wip
This commit is contained in:
Kazuhiro Sera
@@ -0,0 +1,235 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import asyncio
|
||||
import io
|
||||
import sys
|
||||
from collections.abc import Sequence
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from agents import Agent, ModelSettings, Runner, flush_traces, gen_trace_id, handoff, trace
|
||||
from agents.items import HandoffOutputItem, ToolCallItem
|
||||
from agents.run import RunConfig
|
||||
from agents.sandbox import Manifest, MemoryReadConfig, SandboxAgent, SandboxRunConfig
|
||||
from agents.sandbox.capabilities import LocalDirLazySkillSource, Memory, Shell, Skills
|
||||
from agents.sandbox.entries import Dir, File, LocalDir
|
||||
from agents.sandbox.sandboxes.unix_local import UnixLocalSandboxClient
|
||||
from agents.sandbox.session.base_sandbox_session import BaseSandboxSession
|
||||
|
||||
if __package__ is None or __package__ == "":
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
|
||||
|
||||
DEFAULT_MODEL = "gpt-5.4"
|
||||
EXAMPLE_DIR = Path(__file__).resolve().parent
|
||||
SKILLS_DIR = EXAMPLE_DIR / "tracing_dashboard_skills"
|
||||
|
||||
DEFAULT_PROMPT = (
|
||||
"Review the Bluebird data export case. The customer wants a production data export "
|
||||
"this week, security review is still open, and finance already mapped SSO attributes."
|
||||
)
|
||||
|
||||
|
||||
def _build_manifest() -> Manifest:
|
||||
return Manifest(
|
||||
entries={
|
||||
"data": Dir(
|
||||
children={
|
||||
"customer_case.md": File(
|
||||
content=(
|
||||
b"# Bluebird Logistics case\n\n"
|
||||
b"- Customer wants a production data export enabled this week.\n"
|
||||
b"- Security review is still open.\n"
|
||||
b"- Finance already mapped the required SSO attributes.\n"
|
||||
)
|
||||
),
|
||||
}
|
||||
),
|
||||
"memories": Dir(
|
||||
children={
|
||||
"memory_summary.md": File(
|
||||
content=(
|
||||
b"# Memory summary\n\n"
|
||||
b"- Bluebird should not get production exports before security review "
|
||||
b"is complete.\n"
|
||||
)
|
||||
),
|
||||
"MEMORY.md": File(
|
||||
content=(
|
||||
b"# Task Group: Bluebird Logistics\n\n"
|
||||
b"## What to remember\n\n"
|
||||
b"- A prior export request was paused because security approval was "
|
||||
b"missing.\n\n"
|
||||
b"## Search keywords\n\n"
|
||||
b"Bluebird, data export, security review\n"
|
||||
)
|
||||
),
|
||||
}
|
||||
),
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
def _build_agents(*, model: str, manifest: Manifest) -> tuple[Agent[Any], SandboxAgent[Any]]:
|
||||
evidence_reviewer = SandboxAgent(
|
||||
name="Evidence Reviewer",
|
||||
model=model,
|
||||
handoff_description="Reviews the Bluebird export case in the sandbox.",
|
||||
instructions=(
|
||||
"Use the sandbox workspace for the case review. First, call only `load_skill` for "
|
||||
"`case-note`. After the skill loads, do not call any more tools. Return the customer "
|
||||
"evidence note using the user-provided case facts and memory if relevant. Include "
|
||||
"exactly these sections: `## Facts`, `## Policy`, and `## Recommendation`. Keep it "
|
||||
"under 90 words."
|
||||
),
|
||||
default_manifest=manifest,
|
||||
capabilities=[
|
||||
Memory(read=MemoryReadConfig(live_update=False), generate=None),
|
||||
Shell(),
|
||||
Skills(
|
||||
lazy_from=LocalDirLazySkillSource(
|
||||
source=LocalDir(src=SKILLS_DIR),
|
||||
)
|
||||
),
|
||||
],
|
||||
model_settings=ModelSettings(tool_choice="load_skill"),
|
||||
)
|
||||
|
||||
account_coordinator = Agent(
|
||||
name="Account Coordinator",
|
||||
model=model,
|
||||
instructions=(
|
||||
"You route customer-account review requests. For this request, hand off to the "
|
||||
"evidence reviewer."
|
||||
),
|
||||
handoffs=[
|
||||
handoff(evidence_reviewer, tool_name_override="transfer_to_evidence_reviewer"),
|
||||
],
|
||||
model_settings=ModelSettings(tool_choice="transfer_to_evidence_reviewer"),
|
||||
)
|
||||
return account_coordinator, evidence_reviewer
|
||||
|
||||
|
||||
async def _prestage_case_note_skill(sandbox: BaseSandboxSession) -> None:
|
||||
skill_markdown = (SKILLS_DIR / "case-note" / "SKILL.md").read_bytes()
|
||||
await sandbox.mkdir(Path(".agents/case-note"), parents=True)
|
||||
await sandbox.write(Path(".agents/case-note/SKILL.md"), io.BytesIO(skill_markdown))
|
||||
|
||||
|
||||
def _raw_attr(raw_item: object, name: str) -> object:
|
||||
if isinstance(raw_item, dict):
|
||||
return raw_item.get(name)
|
||||
return getattr(raw_item, name, None)
|
||||
|
||||
|
||||
def _tool_call_name(item: ToolCallItem) -> str:
|
||||
raw_item = item.raw_item
|
||||
raw_type = _raw_attr(raw_item, "type")
|
||||
name = _raw_attr(raw_item, "name")
|
||||
|
||||
if raw_type == "apply_patch_call":
|
||||
return "apply_patch"
|
||||
if raw_type == "mcp_call" and isinstance(name, str):
|
||||
return f"mcp:{name}"
|
||||
if isinstance(name, str) and name:
|
||||
return name
|
||||
if isinstance(raw_type, str) and raw_type:
|
||||
return raw_type
|
||||
return ""
|
||||
|
||||
|
||||
def _handoff_names(items: Sequence[object]) -> list[str]:
|
||||
return [
|
||||
f"{item.source_agent.name} -> {item.target_agent.name}"
|
||||
for item in items
|
||||
if isinstance(item, HandoffOutputItem)
|
||||
]
|
||||
|
||||
|
||||
def _validate_trace_features(
|
||||
result_items: Sequence[object],
|
||||
final_output: object,
|
||||
) -> None:
|
||||
handoff_names = _handoff_names(result_items)
|
||||
tool_calls = [item for item in result_items if isinstance(item, ToolCallItem)]
|
||||
tool_names = [_tool_call_name(item) for item in tool_calls]
|
||||
|
||||
if len(handoff_names) != 1:
|
||||
raise RuntimeError(f"Expected exactly one handoff, saw: {handoff_names}")
|
||||
if "load_skill" not in tool_names:
|
||||
raise RuntimeError(f"Expected a load_skill call, saw: {tool_names}")
|
||||
extra_tool_names = [
|
||||
name for name in tool_names if name not in {"transfer_to_evidence_reviewer", "load_skill"}
|
||||
]
|
||||
if extra_tool_names:
|
||||
raise RuntimeError(
|
||||
f"Expected no tool calls except handoff and load_skill, saw: {tool_names}"
|
||||
)
|
||||
if "## Policy" not in str(final_output):
|
||||
raise RuntimeError("Expected final output to include a Policy section.")
|
||||
|
||||
|
||||
async def main(*, model: str, prompt: str) -> None:
|
||||
manifest = _build_manifest()
|
||||
account_coordinator, _evidence_reviewer = _build_agents(
|
||||
model=model,
|
||||
manifest=manifest,
|
||||
)
|
||||
client = UnixLocalSandboxClient()
|
||||
sandbox = await client.create(manifest=manifest)
|
||||
trace_id = gen_trace_id()
|
||||
trace_url = f"https://platform.openai.com/traces/trace?trace_id={trace_id}"
|
||||
|
||||
try:
|
||||
async with sandbox:
|
||||
await _prestage_case_note_skill(sandbox)
|
||||
with trace(
|
||||
workflow_name="Customer evidence review",
|
||||
trace_id=trace_id,
|
||||
group_id="customer-evidence-review-demo",
|
||||
metadata={
|
||||
"example": "examples/sandbox/tracing_dashboard.py",
|
||||
"features": "regular_agent,sandbox_agent,handoff,load_skill,memory",
|
||||
},
|
||||
):
|
||||
print(f"View trace: {trace_url}\n")
|
||||
result = await Runner.run(
|
||||
account_coordinator,
|
||||
prompt,
|
||||
max_turns=4,
|
||||
run_config=RunConfig(
|
||||
sandbox=SandboxRunConfig(session=sandbox),
|
||||
group_id="customer-evidence-review-demo",
|
||||
),
|
||||
)
|
||||
|
||||
result_items = list(result.new_items)
|
||||
_validate_trace_features(result_items, result.final_output)
|
||||
|
||||
print("=== Final output ===")
|
||||
print(result.final_output)
|
||||
print("\n=== Trace features observed ===")
|
||||
for handoff_name in _handoff_names(result_items):
|
||||
print(f"handoff: {handoff_name}")
|
||||
for item in result_items:
|
||||
if isinstance(item, ToolCallItem):
|
||||
print(f"tool: {_tool_call_name(item)}")
|
||||
finally:
|
||||
try:
|
||||
await client.delete(sandbox)
|
||||
finally:
|
||||
flush_traces()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser(
|
||||
description=(
|
||||
"Generate a trace with one handoff from a regular agent to a sandbox agent, "
|
||||
"plus a load_skill call and sandbox memory."
|
||||
)
|
||||
)
|
||||
parser.add_argument("--model", default=DEFAULT_MODEL, help="Model name to use.")
|
||||
parser.add_argument("--prompt", default=DEFAULT_PROMPT, help="Prompt to send to the agent.")
|
||||
args = parser.parse_args()
|
||||
|
||||
asyncio.run(main(model=args.model, prompt=args.prompt))
|
||||
@@ -0,0 +1,20 @@
|
||||
---
|
||||
name: case-note
|
||||
description: Writes a short customer case note from provided facts and memory.
|
||||
---
|
||||
|
||||
# Case Note
|
||||
|
||||
Use this skill when preparing a short customer case note.
|
||||
|
||||
## Steps
|
||||
|
||||
1. Use the case facts already provided in the conversation.
|
||||
2. Do not call additional tools after loading this skill.
|
||||
3. Include one memory signal if it applies.
|
||||
4. Return a note with these sections:
|
||||
- `## Facts`
|
||||
- `## Policy`
|
||||
- `## Recommendation`
|
||||
|
||||
Keep the note under 90 words.
|
||||
Reference in New Issue
Block a user