@adaptic/utils 0.0.1026 → 0.0.1027

package/dist/index.mjs

@@ -63275,58 +63275,59 @@ function routeSupports(route, needs) {
 var defaults$1 = {
 	_readme: "Applied to any provider whose published ceiling is unknown. Chosen to be comfortably below the slowest plausible published limit: being slower than necessary costs latency, while being faster than permitted costs 429s that the fallback chain will read as provider ill-health and use to open a circuit breaker. The asymmetry is what makes the conservative side the correct default.",
 	basis: "conservative-default",
-	requests_per_minute: 60,
-	max_concurrent: 4,
-	acquire_timeout_ms: 15000
+	requests_per_minute: 120,
+	max_concurrent: 12,
+	acquire_timeout_ms: 15000,
+	note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 };
 var providers$1 = {
 	anthropic: {
 		basis: "conservative-default",
-		requests_per_minute: 50,
-		max_concurrent: 4,
+		requests_per_minute: 120,
+		max_concurrent: 12,
 		acquire_timeout_ms: 15000,
 		source: null,
-		note: "Anthropic publishes tier-dependent limits; the account's tier is not recorded here. Transcribe the real ceiling from the console at W3-07 and set basis to published."
+		note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 	},
 	openai: {
 		basis: "conservative-default",
-		requests_per_minute: 60,
-		max_concurrent: 4,
+		requests_per_minute: 120,
+		max_concurrent: 12,
 		acquire_timeout_ms: 15000,
 		source: null,
-		note: "Tier-dependent. No alias routes here today; the entry exists so a revert to an OpenAI incumbent inherits a bounded client rather than an unbounded one."
+		note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 	},
 	deepseek: {
 		basis: "conservative-default",
-		requests_per_minute: 60,
-		max_concurrent: 4,
+		requests_per_minute: 120,
+		max_concurrent: 12,
 		acquire_timeout_ms: 30000,
 		source: null,
-		note: "Serves llm.extract, a batch-class alias, so a longer acquire timeout is appropriate: a batch caller can afford to queue where a hot-path caller cannot."
+		note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 	},
 	deepinfra: {
 		basis: "conservative-default",
-		requests_per_minute: 60,
-		max_concurrent: 4,
+		requests_per_minute: 240,
+		max_concurrent: 24,
 		acquire_timeout_ms: 15000,
 		source: null,
-		note: "Awaiting W3-02. Transcribe the published ceiling then."
+		note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 	},
 	fireworks: {
 		basis: "conservative-default",
-		requests_per_minute: 60,
-		max_concurrent: 4,
+		requests_per_minute: 120,
+		max_concurrent: 12,
 		acquire_timeout_ms: 15000,
 		source: null,
-		note: "The backlog calls out recording Fireworks' published RPM ceiling specifically (W3-03 -> W4-04). Do that at onboarding."
+		note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 	},
 	zai: {
 		basis: "conservative-default",
-		requests_per_minute: 60,
-		max_concurrent: 4,
+		requests_per_minute: 120,
+		max_concurrent: 12,
 		acquire_timeout_ms: 15000,
 		source: null,
-		note: "Awaiting W3-04."
+		note: "Raised 2026-09-15 after the 4-concurrent ceiling was measured starving the live equity decision path: the alias chain reported 'exhausted its fallback chain' with every leg skipped by this guard (deepinfra primary+secondary and the anthropic incumbent), 173 of 473 signal-coordination calls failed (36.6%), and decisions were lost outright. Corroborating evidence at the time: ZERO 429s observed on any provider, the engine's own global fan-out gate permits 100 concurrent with 17 active, and 877 signal-analysis calls had been admitted to that gate. Concurrency and RPM are raised TOGETHER because they bind in series - lifting max_concurrent alone would only move the bottleneck to the token bucket. STILL conservative-default, NOT published: no provider console was read for these numbers, so they remain a deliberate under-estimate of an unknown ceiling. Transcribe the real tier limits (W3-07) and set basis to published."
 	},
 	groq: {
 		basis: "conservative-default",