[guilyx]

What and why

Fixes for the parser and PDDL 3 review, the last of the full code review. The rule throughout: fix it when a correct fix exists, otherwise refuse the construct by name. Never return a silently wrong plan.

Area	Before	After
Durative actions	At-start and at-end effects were merged with add-wins. (at start (busy)) (at end (not (busy))) left busy true. At-end conditions were folded into the precondition	At-start effects apply first and at-end effects second, so at-end wins a conflict. Over-all and at-end conditions (and always) are checked in the state after the at-start effects. Forms the sequential compilation can't express are refused by name
sometime-after	An obligation set by the last action could never be discharged. A valid [t, r] was reported unsolvable	Goal is ψ ∨ (¬pending ∧ ¬φ)
forall around a constraint	Read as sometime (forall …), i.e. all objects at once. Wrapped around a preference, it crashed	One constraint, or one preference, per object
Quantified preferences	Kept a free variable, so they could never be satisfied (cost 11 vs the optimal 3). A preference under not silently made the goal false	Expanded per object. A preference under not/or/imply/exists is a PDDLError
Timed literals and PDDL 3	Firing a timed literal bypassed forced monitors. A preference could be credited and then undone	Monitors ride on the firing actions. Closing the plan blocks everything, timed literals included. A plan can't end with a due literal unfired
Object fluents	Constraints became bogus numeric comparisons. (assign (loc ?p) (loc ?q)) treated a fluent name as an object	Constraints are rewritten like goals. Assigns from another fluent are refused
Metric	A preference missing from the metric cost 1. maximize was treated as minimise. Without total-cost in the metric, actions still cost something	An unmentioned preference costs 0. Without total-cost, domain actions are free. (total-time) charges durations. maximize and other shapes are refused
ground()	Mutated the parsed Domain/Problem in place: problem 2 inherited problem 1's timed literals	Deep copy first
Invalid PDDL	:preconditon, (:goals …) and a missing :goal were silently accepted. Malformed input raised bare IndexError	PDDLError naming the key. The tokenizer reports line numbers. A 25k-mutation fuzz raised nothing but PDDLError
Objects named nan/inf	Parsed as numbers	Parsed as objects
IPC action costs	(increase (total-cost) (road-length ?a ?b)) was refused, although :action-costs is marked native	Supported over static functions. The 3 IPC transport domains now parse (212 of 213 example files; grid stays intentionally unsupported)
Requirement flags	:ucpop, :domain-axioms and :safety-constraints were reported as unknown	:fluents implies :object-fluents. The three flags are refused by name, so the README reads 20 of 24 flags supported

requirements.py, the README PDDL table (regenerated from the registry), .docs/PUBLISHING.md and AGENTS.md (new notes on durative semantics and the cost model) are updated. tools/make_promo.py now computes the flag count. The promo video has not been re-rendered.

How it was checked

• New tests/test_frontend_semantics.py: 56 cases. 55 fail on the previous code, each for its semantic reason; the 56th is a guard that the fix didn't loosen anything.
• vehicle is still correctly unsolvable, and grid still raises UnsupportedFeatureError.
• Full suite on top of current main (including fix(learn): honest ranking metrics, seed-sensitive generators, distinct benchmark configs #171): 625 passed, 1 skipped. flake8 and black are clean, and web/dist is rebuilt.

Behaviour changes to note

• Domains that used to be accepted can now be refused, if they relied on a misspelt key, a maximize metric, or a form the compilations can't honour. Each refusal names the construct.
• Plans for durative domains with at-start effects can change. They are now correct.
• Known limitation, unchanged: timed literals still put 2^k disjuncts into each domain action's precondition. This predates this PR.

Checklist

• flake8 jupyddl tests tools and pytest pass locally
• python tools/build_web.py re-run; web/dist committed
• requirements.py changed first; README table follows it
• CHANGELOG.md updated

The history has a merge -s ours of the already-merged #171 tip so the push fast-forwards. Squash-merge with the PR title.

---

Generated by Claude Code

[guilyx]

CodeFactor: "3 issues found"

The CodeFactor page lists issues client-side, so I couldn't read the exact three. I measured cyclomatic complexity of the changed files instead (flake8 --select C901 --max-complexity 12) against main. These functions grew:

Function	main	this PR
compile.compile_object_fluents	25	31
parser.parse_domain	15	19
grounding._collect_effect	16	19
compile.compile_constraints, parser._parse_constraints, _metric_terms, _parse_expression, parse_effect, grounding._ground_raw_operators	≤ 12	13–15

That is almost certainly CodeFactor's "Complex Method" findings. The growth comes from the new validation and refusal branches: unknown keys and sections, metric shapes, timed effects.

Decision: not refactoring in this PR. These are maintainability findings, not correctness ones, and the repository's own lint does not gate on complexity. Splitting the parser functions here would mix a structural refactor into a semantic change that 56 new tests pin. It is worth doing as its own refactor: PR with no behaviour change: pull the section-key validation out of parse_domain/parse_problem, and the constraint rewriting out of compile_object_fluents.

---

Generated by Claude Code