@ggui-ai/ui-gen 0.13.0 → 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.
Files changed (62) hide show
  1. package/dist/adapters/generation-dispatch.d.ts +2 -2
  2. package/dist/adapters/generation-dispatch.js +190 -57
  3. package/dist/adapters/generation-dispatch.js.map +1 -1
  4. package/dist/adapters/index.js +5 -8
  5. package/dist/adapters/index.js.map +1 -1
  6. package/dist/advanced/index.d.ts +3 -3
  7. package/dist/advanced/index.js +190 -57
  8. package/dist/advanced/index.js.map +1 -1
  9. package/dist/blueprint-validator.js +2 -7
  10. package/dist/blueprint-validator.js.map +1 -1
  11. package/dist/boilerplate.js +4 -4
  12. package/dist/boilerplate.js.map +1 -1
  13. package/dist/check/index.d.ts +1 -1
  14. package/dist/check/index.js +55 -15
  15. package/dist/check/index.js.map +1 -1
  16. package/dist/{create-ui-generator-CF0EL0Li.d.ts → create-ui-generator-D6TKuv9V.d.ts} +2 -2
  17. package/dist/{dispatch-D75M-OfZ.d.ts → dispatch-BuZfTh9r.d.ts} +1 -1
  18. package/dist/evaluation/axis-checks/index.d.ts +3 -3
  19. package/dist/evaluation/axis-checks/index.js.map +1 -1
  20. package/dist/evaluation/axis-checks/registry.d.ts +1 -1
  21. package/dist/evaluation/index.d.ts +3 -3
  22. package/dist/evaluation/index.js +79 -16
  23. package/dist/evaluation/index.js.map +1 -1
  24. package/dist/harness/check/runtime-render/index.d.ts +1 -1
  25. package/dist/harness/check/runtime-render/index.js +2 -7
  26. package/dist/harness/check/runtime-render/index.js.map +1 -1
  27. package/dist/harness/index.d.ts +2 -2
  28. package/dist/harness/index.js +64 -19
  29. package/dist/harness/index.js.map +1 -1
  30. package/dist/harness/prompts.js +2 -0
  31. package/dist/harness/prompts.js.map +1 -1
  32. package/dist/harness/result-types.d.ts +1 -1
  33. package/dist/harness/runtime.d.ts +2 -2
  34. package/dist/harness/runtime.js +5 -5
  35. package/dist/harness/runtime.js.map +1 -1
  36. package/dist/harness/types-public.d.ts +1 -1
  37. package/dist/harness/types-public.js +4 -4
  38. package/dist/harness/types-public.js.map +1 -1
  39. package/dist/harness/validator-trace-sink.d.ts +1 -1
  40. package/dist/index.d.ts +3 -3
  41. package/dist/index.js +190 -57
  42. package/dist/index.js.map +1 -1
  43. package/dist/{llm-router-BEVtbAb9.d.ts → llm-router-D7fjBhtr.d.ts} +39 -0
  44. package/dist/patch.d.ts +1 -1
  45. package/dist/provider-adapter-contract.js +1017 -731
  46. package/dist/provider-adapter-contract.js.map +1 -1
  47. package/dist/{types-public-DIpjC-OA.d.ts → types-public-CkNWEov9.d.ts} +64 -1
  48. package/dist/validation/index.d.ts +1 -1
  49. package/dist/validation/index.js +1 -1
  50. package/dist/validation/index.js.map +1 -1
  51. package/dist/workflows.d.ts +1 -1
  52. package/package.json +21 -21
  53. package/src/adapters/provider-router.ts +9 -0
  54. package/src/boilerplate/system-prompt.ts +4 -4
  55. package/src/check/run-tier0.ts +109 -22
  56. package/src/evaluation/index.ts +4 -0
  57. package/src/evaluation/llm-evaluator.ts +52 -8
  58. package/src/evaluation/types-public.ts +79 -0
  59. package/src/evaluation/visual-evaluator.ts +27 -5
  60. package/src/harness/coding/run-eval-round.ts +32 -6
  61. package/src/harness/llm-router.ts +128 -2
  62. package/src/validation/primitives.ts +1 -1
@@ -8,6 +8,14 @@ interface AgentConfig {
8
8
  * Sampling temperature. When defined it is threaded into the provider
9
9
  * request; when undefined the provider default is used (unchanged
10
10
  * behavior for all existing callers).
11
+ *
12
+ * Capability is the ROUTER's: Anthropic Opus 4.7+ and the 5-family
13
+ * (`claude-opus-4-7*`, `claude-fable-5*`, `claude-opus-5*`,
14
+ * `claude-sonnet-5*`, incl. Bedrock ids) return HTTP 400 on a
15
+ * non-default `temperature`/`top_p`/`top_k`, so the Anthropic adapter
16
+ * STRIPS the value for those ids (one log line naming model + param)
17
+ * and reports the effective sampling on `LLMResponse.sampling`. Callers
18
+ * keep expressing intent; disclosure reads the applied value.
11
19
  */
12
20
  temperature?: number;
13
21
  /**
@@ -55,10 +63,31 @@ interface ProviderRetryInfo {
55
63
  /** One-line error summary (same shape as the existing `console.error` line). */
56
64
  readonly message: string;
57
65
  }
66
+ /**
67
+ * What sampling the provider ACTUALLY received for a text call — so a
68
+ * caller that asked for `temperature: 0` (a disclosed reproducibility
69
+ * property, e.g. the benchmark judge panel) can record the effective
70
+ * value instead of the requested one. `'provider-default'` means no
71
+ * temperature reached the provider: either the caller passed none, or
72
+ * the adapter stripped it because the model rejects sampling params
73
+ * (Anthropic Opus 4.7+ / 5-family → HTTP 400 on non-default
74
+ * `temperature`/`top_p`/`top_k`); `strippedReason` names the second case.
75
+ */
76
+ interface AppliedSampling {
77
+ readonly temperature: number | 'provider-default';
78
+ readonly strippedReason?: string;
79
+ }
58
80
  interface LLMResponse {
59
81
  text: string;
60
82
  inputTokens: number;
61
83
  outputTokens: number;
84
+ /**
85
+ * Effective sampling for this call (see {@link AppliedSampling}).
86
+ * Every provider's `callText` sets it; absent only on paths that
87
+ * predate the field — treat absence as unknown, never as "as
88
+ * requested".
89
+ */
90
+ sampling?: AppliedSampling;
62
91
  }
63
92
  interface LLMTool {
64
93
  name: string;
@@ -87,6 +116,16 @@ interface LLMToolCallResponse {
87
116
  toolCalls: LLMToolCall[];
88
117
  inputTokens: number;
89
118
  outputTokens: number;
119
+ /**
120
+ * The tool-choice mode the provider ACTUALLY received. Callers ask
121
+ * for `'required'` (Anthropic `tool_choice: any`); Claude Fable 5.1
122
+ * rejects forced tool choice with HTTP 400, so the Anthropic adapter
123
+ * downgrades to `'auto'` for that family and reports it here. A
124
+ * downgraded call may legitimately return zero tool calls — the
125
+ * evaluator's `criteriaCoverage` records that as `skipped`, never as
126
+ * a pass. Absent = predates the field (treat as unknown).
127
+ */
128
+ appliedToolChoice?: 'required' | 'auto';
90
129
  /**
91
130
  * Tokens read from / written to the prompt cache, when the provider
92
131
  * reports them. Optional because not every provider exposes
package/dist/patch.d.ts CHANGED
@@ -1,7 +1,7 @@
1
1
  import { PatchFn } from './harness/types-public.js';
2
2
  import '@ggui-ai/protocol';
3
3
  import './axes-CzLEMDeB.js';
4
- import './types-public-DIpjC-OA.js';
4
+ import './types-public-CkNWEov9.js';
5
5
  import './types-BOvHNG7K.js';
6
6
  import './llm.js';
7
7
  import './policy.js';