dataframe-code-agents
Use ai-data-science-team single-agent pandas code generators for cleaning, wrangling, visualization, and feature engineering workflows.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- dataframe-code-agents
- license
- MIT
- metadata
- { "import": "ai_data_science_team", "package": "ai-data-science-team", "version": "0.0.0.9017", "disco-role": "operating" }
- description
- Use ai-data-science-team single-agent pandas code generators for cleaning, wrangling, visualization, and feature engineering workflows.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md11.2 KB
- 参考文档references/human-in-the-loop-and-logging.md6.5 KB
- 参考文档references/troubleshooting.md8.2 KB
- 参考文档references/workflows.md8.8 KB
- 脚本scripts/inspect_dataframe_agents.py11.9 KB
03 SKILL.md 正文
dataframe-code-agents
Use this sub-skill when the task is to run or explain one ai-data-science-team pandas code-generation agent over already-loaded pandas data: cleaning, wrangling, Plotly visualization, or feature engineering.
The package agents call the caller-provided LangChain-compatible model when invoked. This operating skill does not provide or configure that model and its bundled inspection script performs no model calls.
Route first
| User intent | Use here? | Target |
|---|---|---|
| Load local CSV/Excel/JSON/Parquet/Pickle files, discover files, or summarize a DataFrame without code generation | No | ../data-access-and-eda/SKILL.md |
| Clean one DataFrame with generated pandas code | Yes | DataCleaningAgent |
| Wrangle, join, reshape, aggregate, or transform one or more DataFrames with generated pandas code | Yes | DataWranglingAgent |
| Generate a Plotly chart from one DataFrame with generated code | Yes | DataVisualizationAgent |
| Produce a model-ready feature table with generic encodings and optional target preservation | Yes | FeatureEngineeringAgent |
| Query a SQL database or reason about SQL safety | No | ../sql-analysis/SKILL.md |
| Compose multiple agents, use analyst teams, supervisors, or Streamlit apps | No | ../multiagent-and-app-workflows/SKILL.md |
| Train H2O models or inspect MLflow tools | No | ../modeling-and-mlflow/SKILL.md |
Minimum runtime assumptions
ai_data_science_teamimports successfully in the current Python environment.- The caller already has data in memory as a pandas
DataFrame, a dict convertible to a DataFrame, or for wrangling a list of DataFrames/dicts. - The caller supplies a LangChain-compatible chat model object as
model=llmbefore invoking an agent. - Data loading, EDA-only reporting, SQL access, app launching, model training, and service orchestration are delegated to sibling sub-skills.
Operating checklist
- Identify the single-agent task: cleaning, wrangling, visualization, or feature engineering.
- Verify the data shape is compatible with the selected agent.
- Choose conservative execution settings:
n_sampleslow enough for wide data.max_retriesbounded, usually1to3.log=Trueonly when the user wants generated code/error files, with a user-selected relativelog_path.human_in_the_loop=Trueonly when a review/resume interaction is acceptable.
- Invoke through the class wrapper (
DataCleaningAgent,DataWranglingAgent,DataVisualizationAgent, orFeatureEngineeringAgent) unless the caller specifically needs the lower-levelmake_*_agentgraph. - Retrieve results through getters or
get_response(); inspect generated code before reusing it outside the package workflow. - If generated code fails, use the returned error key, bounded retries, logs, and the troubleshooting table rather than rerunning blindly.
Fast reference map
- API signatures, response keys, getters, and shared graph behavior:
references/api-reference.md - End-to-end cleaning, wrangling, visualization, and feature engineering patterns:
references/workflows.md - Human review, checkpointers, resume commands, and generated-code logging:
references/human-in-the-loop-and-logging.md - Symptom-based fixes for token pressure, sandbox errors, malformed outputs, and Plotly warnings:
references/troubleshooting.md - Safe import/signature/sandbox probe with no model call:
scripts/inspect_dataframe_agents.py
Key safety facts
- Generated pandas/Plotly code is executed by the package through a subprocess sandbox with a timeout, a memory limit, and blocked imports for common filesystem, process, and network modules.
- The sandbox is a guardrail, not a proof of correctness. Tell the user to review generated functions, especially before copying them into production notebooks or pipelines.
- The default report node is deterministic: it packages selected state keys into an
AIMessagewithout another model explanation call. - Human-in-the-loop review pauses after code execution when enabled, so use a checkpointer and stable
thread_idif the run must be resumed.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).