Comparing changes

The report() returned -1 as the exit code when multi-env runs had failures. On Unix, SystemExit(-1) wraps to exit code 255 (unsigned byte), which works fine. But on Windows CMD, the exit code stays -1, and IF ERRORLEVEL 1 checks %ERRORLEVEL% >= 1 — since -1 < 1, the failure was silently treated as success. The only behavioral difference is that multi-env failures now return 1 instead of 255 on Unix — which is actually more idiomatic. The single-env case already returned the actual command exit code (e.g., 1, 2), so this makes the multi-env case consistent. Fixes: #2945  - [x] ran the linter to address style issues (`tox -e fix`) - [x] wrote descriptive pull request text - [x] ensured there are test(s) validating the fix - [x] added news fragment in `docs/changelog` folder - [ ] updated/extended the documentation

A common pain point for tox users is referencing files by wildcard patterns in configuration — uploading all wheels from a dist directory, pointing deps at requirement files matching a pattern, or collecting test artifacts. Until now this required shell expansion or external scripting, which breaks cross-platform compatibility and complicates configuration. This adds a `{glob:PATTERN}` substitution that expands file system glob patterns inline. It works in both INI string syntax (`{glob:dist/*.whl}`, with optional `{glob:dist/*.whl:fallback}` default) and TOML dict syntax (`{ replace = "glob", pattern = "dist/*.whl", extend = true }`). Relative patterns resolve against `tox_root`, results are sorted for deterministic builds, and `**` enables recursive directory matching. The `extend` flag in TOML dict syntax controls whether matched paths are inserted as separate list elements (useful in `commands`) or joined as a single space-separated string (useful in `description` or other string fields). When no files match and no default is provided, the result is an empty string or empty list depending on context. Closes #1571

Adds a new `sdist-wheel` package type that builds a wheel from a source distribution rather than directly from the source tree. This validates that the sdist is complete and that the package can be correctly built from it — catching missing files or packaging errors early. Usage [env_run_base] package = "sdist-wheel" [testenv] package = sdist-wheel How it works 1. Builds an sdist via build_sdist 2. Extracts the sdist to a temporary directory 3. Swaps the package root to the extracted source 4. Builds a wheel via build_wheel from that extracted sdist 5. Cleans up the sdist and extraction directory  - [x] ran the linter to address style issues (`tox -e fix`) - [x] wrote descriptive pull request text - [x] ensured there are test(s) validating the fix - [x] added news fragment in `docs/changelog` folder - [x] updated/extended the documentation --------- Signed-off-by: Bernát Gábor <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Bernát Gábor <[email protected]>

TOML configurations had no way to conditionally select values based on environment variables, forcing users to rely on external scripts or fall back to INI-only factor conditions. This was one of the most frequently requested features (#3650), particularly for CI/CD workflows where behavior needs to vary based on environment context. The `replace = "if"` syntax introduces a `condition` expression field that supports `env.VAR` lookups, `==`/`!=` comparisons, and `and`/`or`/`not` boolean logic. 🔧 Expressions are parsed safely via Python's `ast` module — only a restricted set of node types is evaluated, so arbitrary code execution is not possible. The `env.VAR` lookups read directly from `os.environ` rather than `set_env`, avoiding circular dependency issues when conditionals are used inside `set_env` itself. The expression language was chosen over PEP 508 markers because the `packaging` library's marker evaluator has a hard-coded set of allowed variable names and cannot be extended with `os.environ` lookups. A custom `ast`-based evaluator provides the same familiar Python syntax (`and`, `or`, `not`, `==`, `!=`) while supporting the actual use case of environment variable inspection. Closes #3650

Automatically create a .gitignore file containing * in tox-managed directories so users don't need to manually add .tox (or other work directories) to their project's .gitignore.  - [x] ran the linter to address style issues (`tox -e fix`) - [x] wrote descriptive pull request text - [x] ensured there are test(s) validating the fix - [x] added news fragment in `docs/changelog` folder - [ ] updated/extended the documentation

Users running commands subject to transient failures -- network timeouts, flaky integration tests, intermittent CI infrastructure issues -- had no built-in way to automatically retry. The only workarounds were shell-level retry loops or wrapping commands in helper scripts, both of which add friction and reduce portability across platforms. This adds a `commands_retry` integer configuration option (default `0`) to each environment. Setting it to `N` means each command can be attempted up to `N + 1` times total. The retry applies uniformly to `commands_pre`, `commands`, and `commands_post`. When a command fails and retries remain, tox logs a warning with the attempt count and re-executes. Commands prefixed with `-` (ignore exit code) are never retried since their failures are already ignored. The feature composes naturally with `ignore_errors` -- retries are exhausted first, then the existing error-handling logic takes over. The default of `0` preserves existing behavior. No configuration changes are required for users who don't need retries. Fixes #1578

When `set_env` both inherits via cross-section substitution (e.g., `{[testenv]set_env}`) and explicitly overrides the same variable in the same section, the explicit value is silently replaced by the inherited one. 🐛 This regression was introduced in v4.39.0 and breaks a common configuration pattern where users layer environment variables across sections. The root cause is a timing issue in `_iter_needs_replacement()`. During iteration, `load()` moves explicitly-set keys from `_raw` to `_materialized`. By the time the deferred cross-section substitution expands, the `key not in self._raw` guard no longer sees those keys, so inherited values overwrite explicit ones. The fix tracks originally-defined keys in a separate `_defined_keys` set that persists across `load()` calls, ensuring deferred substitutions never overwrite values the user explicitly configured. All existing precedence behaviors are preserved: defaults still get overridden by cross-section values (#2872), explicit values still win over cross-section values (#2831), and `file|` ordering still respects later inline definitions (#3335). Fixes #3773 --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Comparing changes

Open a pull request

Commits on Feb 19, 2026

This comparison is taking too long to generate.

Uh oh!