Changelog

Newest first, following Keep a Changelog and semantic versioning.

What is versioned is the site and the tooling that builds it, not a library API. Captured traces are never rewritten, so a version bump changes how the atlas is built or presented, never what was observed.

Unreleased

1.3.0 - 2026-09-21

Changed

• The flow, renderer and stylesheet tests move to hook-atlas, which is where that code lives. They were left behind by the extraction, so this project was testing somebody else's package while that package shipped nearly untested.

Added

• A scenario with no pages for the oldest releases says why on its index. The xdist scenario starts at pytest 7.0 because the pytest-xdist it pins needs it; unexplained, that read as a gap rather than a limit.
• The site and the README point at hook-atlas for tracing your own project. This repository is the pytest-specific half - scenarios, traces and pages - and is worth reading rather than installing.
• pytest 6.0 through 7.3 are captured, so the atlas now covers every release from 6.0 onward. They need an interpreter older than this package can run on - 6.0 tops out at Python 3.9 - so capture-missing takes a --base-python to build capture environments from.
• build --majors N chooses how many pytest majors get pages, with 0 for all of them. Traces are never dropped, so the window can be widened and the site rebuilt at any time without re-capturing anything.
• Scenarios declare which pytest releases they support. The xdist scenario needs pytest 7, and is now skipped for older releases deliberately and visibly rather than by accident.

Fixed

• Requires hook-atlas 0.2, for the DocLinks.documented field the fix below needs.

• Pages no longer link hooks their reference does not document. pytest 6.0 and 6.1 have no published documentation, so those pages fall back to stable, where hooks pytest has removed since do not exist - and the links 404ed. linkcheck now asks the same question the build asks, rather than checking links the site does not emit, treats an unreachable page as a fetch failure rather than a dead link, and reads each reference page once instead of once per trace.

• A capture is refused when the environment does not hold the pytest that was asked for. Cheap insurance against a trace filed under one version while describing another - pip itself refuses the obvious version conflicts.

• linkcheck read the pytest version from a trace field that only schema 2 traces have.

1.2.0 - 2026-09-16

Changed

• Breaking: the command is now pytest-hook-atlas build rather than hook-atlas build. hook-atlas belongs to the generic tool, which gained its own trace, draw and check; both packages previously declared the same console script, so installing them together left which one you got undefined. A minor rather than a major version because nothing here has ever been published - the rename reaches no installed copy but our own.
• The tracing and drawing move to hook-atlas, a package that knows nothing about pytest and can be pointed at tox, devpi or anything else built on pluggy. What stays here is what makes this pytest's atlas: the four phases, the documentation links, the scenarios and the pages. The generated site is byte-identical across the move.

1.1.0 - 2026-09-13

Added

• Traces record which plugin declared each set of hookspecs, and at what version, so a capture can say which xdist - or which project - produced it. Read from __version__ or stdlib metadata, never by adding a dependency to the environment being measured.
• An xdist scenario, drawn once per process - the controller never collects or runs a test, and only it sees the report-serialization hooks. Workers are grouped by the flow they produced, not by name: under a splitting scheduler gw0 and gw1 swap flows between captures, so a name is a race, not an identity. Every distinct worker flow is drawn in full, however many there are.
• Overview columns start at the same height. A cluster's label is two lines when its hook has semantics to show and one when it does not, and that difference was landing on the column top.
• Each diagram says which command produced it, per process. A worker's is empty, because xdist starts it over execnet rather than from a command line.
• Long repetitive stretches fold when drawn: one cycle, then a dashed box naming the hooks it stands for. The xdist controller's run loop is a hundred steps of reports arriving in whatever order workers finish, so it collapses to nothing and rendered seven thousand pixels tall. Never fingerprinted.
• Scenarios can run across several processes. Each writes its own trace, named <scenario>.<process>.json after xdist's logical worker names, and a scenario can declare the packages its capture virtualenv needs. Groundwork for the xdist scenario, where the controller and workers see genuinely different things - the controller never collects or runs a test.
• Hookspecs are read from the live plugin manager rather than by importing _pytest.hookspec, so hooks a project declares are described too - pytest contributes 52, pytest-xdist adds 12, and any conftest or plugin calling add_hookspecs contributes its own. Traces record which module declared each hook, and hooks pytest did not declare render without a documentation link rather than with a dead one. Trace schema is now 2.
• Three scenarios reaching hooks no other scenario touched: edge-cases (assertion comparison and assertion-pass, deselection, string skipif conditions, the report header, both debugger hooks), interrupted (a run stopped by Ctrl-C) and internal-error (a hook implementation raising during collection). Coverage rises from 40 of pytest's 52 declared hooks to 49.
• Scenarios declare what they expect - which hooks they exist to reach, whether the session completes, and any minimum hook count. Nothing in the tooling imposes a minimum: a run that dies in its first hookimpl is drawn as it happened.
• Hook tables name the plugins implementing each hook, in pluggy's call order, with full module paths.
• Implementers are reconciled across each page's release range, with footnotes giving the deltas where releases disagree - pytest moves implementations between its own plugins without changing the flow.
• Implementations are annotated with pluggy's ordering markers - [wrapper], [tryfirst], [trylast] - so the call order explains itself rather than looking arbitrary.
• A toggle to hide pytest's own plugins, leaving what the project under test contributes. Only appears where there is something to reveal.
• A sweep over every generated page: all four stages present, each diagram rendered and plausibly sized, links live, footnotes resolved, nothing swallowed as HTML.

Fixed

• Long stretches sometimes folded to nothing, drawing the xdist controller at 14530pt. A hook that ends a stretch rather than belonging to it - pytest_testnodedown, once per worker - kept the whole stretch drawn.
• Worker sections sorted gw10 before gw2.
• The page sweep only looked at single-process pages, so nothing checked the xdist pages it was written to cover.
• every-hook-conftest was silently broken on pytest 8.0.x and 9.0.x. pytest treats its own removal warnings as errors, so the conftest failed to import and the run collapsed to five hook calls while the page still claimed every hook was implemented. It now suppresses those warnings and implements everything the version declares - which revealed that 8.0 still calls the deprecated pytest_cmdline_preparse, too early for a conftest to serve.
• Implementers were listed in pluggy's storage order, which is the reverse of call order, so the table read upside down.
• The plugin filter hid nothing: two rules of equal specificity fought, and the later one un-hid what the first had hidden.
• The plugin registered as <anonymous> was swallowed as an HTML tag.
• Offline builds emitted unverified documentation links. pytest's own 9.1.x docs serve a redirect loop, so they were dead; offline builds now fall back to stable.
• A conftest.py is labelled by its path, since every conftest imports as the module conftest.
• Long hook names no longer break mid-word in tables.
• Tests no longer depend on, or write into, generated files in the repository. build() takes a config path, so a build into a temporary directory is self-contained.

1.0.0 - 2026-09-10

First release. A ground-up rework of doc_pytest_flow_chart, which generated a single flat SVG from a scraped pytest log and had been dormant five years.

Added

• Structural capture. A standalone tracer installs pluggy's add_hookcall_monitoring and records the exact call tree - order, nesting, and which plugin supplied each implementation. Traces are committed JSON.
• A version matrix. Releases are discovered from PyPI and captured in a throwaway virtualenv holding exactly that pytest. Yanked releases are skipped; the Python each needs comes from its Trove classifiers. 29 captured (7.3.2 - 9.1.1); 6.0 - 7.3 await older interpreters.
• Grouping by flow. Releases producing an identical flow share one page, labelled with the range it covers.
• Scenarios as a first-class concept: a small pytest project plus how to invoke it, each page linking to the code that produced it.
• Ordered flow diagrams in Graphviz - phases as columns, steps top to bottom, nesting as clusters. Column headings jump to their detail; every hook links to its documentation.
• A weekly watcher that captures new pytest releases and commits them.
• Pages for every release, what changed, and why the diagrams look as they do.
• A weekly link checker over every hook-to-documentation anchor.

Changed

• Diagrams show order, not just containment. The old design drew only "called inside", hiding that setup, call and teardown are each followed by their own makereport and logreport.
• Hook semantics are written in words rather than encoded in node shapes.
• Packaging modernised to a single PEP 621 pyproject.toml; Python 3.11+.

Fixed

• Documentation links, dead for years. A stale base URL after pytest moved reference.html, plus rstrip("[hook]") stripping characters rather than a suffix, turning pytest_addhooks into pytest_add. Pinned URLs are now verified before use.
• Traces are reproducible as whole files. The throwaway capture directory leaked into the recorded command line and conftest plugin names, so every capture produced a diff even when nothing had changed.
• Diagram colours are computed for contrast in both themes. The conventional palette put two of its four hues at a colour difference of 3.9 under deuteranopia - effectively identical.

Removed

• The pytest --debug log scraper, superseded by structural capture.
• Cookiecutter scaffolding: setup.py, setup.cfg, MANIFEST.in, tox.ini, Makefile, and the unused Sphinx tree.