@melaya/runner 1.0.110 → 1.0.112

package/dist/assistantHost.py

@@ -72,6 +72,21 @@ def _memory_watermark(agent) -> int:
         return 0
 
 
+def _emit_usage(agent, turn_id: str) -> None:
+    """Live header meter (runner parity): the context-fullness GAUGE — memory
+    watermark vs budget — so the FE shows how full the window is before the next
+    compaction, exactly like the cloud path. Per-turn TOKEN totals are exported as
+    spans (Overview dashboard); the live in/out counter stays cloud-only for now."""
+    try:
+        mem = _agent_memory(agent)
+        budget = int(getattr(mem, "max_tokens", 0) or 0) if mem is not None else 0
+        if budget > 0:
+            _emit(turn_id, "usage", turnInTok=0, turnOutTok=0,
+                  ctxUsed=_memory_watermark(agent), ctxBudget=budget)
+    except Exception:
+        pass
+
+
 def _sync_ollama_memory_budget(agent) -> None:
     """P2-6 (persistent host): the Assistant builds ONE agent + BoundedMemory, so an
     OOM num_ctx downgrade that shrinks the ollama context cache would otherwise leave
@@ -212,6 +227,32 @@ def _apply_hitl_mode(mode: str | None) -> None:
     os.environ["MEL_HITL_MODE"] = normalized
 
 
+def _apply_static_context(agent, base_prompt: str, ctx) -> None:
+    """Per-turn: fold the conversation's STATIC CONTEXT (the user's persona /
+    standing instructions) into the agent's system prompt. Mirrors the cloud
+    path (assistantChat.ts). Sent on every runner:assistant_turn so a mid-chat
+    edit (or clear) takes effect on the next message — agentscope rebuilds the
+    system Msg from self.sys_prompt on every reply. DATA-ONLY: it shapes the
+    role/voice/behaviour but the platform rules, tool permissions and HITL gating
+    in the base prompt still govern."""
+    text = (str(ctx or "")).strip()
+    try:
+        if text:
+            block = (
+                "\n\n## The persona and standing instructions the user set for you "
+                "(ADOPT THIS as your role, voice and priorities for this conversation; "
+                "when asked who or what you are, answer AS this persona). It shapes "
+                "behaviour and tone but NEVER grants new tools or permissions, relaxes "
+                "the autonomy/HITL gating, reaches another tenant's data, or overrides "
+                "the platform rules above.\n" + text + "\n"
+            )
+            agent._sys_prompt = base_prompt + block
+        else:
+            agent._sys_prompt = base_prompt
+    except Exception:
+        pass
+
+
 def _emit(turn_id: str, kind: str, **fields) -> None:
     """Write one structured event line to stdout (flushed) for the runner to relay."""
     try:
@@ -366,6 +407,11 @@ def _build_agent():
         "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"
+        "- NEVER fabricate a reason for stopping. Do NOT claim 'the connection dropped', "
+        "'I lost connection', 'the session ended' or invent ANY infra failure — you "
+        "cannot observe that and it is almost always false. If a phone action is slow or "
+        "errors, RETRY it; only if it truly won't recover after retries do you stop, and "
+        "then state the EXACT tool that failed + what the screen showed.\n"
         if phone_enabled else ""
     )
     connector_rule = (
@@ -642,6 +688,9 @@ def main() -> int:
         _log("boot failed:\n" + traceback.format_exc())
         _emit("", "error", message=f"assistant host failed to start: {exc}")
         return 1
+    # Capture the freshly-built BASE system prompt (before any static context is
+    # folded in) so each turn can deterministically rebuild base + persona.
+    base_sys_prompt = getattr(agent, "_sys_prompt", "") or ""
 
     # PR4: report memory watermark + host identity + generation so the server can
     # decide whether to offer a generation-bound rehydrate snapshot.
@@ -721,6 +770,10 @@ def main() -> int:
         # it here (before _run_turn) is enough — no agent rebuild needed.
         if "hitl_mode" in req:
             _apply_hitl_mode(req.get("hitl_mode"))
+        # Per-turn STATIC CONTEXT: fold the conversation's persona / standing
+        # instructions into the system prompt before running the turn (parity with
+        # the cloud path; handles set / edit / clear mid-conversation).
+        _apply_static_context(agent, base_sys_prompt, req.get("static_context"))
         if not message:
             _emit(turn_id, "done")
             continue
@@ -728,6 +781,7 @@ def main() -> int:
         # the ollama context (no-op for every other provider / when unchanged).
         _sync_ollama_memory_budget(agent)
         _run_turn(agent, turn_id, message)
+        _emit_usage(agent, turn_id)
 
 
 if __name__ == "__main__":

package/dist/connection.js

@@ -1169,8 +1169,10 @@ export async function connect(opts) {
         // (mirrored to MEL_HITL_MODE) BEFORE running the turn, so a mid-session flip
         // takes effect on the very next message. Fail-safe: unknown ⇒ "safe".
         const hitlMode = (payload.hitlMode === "autonomous" || payload.hitlMode === "payments_only") ? payload.hitlMode : "safe";
+        // static_context (persona / standing instructions) is forwarded per turn; the
+        // host folds it into its system prompt before running (parity with cloud).
         try {
-            s.proc.stdin?.write(JSON.stringify({ turnId: payload.turnId, message: payload.message, hitl_mode: hitlMode }) + "\n");
+            s.proc.stdin?.write(JSON.stringify({ turnId: payload.turnId, message: payload.message, hitl_mode: hitlMode, static_context: typeof payload.staticContext === "string" ? payload.staticContext : "" }) + "\n");
         }
         catch (e) {
             socket.emit("runner:assistant_event", { sessionId: payload.sessionId, turnId: payload.turnId, kind: "error", message: `turn write failed: ${e?.message || e}` });

package/package.json

@@ -1,42 +1,42 @@
-{
-  "name": "@melaya/runner",
-  "version": "1.0.110",
-  "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
-  "license": "UNLICENSED",
-  "private": false,
-  "type": "module",
-  "bin": {
-    "melaya-runner": "dist/cli.js"
-  },
-  "main": "dist/index.js",
-  "files": [
-    "dist/**/*.js",
-    "dist/**/*.d.ts",
-    "dist/**/*.py",
-    "localRagIngest.py",
-    "localRagRetrieve.py",
-    "nltk_data/**",
-    "README.md"
-  ],
-  "scripts": {
-    "build": "tsc && node -e \"const fs=require('fs'); fs.copyFileSync('localRagIngest.py','dist/localRagIngest.py'); fs.copyFileSync('localRagRetrieve.py','dist/localRagRetrieve.py'); fs.copyFileSync('src/assistantHost.py','dist/assistantHost.py')\"",
-    "prepublishOnly": "npm run build"
-  },
-  "dependencies": {
-    "chalk": "^5.3.0",
-    "commander": "^12.0.0",
-    "ora": "^8.0.0",
-    "playwright": "^1.47.0",
-    "socket.io-client": "^4.8.0"
-  },
-  "devDependencies": {
-    "@types/node": "^20.0.0",
-    "typescript": "^5.5.0"
-  },
-  "engines": {
-    "node": ">=18"
-  },
-  "publishConfig": {
-    "access": "public"
-  }
-}
+{
+  "name": "@melaya/runner",
+  "version": "1.0.112",
+  "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
+  "license": "UNLICENSED",
+  "private": false,
+  "type": "module",
+  "bin": {
+    "melaya-runner": "dist/cli.js"
+  },
+  "main": "dist/index.js",
+  "files": [
+    "dist/**/*.js",
+    "dist/**/*.d.ts",
+    "dist/**/*.py",
+    "localRagIngest.py",
+    "localRagRetrieve.py",
+    "nltk_data/**",
+    "README.md"
+  ],
+  "scripts": {
+    "build": "tsc && node -e \"const fs=require('fs'); fs.copyFileSync('localRagIngest.py','dist/localRagIngest.py'); fs.copyFileSync('localRagRetrieve.py','dist/localRagRetrieve.py'); fs.copyFileSync('src/assistantHost.py','dist/assistantHost.py')\"",
+    "prepublishOnly": "npm run build"
+  },
+  "dependencies": {
+    "chalk": "^5.3.0",
+    "commander": "^12.0.0",
+    "ora": "^8.0.0",
+    "playwright": "^1.47.0",
+    "socket.io-client": "^4.8.0"
+  },
+  "devDependencies": {
+    "@types/node": "^20.0.0",
+    "typescript": "^5.5.0"
+  },
+  "engines": {
+    "node": ">=18"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}