@warmdrift/kgauto-compiler 2.0.0-alpha.95 → 2.0.0-alpha.97
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/{chunk-BJD75WCL.mjs → chunk-77AHSX6Y.mjs} +2 -2
- package/dist/{chunk-IRXC7WEQ.mjs → chunk-AFYBZRQV.mjs} +6 -5
- package/dist/{chunk-D3BVYXNA.mjs → chunk-HHIXHL4O.mjs} +137 -42
- package/dist/{chunk-JFM5A556.mjs → chunk-IEEROFZW.mjs} +1 -1
- package/dist/glassbox-routes/index.js +118 -40
- package/dist/glassbox-routes/index.mjs +2 -2
- package/dist/index.d.mts +1 -1
- package/dist/index.d.ts +1 -1
- package/dist/index.js +143 -47
- package/dist/index.mjs +4 -4
- package/dist/key-health.js +1 -1
- package/dist/key-health.mjs +1 -1
- package/dist/probe.js +142 -46
- package/dist/probe.mjs +3 -3
- package/dist/profiles.d.mts +4 -1
- package/dist/profiles.d.ts +4 -1
- package/dist/profiles.js +137 -42
- package/dist/profiles.mjs +1 -1
- package/package.json +1 -1
package/dist/index.js
CHANGED
|
@@ -1163,18 +1163,22 @@ var PROFILES_RAW = [
|
|
|
1163
1163
|
maxContextTokens: 1e6,
|
|
1164
1164
|
maxOutputTokens: 384e3,
|
|
1165
1165
|
maxTools: 16,
|
|
1166
|
-
|
|
1166
|
+
// alpha.96 — CORRECTED from `false`, which was AI-generated at greenfield
|
|
1167
|
+
// (2026-04-25) and never re-checked. Measured 2026-08-28 direct to
|
|
1168
|
+
// api.deepseek.com, no kgauto in the path, 6 tools offered, 3 trials:
|
|
1169
|
+
// 3 parallel tool calls returned, 3/3. L-040's original evidence said
|
|
1170
|
+
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
1171
|
+
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
1172
|
+
parallelToolCalls: true,
|
|
1167
1173
|
structuredOutput: "native",
|
|
1168
1174
|
systemPromptMode: "inline",
|
|
1169
1175
|
streaming: true,
|
|
1170
|
-
|
|
1171
|
-
|
|
1172
|
-
|
|
1173
|
-
|
|
1174
|
-
|
|
1175
|
-
|
|
1176
|
-
}
|
|
1177
|
-
],
|
|
1176
|
+
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
1177
|
+
// It restated `parallelToolCalls: false` as a threshold and was falsified by
|
|
1178
|
+
// the same measurement. It is not kept-and-flagged (the usual lifecycle)
|
|
1179
|
+
// because it was never a real failure mode: it was a second enforcement of a
|
|
1180
|
+
// wrong belief, and keeping it active would keep amputating tools.
|
|
1181
|
+
cliffs: [],
|
|
1178
1182
|
costInputPer1m: 0.44,
|
|
1179
1183
|
costOutputPer1m: 1.32,
|
|
1180
1184
|
lowering: {
|
|
@@ -1186,7 +1190,12 @@ var PROFILES_RAW = [
|
|
|
1186
1190
|
{ signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
|
|
1187
1191
|
],
|
|
1188
1192
|
strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
|
|
1189
|
-
|
|
1193
|
+
// alpha.96 — 'parallel_tools' REMOVED: measurably false (2026-08-28, direct
|
|
1194
|
+
// to api.deepseek.com, 6 tools, 3 trials, 3 parallel calls, both models).
|
|
1195
|
+
// 'large_tool_sets' KEPT and explicitly unverified — the probe offered 6 of
|
|
1196
|
+
// a declared maxTools 16, so it says nothing about behaviour at 16. Not
|
|
1197
|
+
// falsified is not the same as confirmed.
|
|
1198
|
+
weaknesses: ["large_tool_sets"],
|
|
1190
1199
|
// alpha.47 — explicit slow override. Tag derivation would say 'medium'
|
|
1191
1200
|
// (no 'latency' weakness, no 'speed' strength), but the alpha.46 shadow
|
|
1192
1201
|
// probe MEASURED deepseek-v4-flash at 20485ms served (2026-06-03) — ~2.3×
|
|
@@ -1198,7 +1207,11 @@ var PROFILES_RAW = [
|
|
|
1198
1207
|
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
1199
1208
|
// Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
|
|
1200
1209
|
// classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
|
|
1201
|
-
//
|
|
1210
|
+
// alpha.96 — the hunt half of this sentence is RETRACTED: "sequential tools
|
|
1211
|
+
// — L-040" was falsified 2026-08-28 (direct measurement, 3 parallel tool
|
|
1212
|
+
// calls, both models). Do not use this comment to re-lower `hunt`; it is
|
|
1213
|
+
// the exact prose that justified the wrong score. The reasoning-depth
|
|
1214
|
+
// rationale is untouched and still carries `critique`.
|
|
1202
1215
|
archetypePerf: {
|
|
1203
1216
|
classify: 7,
|
|
1204
1217
|
// brain-validated, 169 rows
|
|
@@ -1210,8 +1223,16 @@ var PROFILES_RAW = [
|
|
|
1210
1223
|
plan: 5,
|
|
1211
1224
|
extract: 5,
|
|
1212
1225
|
critique: 4,
|
|
1213
|
-
|
|
1214
|
-
//
|
|
1226
|
+
// alpha.96 — was 4. Its ONLY written justification was "sequential tool
|
|
1227
|
+
// calls only — L-040", which is falsified. A score derived from a false
|
|
1228
|
+
// premise cannot be kept, and re-guessing a penalty would just relaunder
|
|
1229
|
+
// the same belief. Set to the floor (6) = "no evidence of deficiency",
|
|
1230
|
+
// NOT an endorsement. This is a JUDGMENT, not a measurement, and it is
|
|
1231
|
+
// the honest direction of the two: the only hunt-shaped evidence we have
|
|
1232
|
+
// is positive (a research-agent fixture, 3 of 3 demanded tools correctly
|
|
1233
|
+
// selected from a 6-tool menu, 3/3 trials, both models). Replace with a
|
|
1234
|
+
// real number from the golden-eval hunt corpus — that is what it is for.
|
|
1235
|
+
hunt: 6
|
|
1215
1236
|
},
|
|
1216
1237
|
// alpha.43 — family-level conventions for the `deepseek-chat` family
|
|
1217
1238
|
// (V4-Flash + future non-thinking-mode members). V4-Flash struggles
|
|
@@ -1236,18 +1257,22 @@ var PROFILES_RAW = [
|
|
|
1236
1257
|
maxContextTokens: 1e6,
|
|
1237
1258
|
maxOutputTokens: 384e3,
|
|
1238
1259
|
maxTools: 16,
|
|
1239
|
-
|
|
1260
|
+
// alpha.96 — CORRECTED from `false`, which was AI-generated at greenfield
|
|
1261
|
+
// (2026-04-25) and never re-checked. Measured 2026-08-28 direct to
|
|
1262
|
+
// api.deepseek.com, no kgauto in the path, 6 tools offered, 3 trials:
|
|
1263
|
+
// 3 parallel tool calls returned, 3/3. L-040's original evidence said
|
|
1264
|
+
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
1265
|
+
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
1266
|
+
parallelToolCalls: true,
|
|
1240
1267
|
structuredOutput: "native",
|
|
1241
1268
|
systemPromptMode: "inline",
|
|
1242
1269
|
streaming: true,
|
|
1243
|
-
|
|
1244
|
-
|
|
1245
|
-
|
|
1246
|
-
|
|
1247
|
-
|
|
1248
|
-
|
|
1249
|
-
}
|
|
1250
|
-
],
|
|
1270
|
+
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
1271
|
+
// It restated `parallelToolCalls: false` as a threshold and was falsified by
|
|
1272
|
+
// the same measurement. It is not kept-and-flagged (the usual lifecycle)
|
|
1273
|
+
// because it was never a real failure mode: it was a second enforcement of a
|
|
1274
|
+
// wrong belief, and keeping it active would keep amputating tools.
|
|
1275
|
+
cliffs: [],
|
|
1251
1276
|
// Verified against the live DeepSeek pricing page 2026-06-22: the 75%-off
|
|
1252
1277
|
// launch "promo" ($0.435/$0.87) did NOT revert on 2026-05-31 — it became the
|
|
1253
1278
|
// standing price (docs show $0.435 in / $0.87 out cache-miss, $0.003625
|
|
@@ -1265,7 +1290,12 @@ var PROFILES_RAW = [
|
|
|
1265
1290
|
{ signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
|
|
1266
1291
|
],
|
|
1267
1292
|
strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
|
|
1268
|
-
|
|
1293
|
+
// alpha.96 — 'parallel_tools' REMOVED: measurably false (2026-08-28, direct
|
|
1294
|
+
// to api.deepseek.com, 6 tools, 3 trials, 3 parallel calls, both models).
|
|
1295
|
+
// 'large_tool_sets' KEPT and explicitly unverified — the probe offered 6 of
|
|
1296
|
+
// a declared maxTools 16, so it says nothing about behaviour at 16. Not
|
|
1297
|
+
// falsified is not the same as confirmed.
|
|
1298
|
+
weaknesses: ["large_tool_sets"],
|
|
1269
1299
|
// alpha.47 — explicit slow override. Measured 47722ms on the alpha.46
|
|
1270
1300
|
// shadow probe (2026-06-03) — it's an extended-thinking reasoner, slowest
|
|
1271
1301
|
// of the served set. Tag derivation would say 'medium'; the measurement says
|
|
@@ -1284,14 +1314,24 @@ var PROFILES_RAW = [
|
|
|
1284
1314
|
summarize: 7,
|
|
1285
1315
|
extract: 6,
|
|
1286
1316
|
transform: 6,
|
|
1287
|
-
|
|
1288
|
-
//
|
|
1317
|
+
// alpha.96 — was 4. Its ONLY written justification was "sequential tool
|
|
1318
|
+
// calls only — L-040", which is falsified. A score derived from a false
|
|
1319
|
+
// premise cannot be kept, and re-guessing a penalty would just relaunder
|
|
1320
|
+
// the same belief. Set to the floor (6) = "no evidence of deficiency",
|
|
1321
|
+
// NOT an endorsement. This is a JUDGMENT, not a measurement, and it is
|
|
1322
|
+
// the honest direction of the two: the only hunt-shaped evidence we have
|
|
1323
|
+
// is positive (a research-agent fixture, 3 of 3 demanded tools correctly
|
|
1324
|
+
// selected from a 6-tool menu, 3/3 trials, both models). Replace with a
|
|
1325
|
+
// real number from the golden-eval hunt corpus — that is what it is for.
|
|
1326
|
+
hunt: 6
|
|
1289
1327
|
},
|
|
1290
1328
|
// alpha.43 — family-level conventions for the `deepseek-reasoner` family
|
|
1291
1329
|
// (V4-Pro is the family rep; future thinking-mode members inherit).
|
|
1292
1330
|
//
|
|
1293
|
-
// The decisive-archetype suffix is the load-bearing piece:
|
|
1294
|
-
//
|
|
1331
|
+
// The decisive-archetype suffix is the load-bearing piece. ⚠️ alpha.96:
|
|
1332
|
+
// its evidence was purged as poison on 2026-05-29 (s51) — retained as a
|
|
1333
|
+
// hypothesis, not a finding. Original claim, for the record: 2026-05-28
|
|
1334
|
+
// first real V4-Pro probe on tt-intel/classify
|
|
1295
1335
|
// showed 8/10 judge rationales citing "candidate hedges and fails to
|
|
1296
1336
|
// commit to a single classification" or "candidate refuses to pick
|
|
1297
1337
|
// among the labels." Without the forcing-function suffix, every
|
|
@@ -1312,24 +1352,35 @@ var PROFILES_RAW = [
|
|
|
1312
1352
|
{
|
|
1313
1353
|
archetype: "classify",
|
|
1314
1354
|
promptSuffix: "\n\nOutput exactly one classification label. Do not hedge, do not add preamble, do not explain your reasoning. If uncertain, pick the most likely label and commit to it.",
|
|
1315
|
-
reason:
|
|
1355
|
+
reason: 'Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. UNEVIDENCED HYPOTHESIS as of alpha.96 \u2014 this reason used to cite a 2026-05-28 V4-Pro probe on tt-intel/classify ("8 of 10 judge rationales cited hedging") and named `brain probe_outcomes` as its source. That source is EMPTY: probed 2026-08-28, zero probe_outcomes rows exist before 2026-06-01 and zero `classify` rows have ever existed. The probe was deleted on 2026-05-29 by the s51 poison purge \u2014 it replayed against truncated `prompt_preview`, which s51 established yields meaningless verdicts. The conclusion outlived the purge of its own evidence and kept shaping prompts for three months. The pointer was wrong too: exclusion-finding ID 20 is a zero-calls `recommend-probe` on classify, confidence low, later declined \u2014 the probe TRIGGER, never its result. KEPT because a promptSuffix is capability-PRESERVING (adds an instruction, removes nothing a raw call had) and is visible in mutations_applied. Re-evidence from the golden-eval corpus or drop it.'
|
|
1316
1356
|
},
|
|
1317
1357
|
{
|
|
1318
1358
|
archetype: "extract",
|
|
1319
1359
|
promptSuffix: "\n\nOutput exactly the requested structured fields. Do not hedge, do not add preamble, do not explain your reasoning. If a field is unknown, emit the field with a null/empty value and commit to that \u2014 do not refuse to extract.",
|
|
1320
|
-
reason:
|
|
1360
|
+
reason: 'Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. UNEVIDENCED HYPOTHESIS as of alpha.96 \u2014 the supporting evidence NO LONGER EXISTS. The cited 2026-05-28 V4-Pro classify probe ("8/10 judge rationales") was deleted on 2026-05-29 by the s51 poison purge, because it replayed against truncated `prompt_preview` and s51 established such verdicts are meaningless. Probed: zero probe_outcomes rows exist before 2026-06-01 and zero `classify` rows have EVER existed. The pointer was also wrong \u2014 exclusion-finding ID 20 is a zero-calls `recommend-probe` on classify, confidence low, later declined; it was the trigger for the probe, never its result. KEPT anyway because a promptSuffix is capability-PRESERVING (it adds an instruction and removes nothing a raw call had), so the cost of being wrong is bounded and a consumer can see it in mutations_applied. Re-evidence it from the golden-eval corpus or drop it.'
|
|
1321
1361
|
},
|
|
1322
1362
|
{
|
|
1323
1363
|
archetype: "ask",
|
|
1324
1364
|
promptSuffix: "\n\nAnswer the question directly. Do not hedge, do not add preamble, do not explain your reasoning unless explicitly asked. If uncertain, give your best answer and commit to it.",
|
|
1325
|
-
reason:
|
|
1326
|
-
},
|
|
1327
|
-
{
|
|
1328
|
-
archetype: "hunt",
|
|
1329
|
-
whenToolCountAtLeast: 3,
|
|
1330
|
-
cliffWarning: "Reasoners (deepseek-reasoner family) produce sequential reasoning chains by design; hunt archetypes with 3+ parallel tools lose substantial throughput. Consider gemini-flash family (parallel-strong) instead \u2014 see L-040 / L-041.",
|
|
1331
|
-
reason: "L-040 + L-041 architectural reading; 2026-05-28 V4-Pro probe evidence (exclusion-finding ID 20) confirms reasoner-family hunt fit is structurally weak."
|
|
1365
|
+
reason: 'Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. UNEVIDENCED HYPOTHESIS as of alpha.96 \u2014 the supporting evidence NO LONGER EXISTS. The cited 2026-05-28 V4-Pro classify probe ("8/10 judge rationales") was deleted on 2026-05-29 by the s51 poison purge, because it replayed against truncated `prompt_preview` and s51 established such verdicts are meaningless. Probed: zero probe_outcomes rows exist before 2026-06-01 and zero `classify` rows have EVER existed. The pointer was also wrong \u2014 exclusion-finding ID 20 is a zero-calls `recommend-probe` on classify, confidence low, later declined; it was the trigger for the probe, never its result. KEPT anyway because a promptSuffix is capability-PRESERVING (it adds an instruction and removes nothing a raw call had), so the cost of being wrong is bounded and a consumer can see it in mutations_applied. Re-evidence it from the golden-eval corpus or drop it.'
|
|
1332
1366
|
}
|
|
1367
|
+
// alpha.96 — the `hunt` convention was DELETED, not corrected.
|
|
1368
|
+
// It warned that "hunt archetypes with 3+ parallel tools lose substantial
|
|
1369
|
+
// throughput" and steered callers to gemini-flash. Measured false
|
|
1370
|
+
// 2026-08-28: 6 tools offered, 3 of 3 demanded served in parallel, 3/3
|
|
1371
|
+
// trials, both V4 models, direct to the provider.
|
|
1372
|
+
//
|
|
1373
|
+
// Its cited evidence did not say what it claimed. exclusion-finding ID 20
|
|
1374
|
+
// is `intent_archetype: 'classify'` (not hunt), `verdict:
|
|
1375
|
+
// 'recommend-probe'` (i.e. NOBODY PROBED), `confidence: 'low'`, its
|
|
1376
|
+
// message reads "deepseek-v4-pro has zero calls", it was later
|
|
1377
|
+
// `declined`, and its own `parallel_tool_hint: false` was this profile's
|
|
1378
|
+
// own belief round-tripped back as input. A no-data finding on the wrong
|
|
1379
|
+
// archetype was paraphrased into "probe evidence confirms".
|
|
1380
|
+
//
|
|
1381
|
+
// Deleted rather than re-scoped because a cliffWarning is
|
|
1382
|
+
// capability-REDUCING (it steers traffic away from a model) and nothing
|
|
1383
|
+
// evidenced remains. Re-add only from a measured hunt eval.
|
|
1333
1384
|
]
|
|
1334
1385
|
},
|
|
1335
1386
|
// ── OpenAI ──
|
|
@@ -2228,6 +2279,19 @@ var PROFILES_RAW = [
|
|
|
2228
2279
|
id: "glm-5.2",
|
|
2229
2280
|
verifiedAgainstDocs: "2026-07-18",
|
|
2230
2281
|
provider: "zai",
|
|
2282
|
+
probedLimits: [
|
|
2283
|
+
{
|
|
2284
|
+
metric: "tool_count",
|
|
2285
|
+
testedAt: 6,
|
|
2286
|
+
demanded: 3,
|
|
2287
|
+
served: [3, 3, 3],
|
|
2288
|
+
outcome: "held",
|
|
2289
|
+
trials: 3,
|
|
2290
|
+
lastProbed: "2026-08-29",
|
|
2291
|
+
route: "openrouter:baidu",
|
|
2292
|
+
claim: "Measured 2026-08-29 via OpenRouter (served-by Baidu): 6 tools offered, 3 of 3 demanded served in parallel, 3/3 trials. Route-scoped \u2014 direct z.ai is unmeasured (no key held)."
|
|
2293
|
+
}
|
|
2294
|
+
],
|
|
2231
2295
|
status: "preview",
|
|
2232
2296
|
maxContextTokens: 1048576,
|
|
2233
2297
|
maxOutputTokens: 131072,
|
|
@@ -2238,7 +2302,8 @@ var PROFILES_RAW = [
|
|
|
2238
2302
|
// NOTHING about multiple tool_calls per response. Conservative false so
|
|
2239
2303
|
// hunt-shaped ranking doesn't over-promise; flip on doc or probe
|
|
2240
2304
|
// evidence (L-081: undocumented ≠ true).
|
|
2241
|
-
parallelToolCalls:
|
|
2305
|
+
parallelToolCalls: true,
|
|
2306
|
+
// alpha.96 — see weaknesses note; was an unevidenced blanket false
|
|
2242
2307
|
structuredOutput: "native",
|
|
2243
2308
|
systemPromptMode: "inline",
|
|
2244
2309
|
streaming: true,
|
|
@@ -2261,7 +2326,17 @@ var PROFILES_RAW = [
|
|
|
2261
2326
|
{ signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
|
|
2262
2327
|
],
|
|
2263
2328
|
strengths: ["cost", "1m_context", "code", "tool_use", "json_output", "reasoning"],
|
|
2264
|
-
|
|
2329
|
+
// alpha.96 — 'parallel_tools' REMOVED. Measured 2026-08-29 via OpenRouter
|
|
2330
|
+
// (served-by Baidu), 6 tools offered, 3 of 3 demanded served in PARALLEL,
|
|
2331
|
+
// 3/3 trials. ROUTE-SCOPED: per the alpha.92 execute-leg contract there is
|
|
2332
|
+
// no cross-route verdict transfer, and direct z.ai is UNMEASURED (we hold
|
|
2333
|
+
// no zai key). But `parallelToolCalls: false` was a blanket claim with no
|
|
2334
|
+
// route qualification and no evidence at all, and it is capability-
|
|
2335
|
+
// REDUCING — it drives requires-adapter/reject. An unsubstantiated reducer
|
|
2336
|
+
// must not reduce. Contradicted on the one route we can reach is enough to
|
|
2337
|
+
// stop enforcing it; it is not enough to claim the opposite, which is why
|
|
2338
|
+
// probedLimits below carries the route.
|
|
2339
|
+
weaknesses: [],
|
|
2265
2340
|
notes: "GLM-5.2 (Z.ai, 2026-06): agentic-coding flagship. 1,048,576 ctx / 131,072 max out. $1.40/$4.40 per 1M; cached input $0.26/M (automatic caching, no marker; storage limited-time free as of 2026-07-18). Thinking on by default (`thinking.type`), `reasoning_effort` max..none. Parallel tool calls UNDOCUMENTED on the hosted API \u2014 profile says false until doc or probe evidence. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
|
|
2266
2341
|
// Starter hypotheses — verify with telemetry/probes; NO brain evidence
|
|
2267
2342
|
// yet. Anchored on the vendor's agentic-coding positioning at a
|
|
@@ -2279,8 +2354,11 @@ var PROFILES_RAW = [
|
|
|
2279
2354
|
extract: 6,
|
|
2280
2355
|
critique: 6,
|
|
2281
2356
|
classify: 6,
|
|
2282
|
-
|
|
2283
|
-
//
|
|
2357
|
+
// alpha.96 — was 5, discounted for "parallel unproven". Parallel is no
|
|
2358
|
+
// longer unproven on the reachable route (3/3, 2026-08-29). The discount
|
|
2359
|
+
// had no other stated basis, so it is withdrawn to the neutral floor.
|
|
2360
|
+
// Judgment, not measurement — same standard as the DeepSeek correction.
|
|
2361
|
+
hunt: 6
|
|
2284
2362
|
}
|
|
2285
2363
|
},
|
|
2286
2364
|
// ── Moonshot (Kimi) ──
|
|
@@ -2413,7 +2491,24 @@ function bestEffortProfile(id) {
|
|
|
2413
2491
|
maxContextTokens: 128e3,
|
|
2414
2492
|
maxOutputTokens: 4096,
|
|
2415
2493
|
maxTools: donor.maxTools,
|
|
2416
|
-
|
|
2494
|
+
// alpha.97 — was `false`. That was the SAME MISTAKE as the three profiles
|
|
2495
|
+
// corrected in alpha.96, in the one place it does the most damage: applied
|
|
2496
|
+
// to models nobody has measured at all.
|
|
2497
|
+
//
|
|
2498
|
+
// `false` here is capability-REDUCING and unsubstantiated by construction —
|
|
2499
|
+
// a synthesized profile exists precisely because we know nothing. It made
|
|
2500
|
+
// `getModelCompatibility(<any unprofiled model>, {archetype:'hunt'})`
|
|
2501
|
+
// return `requires-adapter` (verified: raw score 5 vs floor 6, adapter
|
|
2502
|
+
// offered), so every consumer on `onUnprofiledModel:'best-effort'` was told
|
|
2503
|
+
// to serialize tool calls on a model we had never touched.
|
|
2504
|
+
//
|
|
2505
|
+
// The never-worse-than-raw rule decides it: for an unknown model kgauto
|
|
2506
|
+
// must behave exactly as a direct call would, and a direct call passes
|
|
2507
|
+
// every tool and lets the provider default apply. `true` is not a claim
|
|
2508
|
+
// that the model parallelises — it is the absence of a claim, expressed as
|
|
2509
|
+
// non-interference. The honest long-term fix is a third state (unknown)
|
|
2510
|
+
// rather than a boolean; that is P4 in the rebuild contract.
|
|
2511
|
+
parallelToolCalls: true,
|
|
2417
2512
|
structuredOutput: "none",
|
|
2418
2513
|
systemPromptMode: donor.systemPromptMode,
|
|
2419
2514
|
streaming: donor.streaming,
|
|
@@ -2423,7 +2518,7 @@ function bestEffortProfile(id) {
|
|
|
2423
2518
|
lowering: donor.lowering,
|
|
2424
2519
|
recovery: [],
|
|
2425
2520
|
strengths: [],
|
|
2426
|
-
weaknesses: ["best-effort synthesized profile \u2014 no verified capability data"],
|
|
2521
|
+
weaknesses: ["best-effort synthesized profile \u2014 no verified capability data; kgauto does not reduce capability for a model it has never measured (alpha.97)"],
|
|
2427
2522
|
notes: `Synthesized by onUnprofiledModel:'best-effort' (provider ${provider} inferred from id; wire mechanics borrowed from ${donor.id}). No cliffs, no measured knowledge, cost UNKNOWN (recorded as 0). Replace via registerProfiles() for real guards.`
|
|
2428
2523
|
};
|
|
2429
2524
|
CONSUMER_INDEX.set(canonical, synth);
|
|
@@ -2624,7 +2719,7 @@ _setProfileBrainHook({
|
|
|
2624
2719
|
var ARCHETYPE_FLOOR_DEFAULT = 6;
|
|
2625
2720
|
var ABSOLUTE_FLOOR = 4;
|
|
2626
2721
|
function rawArchetypePerf(profile, archetype) {
|
|
2627
|
-
return profile.archetypePerf?.[archetype] ??
|
|
2722
|
+
return profile.archetypePerf?.[archetype] ?? ARCHETYPE_FLOOR_DEFAULT;
|
|
2628
2723
|
}
|
|
2629
2724
|
function hasSequentialToolCliffForHunt(profile) {
|
|
2630
2725
|
if (profile.parallelToolCalls !== false) return false;
|
|
@@ -3775,7 +3870,7 @@ function lower(ir, profile, hints = {}) {
|
|
|
3775
3870
|
case "openai":
|
|
3776
3871
|
return lowerOpenAI(ir, profile, hints);
|
|
3777
3872
|
case "deepseek":
|
|
3778
|
-
return lowerDeepSeek(ir, profile);
|
|
3873
|
+
return lowerDeepSeek(ir, profile, hints);
|
|
3779
3874
|
case "zai":
|
|
3780
3875
|
return lowerZai(ir, profile, hints);
|
|
3781
3876
|
case "moonshot":
|
|
@@ -4036,7 +4131,7 @@ function toOpenAITools(tools) {
|
|
|
4036
4131
|
}
|
|
4037
4132
|
}));
|
|
4038
4133
|
}
|
|
4039
|
-
function lowerDeepSeek(ir, profile) {
|
|
4134
|
+
function lowerDeepSeek(ir, profile, hints) {
|
|
4040
4135
|
const ordered = sortSections(ir.sections);
|
|
4041
4136
|
const systemText = ordered.map((s) => s.text).join("\n\n");
|
|
4042
4137
|
const messages = systemText ? [{ role: "system", content: systemText }] : [];
|
|
@@ -4058,7 +4153,8 @@ function lowerDeepSeek(ir, profile) {
|
|
|
4058
4153
|
provider: "deepseek",
|
|
4059
4154
|
model: profile.id,
|
|
4060
4155
|
messages,
|
|
4061
|
-
|
|
4156
|
+
// Every tool the IR carries. Never fewer — see the note above.
|
|
4157
|
+
tools: ir.tools && ir.tools.length > 0 ? ir.tools.map((t) => ({
|
|
4062
4158
|
type: "function",
|
|
4063
4159
|
function: {
|
|
4064
4160
|
name: t.name,
|
|
@@ -6574,7 +6670,7 @@ function validateFinalFit(ir, profile, tokens) {
|
|
|
6574
6670
|
}
|
|
6575
6671
|
|
|
6576
6672
|
// src/version.ts
|
|
6577
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
6673
|
+
var LIBRARY_VERSION = "2.0.0-alpha.97";
|
|
6578
6674
|
|
|
6579
6675
|
// src/pricing-brain.ts
|
|
6580
6676
|
function isPricingRow(x) {
|
package/dist/index.mjs
CHANGED
|
@@ -7,7 +7,7 @@ import {
|
|
|
7
7
|
LIBRARY_VERSION,
|
|
8
8
|
createKeyHealthRoute,
|
|
9
9
|
keyFingerprint
|
|
10
|
-
} from "./chunk-
|
|
10
|
+
} from "./chunk-IEEROFZW.mjs";
|
|
11
11
|
import {
|
|
12
12
|
ARCHETYPE_FAMILY_FITS,
|
|
13
13
|
BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE,
|
|
@@ -63,7 +63,7 @@ import {
|
|
|
63
63
|
runAdvisor,
|
|
64
64
|
setTokenizer,
|
|
65
65
|
wilsonLowerBound
|
|
66
|
-
} from "./chunk-
|
|
66
|
+
} from "./chunk-AFYBZRQV.mjs";
|
|
67
67
|
import {
|
|
68
68
|
ABSOLUTE_FLOOR,
|
|
69
69
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
@@ -100,7 +100,7 @@ import {
|
|
|
100
100
|
policySetHas,
|
|
101
101
|
readBrainReadEnv,
|
|
102
102
|
resolveProviderKey
|
|
103
|
-
} from "./chunk-
|
|
103
|
+
} from "./chunk-77AHSX6Y.mjs";
|
|
104
104
|
import {
|
|
105
105
|
ALL_ARCHETYPES,
|
|
106
106
|
DIALECT_VERSION,
|
|
@@ -127,7 +127,7 @@ import {
|
|
|
127
127
|
registerProfiles,
|
|
128
128
|
resolveModelAlias,
|
|
129
129
|
tryGetProfile
|
|
130
|
-
} from "./chunk-
|
|
130
|
+
} from "./chunk-HHIXHL4O.mjs";
|
|
131
131
|
import {
|
|
132
132
|
emitAdvisoryFired,
|
|
133
133
|
emitCompileDone,
|
package/dist/key-health.js
CHANGED
|
@@ -28,7 +28,7 @@ __export(key_health_exports, {
|
|
|
28
28
|
module.exports = __toCommonJS(key_health_exports);
|
|
29
29
|
|
|
30
30
|
// src/version.ts
|
|
31
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
31
|
+
var LIBRARY_VERSION = "2.0.0-alpha.97";
|
|
32
32
|
|
|
33
33
|
// src/key-health.ts
|
|
34
34
|
var JSON_HEADERS = { "Content-Type": "application/json" };
|