Upload 6 files

src/worker_payload.py

@@ -412,10 +412,10 @@ if __name__ == "__main__":
     demo.launch()
 """
     (workspace / "app.py").write_text(app_py, encoding="utf-8")
-    req = """gradio>=5.0.0
-huggingface_hub>=0.34.0,<1.0.0
+    req = """gradio>=6.0.0
+huggingface_hub>=0.34.0,<2.0.0
 spaces
-transformers>=4.45.0
+transformers>=4.45.0,<6.0.0
 diffusers
 accelerate
 safetensors
@@ -465,7 +465,7 @@ Non-negotiable safety and product constraints:
 - Work only inside the current workspace.
 - The wrapper will create the private Space, request hardware best-effort, upload files, and validate the live app. Do not create/delete repos yourself in this builder worker.
 - Preserve a cheap health endpoint named `health` with `api_name="health"`. It must not load weights, run GPU work, or download large files.
-- Keep the huggingface_hub pin in requirements.txt: huggingface_hub>=0.34.0,<1.0.0.
+- Do not pin huggingface_hub below 1.0. Use huggingface_hub>=0.34.0,<2.0.0 unless the model card requires a narrower compatible range. If transformers>=5 is used, keep huggingface_hub compatible with it, for example huggingface_hub>=1.5.0,<2.0.0.
 - README.md frontmatter must remain valid; if it uses short_description, it must be 60 characters or fewer.
 
 Implementation contract:
@@ -519,8 +519,111 @@ def sanitize_readme_metadata(workspace: Path, events_path: Path):
         append_event(events_path, "metadata_sanitize", "success", "Sanitized README metadata", {"short_description": "Generated model demo"})
 
 
+
+def normalize_requirements_for_modern_hub(workspace: Path, events_path: Path):
+    """Prevent a known resolver conflict in generated Spaces.
+
+    Older builder versions forced `huggingface_hub<1.0.0` to avoid old Gradio
+    import issues. Modern Spaces can use Gradio 6 and recent Transformers;
+    Transformers 5.x requires huggingface-hub >=1.5.0, so the old pin breaks
+    builds. Keep the constraint broad and modern unless Pi intentionally uses a
+    different compatible stack.
+    """
+    req_path = workspace / "requirements.txt"
+    if not req_path.exists():
+        return
+    raw = req_path.read_text(encoding="utf-8", errors="ignore")
+    lines = [line.rstrip() for line in raw.splitlines()]
+    changed = False
+    filtered = []
+    transformers_needs_hub_15 = False
+    for line in lines:
+        stripped = line.strip()
+        low = stripped.lower().replace("_", "-")
+        if low.startswith("huggingface-hub"):
+            if "<1" in low or "< 1" in low or ",<1" in low:
+                changed = True
+            # Always replace with the policy line to avoid duplicate/conflicting pins.
+            changed = True
+            continue
+        if low.startswith("transformers") and (">=5" in low or "==5" in low or "~=5" in low):
+            transformers_needs_hub_15 = True
+        filtered.append(line)
+    hub_line = "huggingface_hub>=1.5.0,<2.0.0" if transformers_needs_hub_15 else "huggingface_hub>=0.34.0,<2.0.0"
+    # Put hub near the top, after any --extra-index-url lines.
+    insert_at = 0
+    while insert_at < len(filtered) and filtered[insert_at].strip().startswith("--"):
+        insert_at += 1
+    filtered.insert(insert_at, hub_line)
+    new = "\n".join(line for line in filtered if line.strip()) + "\n"
+    if new != raw:
+        req_path.write_text(new, encoding="utf-8")
+        append_event(events_path, "requirements_sanitize", "success", "Normalized huggingface_hub requirement for modern dependency resolution", {"huggingface_hub": hub_line})
+
+
+def repair_workspace_with_pi(workspace: Path, run_dir: Path, events_path: Path, pi_model: str, target_space_id: str, model_id: str, failure_reason: str):
+    """Ask Pi for one minimal build/runtime repair pass based on collected logs."""
+    logs_dir = run_dir / "logs"
+    build_log = (logs_dir / "space_logs_build.txt").read_text(encoding="utf-8", errors="ignore") if (logs_dir / "space_logs_build.txt").exists() else ""
+    runtime_log = (logs_dir / "space_logs_runtime.txt").read_text(encoding="utf-8", errors="ignore") if (logs_dir / "space_logs_runtime.txt").exists() else ""
+    repair_dir = run_dir / "repair"
+    before_dir = repair_dir / "before"
+    after_dir = repair_dir / "after"
+    if before_dir.exists():
+        shutil.rmtree(before_dir)
+    shutil.copytree(workspace, before_dir, ignore=shutil.ignore_patterns(".git", "node_modules", "__pycache__", "*.pyc"))
+    goal = f"""You are Pi repairing a Hugging Face Space generated by Agentic Space Factory.
+
+MODEL_ID: {model_id}
+TARGET_SPACE_ID: {target_space_id}
+
+The first build/runtime validation failed.
+
+Failure summary:
+{failure_reason[:4000]}
+
+Build log tail:
+```text
+{build_log[-12000:]}
+```
+
+Runtime log tail:
+```text
+{runtime_log[-12000:]}
+```
+
+Repair contract:
+- Make the smallest patch possible.
+- Prefer fixing dependency resolver conflicts, missing imports, invalid metadata, Gradio endpoint bugs, and import-order issues.
+- Do not replace real inference with a placeholder unless TECHNICAL_BLOCKERS.json clearly explains why full inference is impossible.
+- Preserve a cheap health endpoint with api_name="health".
+- Keep README frontmatter valid, short_description <= 60 chars.
+- Do not pin huggingface_hub below 1.0. For modern generated Spaces use huggingface_hub>=0.34.0,<2.0.0. If transformers>=5 is present, use huggingface_hub>=1.5.0,<2.0.0.
+- Do not delete the app. Do not publish anything. Work only in the current workspace.
+
+Deliverables:
+- patched app.py / requirements.txt / README.md as needed
+- REPAIR_SUMMARY.md explaining the patch
+- keep or update INFERENCE_CONTRACT.json if the inference contract changed
+"""
+    (workspace / "REPAIR_GOAL.md").write_text(goal, encoding="utf-8")
+    append_event(events_path, "repair", "started", "Running Pi repair pass using build/runtime logs", {"model": pi_model})
+    code, out = run_cmd(["pi", "-p", goal], cwd=workspace, timeout=1500)
+    logs_dir.mkdir(parents=True, exist_ok=True)
+    (logs_dir / "pi_repair_output.txt").write_text(out, encoding="utf-8")
+    if code != 0:
+        append_event(events_path, "repair", "failed", "Pi repair returned a non-zero exit code", {"returncode": code, "output_tail": out[-3000:]})
+        return False
+    normalize_requirements_for_modern_hub(workspace, events_path)
+    if after_dir.exists():
+        shutil.rmtree(after_dir)
+    shutil.copytree(workspace, after_dir, ignore=shutil.ignore_patterns(".git", "node_modules", "__pycache__", "*.pyc"))
+    append_event(events_path, "repair", "success", "Pi repair pass completed", {"output_tail": out[-3000:]})
+    return True
+
 def upload_workspace(api, workspace: Path, target_space_id: str, token: str, run_dir: Path, events_path: Path):
     sanitize_readme_metadata(workspace, events_path)
+    normalize_requirements_for_modern_hub(workspace, events_path)
     append_event(events_path, "upload_files", "started", "Uploading generated universal model-card workspace recursively")
     gen_dir = run_dir / "generated"
     if gen_dir.exists():
@@ -728,7 +831,16 @@ def main():
             selected_hardware = "default-cpu-or-existing"
         write_json(run_dir / "hardware_attempts.json", {"selected_hardware": selected_hardware, "requested_sequence": requested_hardware_sequence, "attempts": hardware_attempts})
 
-        validation = validate_live_api(api, target_space_id, token, run_dir, events_path, timeout_s=1200)
+        try:
+            validation = validate_live_api(api, target_space_id, token, run_dir, events_path, timeout_s=1200)
+        except Exception as validation_error:
+            append_event(events_path, "repair", "started", "Initial live validation failed; attempting one repair pass", {"error": str(validation_error)[:2000]})
+            collect_space_logs(target_space_id, token, run_dir, events_path)
+            repaired = repair_workspace_with_pi(workspace, run_dir, events_path, pi_model, target_space_id, model_id, str(validation_error))
+            if not repaired:
+                raise
+            upload_workspace(api, workspace, target_space_id, token, run_dir, events_path)
+            validation = validate_live_api(api, target_space_id, token, run_dir, events_path, timeout_s=1200)
         inference_gate = infer_generation_gate(workspace, implementation_mode, validation, run_dir, events_path)
 
         # If the generated app looks like real GPU inference but automatic