evalshift-sdk

In-process capture SDK for EvalShift.

Install it inside your agent process to record what the agent does — model calls, tool calls, retrievals — and write CLI-valid traces to .evalshift/captures/. The evalshift CLI reads those captures from disk; the SDK and CLI never call each other.

• Distribution: evalshift-sdk · import name: evalshift
• Runtime deps: none (stdlib-only)
• Python: >= 3.10
• Capture is off by default — set EVALSHIFT_CAPTURE=1 to record.
• License: MIT

For AI coding agents

Point your coding agent at the dense, single-file reference for the piece it is working on:

• EvalShift CLI: https://www.evalshift.dev/cli-llms-full.txt
• EvalShift SDK: https://www.evalshift.dev/sdk-llms-full.txt (source of truth: llms-full.txt in this repo)
• EvalShift GitHub Action (CI): https://www.evalshift.dev/ci-llms-full.txt

Install

pip install evalshift-sdk
# or
uv add evalshift-sdk

Optional LangChain integration (EvalShiftCallbackHandler):

pip install "evalshift-sdk[langchain]"   # adds langchain-core>=0.2

Optional provider client wrappers (wrap_openai / wrap_anthropic / wrap_genai):

pip install "evalshift-sdk[openai]"        # openai>=1.40
pip install "evalshift-sdk[anthropic]"     # anthropic>=0.40
pip install "evalshift-sdk[google-genai]"  # google-genai>=1.0

Every adapter module is import-guarded, so the SDK stays dependency-free at runtime unless you opt in.

Co-install note: the EvalShift CLI (PyPI evalshift, import package evalshift_cli) depends on this SDK, so both live in one environment and pip install evalshift brings the SDK with it. Production agents that only record captures install evalshift-sdk alone.

Usage

from evalshift import capture

@capture.agent(suite="support_agent", redact=True, tools=[])   # no-op unless EVALSHIFT_CAPTURE=1
def handle_ticket(query): ...

Already calling a provider SDK directly? Wrap the client once and every call inside the agent records itself — model, tools offered, tool calls requested, usage, latency:

from openai import OpenAI
from evalshift.adapters.openai import wrap_openai      # also: wrap_anthropic, wrap_genai

client = wrap_openai(OpenAI())                          # OpenAI(base_url=...) covers DeepSeek, Ollama, vLLM, Groq, ...

Full guide: DOCS.md · dense LLM reference: https://www.evalshift.dev/sdk-llms-full.txt · locked design decisions: docs/DECISIONS.md

Keeping captures/ bounded

Capture writes one JSON file per sampled invocation, so .evalshift/captures/ is kept bounded by default (no configuration needed): identical-input re-runs are de-duplicated, and each suite directory is capped at the 200 newest captures (oldest evicted). Tune it with env vars — no code change required (precedence: an explicit configure(...) call > env var > built-in default):

A malformed value falls back to the default (capture never crashes). To restore fully unbounded capture: EVALSHIFT_MAX_CAPTURES=0 EVALSHIFT_DEDUP=off. Disable dedup with off (or 0/none); false/no are not recognised and leave dedup on. EVALSHIFT_SAMPLE_RATE=0 means sampling off (capture every run) — it is not the same as configure(sample_rate=0.0), which captures nothing. The same knobs are available in code via configure(max_captures=..., dedup=..., capture_ttl=..., sample_rate=...).

Further reading

• Build a golden eval suite from production traffic — the capture → promote loop this SDK feeds.
• Evaluating agent tool calls: what text evals can't see — why the captured tool calls matter more than the output text.
• Capture SDK docs