promptup-plugin 0.1.6 → 0.1.8

package/dist/evaluator.js

@@ -180,7 +180,7 @@ ${convo}
 Return ONLY valid JSON with no markdown formatting, no code fences, no extra text:
 {"dimensions":[{"key":"task_decomposition","score":0,"reasoning":"..."},{"key":"prompt_specificity","score":0,"reasoning":"..."},{"key":"output_validation","score":0,"reasoning":"..."},{"key":"iteration_quality","score":0,"reasoning":"..."},{"key":"strategic_tool_usage","score":0,"reasoning":"..."},{"key":"context_management","score":0,"reasoning":"..."}],"domain_dimensions":[{"key":"architectural_awareness","score":0,"reasoning":"..."},{"key":"error_anticipation","score":0,"reasoning":"..."},{"key":"technical_vocabulary","score":0,"reasoning":"..."},{"key":"dependency_reasoning","score":0,"reasoning":"..."},{"key":"tradeoff_articulation","score":0,"reasoning":"..."}],"tech_expertise":[{"roadmap":"typescript","score":75,"competencies":{"type_system":80,"generics":70}}],"recommendations":[{"dimension_key":"...","priority":"high","recommendation":"Add context to prompts","suggestions":["Instead of 'no', try 'no — terminal shows nothing after response'","Instead of 'yep', try 'yes, use the Stop hook approach'"]}],"activity_log":["Did X","Did Y","Fixed Z"],"decisions":[{"type":"steer","summary":"Chose bcrypt over argon2 — simpler dependency","signal":"high"},{"type":"validate","summary":"Ran integration tests after auth implementation","signal":"medium"}]}`;
 }
-function runClaudeCode(prompt, timeoutMs = 120_000) {
+function runClaudeCode(prompt, timeoutMs = 180_000) {
     return new Promise((resolve, reject) => {
         // Strip CLAUDECODE env var to allow spawning from within a Claude Code session
         const env = { ...process.env };
@@ -196,7 +196,7 @@ function runClaudeCode(prompt, timeoutMs = 120_000) {
         proc.stderr.on('data', (chunk) => { stderr += chunk.toString(); });
         const timer = setTimeout(() => {
             proc.kill('SIGTERM');
-            reject(new Error(`Claude Code timed out after ${timeoutMs}ms`));
+            reject(new Error(`[timeout] Claude Code timed out after ${timeoutMs}ms (prompt size: ${prompt.length} chars)`));
         }, timeoutMs);
         proc.on('close', (code) => {
             clearTimeout(timer);
@@ -204,16 +204,22 @@ function runClaudeCode(prompt, timeoutMs = 120_000) {
                 resolve(stdout.trim());
             }
             else {
-                reject(new Error(`Claude Code exited with code ${code}: ${stderr.slice(0, 500)}`));
+                reject(new Error(`[exit] Claude Code exited with code ${code}: ${stderr.slice(0, 1000)}`));
             }
         });
         proc.on('error', (err) => {
             clearTimeout(timer);
-            reject(err);
+            reject(new Error(`[spawn] Could not start claude: ${err.message}`));
         });
-        // Write prompt to stdin and close
-        proc.stdin.write(prompt);
-        proc.stdin.end();
+        // Write prompt to stdin with backpressure handling
+        const ok = proc.stdin.write(prompt);
+        if (!ok) {
+            // Buffer is full — wait for drain before closing
+            proc.stdin.once('drain', () => { proc.stdin.end(); });
+        }
+        else {
+            proc.stdin.end();
+        }
     });
 }
 function parseClaudeResponse(raw) {
@@ -250,9 +256,11 @@ export async function evaluateSession(sessionId, messages, triggerType, weightPr
     let recommendations = [];
     let usedClaude = false;
     try {
-        console.log(`[eval] Running Claude Code evaluation for session ${sessionId.slice(0, 8)}...`);
         const prompt = buildEvalPrompt(messages);
-        const rawOutput = await runClaudeCode(prompt);
+        // Scale timeout: 180s base + 1s per message over 100
+        const timeoutMs = 180_000 + Math.max(0, messages.length - 100) * 1000;
+        console.log(`[eval] Running Claude evaluation for ${sessionId.slice(0, 8)} (${messages.length} msgs, ${prompt.length} chars, timeout ${Math.round(timeoutMs / 1000)}s)...`);
+        const rawOutput = await runClaudeCode(prompt, timeoutMs);
         const result = parseClaudeResponse(rawOutput);
         usedClaude = true;
         // Store structured data in raw_evaluation (activity log + decisions + raw text)
@@ -340,7 +348,14 @@ export async function evaluateSession(sessionId, messages, triggerType, weightPr
         console.log(`[eval] Claude Code evaluation complete for ${sessionId.slice(0, 8)}`);
     }
     catch (err) {
-        console.warn(`[eval] Claude Code unavailable, using heuristic fallback:`, err.message);
+        const msg = err.message || String(err);
+        const category = msg.startsWith('[timeout]') ? 'TIMEOUT'
+            : msg.startsWith('[spawn]') ? 'SPAWN_FAILED'
+            : msg.startsWith('[exit]') ? 'PROCESS_ERROR'
+            : msg.includes('No JSON object found') ? 'PARSE_FAILED'
+            : msg.includes('Missing dimensions') ? 'INVALID_RESPONSE'
+            : 'UNKNOWN';
+        console.warn(`[eval] Claude failed (${category}), using heuristic fallback: ${msg}`);
         // Fall back to heuristic — generate basic activity log from messages
         const heuristic = heuristicEvaluate(messages, profile);
         dimensionScores = heuristic.dimensionScores;
@@ -445,10 +460,23 @@ function heuristicEvaluate(messages, profile) {
         if (!def)
             continue;
         const next = def.ranges.find(r => r.min > dim.score);
+        const tipMap = {
+            task_decomposition: 'Break your next task into 2-3 explicit steps before starting',
+            prompt_specificity: 'Add one constraint or example to your next prompt',
+            output_validation: 'Check one output against your expectation before moving on',
+            iteration_quality: 'When something doesn\'t work, name what failed before retrying',
+            strategic_tool_usage: 'Try a different tool or approach for your next task',
+            context_management: 'Summarize where you are before switching topics',
+            architectural_awareness: 'Name one system-level concern before making a change',
+            error_anticipation: 'Ask "what could break?" once before implementing',
+            technical_vocabulary: 'Use the precise term for what you\'re describing',
+            dependency_reasoning: 'Trace one data flow before changing it',
+            tradeoff_articulation: 'Name one alternative you considered and why you didn\'t pick it',
+        };
         recommendations.push({
             dimension_key: dim.key,
             priority: dim.score < 35 ? 'high' : dim.score < 55 ? 'medium' : 'low',
-            recommendation: next ? `Aim for: ${next.description}` : 'Continue current approach',
+            recommendation: tipMap[dim.key] || (next ? `Build toward: ${next.label}` : 'Continue current approach'),
             suggestions: def.signals.slice(0, 2),
         });
     }

package/dist/pr-report-generator.js

@@ -254,10 +254,7 @@ export async function generatePRReport(options) {
     // We need the repo for the cache key — get it first
     const ghAvailable = await checkGhAvailable();
     const repo = ghAvailable ? await getRepo(projectPath) : '';
-    const cached = getPRReportByBranch(branch, repo);
-    if (cached) {
-        return { report: cached, isNew: false };
-    }
+    // Always regenerate — no cache. Scores evolve as more evals run.
     // 3. Get PR info
     let prInfo = null;
     if (ghAvailable) {
@@ -295,19 +292,8 @@ export async function generatePRReport(options) {
             }
         }
     }
-    // 6. Gather decisions
-    const decisions = gatherDecisions(sessionIds);
-    // 7. Compute DQS — use validate decisions as proxy for validation rate
-    const validateCount = decisions.filter(d => d.type === 'validate').length;
-    const validationRate = decisions.length > 0 ? validateCount / decisions.length : 0;
-    const dqs = computeDQS(decisions, validationRate);
-    // 8. Build decision breakdown
-    const breakdown = {};
-    for (const d of decisions) {
-        const t = d.type;
-        breakdown[t] = (breakdown[t] ?? 0) + 1;
-    }
-    // 9. Auto-eval sessions that haven't been evaluated yet
+    // 6. Auto-eval sessions FIRST so decisions get extracted before DQS
+    //    This makes /pr-report self-contained — no need to run /eval first
     //    This makes /pr-report self-contained — no need to run /eval first
     for (const sid of sessionIds) {
         const existingEval = getLatestEvaluation(sid);
@@ -333,6 +319,18 @@ export async function generatePRReport(options) {
             }
         }
     }
+    // 7. Now gather decisions (AFTER auto-eval extracted them)
+    const decisions = gatherDecisions(sessionIds);
+    // 8. Compute DQS
+    const validateCount = decisions.filter(d => d.type === 'validate').length;
+    const validationRate = decisions.length > 0 ? validateCount / decisions.length : 0;
+    const dqs = computeDQS(decisions, validationRate);
+    // 9. Build decision breakdown
+    const breakdown = {};
+    for (const d of decisions) {
+        const t = d.type;
+        breakdown[t] = (breakdown[t] ?? 0) + 1;
+    }
     // 10. Fetch evaluations (averaged across all evals) + message counts
     let compositeScore = null;
     let dimensionScores;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "promptup-plugin",
-  "version": "0.1.6",
+  "version": "0.1.8",
   "description": "AI coding skill evaluator for Claude Code — 11-dimension scoring, decision intelligence, PR reports",
   "type": "module",
   "main": "./dist/index.js",