@melaya/runner 1.1.28 → 1.1.29

package/dist/assistantHost.py

@@ -889,7 +889,32 @@ def _register_stream_hooks(agent) -> None:
     _log(f"stream hooks registered (delta={ok_a} tool={ok_b})")
 
 
-def _run_turn(agent, turn_id: str, message: str, browser_turn: bool = False) -> None:
+def _build_user_msg(message: str, images=None):
+    """Build the user Msg. With images (base64 blocks from the turn frame) and a
+    vision-capable model, this is a multimodal Msg ([text?] + image blocks in the
+    Anthropic-style `source.base64` shape agentscope formats per provider). Any
+    problem building the blocks falls back to a plain text Msg so a turn NEVER
+    breaks on a text-only model / older agentscope."""
+    from agentscope.message import Msg
+    if not images:
+        return Msg("user", message, "user")
+    try:
+        blocks = []
+        if message:
+            blocks.append({"type": "text", "text": message})
+        for im in images:
+            data = str((im or {}).get("data") or "")
+            mt = str((im or {}).get("media_type") or "")
+            if data and mt:
+                blocks.append({"type": "image", "source": {"type": "base64", "media_type": mt, "data": data}})
+        if len(blocks) > (1 if message else 0):
+            return Msg("user", blocks, "user")
+    except Exception:
+        pass
+    return Msg("user", message, "user")
+
+
+def _run_turn(agent, turn_id: str, message: str, browser_turn: bool = False, images=None) -> None:
     import asyncio
     from agentscope.message import Msg
 
@@ -916,7 +941,7 @@ def _run_turn(agent, turn_id: str, message: str, browser_turn: bool = False) ->
     # unwinds the agent loop wherever it is awaiting — including mid phone-command
     # HTTP wait — so the run stops promptly instead of finishing the step first.
     async def _go():
-        task = asyncio.ensure_future(agent(Msg("user", message, "user")))
+        task = asyncio.ensure_future(agent(_build_user_msg(message, images)))
         while not task.done():
             if _stream.get("cancel"):
                 task.cancel()
@@ -1108,6 +1133,10 @@ def main() -> int:
             continue
         turn_id = str(req.get("turnId") or "")
         message = str(req.get("message") or "")
+        # Vision: base64 image blocks for THIS turn (validated server-side). Used
+        # to build a multimodal Msg when the model supports it; text-only otherwise.
+        _images = req.get("images")
+        images = _images if isinstance(_images, list) and _images else None
         # Per-turn autonomy mode: the TS side carries an updated `hitl_mode`
         # on runner:assistant_turn so a mid-session flip takes effect next
         # turn. Absent ⇒ keep the current (spawn / previous-turn) mode. The
@@ -1123,7 +1152,7 @@ def main() -> int:
         # 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:
+        if not message and not images:
             _emit(turn_id, "done")
             continue
         # P2-6: nudge the live memory budget down if a prior turn's OOM downgraded
@@ -1135,7 +1164,7 @@ def main() -> int:
         # ever survives in the warm host environment.
         browser_turn = _apply_browser_turn_grant(req)
         try:
-            _run_turn(agent, turn_id, message, browser_turn=browser_turn)
+            _run_turn(agent, turn_id, message, browser_turn=browser_turn, images=images)
         finally:
             _purge_browser_turn_grant()
         _emit_usage(agent, turn_id)

package/dist/detect.js

@@ -8,7 +8,9 @@ import { readFile } from "fs/promises";
 import { homedir } from "os";
 import { join } from "path";
 const execFileAsync = promisify(execFile);
-const COPILOT_MODELS_FALLBACK = ["gpt-4o", "gpt-4.1", "o3-mini", "claude-3.5-sonnet", "gemini-2.0-flash"];
+// Fallback ONLY when the live /models fetch fails. Must be ENABLED-by-default
+// Copilot models (policy!=disabled) — Claude/Gemini default to disabled and 400.
+const COPILOT_MODELS_FALLBACK = ["gpt-4.1", "gpt-4o", "gpt-5-mini"];
 /** Fallbacks used ONLY when live discovery fails (offline runner, transient
  *  error, missing cache). The REAL lists are fetched dynamically below so a new
  *  claude/codex release is available on Melaya automatically — no code change,
@@ -208,11 +210,18 @@ async function detectGithubCopilot() {
             });
             if (r.ok) {
                 const d = (await r.json());
-                // Keep ONLY real chat models the agent runtime can drive (tool calling
-                // is required), and drop Copilot-internal routing/search/compaction ids
-                // and embedding models that the /models list also advertises.
+                // Keep ONLY real chat models the agent runtime can DRIVE + CALL:
+                //  - tool calling required (agent loop),
+                //  - `policy.state !== "disabled"` — Copilot's /models over-reports:
+                //    Gemini/Claude/kimi/gpt-5.4-5.6 default to policy=disabled and a
+                //    chat/completions call to one 400s `model_not_supported` until the
+                //    user enables it at github.com/settings/copilot/features. `enabled`
+                //    and absent-policy (base gpt-4o/4.1) are callable. Enabling a model
+                //    there flips it to `enabled` and it auto-appears on the next detect.
+                //  - drop Copilot-internal routing/search/compaction/embedding ids.
                 const ids = (d.data ?? [])
                     .filter((m) => m?.capabilities?.type === "chat" && m?.capabilities?.supports?.tool_calls === true)
+                    .filter((m) => (m?.policy?.state ?? "enabled") !== "disabled")
                     .map((m) => m?.id ?? "")
                     .filter((s) => s && !/^copilot-search|^exec-agent|^trajectory-compaction|-free-auto$|-flash-(picker|secondary|tertiary)$|embedding/i.test(s));
                 if (ids.length)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.1.28",
+  "version": "1.1.29",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,