max_tokens per vaihe: manageri 200, koodari 512, testaaja 200, QA 512, DevOps 256

Hub ja natiivisolmu tukevat nyt max_tokens-kenttää API-pyynnöissä.
Pipeline-vaiheet käyttävät sopivan kokoisia token-rajoja.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

network-poc/hub/src/main.rs

@@ -947,6 +947,8 @@ struct ChatCompletionRequest {
     model: String,
     prompt: String,
     task_id: String,
+    #[serde(default)]
+    max_tokens: Option<u64>,
 }
 
 #[derive(serde::Serialize)]
@@ -1069,12 +1071,15 @@ async fn api_chat_completions(
     state.node_busy.lock().unwrap().insert(target_node_id);
     state.pending_task_ids.lock().unwrap().insert(payload.task_id.clone());
 
-    let msg = serde_json::json!({
+    let mut msg = serde_json::json!({
         "type": "llm_prompt",
         "prompt": payload.prompt,
         "model": payload.model,
         "task_id": payload.task_id,
     });
+    if let Some(mt) = payload.max_tokens {
+        msg.as_object_mut().unwrap().insert("max_tokens".to_string(), serde_json::json!(mt));
+    }
 
     // Odotuskanava valmiiksi (solmu palauttaa tuloksen stats_tx kautta)
     let mut rx = state.stats_tx.subscribe();

network-poc/native-node/src/main.rs

@@ -326,9 +326,10 @@ async fn main() {
 
                                     if let Some(ref mut engine) = llm {
                                         busy = true;
-                                        tracing::info!("Generoidaan (task_id: {}): \"{}\"", task_id, prompt);
+                                        let max_tokens = task.get("max_tokens").and_then(|v| v.as_u64()).unwrap_or(512) as usize;
+                                        tracing::info!("Generoidaan (task_id: {}, max_tokens: {}): \"{}\"", task_id, max_tokens, prompt);
 
-                                        match engine.generate(prompt, 512) {
+                                        match engine.generate(prompt, max_tokens) {
                                             Ok(result) => {
                                                 tracing::info!(
                                                     "Tulos: {} tokenia | {:.0}ms | {:.1} tok/s | \"{}\"",

network-poc/static/index.html

@@ -1767,7 +1767,7 @@
         const activeStreams = {};
 
         // Lähettää promptin mallille ja palauttaa vastauksen (tai null virhetilanteessa)
-        async function kpnRun(model, prompt, silent) {
+        async function kpnRun(model, prompt, silent, maxTokens) {
             const taskId = crypto.randomUUID();
             // Yksittäinen status-rivi jota päivitetään läpi pyynnön elinkaaren
             const statusDiv = document.createElement('div');
@@ -1801,7 +1801,7 @@
                 const res = await fetch('/api/v1/chat/completions', {
                     method: 'POST',
                     headers: { 'Content-Type': 'application/json' },
-                    body: JSON.stringify({ model, prompt: fullPrompt, task_id: taskId }),
+                    body: JSON.stringify({ model, prompt: fullPrompt, task_id: taskId, ...(maxTokens ? { max_tokens: maxTokens } : {}) }),
                 });
 
                 if (!res.ok) {
@@ -2066,7 +2066,7 @@ CONSTRAINTS — the coder can only generate ~400 tokens per file:
 - Prefer fewer, focused files over many small ones
 
 Project: ${task}`;
-            const plan = await kpnRun(agentPrompts.manager.model, managerPrompt);
+            const plan = await kpnRun(agentPrompts.manager.model, managerPrompt, false, 200);
             if (!plan) { termLog('  ✗ Pipeline keskeytyi (manageri)', '#f85149'); return; }
             pipelineStep('manager', 'Suunnittelu', 'done', task, plan);
 
@@ -2155,7 +2155,7 @@ IMPORTANT: Keep the code SHORT and focused. Max ~50 lines. No comments, no docst
 If the code is correct, say "LGTM".
 
 ${allCode}`;
-            const review = await kpnRun(agentPrompts.tester.model, reviewPrompt);
+            const review = await kpnRun(agentPrompts.tester.model, reviewPrompt, false, 200);
             pipelineStep('tester', 'Review', 'done', `${Object.keys(generatedFiles).length} tiedostoa`, review);
 
             // Vaihe 4: Korjausluuppi — jos testaaja löysi ongelmia
@@ -2174,7 +2174,7 @@ Write the corrected code.`;
                 if (fixedCode) {
                     termLog(`\n<span style="color:#58a6ff;font-weight:bold">[${fileList.length + 4}] Testaaja</span> — uudelleenarviointi`);
                     pipelineStep('tester', 'Re-review', 'active', fixedCode);
-                    const reReview = await kpnRun(agentPrompts.tester.model, `Review the corrected code briefly:\n${fixedCode}`);
+                    const reReview = await kpnRun(agentPrompts.tester.model, `Review the corrected code briefly:\n${fixedCode}`, false, 128);
                     pipelineStep('tester', 'Re-review', 'done', fixedCode, reReview);
                 }
             }
@@ -2186,7 +2186,7 @@ Write the corrected code.`;
             const qaPrompt = `Write a short test file (test_app.py) for this project. Use pytest. Max 3 test functions. Keep it minimal.
 
 ${Object.entries(generatedFiles).map(([n, c]) => `--- ${n} ---\n${c}`).join('\n\n')}`;
-            const tests = await kpnRun(agentPrompts.qa.model, qaPrompt);
+            const tests = await kpnRun(agentPrompts.qa.model, qaPrompt, false, 512);
             if (tests) generatedFiles['test_app.py'] = tests;
             pipelineStep('qa', 'Testit', 'done', 'test_app.py', tests);
 
@@ -2202,7 +2202,7 @@ ${Object.entries(generatedFiles).map(([n, c]) => `--- ${n} ---\n${c}`).join('\n\
 Max 15 lines. No badges, no license, no contributing section.
 
 Files: ${Object.keys(generatedFiles).join(', ')}`;
-            const readme = await kpnRun(agentPrompts.tester.model, devopsPrompt);
+            const readme = await kpnRun(agentPrompts.tester.model, devopsPrompt, false, 256);
             if (readme) generatedFiles['README.md'] = readme;
             pipelineStep('tester', 'DevOps', 'done', 'README.md', readme);