Environment guide¶
python -m tscfbench is intentionally usable in several environments. The right entry point depends on how formal, reproducible, and automated your workflow needs to be.
Notebook research¶
Best for: Exploration, pedagogy, and first-pass method debugging.
What works well here
- Wrap your data in ImpactCase or PanelCase and inspect results interactively.
- Prototype on built-in baselines before installing heavier optional dependencies.
- Render Markdown reports once the notebook logic stabilizes.
Recommended APIs: ImpactCase, PanelCase, benchmark, benchmark_panel, render_panel_markdown
Recommended CLI commands
python -m tscfbench demo
Install extras: core
Cautions
- Notebook state can drift; switch to JSON specs when you want a reproducible benchmark.
CLI-first research workflow¶
Best for: Reproducible scripts, shared repositories, and low-friction onboarding.
What works well here
- Create specs and reports without writing orchestration code.
- Share the same benchmark recipe across machines and collaborators.
- Use canonical studies as the public entry point for the project.
Recommended APIs: PanelExperimentSpec, CanonicalBenchmarkSpec, SweepMatrixSpec
Recommended CLI commands
python -m tscfbench make-canonical-spec
python -m tscfbench run-canonical
python -m tscfbench run-sweep
Install extras: core, research
Cautions
- Keep output files under version control if they are part of a paper companion or release process.
CI and release engineering¶
Best for: Regression checks, optional-dependency smoke tests, and repeatable releases.
What works well here
- Use snapshot-backed canonical studies to avoid flaky network-bound tests.
- Separate core tests from optional third-party install jobs.
- Render reports as artifacts in CI for easier inspection of failures.
Recommended APIs: CanonicalBenchmarkSpec, run_canonical_benchmark, install_matrix
Recommended CLI commands
python -m tscfbench make-canonical-spec --data-source snapshot
python -m tscfbench run-canonical
Install extras: dev, research
Cautions
- Do not assume every optional third-party package is available on every platform; keep install tiers explicit.
Agent-assisted coding environment¶
Best for: Token-aware research automation, code navigation, and structured tool use.
What works well here
- Build a small bundle and let the agent read manifest-backed artifacts on demand.
- Use function-tool or MCP exports when you want the package to explain its own surface to the agent.
- Keep large files out of chat and let context plans decide what enters the window.
Recommended APIs: AgentResearchTaskSpec, build_panel_agent_bundle, build_context_plan, export_openai_function_tools, TSCFBenchMCPServer
Recommended CLI commands
python -m tscfbench make-agent-spec
python -m tscfbench build-agent-bundle
python -m tscfbench export-openai-tools
python -m tscfbench mcp-server
Install extras: core
Cautions
- Separate planning turns from editing turns if you want tighter control over token usage and tool availability.
Shared server or HPC-style batch environment¶
Best for: Larger sweeps, heavier optional models, and scheduled benchmarks.
What works well here
- Use sweep specs to make runs explicit and easy to rerun on another machine.
- Install only the extras you need for a given benchmark battery.
- Store JSON results and Markdown summaries as durable artifacts rather than relying on notebook state.
Recommended APIs: SweepMatrixSpec, run_sweep, render_sweep_markdown, install_matrix
Recommended CLI commands
python -m tscfbench make-sweep-spec
python -m tscfbench run-sweep
Install extras: research, forecast
Cautions
- Third-party deep-learning or Bayesian stacks may have platform-specific dependency constraints.
Teaching and workshop environment¶
Best for: Live demos, student assignments, and reproducible teaching materials.
What works well here
- Use canonical snapshot studies so every participant sees the same results.
- Prefer the CLI and small example scripts for classrooms with mixed Python skill levels.
- Use docs pages and case studies as the main narrative surface, not only raw API references.
Recommended APIs: list_canonical_studies, CanonicalBenchmarkSpec, render_canonical_markdown
Recommended CLI commands
python -m tscfbench list-canonical-studies
python -m tscfbench run-canonical
Install extras: core
Cautions
- Keep classroom exercises small and deterministic; save external-package comparisons for advanced sessions.