@bitbaum/ai-kit 0.12.0 → 0.14.0

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/dist/complete.js CHANGED
@@ -291,6 +291,25 @@ export async function complete(options) {
291
291
  options.onLinkFailure?.(link, failure);
292
292
  if (failure.kind === "daily")
293
293
  deadProviders.add(link.provider.id);
294
+ // A REJECTED KEY is a verdict about the VENDOR, not the model.
295
+ //
296
+ // 401/403 says "not you". Every remaining link at this provider presents
297
+ // the identical credential, so walking them spends a request each to be
298
+ // told the same thing — and then reports "all 5 links failed", which
299
+ // reads as an outage at someone else's shop and sends the reader looking
300
+ // for one. The fact worth surfacing is that a key this app holds was
301
+ // refused.
302
+ //
303
+ // Crossing to the NEXT vendor still happens: that is a different key, and
304
+ // the whole reason the chain spans vendors.
305
+ //
306
+ // Deliberately narrow. A 404 is a retired id, a 5xx is a vendor being
307
+ // unwell, a capacity 429 is a busy minute — all three are answered by
308
+ // asking a different model, and widening this skip to cover them would
309
+ // quietly turn the chain back into the pin it replaced.
310
+ if (failure.status === 401 || failure.status === 403) {
311
+ deadProviders.add(link.provider.id);
312
+ }
294
313
  // The caller cancelled — the request they were waiting on is gone. Walking
295
314
  // the rest of the chain now would spend their daily budget on an answer
296
315
  // nobody will read, and would report "every vendor failed" about vendors
@@ -94,6 +94,36 @@ export interface LivenessOptions extends Omit<CompleteOptions, "messages" | "max
94
94
  chain: Link[];
95
95
  env?: Env;
96
96
  }>;
97
+ /**
98
+ * Make the call YOURSELF, using the app's own path. Takes precedence over
99
+ * `chain` and `resolveChain`.
100
+ *
101
+ * Not every app can hand over a `Link[]`. Four in this fleet cannot: one
102
+ * builds its chain from its own provider CLASSES, one from a registry it
103
+ * deliberately owns (its BYOK list includes paid ids), one talks to a single
104
+ * operator-configured endpoint chosen for data residency, and one keeps a
105
+ * provider layer that predates this package. Requiring a chain would have
106
+ * meant either rewriting those or leaving them with no probe at all — and
107
+ * "no probe" is what left seven apps unable to answer "does the AI work?"
108
+ * after a fleet-wide refactor.
109
+ *
110
+ * There is a second, better reason. A probe built from a chain this module
111
+ * assembles tests A path; `ask` tests THE path — the same function the app's
112
+ * real features call. That is strictly stronger evidence, and it means the
113
+ * probe cannot quietly drift away from the code it is meant to vouch for.
114
+ *
115
+ * Everything else still applies: it runs only on an explicit, authorised
116
+ * probe, a success is cached, and a failure never is. Return the text the
117
+ * model produced and, if you have it, the `provider/model` that served it.
118
+ *
119
+ * An empty or whitespace-only `text` is treated as a FAILURE, for the same
120
+ * reason `complete()` treats it as one: a 200 carrying nothing is the
121
+ * failure most likely to be reported as success.
122
+ */
123
+ ask?: () => Promise<{
124
+ text: string;
125
+ id?: string;
126
+ }>;
97
127
  /** Injected for tests. Defaults to `Date.now`. */
98
128
  now?: () => number;
99
129
  }
package/dist/liveness.js CHANGED
@@ -90,6 +90,26 @@ export function createLivenessProbe(options = {}) {
90
90
  }
91
91
  const started = now();
92
92
  try {
93
+ if (options.ask) {
94
+ const asked = await options.ask();
95
+ const text = asked.text.trim();
96
+ // Same rule as `complete()`: a 200 carrying nothing is not an answer.
97
+ // Without this, an app whose own path returns "" on failure — several
98
+ // do, by design, so callers can degrade — would report itself healthy
99
+ // on exactly the outage this route exists to catch.
100
+ if (text === "") {
101
+ throw new Error("the app's own path returned empty content — no output was produced");
102
+ }
103
+ const fresh = {
104
+ ok: true,
105
+ ...(asked.id ? { servedBy: asked.id } : {}),
106
+ answer: text,
107
+ ms: now() - started,
108
+ cached: false,
109
+ };
110
+ lastOk = { at: now(), result: fresh };
111
+ return fresh;
112
+ }
93
113
  // Resolved here, not at construction, and only on a real probe — so a
94
114
  // monitor polling this route does not also poll whatever backs it.
95
115
  const resolved = options.resolveChain ? await options.resolveChain() : options.chain;
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@bitbaum/ai-kit",
3
- "version": "0.12.0",
3
+ "version": "0.14.0",
4
4
  "description": "One install for the AI layer of an app: which model to call, what to do when the vendor retires it, how to walk the fallback chain and know when none of it worked, how to read the three kinds of 429, a fair daily budget across users, headless AI form filling — and now the model registry (one SSOT for every callable id, with the paid/free boundary as a field) and the grounding harness (facts, contract, deterministic fabrication check).",
5
5
  "license": "MIT",
6
6
  "author": "Mao Nakamoto",
package/src/complete.ts CHANGED
@@ -449,6 +449,26 @@ export async function complete(options: CompleteOptions): Promise<CompleteResult
449
449
 
450
450
  if (failure.kind === "daily") deadProviders.add(link.provider.id);
451
451
 
452
+ // A REJECTED KEY is a verdict about the VENDOR, not the model.
453
+ //
454
+ // 401/403 says "not you". Every remaining link at this provider presents
455
+ // the identical credential, so walking them spends a request each to be
456
+ // told the same thing — and then reports "all 5 links failed", which
457
+ // reads as an outage at someone else's shop and sends the reader looking
458
+ // for one. The fact worth surfacing is that a key this app holds was
459
+ // refused.
460
+ //
461
+ // Crossing to the NEXT vendor still happens: that is a different key, and
462
+ // the whole reason the chain spans vendors.
463
+ //
464
+ // Deliberately narrow. A 404 is a retired id, a 5xx is a vendor being
465
+ // unwell, a capacity 429 is a busy minute — all three are answered by
466
+ // asking a different model, and widening this skip to cover them would
467
+ // quietly turn the chain back into the pin it replaced.
468
+ if (failure.status === 401 || failure.status === 403) {
469
+ deadProviders.add(link.provider.id);
470
+ }
471
+
452
472
  // The caller cancelled — the request they were waiting on is gone. Walking
453
473
  // the rest of the chain now would spend their daily budget on an answer
454
474
  // nobody will read, and would report "every vendor failed" about vendors
package/src/liveness.ts CHANGED
@@ -96,6 +96,33 @@ export interface LivenessOptions extends Omit<
96
96
  */
97
97
  resolveChain?: () =>
98
98
  Link[] | { chain: Link[]; env?: Env } | Promise<Link[] | { chain: Link[]; env?: Env }>;
99
+ /**
100
+ * Make the call YOURSELF, using the app's own path. Takes precedence over
101
+ * `chain` and `resolveChain`.
102
+ *
103
+ * Not every app can hand over a `Link[]`. Four in this fleet cannot: one
104
+ * builds its chain from its own provider CLASSES, one from a registry it
105
+ * deliberately owns (its BYOK list includes paid ids), one talks to a single
106
+ * operator-configured endpoint chosen for data residency, and one keeps a
107
+ * provider layer that predates this package. Requiring a chain would have
108
+ * meant either rewriting those or leaving them with no probe at all — and
109
+ * "no probe" is what left seven apps unable to answer "does the AI work?"
110
+ * after a fleet-wide refactor.
111
+ *
112
+ * There is a second, better reason. A probe built from a chain this module
113
+ * assembles tests A path; `ask` tests THE path — the same function the app's
114
+ * real features call. That is strictly stronger evidence, and it means the
115
+ * probe cannot quietly drift away from the code it is meant to vouch for.
116
+ *
117
+ * Everything else still applies: it runs only on an explicit, authorised
118
+ * probe, a success is cached, and a failure never is. Return the text the
119
+ * model produced and, if you have it, the `provider/model` that served it.
120
+ *
121
+ * An empty or whitespace-only `text` is treated as a FAILURE, for the same
122
+ * reason `complete()` treats it as one: a 200 carrying nothing is the
123
+ * failure most likely to be reported as success.
124
+ */
125
+ ask?: () => Promise<{ text: string; id?: string }>;
99
126
  /** Injected for tests. Defaults to `Date.now`. */
100
127
  now?: () => number;
101
128
  }
@@ -165,6 +192,27 @@ export function createLivenessProbe(options: LivenessOptions = {}): LivenessProb
165
192
 
166
193
  const started = now();
167
194
  try {
195
+ if (options.ask) {
196
+ const asked = await options.ask();
197
+ const text = asked.text.trim();
198
+ // Same rule as `complete()`: a 200 carrying nothing is not an answer.
199
+ // Without this, an app whose own path returns "" on failure — several
200
+ // do, by design, so callers can degrade — would report itself healthy
201
+ // on exactly the outage this route exists to catch.
202
+ if (text === "") {
203
+ throw new Error("the app's own path returned empty content — no output was produced");
204
+ }
205
+ const fresh: LivenessResult = {
206
+ ok: true,
207
+ ...(asked.id ? { servedBy: asked.id } : {}),
208
+ answer: text,
209
+ ms: now() - started,
210
+ cached: false,
211
+ };
212
+ lastOk = { at: now(), result: fresh };
213
+ return fresh;
214
+ }
215
+
168
216
  // Resolved here, not at construction, and only on a real probe — so a
169
217
  // monitor polling this route does not also poll whatever backs it.
170
218
  const resolved = options.resolveChain ? await options.resolveChain() : options.chain;