@kognai/orchestrator-core 0.2.9 → 0.2.11

package/dist/lib/engine-coding-agent.js

@@ -49,7 +49,11 @@ class CodingAgent {
     constructor(name, systemPrompt) { this.name = name; this.systemPrompt = systemPrompt; }
     async execute(task, previousReview) {
         (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.cyan, `\n[${this.name}] Executing: ${task.id} (${task.priority})`);
-        const deliverables = [...(task.deliverables.code || []), ...(task.deliverables.tests || []), ...(task.deliverables.docs || [])];
+        // TICKET-353: TEST-FIRST ordering. Generate paired tests BEFORE their code so
+        // the test is authored from the SPEC (contract-first), frozen, then the code
+        // is written to pass it — instead of one model writing code+test together and
+        // letting both agree on the same wrong understanding (#2 separate author).
+        const deliverables = [...(task.deliverables.tests || []), ...(task.deliverables.code || []), ...(task.deliverables.docs || [])];
         // Complexity-aware model routing:
         // Claude Sonnet → complex tasks (many files, complex keywords, large existing files)
         // MiniMax M2.5  → simple tasks (small edits, config, stubs) + truncation retry as safety net
@@ -233,15 +237,31 @@ class CodingAgent {
             const existingContent = (0, fs_1.existsSync)(filepath)
                 ? `\n\n## EXISTING FILE — SURGICAL EDIT ONLY\nDo NOT rewrite the entire file. Output the COMPLETE updated file with your changes merged in.\nIf you add a function, append it. If you edit a line, change only that line.\nFile has ${existingLines} lines — preserve ALL existing code.\n\n### Current Content\n\`\`\`typescript\n${(0, fs_1.readFileSync)(filepath, 'utf-8').substring(0, 3000)}\n\`\`\`\n`
                 : `\n\n## Note: This is a NEW file — create it from scratch.\n`;
+            // TICKET-353 (#1 examples + #2 separate author): for a test file, you ARE
+            // the test author and this test DEFINES correctness. The implementation is
+            // written separately to PASS this test and cannot change it.
             const testConstraint = isTestFile
-                ? `\n\n## CRITICAL: TEST FILE SIZE LIMIT
-This is a test file. You MUST keep it SHORT to avoid truncation:
-- Maximum 5-6 test cases (describe + it blocks)
-- Maximum 80 lines total
-- NO verbose setup — use inline mocks
-- NO redundant tests — one test per behavior
-- Cover: happy path, error case, edge case, defaults — that's it
-- If you write more than 80 lines, the file WILL be truncated and REJECTED\n`
+                ? `\n\n## YOU ARE THE TEST AUTHOR — this test DEFINES correctness
+The implementation file this pairs with will be written SEPARATELY to PASS this test and CANNOT modify it. So this test is the CONTRACT.
+- Derive 3–5 CONCRETE golden input→output examples from the spec/description above and encode them as assertions with EXACT expected values (do the arithmetic yourself; don't assert on whatever the code returns).
+- Use Node's built-in runner: \`import { test } from 'node:test'\` and \`import assert from 'node:assert/strict'\`.
+- Import the real symbols from the implementation file (correct relative path) and exercise ACTUAL behavior: happy path, an error/edge case, and any invariant ("never exceeds capacity").
+- It WILL be executed; assertions must reflect the SPEC's truth, not a guess. Keep it ≤80 lines, no verbose setup.\n`
+                : '';
+            // TICKET-353 (#2): author tests with a stronger model than the coder
+            // default; code files use the assessed/escalated coder model.
+            const TEST_AUTHOR_MODEL = process.env.KOGNAI_TEST_AUTHOR_MODEL || 'anthropic/claude-sonnet-4.6';
+            const fileProvider = isTestFile ? 'clawrouter' : provider;
+            const fileModel = isTestFile ? TEST_AUTHOR_MODEL : model;
+            if (isTestFile)
+                (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.magenta, `  ✎ [TEST-AUTHOR] ${filepath} via ${fileModel} (contract-first, examples)`);
+            // For a CODE file, inject its already-authored paired test (tests-first
+            // ordering placed it in createdFiles) as the frozen contract to satisfy.
+            const pairedTest = !isTestFile
+                ? createdFiles.find(f => f.path === filepath.replace(/\.([tj]sx?)$/, '.test.$1'))
+                : undefined;
+            const contractBlock = pairedTest
+                ? `\n\n## CONTRACT — your code MUST make this test pass (already written; DO NOT modify it; it defines correctness):\n\`\`\`typescript\n${pairedTest.content.substring(0, 3000)}\n\`\`\`\n`
                 : '';
             // EXACT CONTENT mode: task description contains code block(s) with the exact file content.
             // Extract them deterministically and bypass LLM to prevent model hallucination.
@@ -270,7 +290,7 @@ ${task.context}
 ${fileList}
 
 ## Generate ONLY: ${filepath}
-${existingContent}${priorCtx}${testConstraint}
+${existingContent}${priorCtx}${testConstraint}${contractBlock}
 Write ONLY the content for "${filepath}". Rules:
 - S64-001: Output the raw file content using FILE: format as described in the system prompt
 - Do NOT wrap output in markdown code fences (\`\`\`) — for .md files especially, output RAW markdown text, NOT inside a \`\`\`markdown or \`\`\`typescript block
@@ -281,7 +301,7 @@ Write ONLY the content for "${filepath}". Rules:
 - No explanatory text — output file content only`;
             try {
                 const startTime = Date.now();
-                const response = await (0, orchestrate_engine_1.callLLM)(provider, model, this.systemPrompt, userPrompt, 480000, this.name, task.id); // 8 min — qwen3:14b needs time for large files
+                const response = await (0, orchestrate_engine_1.callLLM)(fileProvider, fileModel, this.systemPrompt, userPrompt, 480000, this.name, task.id); // 8 min — qwen3:14b needs time for large files
                 const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
                 let content = response.choices?.[0]?.message?.content || '';
                 // Strip MiniMax <think>...</think> tags that leak into responses
@@ -358,7 +378,7 @@ ${fileContent.substring(fileContent.length - 1500)}
 Continue EXACTLY from where it left off and output ONLY the remaining code (no duplicated content, no preamble, no markdown fences). End the file with its final closing brace.`;
                     let grew = false;
                     try {
-                        const contResponse = await (0, orchestrate_engine_1.callLLM)(provider, model, this.systemPrompt, continuationPrompt, 120000, this.name, `${task.id}_continuation_${chunk}`);
+                        const contResponse = await (0, orchestrate_engine_1.callLLM)(fileProvider, fileModel, this.systemPrompt, continuationPrompt, 120000, this.name, `${task.id}_continuation_${chunk}`);
                         let contContent = contResponse.choices?.[0]?.message?.content || '';
                         contContent = contContent.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
                         const contBlocks = this.extractCodeBlocks(contContent);

package/dist/lib/engine-primitives.js

@@ -454,9 +454,13 @@ function mapLegacyToComplexity(provider, model) {
     // Anthropic Haiku → EXEC (cloud, not constitutional)
     if (provider === 'anthropic' && model.includes('haiku'))
         return 'exec';
-    // ClawRouter DeepSeek → EXEC
+    // TICKET-350 honest routing: routine clawrouter calls → EXEC (a cheap+fast
+    // model run DIRECTLY by the router); escalated/hard ones (sonnet/opus) → APEX
+    // (T3, the OpenClaw gateway = opus). Previously ALL clawrouter calls collapsed
+    // to EXEC and the gateway then flattened them to opus/main regardless — so the
+    // "deepseek" label AND the sonnet escalation were cosmetic (everything was opus).
     if (provider === 'clawrouter')
-        return 'exec';
+        return /sonnet|opus/i.test(model) ? 'apex' : 'exec';
     // MiniMax → EXEC (cloud)
     if (provider === 'minimax')
         return 'exec';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kognai/orchestrator-core",
-  "version": "0.2.9",
+  "version": "0.2.11",
   "description": "Kognai sovereign orchestrator — core engine (template-agnostic). Shared by all products (Kognai/coding, Voxight/market-intel, Invoica/fin-compliance); each supplies only its template. Replaces per-repo forks of orchestrate-agents-v2 / sprint-runner / lib.",
   "license": "MIT",
   "author": "SkinGem",