@warmdrift/kgauto-compiler 2.0.0-alpha.8 → 2.0.0-alpha.9

package/dist/profiles.js

@@ -79,7 +79,24 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
     weaknesses: ["cost", "latency"],
-    notes: "Frontier (2026-05). Step-change improvement over 4.6 in agentic coding. Adaptive thinking only \u2014 no extended-thinking toggle. 1M context, 128k max output."
+    notes: "Frontier (2026-05). Step-change improvement over 4.6 in agentic coding. Adaptive thinking only \u2014 no extended-thinking toggle. 1M context, 128k max output.",
+    // Frontier perf. Drops on archetypes where parallel-tool throughput
+    // (hunt) or low-budget cost-sensitivity (classify/summarize) matters
+    // more than reasoning depth.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // strong but Flash dominates parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype; cost-out of frontier
+      classify: 8
+      // overkill; brain-validated cheaper models cover this
+    }
   },
   {
     id: "claude-opus-4-6",
@@ -111,7 +128,20 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "long_context", "reliable_tool_use", "structured_output", "extended_thinking"],
     weaknesses: ["cost", "latency"],
-    notes: "Predecessor to 4.7. Still current in Anthropic legacy table. Same pricing as 4.7 \u2014 choose 4.7 unless you need extended-thinking budget control (4.7 is adaptive-only)."
+    notes: "Predecessor to 4.7. Still current in Anthropic legacy table. Same pricing as 4.7 \u2014 choose 4.7 unless you need extended-thinking budget control (4.7 is adaptive-only).",
+    // One notch below 4.7 across the board — extended-thinking edge does
+    // not flip any archetype ranking. Legacy: chains should prefer 4.7.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      summarize: 8,
+      classify: 8
+    }
   },
   {
     id: "claude-sonnet-4-6",
@@ -135,7 +165,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
     weaknesses: [],
-    notes: "Workhorse. Best price/quality for most multi-turn agentic work. 1M context, 64k max output."
+    notes: "Workhorse. Best price/quality for most multi-turn agentic work. 1M context, 64k max output.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   },
   {
     id: "claude-haiku-4-5",
@@ -165,7 +211,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["speed", "cost", "classification", "cache_friendly", "extended_thinking"],
     weaknesses: ["complex_reasoning", "large_tool_sets"],
-    notes: "Cheapest Anthropic. Great for classify, summarize, ask shapes. 200k context, 64k max output. API alias `claude-haiku-4-5` resolves to dated snapshot `claude-haiku-4-5-20251001`."
+    notes: "Cheapest Anthropic. Great for classify, summarize, ask shapes. 200k context, 64k max output. API alias `claude-haiku-4-5` resolves to dated snapshot `claude-haiku-4-5-20251001`.",
+    // Tier 1 cross-provider anchor for short-output chains (classify/
+    // summarize/extract/transform). Falls off on plan/critique where
+    // reasoning depth matters; competes with Pro on cost+latency.
+    archetypePerf: {
+      classify: 8,
+      summarize: 8,
+      ask: 7,
+      transform: 7,
+      extract: 7,
+      hunt: 6,
+      // tool reliability drops at 16 — cliff guard fires
+      generate: 6,
+      plan: 5,
+      critique: 4
+      // reasoning depth gap vs Sonnet/Opus
+    }
   },
   // ── Google ──
   {
@@ -243,7 +305,131 @@ var PROFILES_RAW = [
     ],
     strengths: ["speed", "volume", "classification", "1m_context", "cost"],
     weaknesses: ["complex_schemas", "large_tool_sets", "high_context_quality"],
-    notes: "Fast and cheap with 1M context. Quality cliffs at 8K context and 20 tools \u2014 guard with cliffs."
+    notes: "Fast and cheap with 1M context. Quality cliffs at 8K context and 20 tools \u2014 guard with cliffs.",
+    // Master plan §6.2 anchor. Tier 0 for hunt (parallel tool throughput
+    // 15-75 calls/step beats Sonnet — L-040), summarize, classify.
+    archetypePerf: {
+      hunt: 9,
+      // L-040: parallel tool throughput 15-75/step
+      classify: 7,
+      // brain-validated, 218 rows
+      summarize: 7,
+      // brain-validated; cliff strips tools when present
+      transform: 7,
+      ask: 7,
+      generate: 6,
+      plan: 5,
+      extract: 6,
+      // alpha.8 MAX_TOKENS history on structured output
+      critique: 4
+      // reasoning shallower than Sonnet/Opus
+    }
+  },
+  {
+    // ── Gemini 2.5 Flash-Lite ──
+    // Onboarded 2026-05-13 (s22) after the model-release watcher surfaced
+    // it as a UNREGISTERED + NEW candidate. Released by Google July 2025,
+    // stable. Positioned BELOW Flash on the cost/perf frontier:
+    //   input  $0.10/M (Flash $0.30/M)   — 3× cheaper
+    //   output $0.40/M (Flash $2.50/M)   — 6× cheaper
+    //   cache  $0.01/M                    — 1/10 of input (vs Flash 0.25 discount)
+    // Cliffs are HYPOTHESIZED from Flash's known failure modes — Flash-Lite
+    // is a smaller sibling, so we inherit Flash's cliffs at equal-or-tighter
+    // thresholds. The brain will validate/relax these as evidence accumulates
+    // per (archetype, model) tuple. Currently ZERO brain rows for this model.
+    id: "gemini-2.5-flash-lite",
+    verifiedAgainstDocs: "2026-05-13",
+    provider: "google",
+    status: "current",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from Flash: quality degrades above ~8K. Smaller model \u2014 likely degrades faster. Re-tune from brain after n\u226520."
+      },
+      {
+        metric: "tool_count",
+        threshold: 10,
+        action: "drop_to_top_relevant",
+        reason: "Conservative: Flash drops at 20, Flash-Lite is smaller \u2014 assume tighter ceiling until brain proves otherwise."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking enabled per Google API (thinking: true). Same drain risk as Flash \u2014 thinking tokens consume maxOutputTokens."
+      },
+      {
+        // Strong prior: Flash hit 5/5 empty rate on summarize+tools (s11
+        // trust artifact, kgauto commit 3872832). Flash-Lite shares the
+        // same architectural family — almost certainly inherits this cliff.
+        // Ship the guard preemptively; brain telemetry confirms or relaxes.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from Flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on Flash-Lite specifically."
+      }
+    ],
+    costInputPer1m: 0.1,
+    costOutputPer1m: 0.4,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× (vs Flash 4×) — Google's spec is $0.01/M cache vs
+      // $0.10/M input. Material for repeat-prompt workloads (classify shape).
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known on Flash family: empty after tool result \u2014 retry with thinking off."
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off."
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target."
+      }
+    ],
+    strengths: ["lowest_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: "Bottom-frontier anchor on cost: $0.10/$0.40 per 1M tokens, 1M context, 65K max output. Released July 2025 (stable). Positioned for classify / summarize / transform archetypes where quality bar is forgiving. Cliffs inherited from Flash at equal-or-tighter thresholds \u2014 re-tune per (archetype) once brain has n\u226520 rows. Alpha.8 contract layer handles MAX_TOKENS-on-structured-output via fallback chain, so structuredOutput=native is safe to declare even though Flash had alpha.8 history. Cache discount in spec: $0.01/M = 1/10 of input (richer than Flash 25%) \u2014 meaningful for repeat-prompt workloads.",
+    // Tier 3 emergency floor for summarize/classify chains. ZERO brain
+    // rows — all values are starter hypotheses anchored to "smaller
+    // sibling of Flash, at-or-below Flash perf on every archetype." The
+    // first 50 brain rows per archetype will validate or relax these.
+    archetypePerf: {
+      classify: 6,
+      // starter hypothesis — verify (Flash is 7, lite likely ≤)
+      summarize: 6,
+      // starter hypothesis — verify; cliff strips tools
+      transform: 6,
+      // starter hypothesis — verify
+      ask: 5,
+      hunt: 5,
+      generate: 4,
+      extract: 4,
+      plan: 3,
+      critique: 3
+    }
   },
   {
     id: "gemini-2.5-pro",
@@ -279,7 +465,21 @@ var PROFILES_RAW = [
       }
     ],
     strengths: ["reasoning", "1m_context", "structured_output", "tool_use"],
-    weaknesses: ["pricing_above_200k"]
+    weaknesses: ["pricing_above_200k"],
+    // Master plan §3.3 anchor: tier-2 cross-provider in almost every chain.
+    // Sits on the frontier at perf-9 — close to Sonnet but cheaper input.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      ask: 8,
+      generate: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 8,
+      // tier 1 cross-provider for hunt chain
+      summarize: 7,
+      classify: 7
+    }
   },
   {
     id: "gemini-3.1-pro-preview",
@@ -317,7 +517,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "1m_context", "agentic_coding", "structured_output", "tool_use"],
     weaknesses: ["cost", "preview_status", "pricing_above_200k"],
-    notes: "Frontier Gemini (preview, 2026-Q2). Step-change agentic coding per Google. Cache discount 10\xD7 (vs 4\xD7 for 2.5 Pro). Use status=preview to flag rollback path until GA."
+    notes: "Frontier Gemini (preview, 2026-Q2). Step-change agentic coding per Google. Cache discount 10\xD7 (vs 4\xD7 for 2.5 Pro). Use status=preview to flag rollback path until GA.",
+    // Frontier-Gemini preview — bumped one notch over 2.5 Pro on agentic
+    // coding / reasoning per Google's release notes. Preview status:
+    // chains should stay on 2.5 Pro until GA. Starter hypothesis.
+    archetypePerf: {
+      critique: 10,
+      // Google claims step-change on reasoning
+      plan: 10,
+      ask: 9,
+      generate: 9,
+      extract: 9,
+      transform: 8,
+      hunt: 9,
+      // step-change agentic per Google
+      summarize: 8,
+      classify: 7
+    }
   },
   // ── DeepSeek ──
   // 2026-05-08 audit (L-073): DeepSeek's `deepseek-chat` was silently aliased
@@ -357,7 +573,24 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "large_tool_sets"],
-    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."
+    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.",
+    // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
+    // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
+    // Falls off on hunt (sequential tools — L-040) and reasoning depth.
+    archetypePerf: {
+      classify: 7,
+      // brain-validated, 169 rows
+      summarize: 7,
+      // archetype-tolerant, no brain evidence yet
+      ask: 6,
+      transform: 6,
+      generate: 5,
+      plan: 5,
+      extract: 5,
+      critique: 4,
+      hunt: 4
+      // sequential tool calls only — L-040
+    }
   },
   {
     id: "deepseek-v4-pro",
@@ -393,7 +626,22 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "large_tool_sets"],
-    notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking."
+    notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
+    // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
+    // bumped one notch over V4-Flash; same parallel-tool ceiling.
+    archetypePerf: {
+      plan: 7,
+      // §3.3 tier 3 for plan
+      critique: 6,
+      ask: 7,
+      generate: 6,
+      classify: 7,
+      summarize: 7,
+      extract: 6,
+      transform: 6,
+      hunt: 4
+      // sequential tools — same as V4-Flash
+    }
   }
 ];
 var ALIASES = {

package/dist/profiles.mjs

@@ -4,7 +4,7 @@ import {
   getProfile,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-MBEI5UOM.mjs";
+} from "./chunk-3KVKELZN.mjs";
 export {
   ALIASES,
   allProfiles,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.8",
+  "version": "2.0.0-alpha.9",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",