deploy: update Space from deploy_preflight --push

.codeboarding/logs/wrapper-server.log

@@ -122,3 +122,12 @@ INFO:     127.0.0.1:44498 - "GET /health HTTP/1.1" 200 OK
 [stderr] 2026-06-14 04:57:18 INFO     [watchfiles.main:308] 2 changes detected
 [stderr] 2026-06-14 04:57:18 WARNING  [codeboarding_pro.analysis_controller:506] Skipping incremental update: full analysis artifacts not ready yet
 [stderr] 2026-06-14 04:57:18 INFO     [codeboarding_pro.analysis_controller:314] No cached analysis results; skipping health report
+[stderr] 2026-06-14 04:58:20 INFO     [watchfiles.main:308] 1 change detected
+[stderr] 2026-06-14 04:58:20 WARNING  [codeboarding_pro.analysis_controller:506] Skipping incremental update: full analysis artifacts not ready yet
+[stderr] 2026-06-14 04:58:20 INFO     [codeboarding_pro.analysis_controller:314] No cached analysis results; skipping health report
+[stderr] 2026-06-14 05:02:51 INFO     [watchfiles.main:308] 1 change detected
+[stderr] 2026-06-14 05:02:51 WARNING  [codeboarding_pro.analysis_controller:506] Skipping incremental update: full analysis artifacts not ready yet
+[stderr] 2026-06-14 05:02:51 INFO     [codeboarding_pro.analysis_controller:314] No cached analysis results; skipping health report
+[stderr] 2026-06-14 05:03:14 INFO     [watchfiles.main:308] 1 change detected
+[stderr] 2026-06-14 05:03:15 WARNING  [codeboarding_pro.analysis_controller:506] Skipping incremental update: full analysis artifacts not ready yet
+[stderr] 2026-06-14 05:03:15 INFO     [codeboarding_pro.analysis_controller:314] No cached analysis results; skipping health report

core/llm.py

@@ -14,6 +14,8 @@ import json
 import os
 
 MODEL = os.environ.get("CHIEF_ENGINEER_MODEL", "gemma4:e4b")
+MODAL_API_URL = os.environ.get("CHIEF_ENGINEER_MODAL_URL",
+    "https://kylebrodeur--microfactory-node-inference-serve.modal.run/v1/chat/completions")
 
 # Backend select. Default "ollama" keeps local/recording behavior IDENTICAL.
 # Set CHIEF_ENGINEER_BACKEND=zerogpu ONLY on the published Space (see
@@ -37,6 +39,13 @@ def _zerogpu():
         return None
 
 
+def _modal_api():
+    """Lazily check if Modal API backend is selected."""
+    if BACKEND != "modal":
+        return None
+    return True  # Modal API is always available (HTTP endpoint)
+
+
 def _forced_offline() -> bool:
     """Force the deterministic fallback path regardless of any daemon/backend.
     Read dynamically (not cached) so tests can toggle it. Used by the offline
@@ -51,6 +60,8 @@ def is_available() -> bool:
     zg = _zerogpu()
     if zg is not None:
         return zg.is_available()
+    if _modal_api():
+        return True
     if ollama is None:
         return False
     try:
@@ -64,6 +75,8 @@ def backend_status() -> str:
     zg = _zerogpu()
     if zg is not None:
         return zg.backend_status()
+    if _modal_api():
+        return f"<span style='color:var(--ao-green);'>●</span> live · Modal API (remote GPU)"
     return (f"<span style='color:var(--ao-green);'>●</span> live · {MODEL} (local Ollama)"
             if is_available() else
             f"<span style='color:var(--ao-yellow);'>●</span> offline fallback · "
@@ -73,7 +86,8 @@ def backend_status() -> str:
 def warm_up() -> str:
     """Pay the model's cold start now (off-camera), so the first real BUILD is fast.
     On ZeroGPU this enters the GPU window and loads the model; on Ollama/fallback it is
-    a cheap no-op. Returns the (post-load) backend status. Never raises."""
+    a cheap no-op. On Modal API it's a no-op (Modal handles its own warm-up).
+    Returns the (post-load) backend status. Never raises."""
     zg = _zerogpu()
     if zg is not None:
         try:
@@ -89,13 +103,30 @@ def chat_json(system: str, user: str, temperature: float = 0.4) -> dict | None:
         return None
     zg = _zerogpu()
     if zg is not None:
-        # ANY ZeroGPU failure (quota exhausted, model load, OOM, timeout) must fall
-        # back to the deterministic advisor — never surface "Error" on the Space.
-        # This is what keeps the "demo never crashes" guarantee true in production.
         try:
             return zg.chat_json(system, user, temperature)
         except Exception:
             return None
+    if _modal_api():
+        try:
+            import urllib.request
+            body = json.dumps({
+                "messages": [{"role": "user", "content": f"{system}\n\n{user}"}],
+                "max_tokens": 512,
+                "temperature": temperature,
+            }).encode()
+            req = urllib.request.Request(MODAL_API_URL, data=body,
+                headers={"Content-Type": "application/json"})
+            with urllib.request.urlopen(req, timeout=120) as resp:
+                data = json.loads(resp.read())
+                text = data["choices"][0]["message"]["content"].strip()
+                if text.startswith("```"):
+                    text = text.strip("`").lstrip()
+                    if text[:4].lower() == "json":
+                        text = text[4:]
+                return json.loads(text)
+        except Exception:
+            return None
     if not is_available():
         return None
     try:

learn/finetune/activity.jsonl

@@ -31,3 +31,5 @@
 {"timestamp":"2026-06-14T05:55:00-05:00","action":"serving","event":"all_three_complete","details":"All three serving items implemented: 1) gguf_pipeline_modal.py for Ollama GGUF on Modal, 2) modal_serve.py for Modal inference API, 3) core/llm_zerogpu_lora.py + app.py backend for Gradio model switcher. UI placement deferred to other agent per user request."}
 {"timestamp":"2026-06-14T06:00:00-05:00","action":"serving","event":"modal_api_deployed","details":"Modal inference API deployed successfully at https://kylebrodeur--microfactory-node-inference-serve.modal.run. Image built in 71s, app deployed in 75s."}
 {"timestamp":"2026-06-14T06:10:00-05:00","action":"deploy","event":"space_updated","details":"Deploy preflight 10/10 green. Pushed to build-small-hackathon/microfactory-lab. Factory reboot triggered. Space rebuilding. Note: Gradio warns about build_job arg count (model_choice has default, UI agent will wire dropdown)."}
+{"timestamp":"2026-06-14T06:30:00-05:00","action":"serving","event":"gguf_both_launched","details":"Killed old v2-only GGUF run. Launched both v2 and v3 GGUF pipelines in parallel with distinct --name flags (microfactory-node-v2, microfactory-node-v3-qat). Reusing cached llama.cpp image — no rebuild."}
+{"timestamp":"2026-06-14T06:30:00-05:00","action":"deploy","event":"space_v3_default","details":"Space now defaults to LoRA v3 (QAT E4B) — best training loss (1.75), best quantization quality. Pre-warmed at startup. Model options: v3 → v2 → Base → Modal API."}

learn/finetune/gguf_pipeline_modal.py

@@ -45,7 +45,7 @@ if modal is not None:
                      "accelerate>=0.34", "huggingface_hub")
     )
 
-    # CPU image for llama.cpp build + GGUF conversion
+    # CPU image for GGUF conversion — llama.cpp pre-built in the image (reused across runs)
     cpu_image = (
         modal.Image.debian_slim(python_version="3.12")
         .apt_install("git", "build-essential", "cmake")