Upload agents.md with huggingface_hub
Browse files
agents.md
ADDED
|
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Integration Guide – **openai‑agents‑python** with **Hugging Face CLI**
|
| 2 |
+
|
| 3 |
+
This guide demonstrates idiomatic patterns for invoking `huggingface-cli` commands *from* an **OpenAI Agents SDK** tool, including best practices for streaming, error handling, sandboxing, and leveraging Codex‑like code‑execution helpers.
|
| 4 |
+
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
## 1. Quick Start Snippet
|
| 8 |
+
|
| 9 |
+
```python
|
| 10 |
+
from agents import Agent
|
| 11 |
+
from agents.tool import function_tool
|
| 12 |
+
from agents.run import Runner
|
| 13 |
+
import asyncio, json, textwrap
|
| 14 |
+
|
| 15 |
+
HF_TIMEOUT = 30 # seconds
|
| 16 |
+
|
| 17 |
+
class HFCLIExecutor(Agent):
|
| 18 |
+
name = "AG-EXEC-CLI"
|
| 19 |
+
|
| 20 |
+
@function_tool(name="run_hf_cli", description="Run huggingface-cli commands")
|
| 21 |
+
async def run_hf_cli(self, cmd: str, args: list[str] = []) -> dict:
|
| 22 |
+
"""Execute huggingface‑cli <cmd> <args> and return structured result"""
|
| 23 |
+
proc = await asyncio.create_subprocess_exec(
|
| 24 |
+
"huggingface-cli", cmd, *args,
|
| 25 |
+
stdout=asyncio.subprocess.PIPE,
|
| 26 |
+
stderr=asyncio.subprocess.PIPE,
|
| 27 |
+
)
|
| 28 |
+
try:
|
| 29 |
+
stdout, stderr = await asyncio.wait_for(proc.communicate(), HF_TIMEOUT)
|
| 30 |
+
except asyncio.TimeoutError:
|
| 31 |
+
proc.kill()
|
| 32 |
+
return {"exit": -1, "error": "timeout"}
|
| 33 |
+
exit_code = proc.returncode
|
| 34 |
+
# Attempt to parse JSON block if present
|
| 35 |
+
stdout_str = stdout.decode()
|
| 36 |
+
if "{" in stdout_str and "}" in stdout_str:
|
| 37 |
+
try:
|
| 38 |
+
json_block = json.loads(stdout_str[stdout_str.find("{"): stdout_str.rfind("}") + 1])
|
| 39 |
+
parsed = json_block
|
| 40 |
+
except json.JSONDecodeError:
|
| 41 |
+
parsed = stdout_str.strip()
|
| 42 |
+
else:
|
| 43 |
+
parsed = stdout_str.strip()
|
| 44 |
+
return {
|
| 45 |
+
"exit": exit_code,
|
| 46 |
+
"stdout": parsed,
|
| 47 |
+
"stderr": stderr.decode().strip(),
|
| 48 |
+
}
|
| 49 |
+
|
| 50 |
+
if __name__ == "__main__":
|
| 51 |
+
# ad‑hoc run for local dev
|
| 52 |
+
print(asyncio.run(Runner.run(agent=HFCLIExecutor(), input="run_hf_cli cmd='whoami'")))
|
| 53 |
+
```
|
| 54 |
+
|
| 55 |
+
### Why this pattern?
|
| 56 |
+
|
| 57 |
+
* **asyncio.create\_subprocess\_exec** avoids blocking the event loop used by other agents.
|
| 58 |
+
* **wait\_for** enforces a hard timeout to keep MCP requests bounded.
|
| 59 |
+
* Outputs are *attempt‑parsed* to JSON if the CLI supports `--json` (e.g., `models list --json`).
|
| 60 |
+
|
| 61 |
+
---
|
| 62 |
+
|
| 63 |
+
## 2. Hugging Face CLI JSON‑Friendly Flags
|
| 64 |
+
|
| 65 |
+
| Command | Recommended Flags | Returns JSON? | |
|
| 66 |
+
| --------------- | ---------------------------------------- | --------------- | ------- |
|
| 67 |
+
| `models list` | `--author <org> --sort downloads --json` | ✅ | |
|
| 68 |
+
| `datasets list` | `--json` | ✅ | |
|
| 69 |
+
| `spaces list` | `--json` | ✅ | |
|
| 70 |
+
| `whoami` | *(none)* | ❌ (parse stderr | stdout) |
|
| 71 |
+
|
| 72 |
+
Always pass `--json` when available to make parsing trivial.
|
| 73 |
+
|
| 74 |
+
---
|
| 75 |
+
|
| 76 |
+
## 3. Codex Tooling Synergy
|
| 77 |
+
|
| 78 |
+
The **OpenAI Codex Playground** or GPT‑4o‑code model can auto‑generate wrappers for new CLI commands:
|
| 79 |
+
|
| 80 |
+
```python
|
| 81 |
+
# in a dev notebook
|
| 82 |
+
from openai import OpenAI
|
| 83 |
+
client = OpenAI()
|
| 84 |
+
|
| 85 |
+
generated = client.chat.completions.create(
|
| 86 |
+
model="gpt-4o-code", # Codex style
|
| 87 |
+
messages=[
|
| 88 |
+
{"role": "user", "content": "Write an asyncio Python wrapper for 'huggingface-cli datasets list --json'"}
|
| 89 |
+
],
|
| 90 |
+
)
|
| 91 |
+
print(generated.choices[0].message.content)
|
| 92 |
+
```
|
| 93 |
+
|
| 94 |
+
**Tip**: integrate this generation step into `AG-CORE-ORCH` to auto‑scaffold wrappers for infrequent commands.
|
| 95 |
+
|
| 96 |
+
---
|
| 97 |
+
|
| 98 |
+
## 4. Authentication Lifecycle Pattern
|
| 99 |
+
|
| 100 |
+
```python
|
| 101 |
+
class HFAuthMixin:
|
| 102 |
+
async def ensure_login(self, token: str):
|
| 103 |
+
# cache token in Redis or Agent memory
|
| 104 |
+
cached = self.memory.get("hf_token")
|
| 105 |
+
if cached == token:
|
| 106 |
+
return
|
| 107 |
+
await self.run_hf_cli("login", ["--token", token])
|
| 108 |
+
self.memory["hf_token"] = token
|
| 109 |
+
```
|
| 110 |
+
|
| 111 |
+
Call `ensure_login()` before executing any command requiring auth.
|
| 112 |
+
|
| 113 |
+
---
|
| 114 |
+
|
| 115 |
+
## 5. Streaming Large Outputs
|
| 116 |
+
|
| 117 |
+
For listings returning thousands of models, stream lines to the bus:
|
| 118 |
+
|
| 119 |
+
```python
|
| 120 |
+
async for line in proc.stdout:
|
| 121 |
+
await ctx.emit("stream", line.decode())
|
| 122 |
+
```
|
| 123 |
+
|
| 124 |
+
Down‑stream agent (`AG-OBS-MON`) or the gateway can chunk‑forward to HTTP clients (Server‑Sent Events or chunked‑transfer).
|
| 125 |
+
|
| 126 |
+
---
|
| 127 |
+
|
| 128 |
+
## 6. Error Classification Matrix
|
| 129 |
+
|
| 130 |
+
| Exit Code | Likely Cause | Retry? |
|
| 131 |
+
| --------- | ------------------------------ | ------------- |
|
| 132 |
+
| `0` | success | ─ |
|
| 133 |
+
| `1` | bad flag / auth error | ❌ |
|
| 134 |
+
| `126/127` | command not found / permission | ❌ (ops alert) |
|
| 135 |
+
| `-1` | timeout (killed) | ✅ once |
|
| 136 |
+
|
| 137 |
+
Map these statuses in `AG-CORE-ORCH` to the Failure & Retry Policy (see `agents.md`).
|
| 138 |
+
|
| 139 |
+
---
|
| 140 |
+
|
| 141 |
+
## 7. Embedding CLI Logic as Tools vs. External Subprocess
|
| 142 |
+
|
| 143 |
+
* **Pros of subprocess**: decoupled from Python API stability; mirrors real user workflows.
|
| 144 |
+
* **Cons**: slower than `huggingface_hub` SDK for bulk fetches; parsing required.
|
| 145 |
+
|
| 146 |
+
*Recommendation*: keep CLI path for *agentic text* fidelity but optionally fall back to direct SDK calls where performance critical (future optimisation ticket).
|
| 147 |
+
|
| 148 |
+
---
|
| 149 |
+
|
| 150 |
+
## 8. Unit Test Template (pytest‑asyncio)
|
| 151 |
+
|
| 152 |
+
```python
|
| 153 |
+
import pytest, asyncio
|
| 154 |
+
from agents.exec_cli import HFCLIExecutor
|
| 155 |
+
|
| 156 |
+
@pytest.mark.asyncio
|
| 157 |
+
async def test_models_list_json():
|
| 158 |
+
agent = HFCLIExecutor()
|
| 159 |
+
out = await agent.run_hf_cli("models", ["list", "--limit", "1", "--json"])
|
| 160 |
+
assert out["exit"] == 0
|
| 161 |
+
assert isinstance(out["stdout"], list)
|
| 162 |
+
```
|
| 163 |
+
|
| 164 |
+
## 9. Delegating via Handoff
|
| 165 |
+
|
| 166 |
+
|
| 167 |
+
Use the `handoff()` helper to create a chat agent that delegates CLI work to another agent. The SDK's `run_demo_loop` utility preserves conversation history between turns.
|
| 168 |
+
|
| 169 |
+
```python
|
| 170 |
+
from agents.repl import run_demo_loop
|
| 171 |
+
class HFChatAgent(Agent):
|
| 172 |
+
def __init__(self, cli_agent: HFCLIExecutor):
|
| 173 |
+
super().__init__(name="HF-CHAT", tools=[
|
| 174 |
+
handoff(cli_agent, tool_name_override="hf_cli_agent")
|
| 175 |
+
])
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
# example interactive session with memory
|
| 179 |
+
asyncio.run(run_demo_loop(HFChatAgent(HFCLIExecutor())))
|
| 180 |
+
```
|
| 181 |
+
|
| 182 |
+
---
|
| 183 |
+
|
| 184 |
+
*Last updated: 2025‑06‑28*
|