@ddtcorex/dsh-maestro-supervisor 0.6.6 → 0.6.8

package/lib/debug-agent.js

@@ -217,20 +217,40 @@ function defaultWriteFile(p, c) {
     writeFileSync(p, c, 'utf-8');
 }
 async function defaultDryBoot() {
+    const { execSync } = await import('node:child_process');
+    const { resolveDeepseekHarnessDir } = await import('./paths.js');
+    const { buildKillStalePortsCommand } = await import('./restart-guards.js');
+    const port = Math.floor(19000 + Math.random() * 1000);
+    let tmp = '';
     try {
-        const { execSync } = await import('node:child_process');
-        const { resolveDeepseekHarnessDir } = await import('./paths.js');
         const deepseekDir = resolveDeepseekHarnessDir();
-        const tmp = execSync('mktemp -d', { encoding: 'utf-8' }).trim();
-        const port = Math.floor(19000 + Math.random() * 1000);
+        tmp = execSync('mktemp -d', { encoding: 'utf-8' }).trim();
         // Use spawn-based dry-boot to avoid nested quoting hell (prev \\$! / \\$(seq) caused syntax error)
-        const out = execSync(`timeout 8 bash -c 'DSH_HOME=${tmp} pnpm --dir ${deepseekDir} dsh web --port ${port} --no-open >${tmp}/dsh.log 2>&1 & pid=$!; for i in $(seq 1 5); do sleep 1; if curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:${port}/ 2>&1 | grep -q 200; then kill $pid 2>/dev/null || true; wait $pid 2>/dev/null || true; echo ok; exit 0; fi; done; kill $pid 2>/dev/null || true; wait $pid 2>/dev/null || true; echo fail; exit 1'`, { encoding: 'utf-8', timeout: 12000 });
-        execSync(`rm -rf ${tmp}`);
+        const out = execSync(`timeout 8 bash -c 'DSH_HOME=${tmp} pnpm --dir ${deepseekDir} dsh web --port ${port} --no-open >${tmp}/dsh.log 2>&1 & pid=$!; for i in $(seq 1 5); do sleep 1; if curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:${port}/ 2>&1 | grep -q 200; then echo ok; exit 0; fi; done; echo fail; exit 1'`, { encoding: 'utf-8', timeout: 12000 });
         return out.includes('ok');
     }
     catch {
         return false;
     }
+    finally {
+        // `pnpm dsh web` is a pnpm→sh→node tree (see AGENTS.md Known Issues); the
+        // backgrounded job's `$!` above is only the pnpm wrapper, so `kill $pid`
+        // never reached the real node process — it kept running detached,
+        // holding the ephemeral port. Confirmed live 2026-08-31: 3 failed
+        // attempts left 3 orphaned "MainThread" node processes eating 6.4G RAM
+        // for 7+ hours. Resolve the real pid by the port it is actually
+        // listening on instead (same technique as the live-restart kill step).
+        try {
+            execSync(buildKillStalePortsCommand([port]), { timeout: 5000, stdio: 'pipe' });
+        }
+        catch { }
+        if (tmp) {
+            try {
+                execSync(`rm -rf ${tmp}`);
+            }
+            catch { }
+        }
+    }
 }
 async function defaultFetchLLM(prompt) {
     const cfg = await resolveLLMConfig();
@@ -248,6 +268,7 @@ async function defaultFetchLLM(prompt) {
             body: JSON.stringify({
                 model: cfg.model,
                 temperature: 0.2,
+                ...(cfg.reasoningEffort ? { reasoning_effort: cfg.reasoningEffort, reasoningEffort: cfg.reasoningEffort } : {}),
                 messages: [
                     { role: 'system', content: 'You are a systematic-debugging agent for DSH Web resilience. Follow the 4 phases: root cause, pattern analysis, hypothesis, implementation. Always propose minimal single-file fix.' },
                     { role: 'user', content: prompt },
@@ -292,7 +313,7 @@ async function defaultFetchLLM(prompt) {
                 { role: 'user', content: prompt },
             ],
             max_output_tokens: 4000,
-            reasoning: { effort: 'low' },
+            ...(cfg.reasoningEffort ? { reasoning: { effort: cfg.reasoningEffort } } : { reasoning: { effort: 'low' } }),
         }),
     });
     if (!res2.ok)
@@ -319,7 +340,9 @@ async function resolveLLMConfig() {
     // Model: AI_MODEL / DEEPSEEK_MODEL / settings.json domains.supervisor.model -> domains.review.model / default
     // Supervisor has its own picker; falls back to review model for backward compat, then DSH default.
     let model = process.env.AI_MODEL ?? process.env.DEEPSEEK_MODEL ?? process.env.OPENAI_MODEL ?? null;
-    if (!model) {
+    // reasoningEffort: env first, then supervisor -> review, then settings.yaml fallback
+    let reasoningEffort = (process.env.AI_REASONING_EFFORT ?? process.env.REASONING_EFFORT ?? null) ?? undefined;
+    if (!model || !reasoningEffort) {
         try {
             const { readFileSync } = await import('node:fs');
             const { homedir } = await import('node:os');
@@ -327,24 +350,57 @@ async function resolveLLMConfig() {
             const raw = readFileSync(settingsPath, 'utf-8');
             const j = JSON.parse(raw);
             // Prefer supervisor model, fall back to review model (so old installs keep working)
-            const sup = j?.domains?.supervisor?.model?.model ?? j?.domains?.supervisor?.model;
-            const rev = j?.domains?.review?.model?.model ?? j?.domains?.review?.model;
-            let m = null;
-            if (typeof sup === 'string')
-                m = sup;
-            else if (sup?.model && typeof sup.model === 'string')
-                m = sup.model;
-            else if (typeof rev === 'string')
-                m = rev;
-            else if (rev?.model && typeof rev.model === 'string')
-                m = rev.model;
-            if (typeof m === 'string' && m.trim() !== '')
-                model = m;
+            if (!model) {
+                const sup = j?.domains?.supervisor?.model?.model ?? j?.domains?.supervisor?.model;
+                const rev = j?.domains?.review?.model?.model ?? j?.domains?.review?.model;
+                let m = null;
+                if (typeof sup === 'string')
+                    m = sup;
+                else if (sup?.model && typeof sup.model === 'string')
+                    m = sup.model;
+                else if (typeof rev === 'string')
+                    m = rev;
+                else if (rev?.model && typeof rev.model === 'string')
+                    m = rev.model;
+                if (typeof m === 'string' && m.trim() !== '')
+                    model = m;
+            }
+            if (!reasoningEffort) {
+                // config-lib validator: model is {provider, model, reasoningEffort?} inside domains.supervisor.model / domains.review.model
+                // Support both nested object and flat legacy fallback for future compat
+                const supObj = j?.domains?.supervisor?.model;
+                const revObj = j?.domains?.review?.model;
+                const supEff = typeof supObj === 'object' && supObj !== null ? supObj.reasoningEffort : undefined;
+                const supFlat = j?.domains?.supervisor?.reasoningEffort;
+                const revEff = typeof revObj === 'object' && revObj !== null ? revObj.reasoningEffort : undefined;
+                const revFlat = j?.domains?.review?.reasoningEffort;
+                const candidate = supEff ?? supFlat ?? revEff ?? revFlat;
+                if (typeof candidate === 'string' && candidate.trim() !== '')
+                    reasoningEffort = candidate.trim();
+            }
         }
         catch { }
     }
     if (!model)
         model = 'deepseek-chat';
+    // Also try settings.yaml for reasoningEffort if still missing (e.g., llm-pi-ai default reasoning)
+    if (!reasoningEffort) {
+        try {
+            const { readFileSync } = await import('node:fs');
+            const { homedir } = await import('node:os');
+            const yamlPath = `${homedir()}/.dsh/settings.yaml`;
+            const yaml = readFileSync(yamlPath, 'utf-8');
+            // Look for reasoningEffort near the model id in provider blocks (future-proof; currently not in yaml)
+            const re = new RegExp(`id:\\s*${model}[\\s\\S]{0,200}reasoningEffort:\\s*(\\S+)`, 'm');
+            const m = yaml.match(re);
+            if (m) {
+                const v = m[1].trim().replace(/^["']|["']$/g, '');
+                if (v)
+                    reasoningEffort = v;
+            }
+        }
+        catch { }
+    }
     // If url not set via env, try to resolve from ~/.dsh/settings.yaml llm-pi-ai providers (DeepSeek suggested setup)
     if (!url) {
         try {
@@ -412,7 +468,7 @@ async function resolveLLMConfig() {
     if (!url.includes('/v1/') && !url.includes('/chat/completions')) {
         finalUrl = url.replace(/\/$/, '') + '/v1/chat/completions';
     }
-    return { key, url: finalUrl, model };
+    return { key, url: finalUrl, model, ...(reasoningEffort ? { reasoningEffort } : {}) };
 }
 async function resolveApiKey() {
     const cfg = await resolveLLMConfig();

package/lib/health-poller.js

@@ -50,6 +50,10 @@ function isRecentlyStarted(opts, wallMs) {
         return true;
     return false;
 }
+// Specific parse/boot-failure markers only. Bare 'JSON'/'YAML' were removed
+// (2026-08-31): they matched any line whose payload merely *contained* those
+// substrings — e.g. maestro-sync's status JSON listing session.jsonl.zstd /
+// settings.json paths — turning a healthy 401 into a rollback + restart.
 const ERROR_PATTERNS = [
     'ERR_MODULE_NOT_FOUND',
     'ERR_PNPM',
@@ -58,8 +62,6 @@ const ERROR_PATTERNS = [
     'SyntaxError',
     'YAMLParseError',
     'ParseError',
-    'YAML',
-    'JSON',
     'corrupted',
     'allowBuilds',
     'Cannot find module',
@@ -198,6 +200,30 @@ export async function pollHealth(opts = {}) {
                 logTail: logContent.slice(-5000),
             };
         }
+        // Corroborate with a cheap port-liveness check before declaring a crash.
+        // The HTTP fetch shares dsh-web's own event loop, so a busy-but-alive
+        // process (e.g. heavy GitLab-webhook-triggered review work) times out
+        // the same way a genuinely dead one does — but a real crash always frees
+        // the port, while `ss` does not depend on the contended event loop to
+        // answer. Downgrade to DEGRADED (still visible, still escalates after
+        // repeated ticks) instead of forcing an immediate rollback + restartWeb
+        // that would kill an in-flight review for nothing.
+        let portAlive = false;
+        try {
+            portAlive = await psAliveFn();
+        }
+        catch {
+            portAlive = false;
+        }
+        if (portAlive) {
+            return {
+                up: true,
+                httpCode,
+                error: logError ? `${fetchError} + ${logError}` : fetchError,
+                degraded: true,
+                logTail: logContent.slice(-5000),
+            };
+        }
         return {
             up: false,
             httpCode,

package/lib/supervisor.js

@@ -197,7 +197,7 @@ export class Supervisor {
                 return cfg.downThreshold;
         }
         catch { }
-        return 3;
+        return 5;
     }
     async findInterruptedRecent(withinMs) {
         const ms = withinMs ?? this.getResumeWithinMs();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ddtcorex/dsh-maestro-supervisor",
-  "version": "0.6.6",
+  "version": "0.6.8",
   "description": "Supervisor daemon for DSH Web resilience — auto-detect crashes, rollback to LKG, report",
   "type": "module",
   "main": "./lib/index.js",