[renovate]

This PR contains the following updates:

Package	Change	Age	Confidence
jupyddl (changelog)	==2.3.0 → ==2.5.0

---

Release Notes

openplan-labs/PythonPDDL (jupyddl)

v2.5.0

Compare Source

Added

• Every merge is a release, decided by what was merged. Pull request
  titles are Conventional Commits (a pr-title check enforces it), and on
  each merge release.yml takes the largest bump the PR's title and commits
  ask for — fix → patch, feat → minor, feat! → major — bumps the version
  everywhere, dates this changelog, commits and tags main, and publishes to
  PyPI. No release PR and nothing to approve. A release:* label overrides
  the bump. See .docs/RELEASING.md.
• The intent to publish, written down. .docs/PUBLISHING.md says why and
  where we mean to publish jupyddl as citable research software — a Zenodo DOI
  per release, a JOSS paper, ICAPS workshops for the research results — with a
  readiness checklist, the paper outline, the AI-usage disclosure JOSS requires
  and how contributors can be considered for authorship. The README states the
  project's aim (a framework to build on and to explain algorithms with) and
  its Cite section says which version the 2021 preprint describes.
• Plugins. A package can now add planners, heuristics and generators by
  declaring jupyddl.planners, jupyddl.heuristics or jupyddl.generators
  entry points; once it is installed, jupyddl solve -s <its planner> works
  without anyone importing it. Discovery is lazy (the CLI runs it before
  building its parser, the library only when asked for a name it does not
  know), a built-in name always wins, a plugin that fails to import is skipped
  with a warning, and JUPYDDL_NO_PLUGINS=1 turns it off. jupyddl plugins
  lists what was found.
• .docs/EXTENDING.md, a guide to building on the framework: a heuristic,
  a planner from a priority function, a planner from scratch with the full
  budget/observer contract, a generator, how to test the guarantees you claim,
  and how to ship it all as a plugin. Every example runs in the test suite, so
  the guide cannot drift from the API.
• best_first accepts a tuple priority for tie-breaking
  (lambda g, h: (g + h, h)); observers are given its first element as f.
• Community files for outside contributors: a security policy, a
  pull-request template, a Propose an algorithm issue template, and an
  "Adding an algorithm" section in CONTRIBUTING.md.

Changed

• CI now runs on pull requests, including from forks. tests and build
  were on: [push], which never fires for a fork, so an outside contributor's
  PR showed no checks at all. They now run on every pull request and on pushes
  to main.

• Documentation lives in one place: .docs/. It was spread over three
  directories — docs/ for the contributor files and the README charts,
  promo/ for the videos and screenshots, .docs/ for the research notes —
  with no rule saying which took what, so every new file was a guess. The
  research notes, the release runbook and every image, video and measurement
  cache are now under .docs/, with binaries in .docs/assets/. This also
  matches pymapf, the sibling project.

  Paths quoted in the released sections below are the ones that were correct
  at the time. promo/rl-data.json is now .docs/assets/rl-data.json, and
  docs/images/ is now .docs/assets/.

  CONTRIBUTING.md and CODE_OF_CONDUCT.md moved to .github/ rather than
  .docs/: GitHub only recognises those two files in the repository root,
  .github/ or docs/, so filing them under .docs/ would have silently
  dropped the contributing link on the issue and pull-request forms.

Fixed

Wrong answers from the task model and the heuristics. Each of these returned a
plan, a cost or a value without complaint, and validate_plan accepted the
plans because it replays them through the same model. They are now pinned by
tests/test_semantics.py on tiny tasks whose right answer is obvious.

• Derived facts were never retracted. An action that deleted the support of
  a derived fact left it true, and (not <derived>) held from the initial state
  on. Axioms are now recomputed from scratch after every step, and a negated
  derived fact is maintained with them. Negating a derived predicate inside
  another rule needs stratification and is now refused by name.
• An atom both added and deleted left its complement true as well, so
  (not (at a)) and (at a) could hold together, and a when that re-added a
  deleted atom did the same. For complement facts the delete now wins.
• Action costs. Under a (total-cost) metric an action with no cost effect
  now costs 0, as in IPC planners, rather than 1 — on pegsol and sokoban the
  "optimal" planners were optimising the wrong objective. A cost inside when
  was charged whether or not the condition held; it is now refused, as is a
  negative cost.
• A numeric fluent that is never defined read as 0, so (distance a c) for
  a road that does not exist cost nothing. Conditions and effects reading one
  now fail, as PDDL 2.1 specifies.
• blind was not admissible when any action is free, which every PDDL 3
  compilation introduces, so astar+blind could return a costlier plan.
• h1/h2/hm crashed on every numeric or timed task, and LM-cut
  crashed on an empty goal, which made the default solve() crash too.
• lmcut and h2 claimed admissibility on tasks with conditional effects,
  where they can overestimate. The instance flag now says so.
• solve_task(task, "astar") used hff, which is not admissible, so the
  default A* was not optimal. Optimal planners now default to lmcut. -H none
  also fell through to hff; it now means blind.
• make_heuristic(LMCutHeuristic, task) returned the class unbuilt, and the
  planner then crashed calling it on a state. Classes and partials are now built.
• Results depended on PYTHONHASHSEED. Fact ids followed string hashing,
  so LM-cut broke ties differently in every process: gripper with astar/lmcut
  expanded 1725, 1758 or 1760 nodes and returned different plans. Ids are now
  assigned in sorted order, so a benchmark reproduces.
• iddfs, idastar and ehc ignored their budget, so --time-limit 0.5
  on blocksworld12 ran for minutes. beam polled once per layer and bnb kept
  evaluating LM-cut on every sibling after the budget fired; both now stop at
  once. A new contract test runs every planner where the budget is the only
  thing that stops it.
• A depth cap was reported as "no plan exists" (iddfs, bnb), and a
  bnb plan found under one as optimal. Hitting the cap now sets truncated,
  as does a search deeper than Python's recursion limit.
• bfs and iddfs were listed as optimal. They minimise plan length;
  on errands they return cost 16 where 10 is optimal. They are no longer in
  OPTIMAL_PLANNERS, and the workbench no longer badges them.
• The live dashboard said "no plan found" when the budget stopped the
  search, smeared every repaint on a narrow terminal (lines were never clipped
  to its width), and logged ten times too often when not on a terminal.
• idastar, ehc, hc and beam without a heuristic raise a clear
  ValueError instead of a TypeError from deep inside the search;
  MultiObserver accepts positional arguments like the base class; the
  SearchObserver docstring no longer claims hooks are called defensively.
• The command line's exit codes say what happened: 0 solved, 1 no plan, 2
  bad input, 3 stopped by a budget, 4 a plan that failed validation. Before,
  unsolvable, out-of-budget and a crash all exited 1. Bad input — a missing
  file, a parse error, an unknown planner — is one line on stderr instead of a
  traceback (JUPYDDL_DEBUG=1 brings the traceback back). --version exists.
• jupyddl demo and jupyddl benchmark demos never finished: neither took
  a budget, and demo's blind runs on blocksworld12 do not end. Both now stop
  each search after 10 s by default; animate and demo accept
  --max-expansions and --time-limit like solve.
• benchmark checks --planners, --heuristic and --metric before running
  rather than reporting nosuch 0/13, and every chart option checks for the
  viz extra before searching instead of failing (or silently writing
  nothing) afterwards.
• A dead end crashed the charts (Axis limits cannot be NaN or Inf) and
  coloured the wavefront NaN; in the workbench, an infinite h or f made the
  progress message invalid JSON and aborted the run.
• A trace's plan showed compiler bookkeeping (__observe-1#1) and
  disjunct tags; it now records the plan the user reads.
• The sdist's own test suite failed: it dropped .github/ entirely while a
  test checks CONTRIBUTING.md is there. Only .github/workflows/ is left
  out now. build runs twine check --strict, like release does.
• black is pinned to one major version in both the dev extra and the
  format workflow, which installed whatever was newest; Renovate now pins
  third-party actions to commit digests.
• Learned-heuristic top-1 was neither held out nor tie-safe. A sibling
  scoring level with the successor the plan took counted as a correct ranking,
  so a model that outputs a constant scored 1.000 — and checkpoint selection,
  which reads top-1, could prefer a collapsed epoch. Corpus.split also gave
  the held-out side no ranking groups, so the "held out" top-1 jupyddl learn
  printed was measured over training instances. A tie with k siblings now
  earns 1/(k+1), a sibling identical to the chosen successor in features and
  step cost is not recorded at all, and selection and the reported top-1 use
  the held-out groups. The published figures came from the old rule; see the
  dated corrections in .docs/learned-heuristics.md and
  .docs/rl-for-search.md.
• A DAgger or bootstrap state could validate against its own instance.
  split treated X#dagger1 as an instance apart from X; it now groups on
  the name before the #.
• summarise_transfer averaged each heuristic over the runs it solved, so
  one that solved only the easy instance looked many times better than one that
  solved both. An unsolved run is now charged twice the budget, as
  search_cost charges it, and plan cost compares only instances every
  configuration with any coverage solved. The imitation-only blocksworld mean
  the docs quote as 366 is 6 329 under that rule.
• A malformed learned heuristic loaded, then gave one answer with NumPy and
  another without. A network whose input width is not the feature vector's,
  a ragged row, a sizes field that does not describe the weights, a
  non-finite value or a non-positive scale is now a ValueError on load.
  MLP.set_flat checks the length before writing anything, instead of leaving
  a wrong-length vector half-applied.
• Every negated fact shared one feature, not. Complement facts now key on
  the predicate they negate (not-sampled), so rovers and workshop get a slot
  per negated predicate. A model trained on a domain with negative
  preconditions before this sees different features now; retrain it.
• train() with a pure ranking objective and no ranking groups returned its
  random initialisation as a trained model; it now raises. learn_heuristic
  spent a generator passed as sizes on its first read, recording [] and
  raising under cem_iterations; size arguments are now read once, on entry.
• gripper and workshop ignored their seed. Twenty seeds gave one
  instance, so a split by seed held out copies of training instances, and the
  RL stage's tuning and validation sets were the same problems. The seed now
  decides which room each ball and the robot start in (every ball still has to
  change rooms), and how far along the pipeline each part starts. Every
  generated gripper and workshop instance changes; demos/workshop is the one
  in which every part starts raw, and generate no longer reproduces it.
• Generators could emit an instance solved before it starts. blocksworld
  compared tower lists in order and skipped the check for two blocks, where 75
  of the first 200 seeds came out trivial; random-strips could draw a goal
  that already held. Both now re-draw only in that case, so every instance that
  was not trivial is byte-identical to before.
• The benchmark harness merged configurations that differed in planner
  arguments — wastar at weights 1 and 5 summarised as one entry with twice
  the instances — and recorded ("gbfs", None) as heuristic "" although it
  ran hff. Rows carry a config label with the arguments
  (wastar/hff(weight=5.0), also a new CSV column), which summarize and the
  dashboard group on; a repeated config is suffixed #2; and heuristic is
  what solve_task actually ran, worked out by the new
  jupyddl.api.resolve_heuristic.

Wrong answers from the PDDL front end. The parser and the PDDL 3 compilations
handed the grounder a different problem from the one written, so plans
validated — against the wrong problem. Each case is now pinned by
tests/test_frontend_semantics.py.

• Durative actions merged their at-start and at-end effects, and an add
  beat a delete, so (at start (busy)) (at end (not (busy))) ended with
  (busy) true. The at-end effects now apply after the at-start ones and win a
  conflict, and over-all and at-end conditions are checked in the state after
  the at-start effects rather than before them. always is checked in that
  state too; the other trajectory constraints are refused when a durative
  action has at-start effects, because their monitors cannot see it.
• sometime-after could not be met by the final state. An obligation
  recorded by the last action's monitor could never be discharged, so the
  plan [t, r] was reported unsolvable; the goal now accepts the response
  holding at the end.
• forall around a trajectory constraint moved inside it:
  (forall (?x) (sometime (lit ?x))) meant one moment with every ?x lit at
  once. It is now one constraint per object. A soft constraint under forall
  crashed with an AttributeError; it is refused by name like the others.
• A goal preference under forall kept its free variable and could never
  be satisfied; it is now one preference per object. Under not, or, imply
  or exists a preference silently made the goal false; it is now an error.
• Timed initial literals slipped past the PDDL 3 machinery. The
  at-most-once and sometime-after monitors ignored the states literals
  fired in, a literal could fire after the preferences had been paid out, and
  a plan could end before a literal that was already due had fired.
• Object fluents in trajectory constraints were read as numeric
  comparisons; they are now compiled as everywhere else. Assigning one object
  fluent's value to another took the function's name for an object; it is
  refused.
• The metric. A preference the metric does not mention cost 1 to violate
  rather than nothing, a metric without (total-cost) still charged for
  actions, maximize was silently minimised, and (total-time) charged an
  instantaneous action 1. maximize, any metric other than (total-cost) or
  (total-time) plus weighted (is-violated p) terms, and a metric pricing a
  preference the goal does not declare are now refused.
• ground() edited the domain it was given, so a second problem grounded
  against the same parsed domain inherited the first one's synthetic actions
  and lost the domain's own constraints.
• Misspelt PDDL was accepted. :preconditon gave an action with no
  precondition and (:goals ...) a problem with an empty goal; unknown keys
  and sections and a missing :goal are now errors that name them. Malformed
  input — (?x -), (when (q)), (:goal) — raised a bare IndexError; it is
  a PDDLError, and an unbalanced parenthesis names its line.
• Objects called nan, inf or infinity were read as numbers.
• (increase (total-cost) (road-length ?a ?b)) was refused although
  :action-costs is listed as native. A cost over static functions is now
  evaluated when grounding, so the IPC transport domains load. :fluents
  now implies :object-fluents, and :ucpop, :domain-axioms and
  :safety-constraints are refused by name rather than as unknown flags.

v2.4.0

Compare Source

Added

• A Research view in the workbench. jupyddl.learn shipped in 2.3.0 with
  its results only in .docs/ and a video, which meant the published site said
  nothing about the biggest addition in the release. The view carries the
  measured run: imitation against hff and goalcount, the per-instance
  spread behind the mean, the logistics loss and why the feature space causes
  it, and the three claims that turned out to be wrong. Its numbers are built
  from promo/rl-data.json — the same cache the RL video renders from — so
  the page and the video cannot drift apart, and a test pins them together.

Fixed

• Reading the workbench no longer costs a 10 MB download. <main> was
  hidden until Pyodide reported ready, so every page of prose, the requirement
  matrix and the new Research view sat behind a full-screen spinner waiting for
  a runtime none of them use. The shell now renders immediately from the
  committed bundle, the boot notice is a status bar rather than a splash
  screen, and only the controls that actually run a planner stay disabled until
  the interpreter arrives.

Changed

• Python 3.10 is the new floor, and the matrix runs 3.10 through 3.14.
  3.9 reached end of life in October 2025; requires-python moved to match
  what is actually tested.
• The pddl-examples submodule advances to a corpus that now grows on its own:
  it imports fresh domains from planning.domains daily.
• All URLs follow the organisation's rename to openplan-labs.

Removed

• Mergify. Its auto-merge rule waited on an approving review that a
  single-maintainer project does not produce, and its reviewer-assign rule
  pinged an inactive account; every recent pull request was merged by hand.

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request