diff --git a/.changeset/rn-renderer-bundler-fields.md b/.changeset/rn-renderer-bundler-fields.md new file mode 100644 index 0000000..6e3d515 --- /dev/null +++ b/.changeset/rn-renderer-bundler-fields.md @@ -0,0 +1,7 @@ +--- +"@mobile-reality/mdma-renderer-react-native": patch +"@mobile-reality/mdma-runtime": patch +"@mobile-reality/mdma-spec": patch +--- + +Add `main`, `module`, and `react-native` entry fields (and a `default` export condition) alongside the existing `exports` map. These packages previously exposed only an `exports` map, so bundlers that don't opt into package `exports` resolution — notably Metro/Snackager (Expo Snack) — couldn't find an entry point and failed with "Can't resolve ''". The added fields make the packages resolvable in any React Native / Metro bundler without enabling `unstable_enablePackageExports`. Fully additive and backwards-compatible. diff --git a/demo-native/App.tsx b/demo-native/App.tsx index 5be321f..d892144 100644 --- a/demo-native/App.tsx +++ b/demo-native/App.tsx @@ -88,7 +88,9 @@ export function App() { // Mutate the blocks of a specific assistant message. const editBlocks = useCallback((botId: number, fn: (blocks: Block[]) => Block[]) => { setMessages((prev) => - prev.map((m) => (m.id === botId && m.role === 'assistant' ? { ...m, blocks: fn(m.blocks) } : m)), + prev.map((m) => + m.id === botId && m.role === 'assistant' ? { ...m, blocks: fn(m.blocks) } : m, + ), ); }, []); @@ -191,7 +193,10 @@ export function App() { // user stopped — leave partial output as-is } else { const message = e instanceof Error ? e.message : String(e); - editBlocks(botId, (blocks) => [...blocks, { id: nextBlockId(), kind: 'text', text: `⚠️ ${message}` }]); + editBlocks(botId, (blocks) => [ + ...blocks, + { id: nextBlockId(), kind: 'text', text: `⚠️ ${message}` }, + ]); } } finally { abortRef.current = null; @@ -224,14 +229,22 @@ export function App() { MDMA Agent - {live ? `${providerLabel} · ${model}` : `No key · set ${PROVIDERS.find((p) => p.id === provider)?.envVar}`} + {live + ? `${providerLabel} · ${model}` + : `No key · set ${PROVIDERS.find((p) => p.id === provider)?.envVar}`} - setShowSettings((v) => !v)} style={[styles.iconBtn, { borderColor: c.border }]}> + setShowSettings((v) => !v)} + style={[styles.iconBtn, { borderColor: c.border }]} + > ⚙️ - setTheme(dark ? 'light' : 'dark')} style={[styles.iconBtn, { borderColor: c.border }]}> + setTheme(dark ? 'light' : 'dark')} + style={[styles.iconBtn, { borderColor: c.border }]} + > {dark ? '☀️' : '🌙'} @@ -242,7 +255,9 @@ export function App() { {/* Settings — provider + model. Keys come from .env (EXPO_PUBLIC_*). */} {showSettings ? ( - + Provider {PROVIDERS.map((p) => { @@ -252,7 +267,11 @@ export function App() { pickProvider(p.id)} - style={[styles.chip, { borderColor: active ? c.accent : c.border }, active && { backgroundColor: c.accent }]} + style={[ + styles.chip, + { borderColor: active ? c.accent : c.border }, + active && { backgroundColor: c.accent }, + ]} > {hasKey ? '● ' : '○ '} @@ -268,11 +287,14 @@ export function App() { onChangeText={setModel} autoCapitalize="none" autoCorrect={false} - style={[styles.settingsInput, { color: c.text, borderColor: c.border, backgroundColor: c.pageBg }]} + style={[ + styles.settingsInput, + { color: c.text, borderColor: c.border, backgroundColor: c.pageBg }, + ]} /> - ● = key present. Set keys in demo-native/.env (EXPO_PUBLIC_…_API_KEY) and rebuild — they are - never entered in the app. + ● = key present. Set keys in demo-native/.env (EXPO_PUBLIC_…_API_KEY) and rebuild — they + are never entered in the app. ) : null} @@ -299,7 +321,10 @@ export function App() { {m.text} ) : ( - + {m.blocks.length === 0 ? ( @@ -312,9 +337,23 @@ export function App() { {b.text} ) : b.doc ? ( - + ) : ( - + {b.text ?? 'Generating document…'} @@ -356,7 +395,10 @@ export function App() { editable={!busy} onSubmitEditing={() => send(input)} returnKeyType="send" - style={[styles.textInput, { color: c.text, borderColor: c.border, backgroundColor: c.pageBg }]} + style={[ + styles.textInput, + { color: c.text, borderColor: c.border, backgroundColor: c.pageBg }, + ]} /> {busy ? ( @@ -415,8 +457,19 @@ const styles = StyleSheet.create({ row: { flexDirection: 'row', flexWrap: 'wrap', gap: 8 }, iconBtn: { paddingVertical: 6, paddingHorizontal: 10, borderRadius: 8, borderWidth: 1 }, settings: { padding: 12, gap: 6, borderBottomWidth: StyleSheet.hairlineWidth }, - settingsLabel: { fontSize: 11, fontWeight: '700', textTransform: 'uppercase', letterSpacing: 0.4 }, - settingsInput: { borderWidth: 1, borderRadius: 8, paddingHorizontal: 10, paddingVertical: 8, fontSize: 14 }, + settingsLabel: { + fontSize: 11, + fontWeight: '700', + textTransform: 'uppercase', + letterSpacing: 0.4, + }, + settingsInput: { + borderWidth: 1, + borderRadius: 8, + paddingHorizontal: 10, + paddingVertical: 8, + fontSize: 14, + }, settingsHint: { fontSize: 12, marginTop: 2, lineHeight: 17 }, empty: { paddingVertical: 48, paddingHorizontal: 12, gap: 8 }, emptyTitle: { fontSize: 16, fontWeight: '700', textAlign: 'center' }, diff --git a/demo-native/llm.ts b/demo-native/llm.ts index 998ee35..b753115 100644 --- a/demo-native/llm.ts +++ b/demo-native/llm.ts @@ -1,11 +1,4 @@ -import { - streamText, - generateText, - tool, - stepCountIs, - jsonSchema, - type ModelMessage, -} from 'ai'; +import { streamText, generateText, tool, stepCountIs, jsonSchema, type ModelMessage } from 'ai'; import { createAnthropic } from '@ai-sdk/anthropic'; import { createOpenAI } from '@ai-sdk/openai'; import { fetch as expoFetch } from 'expo/fetch'; diff --git a/demo/src/agent/AgentMessage.tsx b/demo/src/agent/AgentMessage.tsx index 7432ef7..c7031b2 100644 --- a/demo/src/agent/AgentMessage.tsx +++ b/demo/src/agent/AgentMessage.tsx @@ -298,8 +298,8 @@ export const AgentMessage = memo(function AgentMessage({ const { blocks } = turn as AssistantTurn; - const hasContent = blocks.some( - (b) => (b.type === 'text' || b.type === 'thinking' ? b.content : (b as ToolUseBlock).document), + const hasContent = blocks.some((b) => + b.type === 'text' || b.type === 'thinking' ? b.content : (b as ToolUseBlock).document, ); return ( diff --git a/demo/src/agent/AgentSettings.tsx b/demo/src/agent/AgentSettings.tsx index 06ff712..f5a5df1 100644 --- a/demo/src/agent/AgentSettings.tsx +++ b/demo/src/agent/AgentSettings.tsx @@ -221,8 +221,8 @@ export const AgentSettings = memo(function AgentSettings({ config, onUpdate }: A

The entire agent runs on your self-hosted MDMA model endpoint (OpenAI-compatible, tool-calling enabled) — no third-party model is called. Enter the - deployed model URL above; leave it blank to use the default. The{' '} - /v1 suffix is added automatically. + deployed model URL above; leave it blank to use the default. The /v1{' '} + suffix is added automatically.

)}

diff --git a/demo/src/agent/anthropic-client.ts b/demo/src/agent/anthropic-client.ts index e8ec1cf..ad8b657 100644 --- a/demo/src/agent/anthropic-client.ts +++ b/demo/src/agent/anthropic-client.ts @@ -35,8 +35,7 @@ export interface AnthropicConfig { * VITE_OWN_MODEL_BASE_URL. */ export const OWN_MODEL_DEFAULT_BASE_URL = - import.meta.env.VITE_OWN_MODEL_BASE_URL ?? - 'https://REDACTED.modal.run/v1'; + import.meta.env.VITE_OWN_MODEL_BASE_URL ?? 'https://REDACTED.modal.run/v1'; export interface ToolDefinition { name: string; diff --git a/demo/src/agent/openai-agent-client.ts b/demo/src/agent/openai-agent-client.ts index 5dbf5c9..c59bdd4 100644 --- a/demo/src/agent/openai-agent-client.ts +++ b/demo/src/agent/openai-agent-client.ts @@ -289,8 +289,7 @@ export async function* streamOpenAIAgentMessages( reader.cancel().catch(() => {}); yield { type: 'stream_error', - message: - 'The model got stuck repeating itself and was stopped. Please try again.', + message: 'The model got stuck repeating itself and was stopped. Please try again.', }; return; } diff --git a/demo/src/agent/use-agent.ts b/demo/src/agent/use-agent.ts index 3362a33..42a34e9 100644 --- a/demo/src/agent/use-agent.ts +++ b/demo/src/agent/use-agent.ts @@ -518,7 +518,10 @@ async function chatOnce( const baseUrl = getBaseUrlForProvider(config); const response = await fetch(`${baseUrl}/chat/completions`, { method: 'POST', - headers: { 'content-type': 'application/json', authorization: `Bearer ${getApiKeyForProvider(config)}` }, + headers: { + 'content-type': 'application/json', + authorization: `Bearer ${getApiKeyForProvider(config)}`, + }, body: JSON.stringify({ model: isOwn ? OWN_MODEL_NAME : config.model, messages: [ diff --git a/demo/src/docs/sections/IntegrationAgui.tsx b/demo/src/docs/sections/IntegrationAgui.tsx index 138f1de..f104bfa 100644 --- a/demo/src/docs/sections/IntegrationAgui.tsx +++ b/demo/src/docs/sections/IntegrationAgui.tsx @@ -10,8 +10,8 @@ export function IntegrationAgui() { AG-UI {' '} agent and route the user's decisions back into the run. AG-UI is the{' '} - transport (suspend/resume via its interrupt primitive); MDMA is - the payload (validated, audited, PII-aware components).{' '} + transport (suspend/resume via its interrupt primitive); MDMA + is the payload (validated, audited, PII-aware components).{' '} @mobile-reality/mdma-agui is the seam between them — a community-maintained adapter, not a framework integration.

diff --git a/demo/src/docs/sections/PromptMatrix.tsx b/demo/src/docs/sections/PromptMatrix.tsx index 841b490..e6e43ce 100644 --- a/demo/src/docs/sections/PromptMatrix.tsx +++ b/demo/src/docs/sections/PromptMatrix.tsx @@ -157,14 +157,12 @@ export function PromptMatrix() { variants and Grok 4.3, the fixer would otherwise see visible "Thinking: **Topic**" prose prepended to every response. The eval config sets{' '} passthrough.reasoning.exclude: true (and the demo's{' '} - usePreviewValidation does the same per-provider) to strip reasoning tokens - from the response body at the API layer rather than the prompt layer. + usePreviewValidation does the same per-provider) to strip reasoning tokens from + the response body at the API layer rather than the prompt layer.

In Progress

-

- The following prompt still ships without model-specific variants and is on the roadmap: -

+

The following prompt still ships without model-specific variants and is on the roadmap:

{[ { diff --git a/demo/src/docs/sections/Usage.tsx b/demo/src/docs/sections/Usage.tsx index 4da6181..fe8c44a 100644 --- a/demo/src/docs/sections/Usage.tsx +++ b/demo/src/docs/sections/Usage.tsx @@ -55,10 +55,13 @@ store.dispatch({ symmetric with getState().

- Hydration overlays the AST defaults without emitting audit events or marking fields{' '} - touched, so a restore never looks like fresh user activity in the - tamper-evident log. It applies only to freshly-created components, so a later streamed - re-parse never clobbers an in-flight edit. + Hydration overlays the AST defaults{' '} + + without emitting audit events or marking fields touched + + , so a restore never looks like fresh user activity in the tamper-evident log. It applies + only to freshly-created components, so a later streamed re-parse never clobbers an in-flight + edit.

{`// 1. Persist — snapshot each component's values on the way out const snapshot = Object.fromEntries( @@ -87,8 +90,12 @@ const store = createDocumentStore(ast, { )}

- Using the AG-UI adapter? Pass the same map to <MdmaAgentView initialState={'{…}'} - /> — each replayed message hydrates only the component ids it contains. + Using the AG-UI adapter? Pass the same map to{' '} + + <MdmaAgentView initialState={'{…}'} + /> + {' '} + — each replayed message hydrates only the component ids it contains.

In a Chat

diff --git a/demo/src/styles.css b/demo/src/styles.css index e4181ea..1cb7ca8 100644 --- a/demo/src/styles.css +++ b/demo/src/styles.css @@ -1636,7 +1636,7 @@ body { gap: 10px; } -.ai-setting--toggle input[type='checkbox'] { +.ai-setting--toggle input[type="checkbox"] { appearance: none; -webkit-appearance: none; flex: 0 0 auto; @@ -1651,8 +1651,8 @@ body { transition: background 0.15s; } -.ai-setting--toggle input[type='checkbox']::after { - content: ''; +.ai-setting--toggle input[type="checkbox"]::after { + content: ""; position: absolute; top: 2px; left: 2px; @@ -1663,11 +1663,11 @@ body { transition: transform 0.15s; } -.ai-setting--toggle input[type='checkbox']:checked { +.ai-setting--toggle input[type="checkbox"]:checked { background: #6c5ce7; } -.ai-setting--toggle input[type='checkbox']:checked::after { +.ai-setting--toggle input[type="checkbox"]:checked::after { transform: translateX(16px); } @@ -1835,9 +1835,7 @@ body { border-radius: 5px; padding: 2px 7px; cursor: pointer; - transition: - background 0.15s, - color 0.15s; + transition: background 0.15s, color 0.15s; } .agent-raw-toggle:hover { diff --git a/evals/own-model/results-author.json b/evals/own-model/results-author.json index 5318acc..aa01a9d 100644 --- a/evals/own-model/results-author.json +++ b/evals/own-model/results-author.json @@ -77,9 +77,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -91,9 +89,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -175,18 +171,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -248,9 +240,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -262,9 +252,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -346,18 +334,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -419,9 +403,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -433,9 +415,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } } @@ -496,18 +476,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } ], @@ -558,9 +534,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -581,9 +555,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } } } @@ -644,9 +616,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { @@ -657,9 +627,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } } ], @@ -710,9 +678,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -724,9 +690,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -796,18 +760,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -862,9 +822,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -876,9 +834,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table" - ] + "allowed": ["table"] } } }, @@ -948,18 +904,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table" - ] + "allowed": ["table"] } }, { @@ -1014,9 +966,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1028,9 +978,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart" - ] + "allowed": ["chart"] } } }, @@ -1100,18 +1048,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart" - ] + "allowed": ["chart"] } }, { @@ -1166,9 +1110,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1180,9 +1122,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "approval-gate" - ] + "allowed": ["approval-gate"] } } } @@ -1243,18 +1183,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "approval-gate" - ] + "allowed": ["approval-gate"] } } ], @@ -1305,9 +1241,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1319,9 +1253,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -1403,18 +1335,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -1476,9 +1404,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1499,10 +1425,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -1575,9 +1498,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { @@ -1588,10 +1509,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -1649,9 +1567,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1663,9 +1579,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } }, @@ -1738,18 +1652,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -1807,9 +1717,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1821,9 +1729,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } }, @@ -1896,18 +1802,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -1965,9 +1867,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1979,9 +1879,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } }, @@ -2054,18 +1952,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -2123,9 +2017,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2137,9 +2029,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart" - ] + "allowed": ["chart"] } } }, @@ -2209,18 +2099,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart" - ] + "allowed": ["chart"] } }, { @@ -2275,9 +2161,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2289,9 +2173,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } } }, @@ -2361,18 +2243,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } }, { @@ -2427,9 +2305,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2441,10 +2317,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -2517,19 +2390,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2587,9 +2455,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2601,9 +2467,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table" - ] + "allowed": ["table"] } } }, @@ -2673,18 +2537,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table" - ] + "allowed": ["table"] } }, { @@ -2739,9 +2599,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2753,9 +2611,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -2837,18 +2693,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2910,9 +2762,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2924,9 +2774,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -3011,18 +2859,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -3087,9 +2931,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -3101,10 +2943,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "tasklist", - "button" - ] + "allowed": ["tasklist", "button"] } } } @@ -3165,19 +3004,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "tasklist", - "button" - ] + "allowed": ["tasklist", "button"] } } ], @@ -3228,9 +3062,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -3242,10 +3074,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } } @@ -3306,19 +3135,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } ], @@ -3369,9 +3193,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -3383,9 +3205,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -3464,18 +3284,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -3534,9 +3350,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -3548,9 +3362,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "approval-gate" - ] + "allowed": ["approval-gate"] } } } @@ -3611,18 +3423,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "approval-gate" - ] + "allowed": ["approval-gate"] } } ], @@ -3673,9 +3481,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -3687,9 +3493,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -3771,18 +3575,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -3844,9 +3644,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -3858,10 +3656,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "table" - ] + "allowed": ["chart", "table"] } } }, @@ -3949,19 +3744,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "table" - ] + "allowed": ["chart", "table"] } }, { @@ -4024,9 +3814,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -4038,9 +3826,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -4122,18 +3908,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4195,9 +3977,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -4209,9 +3989,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -4293,18 +4071,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4366,9 +4140,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -4380,9 +4152,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -4461,18 +4231,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4548,9 +4314,7 @@ "config": { "tags": {}, "description": "MDMA Author Eval (DSL) — own model", - "prompts": [ - "file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-author.mjs" - ], + "prompts": ["file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-author.mjs"], "providers": [ { "id": "openai:chat:mdma-26b", @@ -4578,9 +4342,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4606,9 +4368,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4634,9 +4394,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } ] @@ -4655,9 +4413,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } } ] @@ -4672,9 +4428,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4693,9 +4447,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table" - ] + "allowed": ["table"] } }, { @@ -4714,9 +4466,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart" - ] + "allowed": ["chart"] } }, { @@ -4735,9 +4485,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "approval-gate" - ] + "allowed": ["approval-gate"] } } ] @@ -4752,9 +4500,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4784,10 +4530,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -4809,9 +4552,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -4833,9 +4574,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -4857,9 +4596,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -4881,9 +4618,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart" - ] + "allowed": ["chart"] } }, { @@ -4902,9 +4637,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } }, { @@ -4923,10 +4656,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -4948,9 +4678,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table" - ] + "allowed": ["table"] } }, { @@ -4969,9 +4697,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -4997,9 +4723,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -5028,10 +4752,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "tasklist", - "button" - ] + "allowed": ["tasklist", "button"] } } ] @@ -5046,10 +4767,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } ] @@ -5064,9 +4782,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -5089,9 +4805,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "approval-gate" - ] + "allowed": ["approval-gate"] } } ] @@ -5106,9 +4820,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -5134,10 +4846,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "table" - ] + "allowed": ["chart", "table"] } }, { @@ -5164,9 +4873,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -5192,9 +4899,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -5220,9 +4925,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -5244,9 +4947,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5254,9 +4955,7 @@ "options": {}, "metadata": {} }, - "outputPath": [ - "own-model/results-author.json" - ], + "outputPath": ["own-model/results-author.json"], "extensions": [], "metadata": {}, "evaluateOptions": {} @@ -5270,4 +4969,4 @@ "exportedAt": "2026-06-30T12:37:51.375Z", "evaluationCreatedAt": "2026-06-30T12:37:01.882Z" } -} \ No newline at end of file +} diff --git a/evals/own-model/results-custom.json b/evals/own-model/results-custom.json index 875c8de..19b9dd4 100644 --- a/evals/own-model/results-custom.json +++ b/evals/own-model/results-custom.json @@ -77,9 +77,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -91,9 +89,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -197,18 +193,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -282,9 +274,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -296,9 +286,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -402,18 +390,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -487,9 +471,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -501,10 +483,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "chart" - ] + "allowed": ["form", "chart"] } } }, @@ -605,19 +584,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "chart" - ] + "allowed": ["form", "chart"] } }, { @@ -688,9 +662,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -702,9 +674,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -787,18 +757,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -861,9 +827,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -875,9 +839,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -972,18 +934,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -1053,9 +1011,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1067,10 +1023,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table", - "chart" - ] + "allowed": ["table", "chart"] } } }, @@ -1168,19 +1121,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table", - "chart" - ] + "allowed": ["table", "chart"] } }, { @@ -1248,9 +1196,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1262,9 +1208,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -1359,18 +1303,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -1440,9 +1380,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1454,9 +1392,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } } }, @@ -1548,18 +1484,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -1626,9 +1558,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1640,9 +1570,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -1737,18 +1665,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -1818,9 +1742,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1832,9 +1754,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -1926,18 +1846,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2004,9 +1920,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2018,9 +1932,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -2127,18 +2039,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2215,9 +2123,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2229,9 +2135,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } } }, @@ -2329,18 +2233,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } }, { @@ -2408,9 +2308,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2422,9 +2320,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } } }, @@ -2504,18 +2400,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2592,9 +2484,7 @@ "config": { "tags": {}, "description": "MDMA Author + Custom System Prompt Eval — own model", - "prompts": [ - "file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-custom.mjs" - ], + "prompts": ["file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-custom.mjs"], "providers": [ { "id": "openai:chat:mdma-26b", @@ -2623,9 +2513,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2663,9 +2551,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2703,10 +2589,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "chart" - ] + "allowed": ["form", "chart"] } }, { @@ -2741,9 +2624,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2770,9 +2651,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2806,10 +2685,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table", - "chart" - ] + "allowed": ["table", "chart"] } }, { @@ -2841,9 +2717,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2877,9 +2751,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "callout" - ] + "allowed": ["callout"] } }, { @@ -2910,9 +2782,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2946,9 +2816,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -2979,9 +2847,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -3022,9 +2888,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "button" - ] + "allowed": ["button"] } }, { @@ -3056,9 +2920,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form" - ] + "allowed": ["form"] } }, { @@ -3080,9 +2942,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3090,9 +2950,7 @@ "options": {}, "metadata": {} }, - "outputPath": [ - "own-model/results-custom.json" - ], + "outputPath": ["own-model/results-custom.json"], "extensions": [], "metadata": {}, "evaluateOptions": {} @@ -3106,4 +2964,4 @@ "exportedAt": "2026-06-30T12:35:51.175Z", "evaluationCreatedAt": "2026-06-30T12:35:16.976Z" } -} \ No newline at end of file +} diff --git a/evals/own-model/results-fixer.json b/evals/own-model/results-fixer.json index e77e73f..fad86b5 100644 --- a/evals/own-model/results-fixer.json +++ b/evals/own-model/results-fixer.json @@ -99,9 +99,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: button\nid: action-btn\nvariant: primary\n", - "hasFields": [ - "text" - ] + "hasFields": ["text"] } } } @@ -174,9 +172,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: button\nid: action-btn\nvariant: primary\n", - "hasFields": [ - "text" - ] + "hasFields": ["text"] } } ], @@ -249,9 +245,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: callout\nid: status-notice\nvariant: info\ntitle: System Status\n", - "hasFields": [ - "content" - ] + "hasFields": ["content"] } } } @@ -324,9 +318,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: callout\nid: status-notice\nvariant: info\ntitle: System Status\n", - "hasFields": [ - "content" - ] + "hasFields": ["content"] } } ], @@ -565,10 +557,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: callout\nid: welcome-callout\nvariant: info\n", - "hasFields": [ - "title", - "content" - ] + "hasFields": ["title", "content"] } } } @@ -645,10 +634,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: callout\nid: welcome-callout\nvariant: info\n", - "hasFields": [ - "title", - "content" - ] + "hasFields": ["title", "content"] } } ], @@ -865,9 +851,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: button\nid: refresh-btn\nonAction: refresh-stats\n", - "hasFields": [ - "text" - ] + "hasFields": ["text"] } } } @@ -940,9 +924,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: button\nid: refresh-btn\nonAction: refresh-stats\n", - "hasFields": [ - "text" - ] + "hasFields": ["text"] } } ], @@ -1015,9 +997,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: form\nid: profile-form\n", - "hasFields": [ - "onSubmit" - ] + "hasFields": ["onSubmit"] } } } @@ -1090,9 +1070,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: form\nid: profile-form\n", - "hasFields": [ - "onSubmit" - ] + "hasFields": ["onSubmit"] } } ], @@ -1955,9 +1933,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: form\nid: employee-form\n", - "hasFields": [ - "onSubmit" - ] + "hasFields": ["onSubmit"] } } } @@ -2038,9 +2014,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: form\nid: employee-form\n", - "hasFields": [ - "onSubmit" - ] + "hasFields": ["onSubmit"] } } ], @@ -2384,9 +2358,7 @@ "config": { "tags": {}, "description": "MDMA Fixer Eval (capability probe) — own model", - "prompts": [ - "file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-fixer.mjs" - ], + "prompts": ["file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-fixer.mjs"], "providers": [ { "id": "openai:chat:mdma-26b", @@ -2426,9 +2398,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: button\nid: action-btn\nvariant: primary\n", - "hasFields": [ - "text" - ] + "hasFields": ["text"] } } ] @@ -2455,9 +2425,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: callout\nid: status-notice\nvariant: info\ntitle: System Status\n", - "hasFields": [ - "content" - ] + "hasFields": ["content"] } } ] @@ -2518,10 +2486,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: callout\nid: welcome-callout\nvariant: info\n", - "hasFields": [ - "title", - "content" - ] + "hasFields": ["title", "content"] } } ] @@ -2574,9 +2539,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: button\nid: refresh-btn\nonAction: refresh-stats\n", - "hasFields": [ - "text" - ] + "hasFields": ["text"] } } ] @@ -2603,9 +2566,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: form\nid: profile-form\n", - "hasFields": [ - "onSubmit" - ] + "hasFields": ["onSubmit"] } } ] @@ -2786,9 +2747,7 @@ "value": "file://assertions/fixer-contains-component.mjs", "config": { "expected": "type: form\nid: employee-form\n", - "hasFields": [ - "onSubmit" - ] + "hasFields": ["onSubmit"] } } ] @@ -2842,9 +2801,7 @@ ], "scenarios": [], "env": {}, - "outputPath": [ - "own-model/results-fixer.json" - ], + "outputPath": ["own-model/results-fixer.json"], "extensions": [], "metadata": {}, "evaluateOptions": {} @@ -2858,4 +2815,4 @@ "exportedAt": "2026-06-30T12:38:13.905Z", "evaluationCreatedAt": "2026-06-30T12:37:54.142Z" } -} \ No newline at end of file +} diff --git a/evals/own-model/results-flows.json b/evals/own-model/results-flows.json index 19eef94..a5b8608 100644 --- a/evals/own-model/results-flows.json +++ b/evals/own-model/results-flows.json @@ -77,9 +77,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -91,10 +89,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -177,19 +172,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -252,9 +242,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -266,10 +254,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -352,19 +337,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -427,9 +407,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -441,10 +419,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -527,19 +502,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -602,9 +572,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -616,10 +584,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -711,19 +676,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -790,9 +750,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -804,10 +762,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -890,19 +845,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -965,9 +915,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -979,11 +927,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "callout", - "table" - ] + "allowed": ["chart", "callout", "table"] } } }, @@ -1057,20 +1001,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "callout", - "table" - ] + "allowed": ["chart", "callout", "table"] } }, { @@ -1129,9 +1067,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1143,10 +1079,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -1226,19 +1159,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -1298,9 +1226,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1312,10 +1238,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -1398,19 +1321,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -1473,9 +1391,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1487,11 +1403,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "callout", - "form" - ] + "allowed": ["chart", "callout", "form"] } } }, @@ -1565,20 +1477,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "callout", - "form" - ] + "allowed": ["chart", "callout", "form"] } }, { @@ -1637,9 +1543,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1651,13 +1555,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table", - "callout", - "tasklist", - "chart", - "button" - ] + "allowed": ["table", "callout", "tasklist", "chart", "button"] } } }, @@ -1731,22 +1629,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table", - "callout", - "tasklist", - "chart", - "button" - ] + "allowed": ["table", "callout", "tasklist", "chart", "button"] } }, { @@ -1805,9 +1695,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -1819,10 +1707,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -1917,19 +1802,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -1999,9 +1879,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2013,10 +1891,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -2108,19 +1983,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2187,9 +2057,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2201,10 +2069,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -2299,19 +2164,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2381,9 +2241,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2395,10 +2253,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -2490,19 +2345,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2569,9 +2419,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } }, @@ -2583,10 +2431,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } } }, @@ -2678,19 +2523,14 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } }, { "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2774,9 +2614,7 @@ "config": { "tags": {}, "description": "MDMA Example Flows Eval (DSL) — own model", - "prompts": [ - "file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-custom.mjs" - ], + "prompts": ["file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-custom.mjs"], "providers": [ { "id": "openai:chat:mdma-26b", @@ -2805,10 +2643,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2835,10 +2670,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2865,10 +2697,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2895,10 +2724,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2929,10 +2755,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -2959,11 +2782,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "callout", - "table" - ] + "allowed": ["chart", "callout", "table"] } }, { @@ -2986,10 +2805,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3013,10 +2829,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3043,11 +2856,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "chart", - "callout", - "form" - ] + "allowed": ["chart", "callout", "form"] } }, { @@ -3070,13 +2879,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "table", - "callout", - "tasklist", - "chart", - "button" - ] + "allowed": ["table", "callout", "tasklist", "chart", "button"] } }, { @@ -3099,10 +2902,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3136,10 +2936,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3170,10 +2967,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3207,10 +3001,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3241,10 +3032,7 @@ "type": "javascript", "value": "file://assertions/only-components.mjs", "config": { - "allowed": [ - "form", - "callout" - ] + "allowed": ["form", "callout"] } }, { @@ -3273,9 +3061,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3283,9 +3069,7 @@ "options": {}, "metadata": {} }, - "outputPath": [ - "own-model/results-flows.json" - ], + "outputPath": ["own-model/results-flows.json"], "extensions": [], "metadata": {}, "evaluateOptions": {} @@ -3299,4 +3083,4 @@ "exportedAt": "2026-06-30T12:36:59.230Z", "evaluationCreatedAt": "2026-06-30T12:35:53.402Z" } -} \ No newline at end of file +} diff --git a/evals/own-model/results-guidance.json b/evals/own-model/results-guidance.json index 0a4da2e..65b7e6e 100644 --- a/evals/own-model/results-guidance.json +++ b/evals/own-model/results-guidance.json @@ -1798,9 +1798,7 @@ "config": { "tags": {}, "description": "MDMA Agent Guidance Eval (tool-calling) — own model", - "prompts": [ - "file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-guidance.mjs" - ], + "prompts": ["file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt-guidance.mjs"], "providers": [ { "id": "openai:chat:mdma-26b", @@ -1829,9 +1827,7 @@ "description": "The complete MDMA Markdown document." } }, - "required": [ - "document" - ] + "required": ["document"] } } } @@ -2044,9 +2040,7 @@ ], "scenarios": [], "env": {}, - "outputPath": [ - "own-model/results-guidance.json" - ], + "outputPath": ["own-model/results-guidance.json"], "extensions": [], "metadata": {}, "evaluateOptions": {} @@ -2060,4 +2054,4 @@ "exportedAt": "2026-06-30T12:39:15.164Z", "evaluationCreatedAt": "2026-06-30T12:38:16.261Z" } -} \ No newline at end of file +} diff --git a/evals/own-model/results.json b/evals/own-model/results.json index cc0b888..c213166 100644 --- a/evals/own-model/results.json +++ b/evals/own-model/results.json @@ -77,9 +77,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -143,9 +141,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -199,9 +195,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -265,9 +259,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -321,9 +313,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -387,9 +377,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -443,9 +431,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -509,9 +495,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -565,9 +549,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -631,9 +613,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -687,9 +667,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -753,9 +731,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -809,9 +785,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -875,9 +849,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -931,9 +903,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -997,9 +967,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1053,9 +1021,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1119,9 +1085,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1175,9 +1139,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1241,9 +1203,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1297,9 +1257,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1363,9 +1321,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1419,9 +1375,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1485,9 +1439,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1541,9 +1493,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1607,9 +1557,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1663,9 +1611,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1729,9 +1675,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1785,9 +1729,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1851,9 +1793,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -1907,9 +1847,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -1973,9 +1911,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2029,9 +1965,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2095,9 +2029,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2151,9 +2083,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2217,9 +2147,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2273,9 +2201,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2339,9 +2265,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2395,9 +2319,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2461,9 +2383,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2517,9 +2437,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2583,9 +2501,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2639,9 +2555,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2705,9 +2619,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2761,9 +2673,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2827,9 +2737,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -2883,9 +2791,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -2949,9 +2855,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3005,9 +2909,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3071,9 +2973,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3127,9 +3027,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3193,9 +3091,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3249,9 +3145,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3315,9 +3209,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3371,9 +3263,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3437,9 +3327,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3493,9 +3381,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3559,9 +3445,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3615,9 +3499,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3681,9 +3563,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3737,9 +3617,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3803,9 +3681,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3859,9 +3735,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -3925,9 +3799,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -3981,9 +3853,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4047,9 +3917,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4103,9 +3971,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4169,9 +4035,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4225,9 +4089,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4291,9 +4153,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4347,9 +4207,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4413,9 +4271,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4469,9 +4325,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4535,9 +4389,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4591,9 +4443,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4657,9 +4507,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4713,9 +4561,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4779,9 +4625,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4835,9 +4679,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -4901,9 +4743,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -4957,9 +4797,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5023,9 +4861,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5079,9 +4915,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5145,9 +4979,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5201,9 +5033,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5267,9 +5097,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5323,9 +5151,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5389,9 +5215,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5445,9 +5269,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5511,9 +5333,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5567,9 +5387,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5633,9 +5451,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5689,9 +5505,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5755,9 +5569,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5811,9 +5623,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5877,9 +5687,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -5933,9 +5741,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -5999,9 +5805,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6055,9 +5859,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6121,9 +5923,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6177,9 +5977,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6243,9 +6041,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6299,9 +6095,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6365,9 +6159,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6421,9 +6213,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6487,9 +6277,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6543,9 +6331,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6609,9 +6395,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6665,9 +6449,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6731,9 +6513,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6787,9 +6567,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6853,9 +6631,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -6909,9 +6685,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -6975,9 +6749,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7031,9 +6803,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7097,9 +6867,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7153,9 +6921,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7219,9 +6985,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7275,9 +7039,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7341,9 +7103,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7397,9 +7157,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7463,9 +7221,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7519,9 +7275,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7585,9 +7339,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7641,9 +7393,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7707,9 +7457,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7763,9 +7511,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7829,9 +7575,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -7885,9 +7629,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -7951,9 +7693,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8007,9 +7747,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8073,9 +7811,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8129,9 +7865,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8195,9 +7929,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8251,9 +7983,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8317,9 +8047,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8373,9 +8101,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8439,9 +8165,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8495,9 +8219,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8561,9 +8283,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8617,9 +8337,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8683,9 +8401,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8739,9 +8455,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8805,9 +8519,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8861,9 +8573,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -8927,9 +8637,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -8983,9 +8691,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9049,9 +8755,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9105,9 +8809,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9171,9 +8873,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9227,9 +8927,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9293,9 +8991,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9349,9 +9045,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9415,9 +9109,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9471,9 +9163,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9537,9 +9227,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9593,9 +9281,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9659,9 +9345,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9715,9 +9399,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9781,9 +9463,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9837,9 +9517,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -9903,9 +9581,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -9959,9 +9635,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10025,9 +9699,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10081,9 +9753,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10147,9 +9817,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10203,9 +9871,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10269,9 +9935,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10325,9 +9989,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10391,9 +10053,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10447,9 +10107,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10513,9 +10171,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10569,9 +10225,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10635,9 +10289,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10691,9 +10343,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10757,9 +10407,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10813,9 +10461,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -10879,9 +10525,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -10935,9 +10579,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -11001,9 +10643,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -11057,9 +10697,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -11123,9 +10761,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -11179,9 +10815,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -11245,9 +10879,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -11301,9 +10933,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -11367,9 +10997,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -11423,9 +11051,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -11489,9 +11115,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -11545,9 +11169,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } } @@ -11611,9 +11233,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -11684,9 +11304,7 @@ "config": { "tags": {}, "description": "MDMA-IL DSL Holdout Gate — own model", - "prompts": [ - "file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt.mjs" - ], + "prompts": ["file:///Users/marcinsadowski/GIT/mr-mdma/evals/own-model/prompt.mjs"], "providers": [ { "id": "openai:chat:mdma-26b", @@ -12568,9 +12186,7 @@ "type": "javascript", "value": "file://assertions/validate-mdma.mjs", "config": { - "exclude": [ - "flow-ordering" - ] + "exclude": ["flow-ordering"] } } ], @@ -12578,9 +12194,7 @@ "options": {}, "metadata": {} }, - "outputPath": [ - "own-model/results.json" - ], + "outputPath": ["own-model/results.json"], "extensions": [], "metadata": {}, "evaluateOptions": {} @@ -12594,4 +12208,4 @@ "exportedAt": "2026-06-30T12:35:14.118Z", "evaluationCreatedAt": "2026-06-30T12:32:59.728Z" } -} \ No newline at end of file +} diff --git a/evals/own-model/run-conversation.mjs b/evals/own-model/run-conversation.mjs index 85be2b7..4db33c1 100644 --- a/evals/own-model/run-conversation.mjs +++ b/evals/own-model/run-conversation.mjs @@ -34,7 +34,8 @@ const convs = new Map(); for (const t of turns) { const id = t.metadata?.conversationId ?? 'default'; if (!convs.has(id)) convs.set(id, []); - const allowed = t.assert?.find((a) => String(a.value).includes('only-components'))?.config?.allowed ?? []; + const allowed = + t.assert?.find((a) => String(a.value).includes('only-components'))?.config?.allowed ?? []; convs.get(id).push({ desc: t.description, dsl: String(t.vars.message).trim(), @@ -97,10 +98,18 @@ for (const [cid, ts] of convs) { if (ok) pass++; const notes = []; if (!v.ok) - notes.push(`invalid: ${v.issues.filter((i) => i.severity === 'error').map((i) => i.ruleId).join(',')}`); - if (!onlyExpected) notes.push(`emitted: ${types.join(',') || 'none'}; expected ${turn.allowedType}`); + notes.push( + `invalid: ${v.issues + .filter((i) => i.severity === 'error') + .map((i) => i.ruleId) + .join(',')}`, + ); + if (!onlyExpected) + notes.push(`emitted: ${types.join(',') || 'none'}; expected ${turn.allowedType}`); if (contaminated.length) notes.push(`RE-EMITTED prior: ${contaminated.join(',')}`); - console.log(` ${ok ? 'PASS' : 'FAIL'} ${turn.desc}${notes.length ? ` [${notes.join('] [')}]` : ''}`); + console.log( + ` ${ok ? 'PASS' : 'FAIL'} ${turn.desc}${notes.length ? ` [${notes.join('] [')}]` : ''}`, + ); if (turn.id) priorIds.push(turn.id); } } diff --git a/packages/agui/package.json b/packages/agui/package.json index 0f45a1a..ac3caf4 100644 --- a/packages/agui/package.json +++ b/packages/agui/package.json @@ -67,9 +67,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/agui/src/bridge.ts b/packages/agui/src/bridge.ts index de5ec39..657a856 100644 --- a/packages/agui/src/bridge.ts +++ b/packages/agui/src/bridge.ts @@ -198,10 +198,12 @@ export function createMdmaAgentBridge( const subscriber: AguiSubscriber = { // During streaming, AG-UI's content buffer lags one delta behind — good enough for a live // (slightly-behind) render, throttled to avoid re-parsing every token. - onTextMessageContentEvent: (params) => ingest(params.event.messageId, params.textMessageBuffer, false), + onTextMessageContentEvent: (params) => + ingest(params.event.messageId, params.textMessageBuffer, false), // The end buffer is the COMPLETE message, so parse it immediately — this is what guarantees // the final render isn't a truncated tail (content events never carry the last delta). - onTextMessageEndEvent: (params) => ingest(params.event.messageId, params.textMessageBuffer, true), + onTextMessageEndEvent: (params) => + ingest(params.event.messageId, params.textMessageBuffer, true), // Belt-and-suspenders: a run that ends without a text-end (e.g. tool transport) still flushes. onRunFinishedEvent: () => flush(), onRunFailedEvent: () => flush(), @@ -250,7 +252,12 @@ function serializeAction(action: MdmaActionEvent): Record { payload: action.payload, }; case 'APPROVAL_GRANTED': - return { kind: 'approval', decision: 'granted', componentId: action.componentId, actor: action.actor }; + return { + kind: 'approval', + decision: 'granted', + componentId: action.componentId, + actor: action.actor, + }; case 'APPROVAL_DENIED': return { kind: 'approval', diff --git a/packages/agui/src/react/MdmaAgentView.tsx b/packages/agui/src/react/MdmaAgentView.tsx index a48f29f..a9c6143 100644 --- a/packages/agui/src/react/MdmaAgentView.tsx +++ b/packages/agui/src/react/MdmaAgentView.tsx @@ -16,7 +16,12 @@ export interface MdmaAgentViewProps extends MdmaAgentBridgeOptions { * Drop-in view: subscribes to an AG-UI agent and renders every MDMA document it streams, with * approvals/forms wired back into the run. For finer control, use {@link useMdmaAgentStream}. */ -export function MdmaAgentView({ agent, customizations, className, ...options }: MdmaAgentViewProps) { +export function MdmaAgentView({ + agent, + customizations, + className, + ...options +}: MdmaAgentViewProps) { const { documents } = useMdmaAgentStream(agent, options); return (
diff --git a/packages/agui/tests/bridge.test.ts b/packages/agui/tests/bridge.test.ts index 316dc84..9acc560 100644 --- a/packages/agui/tests/bridge.test.ts +++ b/packages/agui/tests/bridge.test.ts @@ -1,12 +1,7 @@ import { describe, it, expect } from 'vitest'; import { createMdmaAgentBridge, type MdmaMessageState } from '../src/bridge.js'; import { parseMdma } from '../src/parse.js'; -import type { - AguiAgent, - AguiMessage, - AguiSubscriber, - AguiSubscription, -} from '../src/types.js'; +import type { AguiAgent, AguiMessage, AguiSubscriber, AguiSubscription } from '../src/types.js'; const FENCE = '```'; const APPROVAL_DOC = [ @@ -68,8 +63,9 @@ class FakeAgent implements AguiAgent { function astBlockTypes(state: MdmaMessageState | undefined): string[] { if (!state) return []; return state.ast.children - .filter((c): c is { type: 'mdmaBlock'; component: { type: string } } => - (c as { type?: string }).type === 'mdmaBlock', + .filter( + (c): c is { type: 'mdmaBlock'; component: { type: string } } => + (c as { type?: string }).type === 'mdmaBlock', ) .map((c) => c.component.type); } @@ -107,7 +103,10 @@ describe('createMdmaAgentBridge', () => { // Simulate AG-UI: content events carry the buffer *before* the latest delta, so the last // chunk (here the tail of `approval-gate` + the rest of the doc) never arrives via content. - const lagging = APPROVAL_DOC.slice(0, APPROVAL_DOC.indexOf('approval-gate') + 'approval-gat'.length); + const lagging = APPROVAL_DOC.slice( + 0, + APPROVAL_DOC.indexOf('approval-gate') + 'approval-gat'.length, + ); agent.emitContent('m1', lagging); await bridge.flush(); @@ -221,7 +220,12 @@ describe('createMdmaAgentBridge', () => { await bridge.flush(); const store = bridge.documents.get('m1')!.store; - store.dispatch({ type: 'APPROVAL_DENIED', componentId: 'gate1', actor: { id: 'u' }, reason: 'no' }); + store.dispatch({ + type: 'APPROVAL_DENIED', + componentId: 'gate1', + actor: { id: 'u' }, + reason: 'no', + }); await Promise.resolve(); expect(agent.runs).toBe(0); diff --git a/packages/attachables-core/package.json b/packages/attachables-core/package.json index e006e9e..f67c6eb 100644 --- a/packages/attachables-core/package.json +++ b/packages/attachables-core/package.json @@ -31,9 +31,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/cli/package.json b/packages/cli/package.json index 5c8bc38..0d17c15 100644 --- a/packages/cli/package.json +++ b/packages/cli/package.json @@ -1,15 +1,7 @@ { "name": "@mobile-reality/mdma-cli", "version": "0.2.8", - "keywords": [ - "mdma", - "markdown", - "ai", - "llm", - "interactive-document", - "cli", - "prompt-builder" - ], + "keywords": ["mdma", "markdown", "ai", "llm", "interactive-document", "cli", "prompt-builder"], "type": "module", "bin": { "mdma": "./dist/bin/mdma.js" @@ -68,10 +60,7 @@ "vite": "^6.0.0", "vitest": "^3.2.0" }, - "files": [ - "dist", - "app-dist" - ], + "files": ["dist", "app-dist"], "publishConfig": { "access": "public" }, diff --git a/packages/mcp/package.json b/packages/mcp/package.json index ddebb7a..acd8261 100644 --- a/packages/mcp/package.json +++ b/packages/mcp/package.json @@ -40,9 +40,7 @@ "vitest": "^3.2.0", "zod": "^3.24.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/parser/package.json b/packages/parser/package.json index f41713e..29003fd 100644 --- a/packages/parser/package.json +++ b/packages/parser/package.json @@ -41,9 +41,7 @@ "vitest": "^3.2.0", "zod": "^3.24.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/prompt-pack/package.json b/packages/prompt-pack/package.json index 1a294a0..04c5671 100644 --- a/packages/prompt-pack/package.json +++ b/packages/prompt-pack/package.json @@ -37,9 +37,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/prompt-pack/src/prompts/mdma-author/registry.ts b/packages/prompt-pack/src/prompts/mdma-author/registry.ts index 8af58b1..7e18bc0 100644 --- a/packages/prompt-pack/src/prompts/mdma-author/registry.ts +++ b/packages/prompt-pack/src/prompts/mdma-author/registry.ts @@ -113,7 +113,7 @@ export const AUTHOR_PROMPT_VARIANTS: AuthorPromptVariant[] = [ id: 'google/gemma', label: 'Google — Gemma', description: - "Open-weights Gemma family (Gemma 4 26B-a4b / 31B, Gemma 3n 4B). Gemini-native Markdown framing with all defensive blocks bundled (fence closing, scope discipline, string select values) for the open-model tier.", + 'Open-weights Gemma family (Gemma 4 26B-a4b / 31B, Gemma 3n 4B). Gemini-native Markdown framing with all defensive blocks bundled (fence closing, scope discipline, string select values) for the open-model tier.', prompt: MDMA_AUTHOR_PROMPT_GEMMA, }, { diff --git a/packages/renderer-react-native/package.json b/packages/renderer-react-native/package.json index ebddce1..9edd577 100644 --- a/packages/renderer-react-native/package.json +++ b/packages/renderer-react-native/package.json @@ -14,10 +14,15 @@ "ui" ], "type": "module", + "main": "./dist/index.js", + "module": "./dist/index.js", + "react-native": "./dist/index.js", + "types": "./dist/index.d.ts", "exports": { ".": { + "types": "./dist/index.d.ts", "import": "./dist/index.js", - "types": "./dist/index.d.ts" + "default": "./dist/index.js" } }, "scripts": { @@ -41,9 +46,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/renderer-react-native/src/components/ButtonRenderer.tsx b/packages/renderer-react-native/src/components/ButtonRenderer.tsx index dc48610..18c2123 100644 --- a/packages/renderer-react-native/src/components/ButtonRenderer.tsx +++ b/packages/renderer-react-native/src/components/ButtonRenderer.tsx @@ -54,7 +54,9 @@ export const ButtonRenderer = memo(function ButtonRenderer({ marginVertical: spacing.xs, }} > - {component.text} + + {component.text} + ); }); diff --git a/packages/renderer-react-native/src/components/CalloutRenderer.tsx b/packages/renderer-react-native/src/components/CalloutRenderer.tsx index 5a405a8..ff91a9f 100644 --- a/packages/renderer-react-native/src/components/CalloutRenderer.tsx +++ b/packages/renderer-react-native/src/components/CalloutRenderer.tsx @@ -39,7 +39,9 @@ export const CalloutRenderer = memo(function CalloutRenderer({ {component.title} ) : null} - + {component.content} diff --git a/packages/renderer-react-native/src/components/ChartRenderer.tsx b/packages/renderer-react-native/src/components/ChartRenderer.tsx index 01f7ad2..43de1b7 100644 --- a/packages/renderer-react-native/src/components/ChartRenderer.tsx +++ b/packages/renderer-react-native/src/components/ChartRenderer.tsx @@ -68,20 +68,44 @@ export const ChartRenderer = memo(function ChartRenderer({ {data.rows.length === 0 ? ( No chart data ) : ( - + {data.headers.map((h) => ( - - {h} + + + {h} + ))} {data.rows.map((row, i) => ( - + {data.headers.map((h) => ( - - {String(row[h] ?? '')} + + + {String(row[h] ?? '')} + ))} diff --git a/packages/renderer-react-native/src/components/FormRenderer.tsx b/packages/renderer-react-native/src/components/FormRenderer.tsx index 01c2255..10c692b 100644 --- a/packages/renderer-react-native/src/components/FormRenderer.tsx +++ b/packages/renderer-react-native/src/components/FormRenderer.tsx @@ -94,7 +94,12 @@ export const FormRenderer = memo(function FormRenderer({ backgroundColor: selected ? colors.primary : colors.background, }} > - + {opt.label} diff --git a/packages/renderer-react-native/src/components/MdastRenderer.tsx b/packages/renderer-react-native/src/components/MdastRenderer.tsx index c2e1c3c..33c5352 100644 --- a/packages/renderer-react-native/src/components/MdastRenderer.tsx +++ b/packages/renderer-react-native/src/components/MdastRenderer.tsx @@ -125,7 +125,9 @@ export const MdastRenderer = memo(function MdastRenderer({ node }: MdastRenderer {node.ordered ? `${i + 1}. ` : '• '} - {item.children?.flatMap((c) => c.children ?? [c]).map((c, j) => renderInline(c, theme, j))} + {item.children + ?.flatMap((c) => c.children ?? [c]) + .map((c, j) => renderInline(c, theme, j))} ))} @@ -148,9 +150,7 @@ export const MdastRenderer = memo(function MdastRenderer({ node }: MdastRenderer ); case 'thematicBreak': return ( - + ); case 'blockquote': return ( @@ -162,7 +162,9 @@ export const MdastRenderer = memo(function MdastRenderer({ node }: MdastRenderer marginVertical: spacing.xs, }} > - {node.children?.map((c, i) => )} + {node.children?.map((c, i) => ( + + ))} ); default: diff --git a/packages/renderer-react-native/src/components/MdmaDocument.tsx b/packages/renderer-react-native/src/components/MdmaDocument.tsx index 87d0e2c..03b1f30 100644 --- a/packages/renderer-react-native/src/components/MdmaDocument.tsx +++ b/packages/renderer-react-native/src/components/MdmaDocument.tsx @@ -40,8 +40,13 @@ export interface MdmaDocumentProps { function isComponentConfig( entry: ComponentEntry, -): entry is { renderer?: ComponentType; elements?: Record } { - return typeof entry === 'object' && entry !== null && ('renderer' in entry || 'elements' in entry); +): entry is { + renderer?: ComponentType; + elements?: Record; +} { + return ( + typeof entry === 'object' && entry !== null && ('renderer' in entry || 'elements' in entry) + ); } /** Reduce the unified `components` map to a plain renderer record. */ diff --git a/packages/renderer-react-native/src/components/TableRenderer.tsx b/packages/renderer-react-native/src/components/TableRenderer.tsx index eaafac7..05a03ab 100644 --- a/packages/renderer-react-native/src/components/TableRenderer.tsx +++ b/packages/renderer-react-native/src/components/TableRenderer.tsx @@ -38,7 +38,11 @@ export const TableRenderer = memo(function TableRenderer({ {component.label} ) : null} - + {/* Header */} diff --git a/packages/renderer-react-native/src/components/TasklistRenderer.tsx b/packages/renderer-react-native/src/components/TasklistRenderer.tsx index e3a6ea0..e3fcf75 100644 --- a/packages/renderer-react-native/src/components/TasklistRenderer.tsx +++ b/packages/renderer-react-native/src/components/TasklistRenderer.tsx @@ -65,7 +65,12 @@ export const TasklistRenderer = memo(function TasklistRenderer({ accessibilityRole="checkbox" accessibilityState={{ checked }} onPress={() => onToggle(!checked)} - style={{ flexDirection: 'row', alignItems: 'center', gap: spacing.sm, paddingVertical: spacing.xs }} + style={{ + flexDirection: 'row', + alignItems: 'center', + gap: spacing.sm, + paddingVertical: spacing.xs, + }} > ; enable the LayoutAnimation collapse on Android. -if ( - Platform.OS === 'android' && - UIManager.setLayoutAnimationEnabledExperimental -) { +if (Platform.OS === 'android' && UIManager.setLayoutAnimationEnabledExperimental) { UIManager.setLayoutAnimationEnabledExperimental(true); } @@ -65,7 +62,9 @@ export const ThinkingRenderer = memo(function ThinkingRenderer({ {!collapsed ? ( - + {component.content} diff --git a/packages/renderer-react/package.json b/packages/renderer-react/package.json index e723dc0..6a57179 100644 --- a/packages/renderer-react/package.json +++ b/packages/renderer-react/package.json @@ -1,16 +1,7 @@ { "name": "@mobile-reality/mdma-renderer-react", "version": "0.4.0", - "keywords": [ - "mdma", - "markdown", - "ai", - "llm", - "interactive-document", - "react", - "forms", - "ui" - ], + "keywords": ["mdma", "markdown", "ai", "llm", "interactive-document", "react", "forms", "ui"], "type": "module", "exports": { ".": { @@ -38,10 +29,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist", - "styles.css" - ], + "files": ["dist", "styles.css"], "publishConfig": { "access": "public" }, diff --git a/packages/runtime/package.json b/packages/runtime/package.json index 63c6f1c..3ab0d7f 100644 --- a/packages/runtime/package.json +++ b/packages/runtime/package.json @@ -12,10 +12,15 @@ "pii" ], "type": "module", + "main": "./dist/index.js", + "module": "./dist/index.js", + "react-native": "./dist/index.js", + "types": "./dist/index.d.ts", "exports": { ".": { + "types": "./dist/index.d.ts", "import": "./dist/index.js", - "types": "./dist/index.d.ts" + "default": "./dist/index.js" } }, "scripts": { @@ -31,9 +36,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/runtime/tests/document-store.test.ts b/packages/runtime/tests/document-store.test.ts index f693f6b..bdf4d03 100644 --- a/packages/runtime/tests/document-store.test.ts +++ b/packages/runtime/tests/document-store.test.ts @@ -254,7 +254,9 @@ describe('DocumentStore.updateAst', () => { // Re-parse of the same component (e.g. a later streamed chunk) must not wipe user input. store.updateAst( - makeAst([{ id: 'f', type: 'form', fields: [{ name: 'x', type: 'text', label: 'X changed' }] }]), + makeAst([ + { id: 'f', type: 'form', fields: [{ name: 'x', type: 'text', label: 'X changed' }] }, + ]), ); expect(store.getComponentState('f')?.values.x).toBe('typed'); }); @@ -265,7 +267,9 @@ describe('DocumentStore.updateAst', () => { expect(store.getComponentState('deploy-gate')?.type).toBe('approval-gat'); // A later parse resolves the real type — the store must adopt it, not freeze the placeholder. - store.updateAst(makeAst([{ id: 'deploy-gate', type: 'approval-gate', title: 'Approve deploy' }])); + store.updateAst( + makeAst([{ id: 'deploy-gate', type: 'approval-gate', title: 'Approve deploy' }]), + ); expect(store.getComponentState('deploy-gate')?.type).toBe('approval-gate'); }); }); diff --git a/packages/spec/package.json b/packages/spec/package.json index ff14e42..a17c6e0 100644 --- a/packages/spec/package.json +++ b/packages/spec/package.json @@ -1,21 +1,16 @@ { "name": "@mobile-reality/mdma-spec", "version": "0.3.0", - "keywords": [ - "mdma", - "markdown", - "ai", - "llm", - "interactive-document", - "schema", - "zod", - "types" - ], + "keywords": ["mdma", "markdown", "ai", "llm", "interactive-document", "schema", "zod", "types"], "type": "module", + "main": "./dist/index.js", + "module": "./dist/index.js", + "react-native": "./dist/index.js", + "types": "./dist/index.d.ts", "exports": { ".": { - "import": "./dist/index.js", "types": "./dist/index.d.ts", + "import": "./dist/index.js", "default": "./dist/index.js" }, "./schemas/*": { @@ -43,9 +38,7 @@ "typescript": "^5.7.0", "vitest": "^3.2.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" }, diff --git a/packages/validator/package.json b/packages/validator/package.json index 81d2227..ec8e1aa 100644 --- a/packages/validator/package.json +++ b/packages/validator/package.json @@ -35,9 +35,7 @@ "vitest": "^3.2.0", "zod": "^3.24.0" }, - "files": [ - "dist" - ], + "files": ["dist"], "publishConfig": { "access": "public" },