Skip to content

CHANGELOG.md

Latest commit

 

History

History
116 lines (103 loc) · 6.09 KB

CHANGELOG.md

File metadata and controls

116 lines (103 loc) · 6.09 KB
Raw

Changelog

All notable changes to ntask are documented in this file.

The format loosely follows Keep a Changelog; this project uses semantic versioning.

[Unreleased]

Added

  • TUI now has a live log pane docked below the DAG tree that tails the currently-running task's .log file. When a task moves to running the pane clears and rebinds; on completion it freezes on the final content so you can still read it. Polls every ~150ms.

Fixed

  • Structured loggers (stdlib logging and structlog configured with the stdlib handler) were writing directly to the original sys.__stderr__ they captured at app-import time, fighting Textual for the same stream and spilling JSON across the live DAG view. The executor now walks all logging.StreamHandler instances at run-start and rebinds those whose .stream is the captured-original stdout/stderr to point at our _LogTee. FileHandler instances and handlers bound to custom streams are left alone. Originals are restored on run-end.
  • _LogTee (the sys.stdout / sys.stderr proxy installed by the executor to capture per-task output) was writing to the underlying stream even under the TUI. Textual owns stdout but not stderr, so library code that used logging to stderr (very common — structured JSON loggers, etc.) bled raw lines straight onto the live DAG view. The tee now consults _current_silent_capture and suppresses the underlying write while the TUI is active. The per-task log file still receives every line, so nothing is dropped — only the terminal stream is quiet.

Documentation

  • docs/design/pre-post-hooks.md captures the deferred design for per-task before/after hooks: why composition (@with_reset) is the recommended pattern today, and the constraints we'd commit to if we ever do build module-level hooks.
  • README now includes a worked --graph --graph-format mermaid example embedded as a real Mermaid block (the project's own DAG) so users can see what to paste into their own design docs. See ISSUES.md for the target-scoping bug found while verifying this.
  • New README section "When NOT to use @cached" documenting the categories of task — side effects, mutating shared state, time- or randomness- dependent, world-probing — where caching is the wrong default.

Added

  • The line-based renderers (LogRenderer, RichRenderer) now prefix each running <fqn> line with [k/n done; next: <fqn>] so multi-task runs no longer feel positionally opaque. Silent for single-task runs. The TUI already provided this implicitly via the live tree.
  • The end-of-run summary now lists the failed tasks by name on a second line — useful for --keep-going runs where there was no fail-fast and the user would otherwise have to scroll back through the log.
  • Shell completion: ntask --completion bash|zsh|fish prints a completion script. The script calls back into ntask --completion-tasks and ntask --completion-flags <task>, so task and per-task flag suggestions stay in sync with the project's actual tasks.py.
  • ntask init scaffolds a minimal tasks.py in the current directory. --template plain|django|fastapi chooses the boilerplate (django/fastapi templates ship the framework bootstrap block users had to hand-roll). Refuses to overwrite an existing tasks.py without --force. init is also reserved, so naming a task init now emits the same shadowing warning as clean / watch.
  • ntask --list now renders each task's CLI signature next to its name — <required: type>, [--name=default], [--flag] for bool default-False, [--no-flag] for default-True. Long signatures are truncated at 40 chars so the table stays readable.
  • @task(deps=...) now accepts a zero-arg callable returning an iterable of task refs — e.g. @task(deps=lambda: [a, b]) — so dependent tasks can sit above their deps in the source file without the linter complaining. The resolver is invoked once at graph-build time and memoised.
  • Task failures now stream a traceback to stderr by default and append the full traceback to the per-task log file. New --tb global flag accepts short (default), long, line, or none — the renderer's one-line x <fqn> FAILED: ... summary on stdout is unchanged.
  • Per-task log files are now written for every run, regardless of which renderer is active. The on-disk layout is .ntask/runs/<UTC-run-id>/<task-fqn>.log plus a run.json manifest with counts, per-task durations, and log paths. Override the location with the global --log-dir DIR flag.
  • Raw print() output from inside a task body is captured to the per-task log file via a context-aware sys.stdout / sys.stderr tee. Previously the TUI absorbed those writes and they could not be recovered after the run.
  • Retention: only the --max-runs most recent run directories are kept; default is 50. --max-runs 0 disables retention.

Fixed

  • Registering a top-level task whose name collides with a built-in subcommand (clean, watch) now emits a UserWarning at registration time. Routing is unchanged — the built-in still wins — but the warning makes it obvious that the user task body will never execute.

Notes

  • shell() subprocess output in the default sequential line renderer is not yet pumped through the Python-level tee, so it appears on the terminal but is not appended to the per-task log file in that mode. The TUI silent-log path and the parallel-prefix path both already capture it correctly; closing the gap in the default path is tracked separately.

[1.1.0]

Changed

  • The TUI no longer auto-closes when the run finishes. It stays mounted with the final DAG state and a summary footer until you press q, Esc, or Ctrl+C. Non-interactive runs (CI, piped output) are unaffected — the TUI is still gated on sys.stdout.isatty().
  • The TUI footer now includes the run summary (counts + log path) plus a press q/esc to quit hint while the app is showing it. This was previously rendered only by the line-based renderers.

[1.0.0]

Initial public release.