Use when adding a new effect, mechanic, or parser capability to the engine. Covers the full lifecycle from types → parser → resolver → targeting → frontend → AI → tests.
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Use when adding a new effect, mechanic, or parser capability to the engine. Covers the full lifecycle from types → parser → resolver → targeting → frontend → AI → tests.
Adding a New Effect to the Engine
This is the authoritative checklist for adding a new effect (mechanic, keyword action, etc.) to the phase.rs engine. Every step has a file path and a what to do. Missing any step causes silent failures — effects parse but don't resolve, resolve but don't target, target but don't animate, etc.
Before you start: Read the existing effect most similar to yours. Trace it through every file below. This is how you learn the patterns — not by guessing.
CR Verification Rule: Every CR number in annotations MUST be verified by grepping docs/MagicCompRules.txt before writing. Do NOT rely on memory — 701.x and 702.x numbers are arbitrary sequential assignments that LLMs consistently hallucinate. Run grep -n "^701.21" docs/MagicCompRules.txt (etc.) for every number. If you cannot find it, do not write the annotation.
Design Philosophy — Composability Over Completeness
The single most important principle: An effect should do ONE thing. Multi-step card abilities are composed from chains of single-purpose effects linked by sub_ability.
Compose, don't monolith
When a card says "Search your library for a creature card, reveal it, put it into your hand, then shuffle your library," that's four effects composed:
SearchLibrary { filter, count, reveal } — find the card
ChangeZone { destination: Hand } — put it into hand (injected as sub_ability)
Shuffle {} — shuffle library (parsed from ", then shuffle")
Each is a reusable building block. SearchLibrary works for tutors, ChangeZone works for bounce/exile/reanimate, Shuffle works independently. The composition unlocks hundreds of cards, not just one.
Anti-pattern: Creating Effect::SearchAndPutInHandAndShuffle — handles exactly one card template, can't be recomposed.
Building blocks unlock future cards
Before implementing a specific card's text, ask: "What is the general pattern here?" Recent examples from this codebase:
contains_possessive() / starts_with_possessive() — match "your"/"their"/"its owner's" variants in Oracle text. One helper, used by SearchLibrary, Shuffle, ChangeZone, and future effects.
contains_object_pronoun() — match "it"/"them"/"that card"/"those cards". Unlocks ~2000+ card patterns.
match_phrase_variants() — shared logic that eliminated duplication across all phrase-matching helpers.
When you build a helper like this, you're not adding complexity — you're reducing the cost of every future effect.
Sub-ability chains: how multi-step effects work
parse_effect_chain() splits Oracle text on sentence boundaries and links each parsed effect as a sub_ability on the previous one. At runtime, resolve_ability_chain() walks this chain, executing each step.
Target propagation: Parent targets flow to sub-abilities automatically. "Exile target creature. Its controller gains life equal to its power" → the sub-ability for GainLife receives the creature target from the parent ChangeZone, without duplicating target data.
Continuation pattern: For interactive effects (reveal hand → choose card → exile it), use pending_continuation. The resolver sets WaitingFor and stashes the remaining sub-ability chain. When the player responds, engine.rs picks up the continuation.
When to extend an effect vs. compose a new one
Extend: The new behavior is a variation of an existing effect (e.g., adding DamageAmount::TargetToughness alongside DamageAmount::TargetPower).
Compose: The new behavior is a sequence of existing effects (e.g., "exile then return" = ChangeZone to exile + ChangeZone to battlefield).
New effect: The behavior is genuinely novel and can't be expressed by existing variants (e.g., Explore, Proliferate).
The Checklist
Phase 1 — Type Definition
Goal: The effect exists as a typed variant in the data model.
crates/engine/src/types/ability.rs — Effect enum
Add the new variant with typed fields. Follow these rules:
Use enum fields for semantically distinct data (e.g., DamageAmount::Fixed(i32) vs DamageAmount::TargetPower). Never use boolean flags as substitutes for proper enum variants.
Mark optional/new fields #[serde(default)] so existing card-data.json stays deserializable.
If the effect targets something, include a target: TargetFilter field.
If the effect has a duration, include a duration: Option<Duration> field.
crates/engine/src/types/ability.rs — effect_variant_name()
Add a match arm returning the variant's string name (e.g., "RevealHand"). Used in GameEvent::EffectResolved and logging.
crates/engine/src/types/ability.rs — EffectKind enum + From<&Effect>
Add an EffectKind variant and the corresponding From<&Effect> match arm. EffectKind is used for effect categorization and coverage tracking.
Does the effect carry an amount that can be X? Use a QuantityExpr field — never a bare i32 you resolve yourself.
X is locked at announcement (CR 601.2b / CR 602.2b), not at resolution: CR 107.3c lets the value drift while the object is on the stack, so re-measuring X in a resolver reads a different, legal board and silently produces a wrong number. publish_announced_x() (game/ability_utils.rs) is the single publish authority; it writes ResolvedAbility::chosen_x, which every QuantityRef::Variable("X") already reads. The three carriers (GameObject::cost_x_paid = CR 107.3m cast-X, GameState::announced_source_x = CR 107.3a/d announce-X, chosen_x = the object's published channel, CR 107.3i) are not interchangeable — see /casting-stack-conditions § "The X Channels". If you are computing X inside an effect handler, you are in the wrong layer.
Phase 2 — Effect Handler (Resolver)
Goal: The engine can execute the effect at runtime.
Extract typed fields from *ability.effect via pattern match (effect is Box<Effect>).
Find targets from ability.targets (resolved TargetRef values).
Mutate state, push GameEvents to events.
Never access card data or parse text in resolvers — only process the typed ResolvedAbility.
crates/engine/src/game/effects/mod.rs — resolve_effect() match arm
Wire the new variant: Effect::YourEffect { .. } => your_module::resolve(state, ability, events).
Also add mod your_module; at the top.
Phase 3 — Targeting Integration
Goal: The engine correctly requests and validates targets for this effect.
Only needed if the effect has a target: TargetFilter field that isn't None/SelfRef/Controller.
crates/engine/src/game/triggers.rs — extract_target_filter_from_effect()
Add the variant to the match arm that extracts &TargetFilter from effects. If you skip this, triggered abilities with this effect will go on the stack without requesting targets — they'll silently fail at resolution.
crates/engine/src/game/targeting.rs — verify find_legal_targets()
Check that the TargetFilter values your effect uses are actually resolvable. Common gaps:
TargetFilter::Typed(TypedFilter { type_filters: vec![], controller: Some(Opponent), .. }) → wants to target an opponent as a player, but typed targeting only searches battlefield objects. You may need to add player resolution for typed filters.
Custom FilterProp values → ensure they're handled in the filter matching logic.
Zone-aware targeting: If the effect targets cards in non-battlefield zones (graveyard, exile, hand, library), the parser should produce FilterProp::InZone { zone } in the target filter properties. When InZone is present, find_legal_targets searches ONLY that zone exclusively. Per MTG rule 702.16a, hexproof/shroud only apply on the battlefield — non-battlefield targeting checks only protection.
Phase 4 — Parser
Goal: Oracle text for this effect produces the correct typed Effect variant.
crates/engine/src/parser/oracle_effect/ — the right parse_*_ast() helper
Most new effect parser work now belongs in a typed imperative-family helper, not directly in a flat verb list. Register the pattern in the smallest matching parser:
parse_numeric_imperative_ast()
parse_zone_counter_ast()
parse_cost_resource_ast()
parse_targeted_action_ast()
parse_search_and_creation_ast()
parse_hand_reveal_ast()
parse_choose_ast()
parse_put_ast()
parse_shuffle_ast()
parse_utility_imperative_ast()
crates/engine/src/parser/oracle_effect/ — subject-preserving helpers (if needed)
If the subject of the sentence carries game-relevant information (who does it, who it targets), preserve it before fallback subject stripping. The current pattern is try_parse_targeted_controller_gain_life() in lower_imperative_clause(). See the /oracle-parser skill § "Subject Stripping" for the full rationale.
crates/engine/src/parser/oracle_effect/ — parse_effect_chain() integration
If the effect commonly appears as part of a multi-sentence ability (e.g., "Look at target opponent's hand. You may choose a nonland card from it. Exile that card."), ensure the sentence splitting and sub_ability chaining produces the right structure. The engine composes effects via sub_ability chains — don't create mega-effects that do multiple things.
Absorption pattern: Some sentences modify a preceding effect rather than creating a new sub_ability. Examples: RevealHand absorbs card filters, SearchLibrary absorbs destinations, Effect::Mana absorbs "Spend this mana only to cast..." clauses as ManaSpendRestriction, and Effect::Counter can absorb follow-up source_static. If your new effect has modifier sentences, add absorption logic in parse_effect_chain() — don't emit them as separate sub_abilities.
crates/engine/src/parser/oracle_effect/ — parser tests
Every new pattern must have a #[test] in the inline test module:
#[test]fneffect_your_new_pattern() {
lete = parse_effect("Oracle text here");
assert!(matches!(e, Effect::YourEffect { field: expected, .. }));
}
Parser Helpers Reference
These helpers live in oracle_effect/, oracle_util.rs, and oracle_nom/. Know them — they eliminate entire classes of edge cases.
Nom combinators (oracle_nom/primitives.rs) — all parser branches delegate to these for atomic operations:
Combinator
What it does
When to use
parse_number
Digits + English + "a"/"an" with word-boundary guard
Number extraction (rejects "another")
parse_number_or_x
Same + "x" → 0
Costs, P/T, counters where X is variable
parse_color
"white"/"blue"/etc. → ManaColor
Color word extraction
parse_mana_cost
{2}{W}{U} → ManaCost
Full mana cost parsing
parse_pt_modifier
"+2/+3" → (i32, i32)
P/T modification
parse_counter_type
"+1/+1", "loyalty", etc.
Counter identification
Mixed-case bridging (oracle_nom/bridge.rs) — Oracle text is mixed-case but nom tag() requires exact matching. Use these when you need to run nom combinators on lowercase input and preserve the original-case remainder:
Function
Signature
When to use
nom_on_lower
(text, lower, parser) -> Option<(T, &str)>
Standard case — returns Some((value, original_case_rest))
Returns the matched TargetFilterand the remaining text after consuming the "target …" phrase. The leftover is discarded or fed to further parsing. Example:
The possessive vs. targeting fork — the single most important parser decision:
"Look at your/their hand" → contains_possessive matches
→ target = Controller or Any (no WaitingFor needed)
"Look at target opponent's hand" → parse_target matches "target opponent"
→ target = TargetFilter::Typed { controller: Opponent }
→ requires targeting phase (triggers.rs + stack.rs)
Getting this wrong produces silent failures: possessive forms that fall through to parse_target will silently produce no target (or Unimplemented); targeting forms that match contains_possessive will skip the targeting phase entirely.
match_phrase_variants() — shared backbone for all phrase helpers. If you need a new phrase helper (e.g., contains_sacrifice_clause), implement it via match_phrase_variants rather than duplicating the normalization logic.
Damage-to-players helper path — for damage phrases, exact player-set text is not a parse_target() job:
Use parse_damage_each_player_scope() in oracle_effect/mod.rs for exact each player, each opponent, or each foe damage text.
Reuse parse_damage_player_scope() when a compound damage parser needs the same noun parse, such as each opponent and each creature they control.
Do not push these phrases into parse_target(). DamageEachPlayer vs. DamageAll is an effect-layer semantic distinction, and moving it into generic target parsing causes player damage to be misclassified as object damage.
Phase 5 — Interactive Effects (if applicable)
Goal: Effects that require player choices (scry, dig, reveal+choose, search) work end-to-end.
Only needed if the effect pauses for player input.
crates/engine/src/types/game_state.rs — WaitingFor enum
Add a variant (e.g., RevealChoice { player, cards, filter }) that carries enough data for the frontend to render the choice UI.
crates/engine/src/game/engine.rs — apply() match arm
Handle the (WaitingFor::YourChoice, GameAction::YourResponse) pair. This is where the player's choice feeds back into the engine.
crates/engine/src/types/game_state.rs — GameAction enum
Add a variant for the player's response if no existing action fits.
client/src/adapter/types.ts — WaitingFor type
Add the TypeScript discriminated union variant. Note: tsify auto-generates types from Rust, but the types.ts file may have manual overrides — check both.
client/src/components/ — UI for the choice
Build (or extend) a component that renders when waitingFor.type === "YourChoice". Current patterns:
client/src/components/modal/CardChoiceModal.tsx for ScryChoice, DigChoice, SurveilChoice, RevealChoice, SearchChoice, DiscardToHandSize
client/src/components/modal/NamedChoiceModal.tsx for NamedChoice
client/src/components/modal/ModeChoiceModal.tsx for ModeChoice and AbilityModeChoice
Route any new frontend-authored text (titles, prompts, buttons) through t() (useTranslation("game"), keys in client/src/i18n/locales/en/game.json). Card/Oracle/enum text stays raw. See $add-frontend-component Phase 2.5 and client/src/i18n/README.md.
Phase 5b — Multiplayer State Filtering (if applicable)
Goal: Hidden information is correctly filtered in multiplayer games.
Only needed if the effect reveals or hides information (hands, face-down cards, library contents).
crates/server-core/src/filter.rs — filter_state_for_player()
Ensure the filter respects the new revealed/hidden state. For example, RevealHand adds card IDs to state.revealed_cards, and filter_state_for_player checks this set before hiding opponent hand cards. If your effect exposes hidden information, it must be visible through the filter.
crates/server-core/src/session.rs
If the new WaitingFor variant should be sent to specific players (not broadcast), handle the routing here.
crates/engine/src/types/events.rs — GameEvent enum
Add an event variant if the effect needs to be visible in the game log or trigger animations.
client/src/adapter/types.ts — GameEvent type
Add the TypeScript variant.
client/src/components/log/ — Game log rendering
Handle the new event in the log component so players see what happened. Translate the log template (t("log.yourEvent", { ... })) and interpolate engine data (object IDs, amounts, enum strings) raw.
Phase 7 — AI (if applicable)
Goal: The AI can make reasonable decisions when this effect is in play.
crates/phase-ai/src/eval.rs — effect evaluation
If the effect materially changes board evaluation (e.g., exile vs. destroy), ensure the AI's state evaluator accounts for it.
crates/engine/src/ai_support/candidates.rs — action generation
If the effect introduces a new WaitingFor / GameAction pair, the AI needs to generate legal responses. Legal action generation lives in engine::ai_support::legal_actions() → candidate_actions().
Phase 8 — Verification
Tilt-preferred / direct-cargo fallback (see CLAUDE.md § "Canonical verification pattern"):
Discriminating runtime test — At least one test drives the real pipeline via the /card-test recipe (GameScenario + GameRunner::cast(..).resolve() + CastOutcome deltas, verbatim Oracle text) and would FAIL if the change were reverted. Parser AST-shape tests (Phase 4) and snapshots do NOT satisfy this — they are shape tests, not regression tests. Negative assertions need a paired positive reach-guard (see /card-test foot-gun 6).
cargo fmt --all — Always direct (Tilt doesn't auto-format).
Clippy + tests — If tilt get uiresource clippy >/dev/null 2>&1 succeeds: ./scripts/tilt-wait.sh --timeout 240 clippy test-engine. Otherwise: cargo clippy --all-targets -- -D warnings followed by cargo test -p engine.
Snapshot test — If the effect changes a card's parsed abilities, update or add an insta snapshot in crates/engine/tests/oracle_parser.rs.
cargo coverage — One-shot binary (always direct). Verifies the new effect reduces Unimplemented count for the target cards.
Common Mistakes
Mistake
Consequence
Fix
Missing extract_target_filter_from_effect arm
Triggered abilities skip targeting — resolve with no targets, hit MissingParam error
Add the variant to the match in extract_target_filter_from_effect() in game/triggers.rs
Missing extract_target_filter_from_effect arm in stack.rs
Spells don't fizzle when targets become illegal
Ensure the effect's TargetFilter is returned by extract_target_filter_from_effect
Creating a mega-effect instead of composing
One-off solution that handles one card pattern
Decompose into building blocks with sub_ability chains
Boolean flags on Effect variants
Undefined combinations, unclear intent
Use enum variants (see DamageAmount, LifeAmount)
Added parser logic in the wrong AST helper
Pattern parses inconsistently or bypasses continuation absorption
Register it in the smallest matching parse_*_ast() helper
Imperative fallback only, no subject preservation
"Its controller gains life equal to its power" loses context
Preserve the subject before fallback subject stripping
Using parse_target for possessive forms ("your hand")
Silently produces no target or Unimplemented
Use contains_possessive → target: Controller instead
Using contains_possessive for targeting forms ("target opponent's hand")
Skips targeting phase, effect fires with no target
Use parse_target → TargetFilter::Typed { controller: Opponent }
Hardcoding amount: 1 as fallback
Silently wrong behavior, not visible in coverage
Return Unimplemented so the gap shows up
Adding TargetFilter::Typed without player resolution
"Target opponent" as a player never resolves
Check find_legal_targets handles your filter
Forgetting #[serde(default)] on new fields
Old card-data.json files fail to deserialize
Always default new optional fields
Effect reveals info but no filter.rs update
Opponent sees hidden cards in multiplayer, or revealed cards stay hidden
Update filter_state_for_player in server-core/src/filter.rs
Building a multi-step mega-effect
One-off variant that only works for one card template
Decompose into sub_ability chain of single-purpose effects
Missing AI legal action generation
AI hangs on WaitingFor with no valid response
Add the WaitingFor variant to engine/src/ai_support/candidates.rs
Shape-only or vacuous-negative tests
Effect parses but is never proven to resolve; negatives pass via Effect::Unimplemented early-returns
Add a /card-test runtime test with a revert-failing assertion; pair every negative with a positive reach-guard
New field on an existing variant dropped at one seam
Field parses but is a silent no-op in production (resume paths, batch handlers, adapter constructors)
Grep every construction/consumption site of the variant; thread the field or justify the default at each
Resolving X inside the effect handler
Rules-wrong (CR 107.3c): X is fixed at announce, and the resolving board differs — a silently wrong amount, not a crash
Carry a QuantityExpr; read the published chosen_x. See /casting-stack-conditions § "The X Channels"
