@trazum/cli 1.50.1 → 1.50.3
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/dist/gateway-server.d.ts +91 -0
- package/dist/gateway-server.d.ts.map +1 -0
- package/dist/gateway-server.js +290 -0
- package/dist/gateway-server.js.map +1 -0
- package/dist/i18n/en.d.ts.map +1 -1
- package/dist/i18n/en.js +59 -1
- package/dist/i18n/en.js.map +1 -1
- package/dist/i18n/es.d.ts.map +1 -1
- package/dist/i18n/es.js +62 -1
- package/dist/i18n/es.js.map +1 -1
- package/dist/i18n/types.d.ts +36 -0
- package/dist/i18n/types.d.ts.map +1 -1
- package/dist/index.js +242 -9
- package/dist/index.js.map +1 -1
- package/package.json +2 -2
- package/src/gateway-server.ts +350 -0
- package/src/i18n/en.ts +74 -1
- package/src/i18n/es.ts +77 -1
- package/src/i18n/types.ts +38 -0
- package/src/index.ts +289 -7
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@trazum/cli",
|
|
3
|
-
"version": "1.50.
|
|
3
|
+
"version": "1.50.3",
|
|
4
4
|
"description": "Trazum CLI: find where your LLM bill goes, price every finding per month, and enforce token budgets in CI.",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"author": "David Mu\u00f1oz Rey",
|
|
@@ -37,7 +37,7 @@
|
|
|
37
37
|
"prepublishOnly": "npm run build && npm test"
|
|
38
38
|
},
|
|
39
39
|
"dependencies": {
|
|
40
|
-
"@trazum/core": "1.50.
|
|
40
|
+
"@trazum/core": "1.50.3"
|
|
41
41
|
},
|
|
42
42
|
"devDependencies": {
|
|
43
43
|
"@types/node": "^26.2.0",
|
|
@@ -0,0 +1,350 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The proxy that stands in the path, and does as little as possible there.
|
|
3
|
+
*
|
|
4
|
+
* The decision lives in `@trazum/core`'s `gatewayDecision`, which never sees a
|
|
5
|
+
* prompt and cannot return a modified request. This file moves bytes: read a
|
|
6
|
+
* body, ask, and either forward it **unchanged** or answer with the refusal.
|
|
7
|
+
* The split is the safety property — everything that could go wrong in a
|
|
8
|
+
* judgement is tested without a socket, and everything that could go wrong on
|
|
9
|
+
* a socket has no judgement in it.
|
|
10
|
+
*
|
|
11
|
+
* **Loopback only, and the address is not a flag.** Same posture as `serve`
|
|
12
|
+
* since 1.44, and more load-bearing here: this thing has somebody's provider
|
|
13
|
+
* credential passing through it. `127.0.0.1` is compiled in.
|
|
14
|
+
*
|
|
15
|
+
* **The credential is not even borrowed.** The caller's own `authorization`
|
|
16
|
+
* and `x-api-key` headers are forwarded untouched and never read, never
|
|
17
|
+
* stored, never logged, and never put in a URL. Trazum holds no key for the
|
|
18
|
+
* gateway and has no way to make a call of its own through it — which is a
|
|
19
|
+
* stronger promise than the connector's *borrowed, never held*, and the right
|
|
20
|
+
* one for a component sitting between somebody and their provider.
|
|
21
|
+
*
|
|
22
|
+
* **The upstream is compiled in.** A flag naming the host would turn this into
|
|
23
|
+
* a credential-forwarding open proxy: anything that could rewrite a config on
|
|
24
|
+
* disk could point a company's API key at a machine it chose. `checkedEndpoint`
|
|
25
|
+
* has guarded Trazum's outbound calls on that principle since 1.14, and here
|
|
26
|
+
* there is no caller-supplied endpoint at all.
|
|
27
|
+
*
|
|
28
|
+
* **Nothing about the payload is written down.** The body is read to count
|
|
29
|
+
* tokens and to find the model, then forwarded and dropped. It is never
|
|
30
|
+
* logged, never stored, and never included in a refusal — the store has held
|
|
31
|
+
* aggregates since 1.42 and standing in the path changes nothing about that.
|
|
32
|
+
*/
|
|
33
|
+
|
|
34
|
+
import { createServer } from 'node:http';
|
|
35
|
+
import type { IncomingMessage, Server, ServerResponse } from 'node:http';
|
|
36
|
+
import { estimateTokens, gatewayDecision, usageFromResponse } from '@trazum/core';
|
|
37
|
+
import type { GatewayDecision, GatewayPolicy, GatewayStanding, PricingCatalogue } from '@trazum/core';
|
|
38
|
+
|
|
39
|
+
/** Compiled in. See the module note. */
|
|
40
|
+
export const BIND_HOST = '127.0.0.1';
|
|
41
|
+
|
|
42
|
+
export const DEFAULT_GATEWAY_PORT = 7318;
|
|
43
|
+
|
|
44
|
+
/**
|
|
45
|
+
* Bodies larger than this are refused unread.
|
|
46
|
+
*
|
|
47
|
+
* Larger than `serve`'s limit because a real request carries a real prompt,
|
|
48
|
+
* and smaller than unbounded because a proxy that buffers whatever it is
|
|
49
|
+
* handed is a memory exhaustion away from taking down the application it was
|
|
50
|
+
* installed to protect.
|
|
51
|
+
*/
|
|
52
|
+
export const MAX_GATEWAY_BODY_BYTES = 8 * 1024 * 1024;
|
|
53
|
+
|
|
54
|
+
/**
|
|
55
|
+
* Where each provider actually is, and the one path this speaks for it.
|
|
56
|
+
*
|
|
57
|
+
* Deliberately narrow. A gateway that forwarded any path would be a general
|
|
58
|
+
* proxy for somebody's API key, and the budget decision only has meaning for
|
|
59
|
+
* the endpoint that spends tokens.
|
|
60
|
+
*/
|
|
61
|
+
export const UPSTREAMS: Readonly<Record<string, { origin: string; path: string }>> = {
|
|
62
|
+
anthropic: { origin: 'https://api.anthropic.com', path: '/v1/messages' },
|
|
63
|
+
openai: { origin: 'https://api.openai.com', path: '/v1/chat/completions' },
|
|
64
|
+
};
|
|
65
|
+
|
|
66
|
+
/**
|
|
67
|
+
* Headers Trazum adds or removes. Everything else the caller sent is forwarded
|
|
68
|
+
* verbatim, including their credential, which this never reads.
|
|
69
|
+
*/
|
|
70
|
+
const HOP_BY_HOP = new Set([
|
|
71
|
+
'connection',
|
|
72
|
+
'keep-alive',
|
|
73
|
+
'proxy-authenticate',
|
|
74
|
+
'proxy-authorization',
|
|
75
|
+
'te',
|
|
76
|
+
'trailer',
|
|
77
|
+
'transfer-encoding',
|
|
78
|
+
'upgrade',
|
|
79
|
+
'host',
|
|
80
|
+
'content-length',
|
|
81
|
+
]);
|
|
82
|
+
|
|
83
|
+
export interface GatewayContext {
|
|
84
|
+
provider: string;
|
|
85
|
+
catalogue: PricingCatalogue;
|
|
86
|
+
policy: GatewayPolicy;
|
|
87
|
+
/** Where the budget stands, refreshed by the caller — never read per request. */
|
|
88
|
+
standing: () => GatewayStanding | null;
|
|
89
|
+
/**
|
|
90
|
+
* Called after a forwarded call returns, with the provider's own counts.
|
|
91
|
+
*
|
|
92
|
+
* Counts only. There is no parameter here that could carry a prompt, which
|
|
93
|
+
* is what makes "nothing about the payload is written down" a fact about the
|
|
94
|
+
* interface rather than a discipline.
|
|
95
|
+
*/
|
|
96
|
+
record: (measured: {
|
|
97
|
+
model: string;
|
|
98
|
+
label: string | null;
|
|
99
|
+
inputTokens: number;
|
|
100
|
+
outputTokens: number;
|
|
101
|
+
cacheReadTokens: number;
|
|
102
|
+
cacheWriteTokens: number;
|
|
103
|
+
substituted: boolean;
|
|
104
|
+
}) => void;
|
|
105
|
+
/** A line for the operator's terminal. Never given a body, ever. */
|
|
106
|
+
note: (line: string) => void;
|
|
107
|
+
/** Injected so the proxy is testable against a stub upstream. */
|
|
108
|
+
fetchImpl?: typeof fetch;
|
|
109
|
+
}
|
|
110
|
+
|
|
111
|
+
async function readBody(request: IncomingMessage): Promise<string | null> {
|
|
112
|
+
const chunks: Buffer[] = [];
|
|
113
|
+
let size = 0;
|
|
114
|
+
for await (const chunk of request) {
|
|
115
|
+
const buffer = Buffer.from(chunk as Buffer);
|
|
116
|
+
size += buffer.length;
|
|
117
|
+
if (size > MAX_GATEWAY_BODY_BYTES) return null;
|
|
118
|
+
chunks.push(buffer);
|
|
119
|
+
}
|
|
120
|
+
return Buffer.concat(chunks).toString('utf8');
|
|
121
|
+
}
|
|
122
|
+
|
|
123
|
+
/**
|
|
124
|
+
* What the request is asking for, without keeping any of it.
|
|
125
|
+
*
|
|
126
|
+
* The token count is the heuristic estimator's — the same one every other
|
|
127
|
+
* estimate in this product uses, with the same documented error band. Counting
|
|
128
|
+
* exactly would mean an API call to count before the API call, which is a
|
|
129
|
+
* round trip in a hot path to make a budget decision marginally sharper.
|
|
130
|
+
*
|
|
131
|
+
* The returned object holds no text. That is the point: everything downstream
|
|
132
|
+
* of here, including the decision and the record, is structurally incapable of
|
|
133
|
+
* carrying a prompt.
|
|
134
|
+
*/
|
|
135
|
+
function describe(body: string, provider: string): {
|
|
136
|
+
model: string;
|
|
137
|
+
inputTokens: number | null;
|
|
138
|
+
maxOutputTokens: number | null;
|
|
139
|
+
label: string | null;
|
|
140
|
+
} | null {
|
|
141
|
+
let parsed: unknown;
|
|
142
|
+
try {
|
|
143
|
+
parsed = JSON.parse(body);
|
|
144
|
+
} catch {
|
|
145
|
+
return null;
|
|
146
|
+
}
|
|
147
|
+
if (typeof parsed !== 'object' || parsed === null || Array.isArray(parsed)) return null;
|
|
148
|
+
const request = parsed as Record<string, unknown>;
|
|
149
|
+
if (typeof request.model !== 'string') return null;
|
|
150
|
+
|
|
151
|
+
// Every text field the wire format puts in front of the model, counted and
|
|
152
|
+
// then dropped. `JSON.stringify` of the messages over-counts by the
|
|
153
|
+
// structural characters, which is the safe direction for a budget: an
|
|
154
|
+
// estimate that runs high refuses slightly early rather than allowing
|
|
155
|
+
// slightly late.
|
|
156
|
+
const parts: string[] = [];
|
|
157
|
+
if (typeof request.system === 'string') parts.push(request.system);
|
|
158
|
+
if (Array.isArray(request.messages)) parts.push(JSON.stringify(request.messages));
|
|
159
|
+
if (Array.isArray(request.input)) parts.push(JSON.stringify(request.input));
|
|
160
|
+
|
|
161
|
+
const text = parts.join('\n');
|
|
162
|
+
const max = provider === 'anthropic' ? request.max_tokens : request.max_completion_tokens ?? request.max_tokens;
|
|
163
|
+
|
|
164
|
+
return {
|
|
165
|
+
model: request.model,
|
|
166
|
+
inputTokens: text === '' ? null : estimateTokens(text),
|
|
167
|
+
maxOutputTokens: typeof max === 'number' && Number.isFinite(max) ? max : null,
|
|
168
|
+
/**
|
|
169
|
+
* `metadata.trazum_label`, and nothing inferred.
|
|
170
|
+
*
|
|
171
|
+
* A label is what makes a per-workload bill possible, and guessing one
|
|
172
|
+
* from a path or a user agent would attribute somebody's spend to a
|
|
173
|
+
* workload they never named.
|
|
174
|
+
*/
|
|
175
|
+
label: labelOf(request),
|
|
176
|
+
};
|
|
177
|
+
}
|
|
178
|
+
|
|
179
|
+
function labelOf(request: Record<string, unknown>): string | null {
|
|
180
|
+
const metadata = request.metadata;
|
|
181
|
+
if (typeof metadata !== 'object' || metadata === null || Array.isArray(metadata)) return null;
|
|
182
|
+
const label = (metadata as Record<string, unknown>).trazum_label;
|
|
183
|
+
return typeof label === 'string' && label.trim() !== '' ? label : null;
|
|
184
|
+
}
|
|
185
|
+
|
|
186
|
+
/** The refusal, as the caller's SDK will receive it. */
|
|
187
|
+
function refusalBody(decision: Extract<GatewayDecision, { kind: 'refuse' }>): string {
|
|
188
|
+
return `${JSON.stringify(
|
|
189
|
+
{
|
|
190
|
+
schemaVersion: 1,
|
|
191
|
+
error: { type: 'trazum_budget_refusal', message: decision.because },
|
|
192
|
+
reason: decision.reason,
|
|
193
|
+
cause: decision.cause,
|
|
194
|
+
restsOn: decision.restsOn,
|
|
195
|
+
standing: decision.standing,
|
|
196
|
+
estimatedUsd: decision.estimatedUsd,
|
|
197
|
+
alternatives: decision.alternatives,
|
|
198
|
+
},
|
|
199
|
+
null,
|
|
200
|
+
2,
|
|
201
|
+
)}\n`;
|
|
202
|
+
}
|
|
203
|
+
|
|
204
|
+
export function buildGateway(context: GatewayContext): Server {
|
|
205
|
+
const upstream = UPSTREAMS[context.provider];
|
|
206
|
+
const doFetch = context.fetchImpl ?? fetch;
|
|
207
|
+
|
|
208
|
+
return createServer((request: IncomingMessage, response: ServerResponse) => {
|
|
209
|
+
void (async () => {
|
|
210
|
+
if (upstream === undefined) {
|
|
211
|
+
response.writeHead(500, { 'content-type': 'application/json' });
|
|
212
|
+
response.end(`${JSON.stringify({ error: 'no upstream configured for this provider' })}\n`);
|
|
213
|
+
return;
|
|
214
|
+
}
|
|
215
|
+
if (request.method !== 'POST' || request.url !== upstream.path) {
|
|
216
|
+
// Only the one path that spends tokens. A gateway forwarding anything
|
|
217
|
+
// else is a general proxy for somebody's API key.
|
|
218
|
+
response.writeHead(404, { 'content-type': 'application/json' });
|
|
219
|
+
response.end(`${JSON.stringify({ error: 'not a path this gateway forwards' })}\n`);
|
|
220
|
+
return;
|
|
221
|
+
}
|
|
222
|
+
|
|
223
|
+
const body = await readBody(request);
|
|
224
|
+
if (body === null) {
|
|
225
|
+
response.writeHead(413, { 'content-type': 'application/json' });
|
|
226
|
+
response.end(`${JSON.stringify({ error: 'request body too large' })}\n`);
|
|
227
|
+
return;
|
|
228
|
+
}
|
|
229
|
+
|
|
230
|
+
const described = describe(body, context.provider);
|
|
231
|
+
if (described === null) {
|
|
232
|
+
response.writeHead(400, { 'content-type': 'application/json' });
|
|
233
|
+
response.end(`${JSON.stringify({ error: 'could not read a model out of this request' })}\n`);
|
|
234
|
+
return;
|
|
235
|
+
}
|
|
236
|
+
|
|
237
|
+
const decision = gatewayDecision(
|
|
238
|
+
{ provider: context.provider, ...described },
|
|
239
|
+
context.standing(),
|
|
240
|
+
{ catalogue: context.catalogue, policy: context.policy },
|
|
241
|
+
);
|
|
242
|
+
|
|
243
|
+
if (decision.kind === 'refuse') {
|
|
244
|
+
/**
|
|
245
|
+
* **402, deliberately, and never 429.**
|
|
246
|
+
*
|
|
247
|
+
* Every provider SDK retries a 429 automatically — that is what the
|
|
248
|
+
* code means to them — so answering a budget refusal with one turns a
|
|
249
|
+
* single refusal into a retry storm against a gateway that will refuse
|
|
250
|
+
* every time. 402 Payment Required is both literally correct and in
|
|
251
|
+
* nobody's default retry list.
|
|
252
|
+
*/
|
|
253
|
+
response.writeHead(402, { 'content-type': 'application/json' });
|
|
254
|
+
response.end(refusalBody(decision));
|
|
255
|
+
context.note(`refused ${described.model}: ${decision.reason}`);
|
|
256
|
+
return;
|
|
257
|
+
}
|
|
258
|
+
|
|
259
|
+
const outgoing = new Headers();
|
|
260
|
+
for (const [name, value] of Object.entries(request.headers)) {
|
|
261
|
+
if (HOP_BY_HOP.has(name.toLowerCase()) || value === undefined) continue;
|
|
262
|
+
outgoing.set(name, Array.isArray(value) ? value.join(', ') : value);
|
|
263
|
+
}
|
|
264
|
+
|
|
265
|
+
/**
|
|
266
|
+
* The body forwarded is the body received, **byte for byte**, except on
|
|
267
|
+
* a configured substitution — which replaces exactly one field and says
|
|
268
|
+
* so in the record.
|
|
269
|
+
*/
|
|
270
|
+
let forwarded = body;
|
|
271
|
+
if (decision.kind === 'substitute') {
|
|
272
|
+
const parsed = JSON.parse(body) as Record<string, unknown>;
|
|
273
|
+
parsed.model = decision.to.id;
|
|
274
|
+
forwarded = JSON.stringify(parsed);
|
|
275
|
+
context.note(`substituted ${described.model} → ${decision.to.id}: ${decision.configuredReason}`);
|
|
276
|
+
} else if (decision.unjudged !== null) {
|
|
277
|
+
context.note(`forwarded unjudged (${decision.unjudged}): fail-open`);
|
|
278
|
+
}
|
|
279
|
+
|
|
280
|
+
let upstreamResponse: Response;
|
|
281
|
+
try {
|
|
282
|
+
upstreamResponse = await doFetch(`${upstream.origin}${upstream.path}`, {
|
|
283
|
+
method: 'POST',
|
|
284
|
+
headers: outgoing,
|
|
285
|
+
body: forwarded,
|
|
286
|
+
});
|
|
287
|
+
} catch (error) {
|
|
288
|
+
/**
|
|
289
|
+
* The upstream is unreachable. This is **not** a budget refusal and
|
|
290
|
+
* must not look like one: the caller needs to tell "your provider is
|
|
291
|
+
* down" from "you are out of money", and a proxy that blurs them sends
|
|
292
|
+
* somebody to fix the wrong thing.
|
|
293
|
+
*/
|
|
294
|
+
response.writeHead(502, { 'content-type': 'application/json' });
|
|
295
|
+
response.end(
|
|
296
|
+
`${JSON.stringify({
|
|
297
|
+
error: { type: 'trazum_upstream_unreachable', message: error instanceof Error ? error.message : String(error) },
|
|
298
|
+
})}\n`,
|
|
299
|
+
);
|
|
300
|
+
return;
|
|
301
|
+
}
|
|
302
|
+
|
|
303
|
+
const text = await upstreamResponse.text();
|
|
304
|
+
|
|
305
|
+
// Measured at the moment of the call, from the provider's own counts —
|
|
306
|
+
// the reason this beats a connector, which reports the runaway after it
|
|
307
|
+
// ran. Counts only reach `record`; the body is dropped here.
|
|
308
|
+
let measured: unknown;
|
|
309
|
+
try {
|
|
310
|
+
measured = JSON.parse(text);
|
|
311
|
+
} catch {
|
|
312
|
+
measured = null;
|
|
313
|
+
}
|
|
314
|
+
const usage = usageFromResponse(context.provider, measured);
|
|
315
|
+
if (usage !== null) {
|
|
316
|
+
context.record({
|
|
317
|
+
model: decision.kind === 'substitute' ? decision.to.id : described.model,
|
|
318
|
+
label: described.label,
|
|
319
|
+
substituted: decision.kind === 'substitute',
|
|
320
|
+
...usage,
|
|
321
|
+
});
|
|
322
|
+
}
|
|
323
|
+
|
|
324
|
+
const back: Record<string, string> = {};
|
|
325
|
+
upstreamResponse.headers.forEach((value, name) => {
|
|
326
|
+
if (!HOP_BY_HOP.has(name.toLowerCase())) back[name] = value;
|
|
327
|
+
});
|
|
328
|
+
response.writeHead(upstreamResponse.status, back);
|
|
329
|
+
response.end(text);
|
|
330
|
+
})();
|
|
331
|
+
});
|
|
332
|
+
}
|
|
333
|
+
|
|
334
|
+
export function listenGateway(
|
|
335
|
+
server: Server,
|
|
336
|
+
where: { port: number } | { socket: string },
|
|
337
|
+
): Promise<string> {
|
|
338
|
+
return new Promise((resolve, reject) => {
|
|
339
|
+
server.once('error', reject);
|
|
340
|
+
if ('socket' in where) {
|
|
341
|
+
server.listen(where.socket, () => resolve(where.socket));
|
|
342
|
+
return;
|
|
343
|
+
}
|
|
344
|
+
server.listen(where.port, BIND_HOST, () => {
|
|
345
|
+
const address = server.address();
|
|
346
|
+
const port = typeof address === 'object' && address !== null ? address.port : where.port;
|
|
347
|
+
resolve(`http://${BIND_HOST}:${port}`);
|
|
348
|
+
});
|
|
349
|
+
});
|
|
350
|
+
}
|
package/src/i18n/en.ts
CHANGED
|
@@ -57,6 +57,9 @@ ${bold('USAGE')}
|
|
|
57
57
|
trazum conform <file|-> [--contract <name>]
|
|
58
58
|
trazum models
|
|
59
59
|
trazum rules
|
|
60
|
+
trazum gateway <anthropic|openai> --on-cannot-tell <fail-open|fail-closed>
|
|
61
|
+
trazum feedback
|
|
62
|
+
trazum --version
|
|
60
63
|
|
|
61
64
|
${bold('OPTIONS FOR init')}
|
|
62
65
|
--dry-run Print the config it would write and write nothing.
|
|
@@ -88,6 +91,33 @@ ${bold('OPTIONS FOR conform')}
|
|
|
88
91
|
The second never gates. Choosing not to log sessions is a decision, not a
|
|
89
92
|
defect. See docs/format.md.
|
|
90
93
|
|
|
94
|
+
${bold('OPTIONS FOR feedback')}
|
|
95
|
+
(none)
|
|
96
|
+
|
|
97
|
+
Prints where to report a wrong optimisation, a bug, a question or a security
|
|
98
|
+
problem — and a blank issue with the version, runtime and platform already
|
|
99
|
+
filled in, printed in full first so nothing travels you have not read.
|
|
100
|
+
|
|
101
|
+
It sends nothing. Trazum has no telemetry: no ping, no install hook, no
|
|
102
|
+
anonymous counter, and a test fails the build if this command ever reaches
|
|
103
|
+
the network.
|
|
104
|
+
|
|
105
|
+
${bold('OPTIONS FOR gateway')}
|
|
106
|
+
--on-cannot-tell <policy> Required, no default: fail-open or fail-closed.
|
|
107
|
+
What happens when the gateway cannot judge a call
|
|
108
|
+
— no budget, nothing measured, an unpriced model.
|
|
109
|
+
--port <n> | --socket <p> Where to listen. Loopback only, always.
|
|
110
|
+
|
|
111
|
+
Stands between your SDK and the provider, speaking their wire format, so no
|
|
112
|
+
code changes. Usage is measured from the provider's own response as it comes
|
|
113
|
+
back — no export, no connector lag, no missing day.
|
|
114
|
+
|
|
115
|
+
It refuses and never substitutes: a call over budget gets HTTP 402 with the
|
|
116
|
+
cheaper alternatives named. Substitution happens only where you wrote it down
|
|
117
|
+
in spend.substitute, with your reason, and every substituted call is marked.
|
|
118
|
+
|
|
119
|
+
Your credential is forwarded untouched and never read. See docs/gateway.md.
|
|
120
|
+
|
|
91
121
|
${bold('OPTIONS FOR prune')}
|
|
92
122
|
--cases <file> One input per line, or a JSON array. Required.
|
|
93
123
|
--yes Actually spend the calls. Without it the estimate is
|
|
@@ -907,7 +937,8 @@ ${bold('EXAMPLES')}
|
|
|
907
937
|
findingRoute: (model) => `The same work fits ${model}, which is cheaper per token.`,
|
|
908
938
|
findingBatch: () => 'The Batch API halves both halves of the bill, for work that can wait.',
|
|
909
939
|
findingTotal: (usd, days) => `Together: $${usd} over the same ${days} days.`,
|
|
910
|
-
findingNext: () =>
|
|
940
|
+
findingNext: () =>
|
|
941
|
+
'trazum plan <your log> ranks every action, not just this one. If a figure here looks wrong, trazum feedback says where to tell us — it sends nothing on its own.',
|
|
911
942
|
|
|
912
943
|
wouldOverwrite: (keys) => `This replaces keys you already set: ${keys}. Pass --yes to write anyway.`,
|
|
913
944
|
nothingToWrite: () => 'No config written: nothing above could be justified from what is here.',
|
|
@@ -919,6 +950,48 @@ ${bold('EXAMPLES')}
|
|
|
919
950
|
`${path} exists and could not be parsed, so nothing was written over it. Fix or move it first.`,
|
|
920
951
|
},
|
|
921
952
|
|
|
953
|
+
gateway: {
|
|
954
|
+
badProvider: (given, known) =>
|
|
955
|
+
given === ''
|
|
956
|
+
? `Name the provider to stand in front of. Known: ${known}.`
|
|
957
|
+
: `"${given}" is not a provider this gateway speaks for. Known: ${known}.`,
|
|
958
|
+
needsPolicy: (policies) =>
|
|
959
|
+
`--on-cannot-tell is required, and there is no default: ${policies}. When the gateway cannot judge a call — no budget, nothing measured, an unpriced model — one of these happens, and only you know which failure your product can survive. fail-open keeps it working and lets the bill run; fail-closed stops the bill and takes it down with it. Picking one for you would be the most consequential decision in your architecture, made silently at install time.`,
|
|
960
|
+
listening: (where, provider) => `Gateway on ${where}, in front of ${provider}`,
|
|
961
|
+
pointYourSdk: (where) =>
|
|
962
|
+
`Point your SDK's base URL at ${where} and change nothing else. It speaks the provider's own wire format, so no code changes and no new client.`,
|
|
963
|
+
credential: () =>
|
|
964
|
+
'Your credential is forwarded untouched and never read, never stored, never logged and never put in a URL. Trazum holds no key here and cannot make a call of its own through this.',
|
|
965
|
+
neverSubstitutes: () =>
|
|
966
|
+
'A call over budget is refused with HTTP 402 and the cheaper alternatives named — never silently swapped, trimmed or downgraded. 402 rather than 429 on purpose: every provider SDK retries a 429, which would turn one refusal into a retry storm.',
|
|
967
|
+
standing: (consumed, limit) =>
|
|
968
|
+
`Judging against ${consumed} of ${limit}, measured, read once at start — a file read in the request path would put this tool's latency between you and your provider on every call.`,
|
|
969
|
+
noStanding: () =>
|
|
970
|
+
'Nothing measured for this period, so every call is unjudged and the failure policy below decides. Set spend.monthlyUsd and pull with trazum connect.',
|
|
971
|
+
policy: (policy) =>
|
|
972
|
+
policy === 'fail-open'
|
|
973
|
+
? 'When it cannot judge: the call goes through, and the record says it was unjudged rather than within budget.'
|
|
974
|
+
: 'When it cannot judge: the call is refused. Nothing gets through unmeasured.',
|
|
975
|
+
measured: (model, label, input, output, substituted) =>
|
|
976
|
+
` ${model}${label === null ? '' : ` [${label}]`}: ${input} in, ${output} out${substituted ? ' (substituted — marked, and never counted as the call that was asked for)' : ''}`,
|
|
977
|
+
},
|
|
978
|
+
|
|
979
|
+
feedback: {
|
|
980
|
+
heading: () => 'Telling us something',
|
|
981
|
+
sendsNothing: () =>
|
|
982
|
+
'This command sends nothing, and neither does anything else here. Trazum has no telemetry: no ping, no install hook, no anonymous counter. A tool whose whole argument is that it reads your bill without uploading it cannot also be quietly reporting on you, and a test fails the build if this command ever reaches the network.',
|
|
983
|
+
whereHeading: () => 'Where',
|
|
984
|
+
wrongOptimisation: () =>
|
|
985
|
+
'A rule changed what a prompt asks for — the report that matters most, and the failure this product exists to avoid:',
|
|
986
|
+
bug: () => 'Anything else that is wrong:',
|
|
987
|
+
question: () => 'A question, or an idea you are not sure about:',
|
|
988
|
+
security: () => 'A security problem — privately, never a public issue:',
|
|
989
|
+
environmentHeading: () => 'What a maintainer will ask for',
|
|
990
|
+
environmentOnly: () =>
|
|
991
|
+
'That is the whole of it. Nothing about your work is here — not the config, not a prompt, not a label, not a figure. Those are what a good report needs and what only you can decide to share.',
|
|
992
|
+
linkHeading: () => 'A blank issue with the above already filled in',
|
|
993
|
+
},
|
|
994
|
+
|
|
922
995
|
conform: {
|
|
923
996
|
noTarget: () =>
|
|
924
997
|
'Pass a file to check — a usage log, or any document Trazum emits. Use "-" to read from stdin.',
|
package/src/i18n/es.ts
CHANGED
|
@@ -44,6 +44,9 @@ ${bold('USO')}
|
|
|
44
44
|
trazum conform <fichero|-> [--contract <nombre>]
|
|
45
45
|
trazum models
|
|
46
46
|
trazum rules
|
|
47
|
+
trazum gateway <anthropic|openai> --on-cannot-tell <fail-open|fail-closed>
|
|
48
|
+
trazum feedback
|
|
49
|
+
trazum --version
|
|
47
50
|
|
|
48
51
|
${bold('OPCIONES DE init')}
|
|
49
52
|
--dry-run Imprime la configuración que escribiría y no escribe nada.
|
|
@@ -75,6 +78,36 @@ ${bold('OPCIONES DE conform')}
|
|
|
75
78
|
La segunda nunca hace fallar nada. Decidir no registrar sesiones es una
|
|
76
79
|
decisión, no un defecto. Ver docs/format.md.
|
|
77
80
|
|
|
81
|
+
${bold('OPCIONES DE feedback')}
|
|
82
|
+
(ninguna)
|
|
83
|
+
|
|
84
|
+
Imprime dónde reportar una optimización incorrecta, un fallo, una pregunta o
|
|
85
|
+
un problema de seguridad — y una incidencia en blanco con la versión, el
|
|
86
|
+
runtime y la plataforma ya puestos, mostrados enteros antes para que no viaje
|
|
87
|
+
nada que no hayas leído.
|
|
88
|
+
|
|
89
|
+
No envía nada. Trazum no tiene telemetría: ni ping, ni hook de instalación, ni
|
|
90
|
+
contador anónimo, y una prueba hace fallar la compilación si este comando
|
|
91
|
+
llega a tocar la red.
|
|
92
|
+
|
|
93
|
+
${bold('OPCIONES DE gateway')}
|
|
94
|
+
--on-cannot-tell <política> Obligatorio, sin valor por defecto: fail-open o
|
|
95
|
+
fail-closed. Qué pasa cuando la pasarela no puede
|
|
96
|
+
juzgar una llamada — sin presupuesto, sin nada
|
|
97
|
+
medido, un modelo sin precio.
|
|
98
|
+
--port <n> | --socket <p> Dónde escuchar. Solo loopback, siempre.
|
|
99
|
+
|
|
100
|
+
Se pone entre tu SDK y el proveedor, hablando su formato, así que no hay
|
|
101
|
+
cambios de código. El consumo se mide desde la propia respuesta del proveedor
|
|
102
|
+
según llega — sin exportación, sin retraso, sin días que falten.
|
|
103
|
+
|
|
104
|
+
Rechaza y nunca sustituye: una llamada por encima del presupuesto recibe un
|
|
105
|
+
HTTP 402 con las alternativas más baratas nombradas. La sustitución solo
|
|
106
|
+
ocurre donde la escribiste en spend.substitute, con tu motivo, y toda llamada
|
|
107
|
+
sustituida queda marcada.
|
|
108
|
+
|
|
109
|
+
Tu credencial se reenvía intacta y nunca se lee. Ver docs/gateway.md.
|
|
110
|
+
|
|
78
111
|
${bold('OPCIONES DE prune')}
|
|
79
112
|
--cases <fichero> Una entrada por línea, o un array JSON. Obligatorio.
|
|
80
113
|
--yes Gasta las llamadas de verdad. Sin él se imprime la
|
|
@@ -940,7 +973,8 @@ ${bold('EJEMPLOS')}
|
|
|
940
973
|
findingRoute: (model) => `El mismo trabajo cabe en ${model}, más barato por token.`,
|
|
941
974
|
findingBatch: () => 'La API de lotes reduce a la mitad ambas mitades de la factura, para trabajo que puede esperar.',
|
|
942
975
|
findingTotal: (usd, days) => `Juntas: $${usd} en esos mismos ${days} días.`,
|
|
943
|
-
findingNext: () =>
|
|
976
|
+
findingNext: () =>
|
|
977
|
+
'trazum plan <tu registro> ordena todas las acciones, no solo esta. Si alguna cifra de aquí parece mal, trazum feedback dice dónde contarlo — no envía nada por su cuenta.',
|
|
944
978
|
|
|
945
979
|
wouldOverwrite: (keys) => `Esto reemplaza claves que ya tenías: ${keys}. Pasa --yes para escribir igualmente.`,
|
|
946
980
|
nothingToWrite: () => 'No se escribió configuración: nada de lo anterior se pudo justificar con lo que hay aquí.',
|
|
@@ -952,6 +986,48 @@ ${bold('EJEMPLOS')}
|
|
|
952
986
|
`${path} existe y no se pudo interpretar, así que no se escribió nada encima. Arréglalo o muévelo primero.`,
|
|
953
987
|
},
|
|
954
988
|
|
|
989
|
+
gateway: {
|
|
990
|
+
badProvider: (given, known) =>
|
|
991
|
+
given === ''
|
|
992
|
+
? `Nombra el proveedor delante del que ponerse. Conocidos: ${known}.`
|
|
993
|
+
: `"${given}" no es un proveedor por el que hable esta pasarela. Conocidos: ${known}.`,
|
|
994
|
+
needsPolicy: (policies) =>
|
|
995
|
+
`--on-cannot-tell es obligatorio, y no hay valor por defecto: ${policies}. Cuando la pasarela no puede juzgar una llamada — sin presupuesto, sin nada medido, un modelo sin precio — pasa una de las dos, y solo tú sabes qué fallo puede sobrevivir tu producto. fail-open lo mantiene funcionando y deja correr la factura; fail-closed para la factura y se lo lleva por delante. Elegir por ti sería tomar la decisión más consecuente de tu arquitectura, en silencio, al instalar.`,
|
|
996
|
+
listening: (where, provider) => `Pasarela en ${where}, delante de ${provider}`,
|
|
997
|
+
pointYourSdk: (where) =>
|
|
998
|
+
`Apunta la URL base de tu SDK a ${where} y no cambies nada más. Habla el formato del propio proveedor, así que no hay cambios de código ni cliente nuevo.`,
|
|
999
|
+
credential: () =>
|
|
1000
|
+
'Tu credencial se reenvía intacta y nunca se lee, ni se guarda, ni se registra, ni se pone en una URL. Trazum no tiene ninguna clave aquí y no puede hacer una llamada propia a través de esto.',
|
|
1001
|
+
neverSubstitutes: () =>
|
|
1002
|
+
'Una llamada por encima del presupuesto se rechaza con HTTP 402 y las alternativas más baratas nombradas — nunca se cambia, recorta ni degrada en silencio. 402 y no 429 a propósito: todos los SDK de proveedor reintentan un 429, lo que convertiría un rechazo en una tormenta de reintentos.',
|
|
1003
|
+
standing: (consumed, limit) =>
|
|
1004
|
+
`Juzgando contra ${consumed} de ${limit}, medido, leído una vez al arrancar — leer un fichero en el camino de la petición pondría la latencia de esta herramienta entre tú y tu proveedor en cada llamada.`,
|
|
1005
|
+
noStanding: () =>
|
|
1006
|
+
'Nada medido para este periodo, así que ninguna llamada se juzga y decide la política de abajo. Configura spend.monthlyUsd y descarga con trazum connect.',
|
|
1007
|
+
policy: (policy) =>
|
|
1008
|
+
policy === 'fail-open'
|
|
1009
|
+
? 'Cuando no puede juzgar: la llamada pasa, y el registro dice que no se juzgó, no que estaba dentro del presupuesto.'
|
|
1010
|
+
: 'Cuando no puede juzgar: la llamada se rechaza. Nada pasa sin medir.',
|
|
1011
|
+
measured: (model, label, input, output, substituted) =>
|
|
1012
|
+
` ${model}${label === null ? '' : ` [${label}]`}: ${input} entrada, ${output} salida${substituted ? ' (sustituida — marcada, y nunca contada como la llamada que se pidió)' : ''}`,
|
|
1013
|
+
},
|
|
1014
|
+
|
|
1015
|
+
feedback: {
|
|
1016
|
+
heading: () => 'Contarnos algo',
|
|
1017
|
+
sendsNothing: () =>
|
|
1018
|
+
'Este comando no envía nada, y nada más aquí lo hace tampoco. Trazum no tiene telemetría: ni ping, ni hook de instalación, ni contador anónimo. Una herramienta cuyo argumento entero es que lee tu factura sin subirla no puede a la vez estar informando sobre ti en silencio, y una prueba hace fallar la compilación si este comando llega a tocar la red.',
|
|
1019
|
+
whereHeading: () => 'Dónde',
|
|
1020
|
+
wrongOptimisation: () =>
|
|
1021
|
+
'Una regla cambió lo que pedía un prompt — el informe que más importa, y el fallo que este producto existe para evitar:',
|
|
1022
|
+
bug: () => 'Cualquier otra cosa que esté mal:',
|
|
1023
|
+
question: () => 'Una pregunta, o una idea de la que no estás seguro:',
|
|
1024
|
+
security: () => 'Un problema de seguridad — en privado, nunca en una incidencia pública:',
|
|
1025
|
+
environmentHeading: () => 'Lo que te va a preguntar quien mantiene esto',
|
|
1026
|
+
environmentOnly: () =>
|
|
1027
|
+
'Y eso es todo. Aquí no hay nada de tu trabajo — ni la configuración, ni un prompt, ni una etiqueta, ni una cifra. Eso es lo que necesita un buen informe y lo que solo tú puedes decidir compartir.',
|
|
1028
|
+
linkHeading: () => 'Una incidencia en blanco con lo de arriba ya puesto',
|
|
1029
|
+
},
|
|
1030
|
+
|
|
955
1031
|
conform: {
|
|
956
1032
|
noTarget: () =>
|
|
957
1033
|
'Pasa un archivo para comprobar — un registro de consumo, o cualquier documento que emita Trazum. Usa "-" para leer de la entrada estándar.',
|
package/src/i18n/types.ts
CHANGED
|
@@ -252,6 +252,44 @@ export interface CliMessages {
|
|
|
252
252
|
* Two halves, and the copy keeps them apart everywhere: problems gate, gaps
|
|
253
253
|
* do not. Choosing not to log sessions is a decision, not a defect.
|
|
254
254
|
*/
|
|
255
|
+
/**
|
|
256
|
+
* Where to say something, and the standing fact that this command sends
|
|
257
|
+
* nothing. The second half is not decoration: a tool that offers to help you
|
|
258
|
+
* file a report is exactly the shape of a tool that phones home, and the
|
|
259
|
+
* only way to be believed is to say so where somebody is looking.
|
|
260
|
+
*/
|
|
261
|
+
/**
|
|
262
|
+
* The proxy in the path. Every line here is either a promise about what it
|
|
263
|
+
* will not do, or the standing it is judging against — because a component
|
|
264
|
+
* between somebody and their provider is trusted on nothing but what it says
|
|
265
|
+
* plainly at start-up.
|
|
266
|
+
*/
|
|
267
|
+
gateway: {
|
|
268
|
+
badProvider(given: string, known: string): string;
|
|
269
|
+
needsPolicy(policies: string): string;
|
|
270
|
+
listening(where: string, provider: string): string;
|
|
271
|
+
pointYourSdk(where: string): string;
|
|
272
|
+
credential(): string;
|
|
273
|
+
neverSubstitutes(): string;
|
|
274
|
+
standing(consumed: string, limit: string): string;
|
|
275
|
+
noStanding(): string;
|
|
276
|
+
policy(policy: string): string;
|
|
277
|
+
measured(model: string, label: string | null, input: number, output: number, substituted: boolean): string;
|
|
278
|
+
};
|
|
279
|
+
|
|
280
|
+
feedback: {
|
|
281
|
+
heading(): string;
|
|
282
|
+
sendsNothing(): string;
|
|
283
|
+
whereHeading(): string;
|
|
284
|
+
wrongOptimisation(): string;
|
|
285
|
+
bug(): string;
|
|
286
|
+
question(): string;
|
|
287
|
+
security(): string;
|
|
288
|
+
environmentHeading(): string;
|
|
289
|
+
environmentOnly(): string;
|
|
290
|
+
linkHeading(): string;
|
|
291
|
+
};
|
|
292
|
+
|
|
255
293
|
conform: {
|
|
256
294
|
noTarget(): string;
|
|
257
295
|
badContract(given: string, known: string): string;
|