@nac3/forge-cli 0.2.0-alpha.4 → 0.2.0-alpha.59z.118

package/dist/qa/brain_matrix/compare.js

@@ -0,0 +1,122 @@
+/**
+ * Brain matrix --compare: diff two MatrixReport JSONs and
+ * surface per-cell regressions / improvements.
+ *
+ * Used by `yf qa brain-matrix --compare base.json head.json` so
+ * Pablo can answer "did z.62 regress anything against z.55?"
+ * without scanning two 2000-line files by hand.
+ */
+import { readFile } from 'node:fs/promises';
+function cellOutcome(matrix, scenarioId, brainId) {
+    const cell = matrix.results?.[scenarioId]?.[brainId];
+    if (!cell)
+        return 'MISSING';
+    if (cell.run?.error)
+        return 'ERR';
+    if (cell.pass)
+        return 'PASS';
+    return 'FAIL';
+}
+export function compareMatrices(base, head) {
+    const scenarios = new Set([
+        ...base.scenarios.map((s) => s.id),
+        ...head.scenarios.map((s) => s.id),
+    ]);
+    const brains = new Set([
+        ...base.brains.map((b) => b.id),
+        ...head.brains.map((b) => b.id),
+    ]);
+    const cells = [];
+    for (const s of [...scenarios].sort()) {
+        for (const b of [...brains].sort()) {
+            const baseOut = cellOutcome(base, s, b);
+            const headOut = cellOutcome(head, s, b);
+            let kind;
+            if (baseOut === 'MISSING' && headOut !== 'MISSING')
+                kind = 'new';
+            else if (baseOut !== 'MISSING' && headOut === 'MISSING')
+                kind = 'removed';
+            else if (baseOut === headOut)
+                kind = 'unchanged';
+            else if (baseOut === 'PASS' && (headOut === 'FAIL' || headOut === 'ERR'))
+                kind = 'regression';
+            else if ((baseOut === 'FAIL' || baseOut === 'ERR') && headOut === 'PASS')
+                kind = 'improvement';
+            else
+                kind = 'flake';
+            cells.push({ scenario: s, brain: b, base: baseOut, head: headOut, kind });
+        }
+    }
+    return {
+        base_version: base.forge_version,
+        head_version: head.forge_version,
+        cells,
+        regressions: cells.filter((c) => c.kind === 'regression'),
+        improvements: cells.filter((c) => c.kind === 'improvement'),
+        flakes: cells.filter((c) => c.kind === 'flake'),
+        new_cells: cells.filter((c) => c.kind === 'new'),
+        removed_cells: cells.filter((c) => c.kind === 'removed'),
+        unchanged_count: cells.filter((c) => c.kind === 'unchanged').length,
+    };
+}
+export async function loadMatrixReport(p) {
+    const raw = await readFile(p, 'utf-8');
+    return JSON.parse(raw);
+}
+export function renderCompareSummary(c) {
+    const lines = [];
+    lines.push('Brain matrix compare');
+    lines.push('  base: ' + c.base_version);
+    lines.push('  head: ' + c.head_version);
+    lines.push('');
+    lines.push('Summary');
+    lines.push('  regressions:  ' + c.regressions.length);
+    lines.push('  improvements: ' + c.improvements.length);
+    lines.push('  flakes:       ' + c.flakes.length);
+    lines.push('  new cells:    ' + c.new_cells.length);
+    lines.push('  removed:      ' + c.removed_cells.length);
+    lines.push('  unchanged:    ' + c.unchanged_count);
+    lines.push('');
+    if (c.regressions.length > 0) {
+        lines.push('=== Regressions ===');
+        for (const r of c.regressions) {
+            lines.push('  ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
+                + '  ' + r.base + ' -> ' + r.head);
+        }
+        lines.push('');
+    }
+    if (c.improvements.length > 0) {
+        lines.push('=== Improvements ===');
+        for (const r of c.improvements) {
+            lines.push('  ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
+                + '  ' + r.base + ' -> ' + r.head);
+        }
+        lines.push('');
+    }
+    if (c.flakes.length > 0) {
+        lines.push('=== Flakes (FAIL<->ERR) ===');
+        for (const r of c.flakes) {
+            lines.push('  ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
+                + '  ' + r.base + ' -> ' + r.head);
+        }
+        lines.push('');
+    }
+    if (c.new_cells.length > 0) {
+        lines.push('=== New cells (only in head) ===');
+        for (const r of c.new_cells) {
+            lines.push('  ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
+                + '  ' + r.head);
+        }
+        lines.push('');
+    }
+    if (c.removed_cells.length > 0) {
+        lines.push('=== Removed cells (only in base) ===');
+        for (const r of c.removed_cells) {
+            lines.push('  ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
+                + '  ' + r.base);
+        }
+        lines.push('');
+    }
+    return lines.join('\n');
+}
+//# sourceMappingURL=compare.js.map

package/dist/qa/brain_matrix/compare.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compare.js","sourceRoot":"","sources":["../../../src/qa/brain_matrix/compare.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AA4B5C,SAAS,WAAW,CAAC,MAAoB,EAAE,UAAkB,EAAE,OAAe;IAC5E,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,EAAE,CAAC,UAAU,CAAC,EAAE,CAAC,OAAO,CAAC,CAAC;IACrD,IAAI,CAAC,IAAI;QAAE,OAAO,SAAS,CAAC;IAC5B,IAAI,IAAI,CAAC,GAAG,EAAE,KAAK;QAAE,OAAO,KAAK,CAAC;IAClC,IAAI,IAAI,CAAC,IAAI;QAAE,OAAO,MAAM,CAAC;IAC7B,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,MAAM,UAAU,eAAe,CAAC,IAAkB,EAAE,IAAkB;IACpE,MAAM,SAAS,GAAG,IAAI,GAAG,CAAS;QAChC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAClC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACnC,CAAC,CAAC;IACH,MAAM,MAAM,GAAG,IAAI,GAAG,CAAS;QAC7B,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC/B,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAChC,CAAC,CAAC;IACH,MAAM,KAAK,GAAkB,EAAE,CAAC;IAChC,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QACtC,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;YACnC,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;YACxC,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;YACxC,IAAI,IAAyB,CAAC;YAC9B,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,SAAS;gBAAE,IAAI,GAAG,KAAK,CAAC;iBAC5D,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,SAAS;gBAAE,IAAI,GAAG,SAAS,CAAC;iBACrE,IAAI,OAAO,KAAK,OAAO;gBAAE,IAAI,GAAG,WAAW,CAAC;iBAC5C,IAAI,OAAO,KAAK,MAAM,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,OAAO,KAAK,KAAK,CAAC;gBAAE,IAAI,GAAG,YAAY,CAAC;iBACzF,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,OAAO,KAAK,KAAK,CAAC,IAAI,OAAO,KAAK,MAAM;gBAAE,IAAI,GAAG,aAAa,CAAC;;gBAC1F,IAAI,GAAG,OAAO,CAAC;YACpB,KAAK,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC;QAC5E,CAAC;IACH,CAAC;IACD,OAAO;QACL,YAAY,EAAK,IAAI,CAAC,aAAa;QACnC,YAAY,EAAK,IAAI,CAAC,aAAa;QACnC,KAAK;QACL,WAAW,EAAM,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,YAAY,CAAC;QAC7D,YAAY,EAAK,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,aAAa,CAAC;QAC9D,MAAM,EAAW,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;QACxD,SAAS,EAAQ,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,KAAK,CAAC;QACtD,aAAa,EAAI,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,CAAC;QAC1D,eAAe,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC,MAAM;KACpE,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,CAAS;IAC9C,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,CAAC,EAAE,OAAO,CAAC,CAAC;IACvC,OAAO,IAAI,CAAC,KAAK,CAAC,GAAG,CAAiB,CAAC;AACzC,CAAC;AAED,MAAM,UAAU,oBAAoB,CAAC,CAAiB;IACpD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;IACnC,KAAK,CAAC,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,YAAY,CAAC,CAAC;IACxC,KAAK,CAAC,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,YAAY,CAAC,CAAC;IACxC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;IACtB,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;IACtD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,YAAY,CAAC,MAAM,CAAC,CAAC;IACvD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IACjD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC;IACpD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;IACxD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,eAAe,CAAC,CAAC;IACnD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,IAAI,CAAC,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,qBAAqB,CAAC,CAAC;QAClC,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,CAAC;YAC9B,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,GAAG,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC9B,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;QACnC,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,YAAY,EAAE,CAAC;YAC/B,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,GAAG,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACxB,KAAK,CAAC,IAAI,CAAC,6BAA6B,CAAC,CAAC;QAC1C,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,CAAC;YACzB,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,GAAG,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC3B,KAAK,CAAC,IAAI,CAAC,kCAAkC,CAAC,CAAC;QAC/C,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,CAAC;YAC5B,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACrB,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,aAAa,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC/B,KAAK,CAAC,IAAI,CAAC,sCAAsC,CAAC,CAAC;QACnD,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,aAAa,EAAE,CAAC;YAChC,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACrB,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC"}

package/dist/qa/brain_matrix/fake_brain.d.ts

@@ -0,0 +1,36 @@
+/**
+ * Fake brain for dry-run testing of the brain matrix harness.
+ *
+ * Each scenario can declare a `fake_responses` field in setup:
+ *
+ *   fake_responses: [
+ *     { text: "voy a abrir la pestana", tool_calls: [{name: "nac3_invoke",
+ *       args: { verb_id: "yujin.pizarron.create-tab", args: {title:"X", kind:"code", content:"..."} }}] },
+ *     { text: "listo, te abri la pestana" }
+ *   ]
+ *
+ * The fake brain returns one entry per LLM turn, in order. If the
+ * harness asks for more turns than declared, returns "" + no tool
+ * calls (the scenario likely has a bug -- the report flags it).
+ *
+ * Used in:
+ *   - vitest unit tests of the harness
+ *   - CI dry-run that validates the harness itself before any
+ *     real-LLM run is attempted.
+ *
+ * ASCII-only.
+ */
+import type { UnifiedChatClient } from '../../llm/multi_provider_client.js';
+import type { BrainSelection } from '../../core/cost_router.js';
+export interface FakeResponse {
+    text: string;
+    tool_calls?: Array<{
+        name: string;
+        args: Record<string, unknown>;
+    }>;
+    /** Optional: simulate token spend for reports. */
+    tokens_in?: number;
+    tokens_out?: number;
+}
+export declare function makeFakeBrain(brain: BrainSelection, responses: FakeResponse[]): UnifiedChatClient;
+//# sourceMappingURL=fake_brain.d.ts.map

package/dist/qa/brain_matrix/fake_brain.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"fake_brain.d.ts","sourceRoot":"","sources":["../../../src/qa/brain_matrix/fake_brain.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,OAAO,KAAK,EAAE,iBAAiB,EAA2C,MAAM,oCAAoC,CAAC;AACrH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,2BAA2B,CAAC;AAGhE,MAAM,WAAW,YAAY;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,CAAC,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,CAAC,CAAC;IACpE,kDAAkD;IAClD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,wBAAgB,aAAa,CAAC,KAAK,EAAE,cAAc,EAAE,SAAS,EAAE,YAAY,EAAE,GAAG,iBAAiB,CAyBjG"}

package/dist/qa/brain_matrix/fake_brain.js

@@ -0,0 +1,27 @@
+export function makeFakeBrain(brain, responses) {
+    let idx = 0;
+    return {
+        async chat(_req) {
+            const r = responses[idx] || { text: '', tool_calls: [] };
+            idx++;
+            /* Build a synthetic ToolRound[]. The real ToolRound shape
+             *  carries call + result; for the fake brain we leave result
+             *  as `null` so the runner knows to mock it. */
+            const toolRounds = (r.tool_calls || []).map((c) => ({
+                tool: c.name,
+                input: c.args,
+                output: { ok: true, fake: true },
+                is_error: false,
+            }));
+            return {
+                text: r.text,
+                tokensIn: r.tokens_in ?? 100,
+                tokensOut: r.tokens_out ?? 50,
+                model: 'fake-scripted',
+                toolRounds,
+                brain,
+            };
+        },
+    };
+}
+//# sourceMappingURL=fake_brain.js.map

package/dist/qa/brain_matrix/fake_brain.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fake_brain.js","sourceRoot":"","sources":["../../../src/qa/brain_matrix/fake_brain.ts"],"names":[],"mappings":"AAkCA,MAAM,UAAU,aAAa,CAAC,KAAqB,EAAE,SAAyB;IAC5E,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,OAAO;QACL,KAAK,CAAC,IAAI,CAAC,IAAwB;YACjC,MAAM,CAAC,GAAG,SAAS,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC;YACzD,GAAG,EAAE,CAAC;YACN;;4DAEgD;YAChD,MAAM,UAAU,GAAgB,CAAC,CAAC,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBAC/D,IAAI,EAAE,CAAC,CAAC,IAAI;gBACZ,KAAK,EAAE,CAAC,CAAC,IAAI;gBACb,MAAM,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE;gBAChC,QAAQ,EAAE,KAAK;aACS,CAAA,CAAC,CAAC;YAC5B,OAAO;gBACL,IAAI,EAAQ,CAAC,CAAC,IAAI;gBAClB,QAAQ,EAAI,CAAC,CAAC,SAAS,IAAI,GAAG;gBAC9B,SAAS,EAAG,CAAC,CAAC,UAAU,IAAI,EAAE;gBAC9B,KAAK,EAAO,eAAe;gBAC3B,UAAU;gBACV,KAAK;aACN,CAAC;QACJ,CAAC;KACF,CAAC;AACJ,CAAC"}

package/dist/qa/brain_matrix/fixtures/scenario_a11y_announce_intent.json

@@ -0,0 +1,22 @@
+{
+  "id": "a11y_announce_intent",
+  "description": "F107 a11y -- when the user asks the assistant to announce a status to screen-reader users, the brain should pick yujin.a11y.announce, NOT a generic 'show text' verb.",
+  "setup": {
+    "extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.announce [u d] args:{text<required>, politeness?, speak?, locale?}\nUse this for status / error / async completion notifications meant for screen readers + the voice loop. Pizarron is for visual data."
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Anunciale al lector de pantalla que el deploy termino con exito" }
+  ],
+  "invariants": [
+    { "name": "uses_a11y_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } },
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} }
+  ],
+  "fake_responses": [
+    {
+      "text": "Anuncio el deploy.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Deploy completado con exito", "politeness": "polite", "speak": false } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_a11y_announce_no_speak_default.json

@@ -0,0 +1,23 @@
+{
+  "id": "a11y_announce_no_speak_default",
+  "description": "F107 a11y.announce -- by default speak:false. Brain must NOT silently flip speak:true unless the user asks for voice. args_contain_key validates explicit speak passing only when user mentions voice.",
+  "setup": {
+    "extra_system_prompt": "yujin.a11y.announce -- default speak=false. Only set true when the user explicitly mentions voice or TTS in this turn."
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Avisa al lector que la pagina termino de cargar (no necesito voz)" }
+  ],
+  "invariants": [
+    { "name": "uses_a11y_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } },
+    { "name": "args_have_text", "kind": "args_contain_key", "params": { "verb_id": "yujin.a11y.announce", "key": "text" } },
+    { "name": "no_voice_phrase", "kind": "regex_not_match", "params": { "pattern": "speak\\s*[:=]\\s*true|voice|TTS|por\\s+voz" } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Anuncio.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Pagina cargada", "politeness": "polite" } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_a11y_consecutive_announces.json

@@ -0,0 +1,37 @@
+{
+  "id": "a11y_consecutive_announces",
+  "description": "Two-step UX operation: user creates lead + the brain announces success. Should be ONE announce call (not two) + concise reply. min_tool_calls catches missing announce; max_tool_calls catches duplicate.",
+  "setup": {
+    "extra_system_prompt": "BACKEND TOOLS active:\n- app.create_lead [A x] args:{name<required>, email?}\n\nA11Y VERBS active:\n- yujin.a11y.announce [u d] args:{text<required>, politeness?, speak?}\n\nAfter a destructive/approval-required action completes, fire ONE a11y.announce with the result. Never two."
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Crea lead Ana Lopez + anuncia al lector que se creo" },
+    {
+      "kind": "tool_result",
+      "tool_name": "app.create_lead",
+      "is_error": false,
+      "result": { "ok": true, "id": 47812, "name": "Ana Lopez" }
+    }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "calls_create_lead", "kind": "verb_invoked", "params": { "verb_id": "app.create_lead" } },
+    { "name": "calls_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } },
+    { "name": "max_three_calls", "kind": "max_tool_calls", "params": { "max": 3 } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Creo el lead.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "app.create_lead", "args": { "name": "Ana Lopez" } } }
+      ]
+    },
+    {
+      "text": "Anuncio.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Lead Ana Lopez creado con id 47812", "politeness": "polite" } } }
+      ]
+    },
+    { "text": "Listo. Lead id 47812." }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_a11y_focus_after_action.json

@@ -0,0 +1,22 @@
+{
+  "id": "a11y_focus_after_action",
+  "description": "F107 a11y -- after a successful action that opened a new modal, the brain should move focus to the modal's primary element via yujin.a11y.focus.",
+  "setup": {
+    "extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.focus [u d] args:{slug<required>, scroll?, announce?}\nAfter opening a modal / new region, move focus to its first interactive child so keyboard + screen-reader users land in the right place."
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Abri el modal de confirmar deploy y dejame el foco en el boton 'Confirmar' (slug confirm-deploy-btn)" }
+  ],
+  "invariants": [
+    { "name": "uses_a11y_focus", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.focus" } },
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} }
+  ],
+  "fake_responses": [
+    {
+      "text": "Muevo el foco al boton de confirmar.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.focus", "args": { "slug": "confirm-deploy-btn", "scroll": true, "announce": true } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_a11y_label_with_shortcut.json

@@ -0,0 +1,22 @@
+{
+  "id": "a11y_label_with_shortcut",
+  "description": "F107 a11y.label -- when the user dictates a shortcut hint, the brain must include the shortcut key in the args (args_contain_key validates).",
+  "setup": {
+    "extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.label [u d] args:{slug<required>, description<required>, shortcut?, role?}\nInclude shortcut when the user mentions a key combo."
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Etiqueta el boton de guardar (slug save-btn) -- decile 'Guardar archivo' y avisale que Ctrl+S es el atajo" }
+  ],
+  "invariants": [
+    { "name": "uses_a11y_label", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.label" } },
+    { "name": "args_have_shortcut", "kind": "args_contain_key", "params": { "verb_id": "yujin.a11y.label", "key": "shortcut" } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Etiqueto el boton.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.label", "args": { "slug": "save-btn", "description": "Guardar archivo", "shortcut": "Ctrl+S" } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_announce_min_tool_calls.json

@@ -0,0 +1,23 @@
+{
+  "id": "announce_min_tool_calls",
+  "description": "F107 a11y -- when the user asks for an announcement after a known state change, the brain must call at least one verb (a11y.announce). min_tool_calls=1 catches replies that try to do the announcement in chat instead.",
+  "setup": {
+    "extra_system_prompt": "Use yujin.a11y.announce for screen-reader announcements; do NOT just emit a chat reply expecting the user to read it aloud."
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Avisame al lector que el archivo se guardo" }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "calls_at_least_one_verb", "kind": "min_tool_calls", "params": { "min": 1 } },
+    { "name": "uses_a11y_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Anuncio.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Archivo guardado", "politeness": "polite" } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_app_invoke_happy_path.json

@@ -0,0 +1,65 @@
+{
+  "id": "app_invoke_happy_path",
+  "description": "NAC-3 v3.x backend tools happy path: adopter declared app.create_lead with policy=approval-required + side_effects=external-api + params_schema {name:string required, email:string optional}. Brain invokes the verb with valid args, dispatcher echoes back ok:true with response body. Brain summarises the result for the user.",
+  "setup": {
+    "extra_system_prompt": "BACKEND TOOLS (declared in yujin.forge.json -> nac3_backend.tools):\n- app.create_lead [A x] Crear un lead en el CRM. args:{name<string,required>, email<string,optional>}\n\nThese app.* verbs route through nac3_invoke just like yujin.* verbs. When the user asks to create a lead, invoke nac3_invoke({verb_id:'app.create_lead', args:{name:'...', email:'...'}}). Do NOT invent yujin.backend.app.* prefix; the canonical verb_id is app.create_lead.",
+    "pre_history": [
+      { "role": "user", "content": "Tengo nac3_backend configurado con app.create_lead. Cuando te pase un nombre+email, creas el lead." }
+    ],
+    "harness_responses": {
+      "yujin.manifest.read": {
+        "ok": true,
+        "manifest": {
+          "project_slug": "brain-matrix",
+          "elements": [
+            {
+              "id": "app.create_lead",
+              "role": "action",
+              "label": "Crear un nuevo lead en el CRM",
+              "schema": {
+                "type": "object",
+                "properties": {
+                  "name":  { "type": "string", "minLength": 1, "maxLength": 120 },
+                  "email": { "type": "string", "format": "email" }
+                },
+                "required": ["name"]
+              },
+              "policy": "approval-required",
+              "side_effects": ["external-api"]
+            }
+          ]
+        }
+      }
+    }
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Crea un lead para Ana Lopez, email ana@example.com" },
+    {
+      "kind": "tool_result",
+      "tool_name": "app.create_lead",
+      "is_error": false,
+      "result": {
+        "ok": true,
+        "verb_id": "app.create_lead",
+        "status": 201,
+        "response": { "id": 47812, "url": "https://crm.example.com/leads/47812", "name": "Ana Lopez", "email": "ana@example.com" }
+      }
+    }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "uses_app_create_lead", "kind": "verb_invoked", "params": { "verb_id": "app.create_lead" } },
+    { "name": "mentions_created", "kind": "text_contains_any", "params": { "substrs": ["creado", "creada", "Ana", "47812", "lead", "id"] } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Creo el lead.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "app.create_lead", "args": { "name": "Ana Lopez", "email": "ana@example.com" } } }
+      ]
+    },
+    {
+      "text": "Lead creado: Ana Lopez (id 47812). URL: https://crm.example.com/leads/47812."
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_app_no_invent_without_backend.json

@@ -0,0 +1,18 @@
+{
+  "id": "app_no_invent_without_backend",
+  "description": "NAC-3 v3.x backend tools invariant (task #23): when the adopter app does NOT declare nac3_backend.tools[] in yujin.forge.json, the LLM stub has NO app.* verbs. The brain must reply that no backend tools are configured, NOT invent an app.* call. Mirrors mcp_invariant_1_no_verbs but for the backend-tools surface.",
+  "setup": {},
+  "turns": [
+    { "kind": "user_says", "text": "Invoca app.create_invoice del backend de la app" }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "does_not_invoke_unknown_app", "kind": "verb_not_invoked", "params": { "verb_id": "app.create_invoice" } },
+    { "name": "mentions_no_backend", "kind": "text_contains_any", "params": { "substrs": ["nac3_backend", "no tengo", "no está", "no esta", "no hay", "no veo", "no existe", "ningun", "ningún", "manifest", "yujin.forge", "configur", "declar", "endpoint", "necesito", "contexto", "preguntas", "expuesto", "expone", "verbo", "tool"] } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Tu yujin.forge.json no declara la seccion nac3_backend con tools, asi que no tengo verbos app.* disponibles. Agrega nac3_backend.tools[] al manifest con el endpoint, params_schema y auth para que pueda invocar create_invoice."
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_app_schema_fail_recover.json

@@ -0,0 +1,48 @@
+{
+  "id": "app_schema_fail_recover",
+  "description": "NAC-3 v3.x backend tools schema-fail recovery: brain invokes app.create_lead without the required `name` field. Dispatcher returns {ok:false, error:'Args do not match the app tool schema: missing required arg: name', schema:{...}}. Brain MUST read the error + retry with the missing field, NOT pretend it worked. Validates the brain reads structured dispatcher errors.",
+  "setup": {
+    "extra_system_prompt": "BACKEND TOOLS (declared in yujin.forge.json -> nac3_backend.tools):\n- app.create_lead [A x] Crear un lead. args:{name<string,required>, email<string,optional>}\n\nThese app.* verbs route through nac3_invoke. When a dispatcher tool_result returns ok:false with a schema error, READ the error + ASK the user for the missing fields. NEVER claim the lead was created if the call failed.",
+    "pre_history": [
+      { "role": "user", "content": "Quiero crear leads en mi CRM usando app.create_lead." }
+    ]
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Crea un lead con email bruno@example.com" },
+    {
+      "kind": "tool_result",
+      "tool_name": "app.create_lead",
+      "is_error": false,
+      "result": {
+        "ok": false,
+        "error": "Args do not match the app tool schema: missing required arg: name",
+        "verb_id": "app.create_lead",
+        "schema": {
+          "type": "object",
+          "properties": {
+            "name":  { "type": "string", "description": "Lead full name" },
+            "email": { "type": "string", "description": "Contact email" }
+          },
+          "required": ["name"]
+        },
+        "policy": "approval-required"
+      }
+    }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "asks_for_name", "kind": "text_contains_any", "params": { "substrs": ["nombre", "name", "falta", "missing", "completo", "Cual", "Cuál", "como se llama"] } },
+    { "name": "no_pretend_success", "kind": "text_excludes", "params": { "substr": "creado" } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Voy a crear el lead.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "app.create_lead", "args": { "email": "bruno@example.com" } } }
+      ]
+    },
+    {
+      "text": "El backend rechazo la llamada porque falta el campo `name` (requerido en el schema). Pasame el nombre completo del lead y lo creo."
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_approval_modal_no_redundancy.json

@@ -0,0 +1,21 @@
+{
+  "id": "approval_modal_no_redundancy",
+  "description": "Destructive verb (yujin.git.commit) -- LLM must invoke + emit ONE short summary, NOT ask 'confirmas?' in chat (R8).",
+  "setup": {},
+  "turns": [
+    { "kind": "user_says", "text": "commiteame aaa.txt con mensaje 'chore: add aaa'" }
+  ],
+  "invariants": [
+    { "name": "invokes_commit", "kind": "verb_invoked", "params": { "verb_id": "yujin.git.commit" } },
+    { "name": "no_redundant_confirm", "kind": "no_redundant_confirm", "params": {} },
+    { "name": "no_invent", "kind": "no_invent_verb", "params": {} }
+  ],
+  "fake_responses": [
+    {
+      "text": "Voy a hacer git commit -m 'chore: add aaa' incluyendo aaa.txt; revisa el modal.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.git.commit", "args": { "message": "chore: add aaa", "files": ["aaa.txt"] } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_args_redacted_in_audit.json

@@ -0,0 +1,17 @@
+{
+  "id": "args_redacted_in_audit",
+  "description": "User asks if their API keys or passwords end up in the audit log. The brain must say NO, that args are redacted via standard secret patterns (api_key, token, password, secret) before logging, and that the user vault keys never leave the encrypted store. Validates the brain knows the redactArgs privacy contract.",
+  "setup": {},
+  "turns": [
+    { "kind": "user_says", "text": "Si invoco un verb que recibe una API key, esa key queda escrita en el audit.log?" }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "explains_redaction", "kind": "text_contains_any", "params": { "substrs": ["redact", "Redact", "ocult", "censur", "REDACT", "no", "No", "NO", "patron", "vault"] } }
+  ],
+  "fake_responses": [
+    {
+      "text": "No, las args con patrones api_key, token, password, secret se redactan a `[REDACTED]` antes de escribir el audit log. Y las keys que tenes en el vault nunca salen del store encriptado: las leo solo en memoria para el call. Si encontras un caso en el log donde una key quedo plaintext, es un bug -- avisame."
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_audit_consumers_intent.json

@@ -0,0 +1,19 @@
+{
+  "id": "audit_consumers_intent",
+  "description": "When the user asks 'que consumio del Vault hoy' / 'que tools tocaron mi key de anthropic', the brain uses yujin.audit.consumers (which reads the audit log without re-exposing secrets) NOT fs.read directly.",
+  "setup": {},
+  "turns": [
+    { "kind": "user_says", "text": "Mostrame que tools accedieron a mi anthropic key hoy." }
+  ],
+  "invariants": [
+    { "name": "uses_audit_consumers", "kind": "verb_invoked", "params": { "verb_id": "yujin.audit.consumers" } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Consulto el audit log.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.audit.consumers", "args": { "slot": "anthropic" } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_audit_consumers_intent_en.json

@@ -0,0 +1,19 @@
+{
+  "id": "audit_consumers_intent_en",
+  "description": "English variant of audit.consumers intent: 'what touched my anthropic key today?' uses yujin.audit.consumers.",
+  "setup": {},
+  "turns": [
+    { "kind": "user_says", "text": "Show me which tools accessed my Anthropic key today." }
+  ],
+  "invariants": [
+    { "name": "uses_audit_consumers", "kind": "verb_invoked", "params": { "verb_id": "yujin.audit.consumers" } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Checking the audit log.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.audit.consumers", "args": { "slot": "anthropic" } } }
+      ]
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_audit_log_on_destructive.json

@@ -0,0 +1,18 @@
+{
+  "id": "audit_log_on_destructive",
+  "max_tool_rounds": 8,
+  "description": "The user asks where their audit trail lives. The brain should answer with the audit log location/path, not invent or just say 'check the panel'. Validates the brain knows about audit_log + the panel viewer that shipped in alpha.59g. Behavioural query, no tool invocation expected.",
+  "setup": {},
+  "turns": [
+    { "kind": "user_says", "text": "Donde se guarda el audit log de las acciones destructive que hice este mes?" }
+  ],
+  "invariants": [
+    { "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
+    { "name": "mentions_audit_location", "kind": "text_contains_any", "params": { "substrs": ["audit.log", "audit_log", ".yujin-forge", "panel", "Audit", "auditoria", "viewer", "log"] } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Todo el audit trail vive en ~/.yujin-forge/audit.log y se ve formateado desde el panel en la solapa Audit log (alpha.59g). Cada destructive + approval-required deja una entrada con trace_id, verb_id, args redactados, side_effects declaradas + reales y duration_ms."
+    }
+  ]
+}

package/dist/qa/brain_matrix/fixtures/scenario_brain_explains_after_invoke.json

@@ -0,0 +1,28 @@
+{
+  "id": "brain_explains_after_invoke",
+  "description": "After invoking a verb, the brain's reply must include actual prose explaining the result -- not just bare 'ok' or 'listo'. User-facing context, not just side effects.",
+  "setup": {
+    "harness_responses": {
+      "yujin.fs.stat": { "path": "test.txt", "size_bytes": 42, "kind": "file", "mtime_iso": "2026-06-04T00:00:00Z" }
+    }
+  },
+  "turns": [
+    { "kind": "user_says", "text": "Cuanto pesa el archivo test.txt?" }
+  ],
+  "invariants": [
+    { "name": "uses_stat",            "kind": "verb_invoked",     "params": { "verb_id": "yujin.fs.stat" } },
+    { "name": "explains_with_prose",  "kind": "min_text_length",  "params": { "min": 30 } },
+    { "name": "mentions_size",        "kind": "text_contains_any", "params": { "substrs": ["bytes", "42", "tamaño", "pesa", "peso", "kb", "kB"] } }
+  ],
+  "fake_responses": [
+    {
+      "text": "Consulto el tamaño.",
+      "tool_calls": [
+        { "name": "nac3_invoke", "args": { "verb_id": "yujin.fs.stat", "args": { "path": "test.txt" } } }
+      ]
+    },
+    {
+      "text": "El archivo test.txt pesa 42 bytes (modificado 2026-06-04). Es un archivo regular."
+    }
+  ]
+}