autofit_workspace

Imported from PyAutoLabs/autofit_workspace (AGENTS.md). Install upstream with npx skills add PyAutoLabs/autofit_workspace. Copyright stays with the author.

PyAutoFit Workspace — Agent Instructions

This is the tutorial and example workspace for PyAutoFit. These are the canonical, agent-agnostic instructions for this repo. PyAutoFit is a general-purpose Python framework for Bayesian model fitting and inference — model composition, priors, non-linear searches, samples, the aggregator, and graphical models. It is not tied to any science domain; the examples here are deliberately domain-agnostic toy problems (fitting 1D profiles such as Gaussians).

Repository Structure

• scripts/ — Runnable Python scripts, organised by PyAutoFit feature:
  • overview/ — Concise tour of the core features (overview_1_the_basics.py first).
  • searches/ — Every non-linear search: nest.py (dynesty, nautilus), mcmc.py (emcee, zeus), mle.py (drawer, LBFGS), start_point.py.
  • model/ — Model composition (priors, collections, custom model objects).
  • features/ — Advanced features: graphical models, model comparison, search chaining, search grid search, sensitivity mapping, interpolation, shared analysis state.
  • cookbooks/ — Focused API cookbooks (model, result, samples, latent variables, …).
  • plot/ — Plotting search results (dynesty, emcee, nautilus, zeus, GetDist).
  • simulators/ — Scripts that simulate the example datasets fitted throughout.
• notebooks/ — Jupyter notebook versions, generated from scripts/ (do not edit directly).
• config/ — PyAutoFit configuration YAML files (config/general.yaml controls search update intervals, output, logging, profiling, parallel warnings).
• dataset/ — Example 1D datasets.
• output/ — Non-linear search results (generated at runtime, not committed).

There is no projects/ directory.

Running Scripts

Scripts are run from the repository root so relative paths to dataset/, config/, and output/ resolve correctly:

python scripts/overview/overview_1_the_basics.py

There is currently no start_here.py in this workspace — scripts/overview/overview_1_the_basics.py is the canonical entry point and the always-current end-to-end reference. (The runner still honours the convention: if a folder contains a start_here.py, it executes that first, before other scripts and subfolders, since some scripts depend on results it produces.)

Standard imports

import autofit as af
import autofit.plot as aplt

For the canonical workflow (compose a model → configure a search → define an analysis → fit → inspect results), read scripts/overview/overview_1_the_basics.py — it is kept current with the API.

Testing

On CI, every PR is gated by three workflows on Python 3.12 and 3.13: smoke_tests.yml (the smoke runner below — the definition of green), navigator_check.yml (PyAutoHands's reusable navigator-catalogue check; see Notebooks vs Scripts), and url_check.yml (link checking). The smoke and navigator jobs check out PyAutoHands as a sibling and run the PyAuto* libraries from the same-named branch of each source repo, so a workspace PR is validated against matching library branches.

Scripts are tested with the smoke runner, run from the repo root — it executes the curated subset:

python .github/scripts/run_smoke.py

It is driven by smoke_tests.txt (scripts) and smoke_notebooks.txt (notebooks) in the workspace root, with per-entry environment from config/build/profile_smoke.yaml. It prints a [PASS] / [FAIL (exit N)] line per entry, ends with a === Smoke test summary: P/T passed === line, and exits non-zero if any entry failed.

PYAUTO_TEST_MODE is PyAutoFit's own environment variable: it tells the non-linear search to skip its actual sampling, turning a run into a fast structural / end-to-end check that the model composes and the script runs to completion — without paying for inference. A script that fails under it signals a real problem (broken import, renamed API, etc.). PYAUTO_TEST_MODE_SAMPLES=<N> (default 4) additionally sets how many fake samples the bypassed search writes — raise it (e.g. 10000) when a run's samples.csv must be size-representative of real sampling, e.g. for resume/load profiling.

Sandboxed / restricted runs

If numba or matplotlib cannot write to their default cache locations, point them at writable directories:

NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib python scripts/overview/overview_1_the_basics.py

Notebooks vs Scripts

Notebooks in notebooks/ are generated from the .py files in scripts/. Always edit the .py scripts, never the .ipynb notebooks directly. The # %% markers alternate between code and markdown cells.

Generating notebooks

After updating scripts, regenerate the notebooks using the PyAutoHands tool (run from the workspace root):

pip install ipynb-py-convert
git clone https://github.com/PyAutoLabs/PyAutoHands.git ../PyAutoHands
PYTHONPATH=../PyAutoHands/autohands python3 ../PyAutoHands/autohands/generate.py autofit

Commit the regenerated notebooks alongside the script changes. The /generate_and_merge skill automates build + commit + PR + merge.

Bulk-edit safety

When editing the same region across many scripts in one pass (adding a section, renaming a symbol, updating an import block), only rewrite the targeted region. Never produce a whole-file write unless you have read the entire current contents of that file — a whole-file write based on a header skim silently deletes every section below the header. Prefer targeted edits.

Methodological context

PyAutoFit is domain-agnostic, so there is no single science field to reference. Any "context" for its examples is statistical / methodological — Bayesian inference, priors, non-linear search, model comparison — not a scientific subject. Do not graft lensing, galaxy, or other domain concepts onto this workspace.

Related Repos

PyAutoFit sits near the base of the PyAuto stack (all on the PyAutoLabs GitHub org):

• https://github.com/PyAutoLabs/PyAutoNerves — configuration handling (the autonerves dependency).
• https://github.com/PyAutoLabs/PyAutoFit — this library: model composition + non-linear search.
• https://github.com/PyAutoLabs/PyAutoHands — notebook generation + CI.
• https://github.com/PyAutoLabs/PyAutoGalaxy — downstream science library built on PyAutoFit.
• https://github.com/PyAutoLabs/PyAutoLens — downstream science library built on PyAutoFit.
• https://github.com/PyAutoLabs/HowToFit — tutorial lecture series that teaches statistical model-fitting and Bayesian inference with PyAutoFit from first principles; the starting point for beginners new to the framework.

For local development these are typically cloned as siblings of this repo (../PyAutoFit, ../PyAutoNerves, ../PyAutoHands, …).

Local agent skills

Canonical local skills live in skills/<name>/SKILL.md. Claude uses the generated .claude/skills/<name> links; Codex uses generated adapters under .codex/skills/autofit-workspace-<skill>/SKILL.md. Both point to the same instructions. Preserve existing Claude names when adding Codex discovery.

After adding or renaming a skill, regenerate and check both surfaces from the PyAutoLabs workspace (or set PYAUTO_BRAIN to the Brain checkout):

bash "${PYAUTO_BRAIN:-organs/PyAutoBrain}/bin/install.sh" --write-project-discovery autofit_workspace
bash "${PYAUTO_BRAIN:-organs/PyAutoBrain}/bin/install.sh" --check-project-discovery autofit_workspace

Commit the generated adapters with the canonical skill. Standalone clones use these committed adapters; regeneration requires the registered workspace.

Task Workflows

API Update tasks

When assigned an issue titled [API Update]:

1. Read the PR diff in the issue body. Identify every renamed, moved, removed, or changed public API (functions, classes, method signatures, parameter names, import paths).
2. Search all .py files in scripts/ for usages of the old API.
3. Update each file to the new API, preserving existing behaviour, docstrings, and comments.
4. Run python .github/scripts/run_smoke.py from the repo root to test the curated subset.
5. Read the [FAIL (exit N)] output and fix the affected scripts.
6. Repeat 4–5 until the suite is clean.
7. If a script cannot be fixed (ambiguous change, missing dependency), leave it unchanged and list it in the PR description under "Could not update" with the reason.
8. After all scripts pass, regenerate the notebooks (see "Generating notebooks").

General Issue tasks

When assigned a general (non-API) issue:

1. Read the issue description and any linked plan or AI prompt.
2. Identify which scripts need to be created or modified.
3. Only edit files in scripts/. Never edit notebooks/ directly.
4. Preserve all docstrings, comments, and tutorial explanations.
5. Test with python .github/scripts/run_smoke.py after changes.
6. Regenerate notebooks after all scripts pass.

PR description

When opening your PR, include:

• A summary of what changed and why.
• A list of all scripts you updated or created.
• Confirmation that notebooks were regenerated.
• A "Could not update" section for any scripts that still fail, with the error and your assessment.

Never rewrite history

Never rewrite pushed history on any repo with a remote — no git init over a tracked repo, no force-push to main, no fresh-start "Initial commit", no filter-repo / filter-branch / rebase -i on pushed branches. To get a clean tree: git fetch origin && git reset --hard origin/main && git clean -fd.

Sessions end at their deliverable

A session ends when it reports its deliverable — never arm anything that outlives the turn to wait for CI, a review or a merge: no send_later, no subscribe_pr_activity, no CronCreate, no ScheduleWakeup, no /loop, no RemoteTrigger create/update/run. Judge once, report, stop; the human re-runs /prm (or the batch review) when it is green. Measured: five batch members armed hourly check-ins on 2026-08-31, and a mobile /prm re-armed a 60-minute send_later hourly all night on 2026-09-03 with no task active, draining usage.