@melaya/runner 1.0.79 → 1.0.82

package/dist/assistantHost.py

@@ -125,6 +125,12 @@ def _build_agent():
         "phone_post_comment / phone_create_post — every publish is approved by the "
         "user ON THEIR PHONE, so never refuse for safety. Only touch apps the user "
         "asked for; the phone enforces an allowlist and blocks the rest.\n"
+        "- You are FULLY AUTONOMOUS on the phone: NEVER ask the user a question, "
+        "never end your turn asking for confirmation or offering a choice — make the "
+        "reasonable decision and DO it. If a feed is algorithmic, use the app's own "
+        "Search / Explore to find the right content yourself. Keep going until the "
+        "task is COMPLETE (e.g. all N comments posted) or you are genuinely blocked; "
+        "only then report what you did and what (if anything) is blocked.\n"
         if phone_enabled else ""
     )
     sys_prompt = (
@@ -150,7 +156,9 @@ def _build_agent():
         model_name=model,
         provider=provider,
         api_key="",  # model.py resolves the provider's own creds (OAuth off disk / localhost)
-        max_iters=6,
+        # Phone tasks are many-step (each comment ≈ screen_tree→tap→type→post), so
+        # give them room like the device template (40); read-only Q&A needs few.
+        max_iters=40 if phone_enabled else 8,
         reliability=True,
         bounded_memory=True,
     )

package/dist/connection.js

@@ -315,6 +315,10 @@ export async function connect(opts) {
                 PYTHONUTF8: "1",
                 ...sslEnv,
                 MEL_RUN_ID: payload.runId,
+                // Pipeline name so phone-driving runs can self-register as the phone's
+                // active run (arms the on-device working overlay + its Kill button)
+                // regardless of launch surface — desktop, native, or assistant.
+                MEL_PIPELINE_NAME: payload.pipelineName,
                 // Point agentscope hooks at the REAL server (not the browser's Vite URL).
                 // Convert wss://api.melaya.org → https://api.melaya.org so pushMessage,
                 // registerRun, and other agentscope HTTP calls reach the tRPC server.
@@ -370,6 +374,10 @@ export async function connect(opts) {
                         MEL_MODEL_PARAMS_B: String(preflight.profile.paramsBillion ?? ""),
                         MEL_MODEL_MAX_TOKENS: String(preflight.profile.recommendedMaxTokens),
                         MEL_MODEL_MAX_ITERS: String(preflight.profile.recommendedMaxIters),
+                        // Real context window → sizes BoundedMemory so compression can't
+                        // evict the task brief on a big local model. Empty when unknown
+                        // (the agent falls back to a safe per-tier floor).
+                        MEL_MODEL_CONTEXT_TOKENS: String(preflight.profile.contextTokens ?? ""),
                         MEL_MODEL_DISABLE_THINKING: preflight.profile.thinkingDefault === "off" ? "1" : "0",
                     }
                     : {}),
@@ -714,6 +722,10 @@ export async function connect(opts) {
                         MEL_MODEL_PARAMS_B: String(preflight.profile.paramsBillion ?? ""),
                         MEL_MODEL_MAX_TOKENS: String(preflight.profile.recommendedMaxTokens),
                         MEL_MODEL_MAX_ITERS: String(preflight.profile.recommendedMaxIters),
+                        // Real context window → sizes BoundedMemory so compression can't
+                        // evict the task brief on a big local model. Empty when unknown
+                        // (the agent falls back to a safe per-tier floor).
+                        MEL_MODEL_CONTEXT_TOKENS: String(preflight.profile.contextTokens ?? ""),
                         MEL_MODEL_DISABLE_THINKING: preflight.profile.thinkingDefault === "off" ? "1" : "0",
                     }
                     : {}),

package/dist/modelLoader.d.ts

@@ -47,6 +47,7 @@ export interface ModelProfile {
     thinkingDefault: "on" | "off" | "n/a";
     recommendedMaxTokens: number;
     recommendedMaxIters: number;
+    contextTokens: number | null;
     detectedAt: string;
     reason: string;
 }

package/dist/modelLoader.js

@@ -120,11 +120,12 @@ function _parseParamCount(id) {
 }
 function _classify(id, lmstudioMeta) {
     const lower = id.toLowerCase();
+    const ctxTokens = lmstudioMeta?.max_context_length ?? null;
     // 1. Family override wins if matched.
     for (const ov of FAMILY_OVERRIDES) {
         if (ov.pattern.test(lower)) {
             const params = _parseParamCount(id);
-            return _buildProfile(id, ov.tier, params, ov.family, ov.thinking, ov.reason);
+            return _buildProfile(id, ov.tier, params, ov.family, ov.thinking, ov.reason, ctxTokens);
         }
     }
     // 2. LM Studio metadata if present.
@@ -160,9 +161,9 @@ function _classify(id, lmstudioMeta) {
         tier = "text-only";
         reason = `${params}B params < ${TIER_PARAM_THRESHOLDS.agenticMarginal}B threshold — text-only, abort if used for tool-using pipeline`;
     }
-    return _buildProfile(id, tier, params, null, "n/a", reason);
+    return _buildProfile(id, tier, params, null, "n/a", reason, ctxTokens);
 }
-function _buildProfile(id, tier, paramsBillion, family, thinkingDefault, reason) {
+function _buildProfile(id, tier, paramsBillion, family, thinkingDefault, reason, contextTokens = null) {
     // Per-tier resource caps. Smaller models get tighter ceilings to
     // bound runaway thinking loops.
     const recommendedMaxTokens = tier === "agentic-capable" ? 8192
@@ -190,6 +191,7 @@ function _buildProfile(id, tier, paramsBillion, family, thinkingDefault, reason)
         thinkingDefault,
         recommendedMaxTokens,
         recommendedMaxIters,
+        contextTokens: contextTokens && contextTokens > 0 ? contextTokens : null,
         detectedAt: new Date().toISOString(),
         reason,
     };
@@ -233,7 +235,7 @@ export function classifyModel(id, lmstudioMeta) {
 // Bump whenever FAMILY_OVERRIDES, tier thresholds, or any classifier
 // logic changes in a way that could re-tier an existing cached model.
 // Cached profiles with a mismatched version are ignored and re-classified.
-const PROFILE_SCHEMA_VERSION = 3;
+const PROFILE_SCHEMA_VERSION = 4;
 const LMSTUDIO_BASE = process.env.LMSTUDIO_BASE_URL || "http://127.0.0.1:1234";
 const OLLAMA_BASE = process.env.OLLAMA_BASE_URL || "http://127.0.0.1:11434";
 // JIT-load timeout. 8B-class quantised models on M-series Macs typically

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.0.79",
+  "version": "1.0.82",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,