@melaya/runner 1.0.57 → 1.0.60

package/dist/connection.js

@@ -578,6 +578,39 @@ export async function connect(opts) {
             mkdirSync(runDir, { recursive: true });
             writeFileSync(join(runDir, "main.py"), mainPyContent, "utf-8");
             writeFileSync(join(runDir, "config.json"), JSON.stringify(cfg, null, 2), "utf-8");
+            // Materialise static-context bag from the cfg into <runDir>/rag/.
+            // The Node server collected these via POST/PUT /pipelines (which
+            // ran _materialise_inline_rag_docs on the builder box) and shipped
+            // them inline so the runner doesn't need a second HTTP fetch.
+            // Without this the spawned crew worker logs "RAG_DIR missing" and
+            // _load_rag_context() returns empty — every wizard Context card
+            // is silently lost.
+            const ragFilesFromCfg = cfg.rag_files;
+            if (ragFilesFromCfg && typeof ragFilesFromCfg === "object") {
+                const ragDir = join(runDir, "rag");
+                mkdirSync(ragDir, { recursive: true });
+                let wrote = 0;
+                for (const [fname, b64] of Object.entries(ragFilesFromCfg)) {
+                    if (typeof b64 !== "string" || !fname)
+                        continue;
+                    // Filename traversal guard — server-side helper already
+                    // sanitises but defense in depth: drop any name containing a
+                    // slash or starting with a dot.
+                    const safe = fname.replace(/[\\/]/g, "").replace(/^\.+/, "");
+                    if (!safe)
+                        continue;
+                    try {
+                        writeFileSync(join(ragDir, safe), Buffer.from(b64, "base64"));
+                        wrote++;
+                    }
+                    catch (e) {
+                        console.log(chalk.yellow(`  ! could not write rag/${safe}: ${e.message}`));
+                    }
+                }
+                if (wrote > 0) {
+                    console.log(chalk.gray(`  [rag] wrote ${wrote} static-context file(s) to ${ragDir}`));
+                }
+            }
             // Preflight the first persona's model so we surface an early
             // failure if (e.g.) LM Studio hasn't loaded the requested
             // qwen3.6-27b. Same flow as pipelines — it just reads the
@@ -670,6 +703,15 @@ export async function connect(opts) {
                         MEL_MODEL_DISABLE_THINKING: preflight.profile.thinkingDefault === "off" ? "1" : "0",
                     }
                     : {}),
+                // Per-pipeline credentials from agents.credentials. Same
+                // delivery channel the regular runner:run path uses at
+                // line 384. Without this, Python tools that read
+                // os.environ.get("FRED_API_KEY") / "EDGAR_API_KEY" / etc.
+                // get None even when the user has the credential connected on
+                // the Connectors page. The Node tRPC server fetches the map
+                // via the same query agentStudio.credentials.getEnvMap uses
+                // and ships it inline on the crew dispatch payload.
+                ...(cfg.credentials ?? {}),
             };
             const proc = spawn(envResult.pythonPath, ["-u", join(runDir, "main.py")], {
                 cwd: runDir,

package/dist/modelLoader.js

@@ -168,8 +168,17 @@ function _buildProfile(id, tier, paramsBillion, family, thinkingDefault, reason)
     const recommendedMaxTokens = tier === "agentic-capable" ? 8192
         : tier === "agentic-marginal" ? 4096
             : 2048;
-    const recommendedMaxIters = tier === "agentic-capable" ? 10
-        : tier === "agentic-marginal" ? 5
+    // Crew personas typically wield 8–15 scoped tools. A ReAct loop has
+    // to spend one iteration per tool call AND one final iteration on the
+    // text summary. 10 was the original ceiling for short pipelines; it
+    // leaves zero headroom for crew personas that loop on fred/edgar/
+    // funding probes before they can write the regime label, and the
+    // persona terminates mid-thinking with no text reply — the downstream
+    // persona then sees `Output from agent N:\n` followed by nothing.
+    // 25 gives a 27B model enough room to canvass its tool set AND wrap
+    // up; smaller capable models cap themselves naturally on context.
+    const recommendedMaxIters = tier === "agentic-capable" ? 25
+        : tier === "agentic-marginal" ? 8
             : 3;
     return {
         schemaVersion: PROFILE_SCHEMA_VERSION,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.0.57",
+  "version": "1.0.60",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,