respan-instrumentation-skyvern

Trace Skyvern browser agents in Respan. Each Skyvern run becomes one trace:

agent.skyvern_task                 Skyvern.run_task (or run_workflow / login / download_files)
├── task  (skyvern.step)           one per agent step: page URL, actions + results, screenshot
│   ├── llm.<model>                extract-actions: the model call that chose the actions
│   ├── tool.click                 one per browser action (click, input_text, select_option, ...)
│   └── tool.input_text
├── llm.<model>                    check-user-goal, the speculative next-step plan, extraction
└── tool.complete                  the action that finished the run

page.act / page.extract / page.validate / page.ai_* calls on a Skyvern browser page are traced the same way, one agent.<ai_action> trace per call.

Install

pip install respan-ai respan-instrumentation-skyvern
pip install "skyvern[local]"         # the in-process engine, for Skyvern.local()
python -m playwright install chromium

Use

from respan import Respan
from respan_instrumentation_skyvern import SkyvernInstrumentor

respan = Respan(api_key="...", instrumentations=[SkyvernInstrumentor()])

from skyvern import Skyvern

skyvern = Skyvern.local()
result = await skyvern.run_task(
    prompt="Find the top post on Hacker News", url="https://news.ycombinator.com",
    wait_for_completion=True,
)

Skyvern's model can go through the Respan gateway (OpenAI-compatible), so one key covers the model and the traces:

from skyvern.schemas.llm import LLMConfig

skyvern = Skyvern.local(llm_config=LLMConfig(
    model_name="openai/gpt-4.1-mini", required_env_vars=[], supports_vision=True,
    add_assistant_prefix=False,
    litellm_params={"api_base": "https://api.respan.ai/api", "api_key": RESPAN_API_KEY},
))

A full runnable example with live checks is in python-sdks/examples/skyvern.

What is captured

• Local and cloud. With Skyvern.local() the whole tree is captured. Against Skyvern Cloud (Skyvern(api_key=...)) the engine runs remotely, so you get the run span with its status, output and recording link.
• Grouping. Thread, session and trace group default to the Skyvern run id, so a run's spans group together; values you set with propagate_attributes win.
• Secrets. Values Skyvern resolves as secrets during a run (credential and secret workflow parameters, one-time codes) are replaced by <secret>name</secret> in every span.
• Content off. SkyvernInstrumentor(capture_content=False) (or Respan's content-tracing switch) keeps the tree, ids, statuses and token counts and drops prompts, outputs, URLs and screenshots. capture_screenshots=False drops only the images.
• No double logging. Skyvern's model calls go through litellm; while a Skyvern chat span is open, the LiteLLM and OpenAI instrumentations are suppressed.

Skyvern's own OpenTelemetry spans (tracer skyvern) are not exported; Respan spans are parented to each other across them. See docs/GAPS.md for known limitations.

Tests

tests/test_baseline.py drives Skyvern's real engine and a headless Chromium on a local page with a scripted model, and checks every exported span (needs tests/requirements.txt and python -m playwright install chromium):

pip install -e . -r tests/requirements.txt pytest pytest-asyncio
python -m pytest tests -q