) calls System One for typed choice, score, and yes/no outputs, with optional image inputs (`--image`) and named image assets (`--assets`). |
85
85
  | **Server tools catalog** | `tools list` / `tools get <name>` - every OpenRouter-run tool (web search, fetch, datetime, ...) with params, engines and pricing. |
86
86
  | **Hosted shell files** | `containers {list,get,download,promote}` retrieves and promotes files produced by OpenRouter's hosted shell tools. |
87
87
  | **Spend + activity + generations** | `credits` for balance, `whoami` for daily/weekly/monthly usage + rate limit, `keys get <hash>` for per-key spend, `activity --api-key-hash <hash>` for daily aggregates, `generation <id> [--content]` for full per-request metadata + input/output. |
@@ -138,6 +138,7 @@ openrouter batches wait <id> && openrouter batches results <id> -o out.jsonl
138
138
  # Discovery
139
139
  openrouter models --filter claude --sort prompt
140
140
  openrouter models user --order newest
141
+ openrouter models user --client codex # only models a coding harness offers
141
142
  openrouter models show anthropic/claude-sonnet-4.5
142
143
  openrouter models endpoints anthropic/claude-sonnet-4.5 --sort throughput --best
143
144
  openrouter providers
@@ -150,6 +151,7 @@ openrouter activity --date 2026-04-28 --api-key-hash <hash>
150
151
  # Management (provisioning key required)
151
152
  openrouter keys create "agent-key" --limit 25 --limit-reset monthly
152
153
  openrouter guardrails create "no-paid-models" --allowed-providers "" --enforce-zdr
154
+ openrouter guardrails update <id> --harness-models # harness clients see only first-party models
153
155
  openrouter workspaces create "staging" --slug staging --default-text-model openai/gpt-4o-mini
154
156
  openrouter workspaces budgets set staging monthly --limit 100
155
157
  openrouter observability create "langfuse" --type langfuse --config @langfuse.json
package/SKILL.md CHANGED
@@ -163,6 +163,9 @@ openrouter chat "summarize this" -m google/gemini-2.5-flash --file ./report.pdf
163
163
  # Prompt caching: automatic (top-level cache_control, optional 1h TTL) or explicit breakpoints
164
164
  openrouter chat "answer from the doc" -m anthropic/claude-sonnet-4.5 -s "$(cat big-context.txt)" --cache --cache-ttl 1h
165
165
  openrouter chat "answer from the doc" -m anthropic/claude-sonnet-4.5 -s "$(cat big-context.txt)" --cache-system
166
+ # OpenAI explicit caching (GPT-5.6+): 30m or 12h TTL, marks the system prompt (else the prompt);
167
+ # works on chat and responses, 12h writes bill at a higher rate
168
+ openrouter chat "answer from the doc" -m openai/gpt-5.6 -s "$(cat big-context.txt)" --cache-ttl 12h
166
169
 
167
170
  # Tracing, sessions, priority: works on chat / messages / responses
168
171
  openrouter chat "..." -m openai/gpt-4o-mini --session-id run-42 --user u_123 \
@@ -203,7 +206,7 @@ openrouter embed -m openai/text-embedding-3-small "hello" "world"
203
206
  openrouter embed -m <multimodal-embed-model> "a cat" --image ./cat.png
204
207
  openrouter rerank -m cohere/rerank-v3.5 --query "italian food" -d pizza -d sushi -d pasta
205
208
 
206
- # Text to speech (mp3 by default; pcm also allowed): tone, multi-speaker, voice cloning
209
+ # Text to speech (mp3 by default, pcm for HeyGen Voice + Gemini TTS which only return pcm): tone, multi-speaker, voice cloning
207
210
  openrouter speech "Hello there" -m elevenlabs/eleven-v4 --voice <voice> -o out.mp3
208
211
  openrouter speech "Welcome back" -m openai/gpt-4o-mini-tts --voice alloy --instructions "warm, slow"
209
212
  openrouter speech -m google/gemini-2.5-flash-preview-tts --turns @dialogue.json # [{text, voice}]
@@ -236,6 +239,7 @@ openrouter models --free # only :free models
236
239
  openrouter models --input-modalities image # image-INPUT (vision) models
237
240
  openrouter models --output-modalities image # image-OUTPUT models
238
241
  openrouter models user --order newest # account-filtered models, server-side sort
242
+ openrouter models user --client codex # models the named harness offers (claude_code|claude_desktop|codex); honors --harness-models guardrails
239
243
  openrouter models --supported tools # filter by capability
240
244
  # Full detail for one model: pricing breakdown, architecture, supported params,
241
245
  # reasoning config, supported voices
@@ -254,6 +258,7 @@ openrouter datasets daily --period week --category programming
254
258
  openrouter analytics meta
255
259
  openrouter decisions --model typesafe/jev-1.13 --state '...' --questions @questions.json # alias: systemone
256
260
  openrouter decisions --model <vision-decider> --state '...' --image ./chart.png --questions @q.json
261
+ openrouter decisions --model typesafe/jev-1.14 --state '... image: logo' --assets '{"logo":"./logo.png"}' --questions @q.json
257
262
  openrouter containers list <container-id>
258
263
 
259
264
  # Supporting metadata
@@ -312,6 +317,7 @@ openrouter guardrails create "no-paid-models" --allowed-providers "" --enforce-z
312
317
  openrouter guardrails update <id> --limit-usd 10 --reset-interval daily --include-byok
313
318
  openrouter guardrails update <id> --regions europe --builtin-filters email:redact,secrets:block --no-paid-training
314
319
  openrouter guardrails update <id> --allow-safety-retention-google --no-allow-safety-retention-other
320
+ openrouter guardrails update <id> --harness-models # codex/claude harnesses only see their vendor's models
315
321
  openrouter guardrails assign-key <id> <key-hash> # up to 500 hashes per call
316
322
 
317
323
  openrouter workspaces list
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@aaronjmars/openroutercli",
3
- "version": "0.5.0",
3
+ "version": "0.7.0",
4
4
  "description": "Command-line interface for the OpenRouter API. Built for humans and LLM agents.",
5
5
  "type": "module",
6
6
  "bin": {
@@ -42,7 +42,9 @@ export const SERVER_TOOL_HELP = ` --server-tool <name>[=<json|@file>] Open
42
42
  `;
43
43
 
44
44
  export const CACHE_HELP = ` --cache Automatic prompt caching (top-level cache_control)
45
- --cache-ttl <ttl> 5m | 1h (implies --cache)
45
+ --cache-ttl <ttl> 5m | 1h (implies --cache). chat/responses also take
46
+ 30m | 12h: OpenAI explicit prompt_cache_options
47
+ (GPT-5.6+; 12h writes bill higher)
46
48
  `;
47
49
 
48
50
  export const ALIGNMENT_HELP = ` --rule <text> Alignment plugin rule (repeatable, max 128, <=200 chars,
@@ -80,11 +82,25 @@ export async function applyToolExtras(body, values) {
80
82
  }
81
83
  }
82
84
 
83
- export function applyCache(body, values) {
84
- if (!values.cache && !values["cache-ttl"]) return;
85
+ const EXPLICIT_CACHE_TTLS = ["30m", "12h"];
86
+
87
+ // 5m | 1h: top-level cache_control (Anthropic style). 30m | 12h: OpenAI
88
+ // explicit prompt_cache_options, which only caches blocks marked with
89
+ // prompt_cache_breakpoint, so the caller must mark one. Returns true for the
90
+ // explicit form. `explicit` = the endpoint accepts it (chat, responses).
91
+ export function applyCache(body, values, { explicit = false } = {}) {
92
+ if (!values.cache && !values["cache-ttl"]) return false;
85
93
  const ttl = values["cache-ttl"];
86
- if (ttl && ttl !== "5m" && ttl !== "1h") throw new Error(`--cache-ttl must be 5m or 1h`);
94
+ if (EXPLICIT_CACHE_TTLS.includes(ttl)) {
95
+ if (!explicit) throw new Error("--cache-ttl 30m|12h works with chat and responses only");
96
+ body.prompt_cache_options = { mode: "explicit", ttl };
97
+ return true;
98
+ }
99
+ if (ttl && ttl !== "5m" && ttl !== "1h") {
100
+ throw new Error(`--cache-ttl must be ${explicit ? "5m, 1h, 30m or 12h" : "5m or 1h"}`);
101
+ }
87
102
  body.cache_control = { type: "ephemeral", ...(ttl ? { ttl } : {}) };
103
+ return false;
88
104
  }
89
105
 
90
106
  // Adds the alignment plugin from --rule/--alignment-*. An alignment entry
@@ -193,11 +209,14 @@ ${SERVER_TOOL_HELP} --reasoning <effort> none|minimal|low|medium|high|xhigh
193
209
  --reasoning-max-tokens <n> Reasoning token budget (Anthropic-style)
194
210
  --reasoning-exclude Reason internally but omit reasoning from the output
195
211
  --provider <json|@file> Provider routing (order, only, sort, ...;
196
- options: per-provider passthrough keyed by slug)
212
+ options: per-provider passthrough keyed by slug).
213
+ Default routing weighs quality-per-cost; sort:price or
214
+ a :floor model suffix keeps the cheapest-first order
197
215
  --modalities <csv> Output modalities: text,image,audio
198
216
  --image-config <json|@file> Image output options, e.g. {"aspect_ratio":"16:9"}
199
217
  --prediction <text|@file> Predicted output (prediction.content) to cut latency
200
- --prompt-cache-key <k> Provider prompt-cache key
218
+ --response-cache X-OpenRouter-Cache: free repeat responses for
219
+ identical requests (separate from --cache prompt caching)
201
220
  --metadata <json|@file> String key/value pairs (max 16)
202
221
  ${CACHE_HELP} --cache-system Mark the system prompt as a cache breakpoint
203
222
  --cache-user Mark the user prompt as a cache breakpoint
@@ -355,15 +374,18 @@ async function buildPlugins(values) {
355
374
  return addAlignmentPlugin(list, values);
356
375
  }
357
376
 
358
- // Wrap a string into a single text content part so a cache_control breakpoint
359
- // can be attached; content already in part form is returned untouched.
360
- function markCache(content) {
361
- if (typeof content === "string") {
362
- return [{ type: "text", text: content, cache_control: { type: "ephemeral" } }];
363
- }
377
+ // Wrap a string into a single text content part so a cache breakpoint can be
378
+ // attached; content already in part form gets it on its last part. Explicit
379
+ // mode uses prompt_cache_breakpoint, which only text parts accept.
380
+ function markCache(content, explicit) {
381
+ const mark = explicit
382
+ ? { prompt_cache_breakpoint: { mode: "explicit" } }
383
+ : { cache_control: { type: "ephemeral" } };
384
+ if (typeof content === "string") return [{ type: "text", text: content, ...mark }];
364
385
  if (Array.isArray(content) && content.length) {
365
386
  const parts = content.map((p) => ({ ...p }));
366
- parts[parts.length - 1].cache_control = { type: "ephemeral" };
387
+ const i = explicit ? parts.findLastIndex((p) => p.type === "text") : parts.length - 1;
388
+ if (i !== -1) Object.assign(parts[i], mark);
367
389
  return parts;
368
390
  }
369
391
  return content;
@@ -427,22 +449,23 @@ async function applyRequestOptions(body, values) {
427
449
  }
428
450
  if (values["prompt-cache-key"]) body.prompt_cache_key = values["prompt-cache-key"];
429
451
  if (values.metadata) body.metadata = await loadJsonOrFile(values.metadata, "--metadata");
430
- applyCache(body, values);
452
+ const explicit = applyCache(body, values, { explicit: true });
431
453
  await applyRequestFields(body, values);
432
454
 
433
455
  const plugins = await buildPlugins(values);
434
456
  if (plugins.length) body.plugins = plugins;
435
457
 
436
- // cache_control breakpoints are attached to message content, so this must
437
- // run after the messages are assembled.
438
- if (values["cache-system"]) {
439
- const sys = body.messages.find((m) => m.role === "system");
440
- if (sys) sys.content = markCache(sys.content);
441
- }
442
- if (values["cache-user"]) {
458
+ // Cache breakpoints are attached to message content, so this must run after
459
+ // the messages are assembled. Explicit mode caches nothing unmarked, so with
460
+ // no --cache-system/--cache-user it marks the system prompt, else the prompt.
461
+ const sys = body.messages.find((m) => m.role === "system");
462
+ const markSystem = values["cache-system"] || (explicit && !values["cache-user"] && sys);
463
+ const markUser = values["cache-user"] || (explicit && !values["cache-system"] && !sys);
464
+ if (markSystem && sys) sys.content = markCache(sys.content, explicit);
465
+ if (markUser) {
443
466
  for (let i = body.messages.length - 1; i >= 0; i--) {
444
467
  if (body.messages[i].role === "user") {
445
- body.messages[i].content = markCache(body.messages[i].content);
468
+ body.messages[i].content = markCache(body.messages[i].content, explicit);
446
469
  break;
447
470
  }
448
471
  }
@@ -475,13 +498,13 @@ function printUsage(model, usage) {
475
498
  );
476
499
  }
477
500
 
478
- async function streamResponse(body, auth) {
501
+ async function streamResponse(body, auth, extraHeaders) {
479
502
  const res = await withAlignmentErrors(
480
503
  api("POST", "/chat/completions", {
481
504
  auth,
482
505
  body: { ...body, stream: true },
483
506
  raw: true,
484
- headers: { Accept: "text/event-stream" },
507
+ headers: { Accept: "text/event-stream", ...extraHeaders },
485
508
  }),
486
509
  );
487
510
  let usage = null;
@@ -574,6 +597,7 @@ export async function chatCommand(argv) {
574
597
  "image-config": { type: "string" },
575
598
  prediction: { type: "string" },
576
599
  "prompt-cache-key": { type: "string" },
600
+ "response-cache": { type: "boolean" },
577
601
  metadata: { type: "string" },
578
602
  image: { type: "string", multiple: true },
579
603
  file: { type: "string", multiple: true },
@@ -623,7 +647,10 @@ export async function chatCommand(argv) {
623
647
  const body = await buildBody(values, prompt);
624
648
 
625
649
  if (!shouldStream(values)) {
626
- const data = await withAlignmentErrors(api("POST", "/chat/completions", { auth, body }));
650
+ const rcHeaders = values["response-cache"] ? { "X-OpenRouter-Cache": "true" } : undefined;
651
+ const data = await withAlignmentErrors(
652
+ api("POST", "/chat/completions", { auth, body, headers: rcHeaders }),
653
+ );
627
654
  if (isJsonMode() || values.raw) {
628
655
  printJSON(data);
629
656
  } else {
@@ -651,7 +678,11 @@ export async function chatCommand(argv) {
651
678
  return 0;
652
679
  }
653
680
 
654
- const meta = await streamResponse(body, auth);
681
+ const meta = await streamResponse(
682
+ body,
683
+ auth,
684
+ values["response-cache"] ? { "X-OpenRouter-Cache": "true" } : undefined,
685
+ );
655
686
  if (values.usage && meta.usage) printUsage(meta.model, meta.usage);
656
687
  if (ERROR_FINISH_REASONS.includes(meta.finishReason)) {
657
688
  process.stderr.write(`error: generation stopped (${meta.finishReason})\n`);
@@ -16,6 +16,9 @@ Options:
16
16
  --questions <json> Questions object or @questions.json
17
17
  --image <url|path> Add an image to the state (repeatable). Local files
18
18
  are sent as data URLs. For multimodal models.
19
+ --assets <json|@file> Named images referenced from state/instructions/
20
+ criteria: {"ref-name": "<local path>"} (png/jpg/jpeg;
21
+ uploaded as data_base64). Every name must be referenced.
19
22
  --body <json|@file> Fully-formed request body
20
23
  --extra <json> Extra fields merged into the request
21
24
  --alpha Use the older /api/alpha/decisions route
@@ -57,6 +60,7 @@ export async function decisionsCommand(argv) {
57
60
  state: { type: "string" },
58
61
  questions: { type: "string" },
59
62
  image: { type: "string", multiple: true },
63
+ assets: { type: "string" },
60
64
  body: { type: "string" },
61
65
  extra: { type: "string" },
62
66
  alpha: { type: "boolean" },
@@ -81,6 +85,26 @@ export async function decisionsCommand(argv) {
81
85
  const parts = await Promise.all(images.map(imagePart));
82
86
  body.state = withImages(body.state, parts);
83
87
  }
88
+ if (values.assets !== undefined) {
89
+ const assets = await loadJson(values.assets, "--assets");
90
+ if (!assets || typeof assets !== "object" || Array.isArray(assets)) {
91
+ throw new Error('--assets must be a JSON object like {"ref": "<path>"}');
92
+ }
93
+ body.assets = {};
94
+ for (const [name, src] of Object.entries(assets)) {
95
+ const url = await mediaUrl(src);
96
+ if (!url.startsWith("data:")) {
97
+ throw new Error(
98
+ `--assets "${name}": the Decisions API only accepts uploaded images. Give a local file (png/jpg/jpeg), not a URL.`,
99
+ );
100
+ }
101
+ const [mime, b64] = url.slice(5).split(";base64,");
102
+ const format = { "image/png": "png", "image/jpeg": "jpeg" }[mime];
103
+ if (!format)
104
+ throw new Error(`--assets "${name}": expected a png/jpg/jpeg image, got ${mime}`);
105
+ body.assets[name] = { type: "image", format, data_base64: b64 };
106
+ }
107
+ }
84
108
  if (values.extra) Object.assign(body, await loadJson(values.extra, "--extra"));
85
109
  const data = values.alpha
86
110
  ? await api("POST", "/decisions", {
@@ -57,6 +57,10 @@ create / update options:
57
57
  Allow ZDR requests to google/* to use endpoints that retain for safety
58
58
  --allow-safety-retention-other / --no-allow-safety-retention-other
59
59
  Allow ZDR requests to other models to use endpoints that retain for safety
60
+ --harness-models / --no-harness-models
61
+ enforce_first_party_harness_models: when a coding
62
+ harness (claude_code, claude_desktop, codex) lists
63
+ models, show only the harness vendor's own
60
64
  --include-byok / --no-include-byok Count BYOK spend toward the budget
61
65
  --workspace <uuid> (create only)
62
66
  --extra <json|@file> Extra body fields, merged last (e.g. enforce_zdr_xai:false)
@@ -64,7 +68,7 @@ create / update options:
64
68
 
65
69
  const ZDR_PROVIDERS = ["anthropic", "openai", "google", "xai", "other"];
66
70
  const REGIONS = ["global", "europe", "us"];
67
- const MAX_KEY_HASHES = 500;
71
+ const MAX_BULK_IDS = 500;
68
72
 
69
73
  function toggle(values, on, off) {
70
74
  if (values[on] && values[off]) throw new Error(`Pass only one of --${on} or --${off}.`);
@@ -151,6 +155,7 @@ async function buildBody(values) {
151
155
  ["enable_paid_model_training", "paid-training"],
152
156
  ["allow_safety_retention_google", "allow-safety-retention-google"],
153
157
  ["allow_safety_retention_other", "allow-safety-retention-other"],
158
+ ["enforce_first_party_harness_models", "harness-models"],
154
159
  ];
155
160
  for (const [field, flag] of toggles) {
156
161
  const v = toggle(values, flag, `no-${flag}`);
@@ -192,6 +197,8 @@ const COMMON_FIELDS = {
192
197
  "no-allow-safety-retention-google": { type: "boolean" },
193
198
  "allow-safety-retention-other": { type: "boolean" },
194
199
  "no-allow-safety-retention-other": { type: "boolean" },
200
+ "harness-models": { type: "boolean" },
201
+ "no-harness-models": { type: "boolean" },
195
202
  extra: { type: "string" },
196
203
  };
197
204
 
@@ -339,9 +346,9 @@ export async function guardrailsCommand(argv) {
339
346
  if (positionals.length < 2) throw new Error("Usage: <id> <hash>...");
340
347
  const id = positionals[0];
341
348
  const hashes = positionals.slice(1);
342
- if (hashes.length > MAX_KEY_HASHES) {
349
+ if (hashes.length > MAX_BULK_IDS) {
343
350
  throw new Error(
344
- `Too many key hashes (${hashes.length}); the API takes at most ${MAX_KEY_HASHES} per call. Split into batches.`,
351
+ `Too many key hashes (${hashes.length}); the API takes at most ${MAX_BULK_IDS} per call. Split into batches.`,
345
352
  );
346
353
  }
347
354
  const path =
@@ -362,6 +369,11 @@ export async function guardrailsCommand(argv) {
362
369
  if (positionals.length < 2) throw new Error("Usage: <id> <user_id>...");
363
370
  const id = positionals[0];
364
371
  const userIds = positionals.slice(1);
372
+ if (userIds.length > MAX_BULK_IDS) {
373
+ throw new Error(
374
+ `Too many user ids (${userIds.length}); the API takes at most ${MAX_BULK_IDS} per call. Split into batches.`,
375
+ );
376
+ }
365
377
  const path =
366
378
  sub === "assign-member"
367
379
  ? `/guardrails/${encodeURIComponent(id)}/assignments/members`
@@ -374,7 +374,22 @@ export async function responsesCommand(argv) {
374
374
  if (values.background) body.background = true;
375
375
  if (values["prompt-cache-key"]) body.prompt_cache_key = values["prompt-cache-key"];
376
376
  if (values["safety-identifier"]) body.safety_identifier = values["safety-identifier"];
377
- applyCache(body, values);
377
+ // Explicit mode caches only marked blocks: mark the prompt, which caches
378
+ // the instructions before it too.
379
+ if (applyCache(body, values, { explicit: true }) && typeof body.input === "string") {
380
+ body.input = [
381
+ {
382
+ role: "user",
383
+ content: [
384
+ {
385
+ type: "input_text",
386
+ text: body.input,
387
+ prompt_cache_breakpoint: { mode: "explicit" },
388
+ },
389
+ ],
390
+ },
391
+ ];
392
+ }
378
393
  await applyRequestFields(body, values);
379
394
  if (values.extra) Object.assign(body, await loadJson(values.extra, "--extra"));
380
395
  }
@@ -426,13 +426,14 @@ export async function modelsCommand(argv) {
426
426
  "output-modalities": { type: "string" },
427
427
  order: { type: "string" },
428
428
  sort: { type: "string" },
429
+ client: { type: "string" },
429
430
  });
430
431
  } catch (err) {
431
432
  // `models user` used to forward every list filter; the endpoint only
432
433
  // takes these, so say so instead of a bare "Unknown option".
433
434
  if (err.code === "ERR_PARSE_ARGS_UNKNOWN_OPTION") {
434
435
  throw new Error(
435
- `${err.message.split(".")[0]}. models user only supports --output-modalities, --order/--sort, --limit and --offset; use \`openrouter models list\` for search filters.`,
436
+ `${err.message.split(".")[0]}. models user only supports --output-modalities, --order/--sort, --client, --limit and --offset; use \`openrouter models list\` for search filters.`,
436
437
  );
437
438
  }
438
439
  throw err;
@@ -447,6 +448,10 @@ export async function modelsCommand(argv) {
447
448
  "Options:\n" +
448
449
  " --output-modalities <csv> text,image,embeddings,audio,video,rerank,decisions,\n" +
449
450
  ' speech,transcription (or "all"). Default: text\n' +
451
+ " --client <harness> Only models the named coding harness offers:\n" +
452
+ " claude_code | claude_desktop | codex. When a\n" +
453
+ " guardrail sets enforce_first_party_harness_models,\n" +
454
+ " only the harness vendor's own models.\n" +
450
455
  " --order <sort> Server-side sort: most-popular|newest|top-weekly|\n" +
451
456
  " pricing-low-to-high|pricing-high-to-low|\n" +
452
457
  " context-high-to-low|throughput-high-to-low|\n" +
@@ -462,6 +467,14 @@ export async function modelsCommand(argv) {
462
467
  const query = paginationQuery(values);
463
468
  if (values["output-modalities"]) query.output_modalities = values["output-modalities"];
464
469
  if (values.order || values.sort) query.sort = values.order || values.sort;
470
+ if (values.client) {
471
+ if (!["claude_code", "claude_desktop", "codex"].includes(values.client)) {
472
+ throw new Error(
473
+ `Unknown --client "${values.client}". Expected: claude_code, claude_desktop, codex`,
474
+ );
475
+ }
476
+ query.client = values.client;
477
+ }
465
478
  const data = await api("GET", "/models/user", {
466
479
  auth: authFromValues(values),
467
480
  query,
@@ -35,6 +35,13 @@ create / update options:
35
35
  {"publicKey":"pk-lf-...","secretKey":"sk-lf-...",
36
36
  "baseUrl":"https://cloud.langfuse.com"}. Clickhouse and
37
37
  Snowflake accept "shouldIncludeCacheWriteTokens": true.
38
+ For s3, key auth: {"accessKeyId":"...",
39
+ "secretAccessKey":"...",...} or IAM role auth:
40
+ {"authMode":"iam_role","provider":"aws",
41
+ "roleArn":"arn:aws:iam::<acct>:role/<name>",
42
+ "region":"<region>","externalId":"<id>"} (GovCloud and
43
+ China ARNs rejected; trust policy must require the
44
+ external id).
38
45
  On update, masked values are ignored and unset fields
39
46
  keep their value.
40
47
  --name <name> (update only)
@@ -26,7 +26,8 @@ Options:
26
26
  or one image (.png/.jpg/.webp) for voice design.
27
27
  Local files are sent as data URLs.
28
28
  --ref-text <t> Transcript of a --ref clip (repeatable, in --ref order)
29
- --format <fmt> mp3 | pcm (default: mp3; pcm is raw 16-bit audio)
29
+ --format <fmt> mp3 | pcm (default: mp3, or pcm for HeyGen Voice and
30
+ Gemini TTS, which only return pcm; pcm is raw 16-bit audio)
30
31
  --speed <n> Speech speed (e.g. 1.0)
31
32
  -o, --out <file> Output audio file (default: out.mp3 / out.pcm, or stdout if -)
32
33
  --provider <json|@file> Routing and provider options, e.g.
@@ -86,8 +87,10 @@ export async function speechCommand(argv) {
86
87
  input = text;
87
88
  }
88
89
 
89
- // The API defaults to raw PCM; ask for mp3 so the default out.mp3 plays.
90
- const format = values.format || "mp3";
90
+ // The API defaults to raw PCM; ask for mp3 so the default out.mp3 plays,
91
+ // except on HeyGen Voice and Gemini TTS, which return only pcm (400 on mp3).
92
+ const pcmOnly = /^heygen\//.test(values.model) || /^google\/.*-tts\b/.test(values.model);
93
+ const format = values.format || (pcmOnly ? "pcm" : "mp3");
91
94
  const body = { model: values.model, input, response_format: format };
92
95
  if (values.voice) body.voice = values.voice;
93
96
  if (values.instructions) body.instructions = values.instructions;