Changelog

Copy Markdown View Source

All notable changes to :wpl_ai.

The format follows Keep a Changelog. This project adheres to Semantic Versioning.

[Unreleased]

[1.13.0] — 2026-05-13

Added — vocabulary expansion (TS parity with @gymbile/wpl-ai@1.13.0)

Mirrors the corpus-driven additions in the TypeScript reference. Every entry below appeared as an unknown_exercise_ref in the wpl-eval v0.2.0 audit and is a genuine (non-typo) emission observed in at least one trial:

  • @upper_body: inverted_row, hangboard
  • @rehab_mobility (new module attribute):
    • Rotator-cuff / shoulder rehab: scapular_retraction, external_rotation, internal_rotation, prone_T, prone_Y, prone_W
    • Pelvic floor / postpartum / pregnancy: pelvic_tilt, diaphragmatic_breathing

exercises_by_category/0 returns the new :rehab_mobility key alongside the existing six.

Internal

  • All 521 tests pass (1 added for v1.13 vocabulary additions).

[1.12.0] — 2026-05-13

Version jump 1.8.0 → 1.12.0 to restore TS/Elixir parity

The Elixir port's version number has been brought into alignment with the TypeScript reference @gymbile/wpl-ai. Versions 1.9.0, 1.10.0, and 1.11.0 are deliberately skipped — there is no Elixir release at those tags. Going forward both implementations will release under matching version numbers so a wpl_ai 1.x.y Hex package and a @gymbile/wpl-ai@1.x.y npm package always implement the same behavioural surface.

1.12.0 is identical in content to 1.8.0 (released 30 minutes earlier). No functional changes; this release exists solely to realign the version stream.

[1.8.0] — 2026-05-13

Fixed — 8 silent-truncation / tolerance bugs (TS parity with @gymbile/wpl-ai@1.12.0)

Compound effect: Lane B served rate moved from 37/80 to 77/80 in the wpl-eval v0.2.0 corpus with no LLM re-calls. The 0/80 safety claim is unchanged.

  • #1 — rpe/rir range modifiers (rpe 7..8, rir 1..2): parse_exercise_modifiers and parse_intensity now accept range syntax in addition to scalar values. Produces rpe_min/rpe_max (resp. rir_min/rir_max) on %AST.Exercise{}; compiler emits target_rpe_min/target_rpe_max accordingly. Previously the range token leaked into downstream parsing and silently truncated subsequent WEEK blocks.

  • #2 — reps time-unit suffix before modifier keyword (3x30s rpe 6): parse_reps_spec now consumes a trailing s/m/seconds/minutes/hours suffix when the next token is a modifier keyword (rpe, rir, rest, tempo, weight, name, to_failure, bodyweight). Applied to both single-number and range (3x20..30s) reps branches via shared @reps_modifier_follow constant and maybe_consume_reps_unit_suffix/1 helper.

  • #3 — long-form duration units in simple activity (cycling 10 minutes): parse_exercise_or_simple_activity accepts keyword units (seconds, minutes, hours, days) in the simple-activity branch in addition to short bare-word forms (s, m, h).

  • #4 — simple activity trailing modifiers leak (cycling 20m rpe 6 rest 30s): New consume_simple_activity_modifiers/1 helper walks and discards any trailing rpe/rir/rest/tempo/weight/name/to_failure/bodyweight/heart_rate_zone/ bpm/pace tokens after a simple activity. Values are intentionally dropped (SimpleActivity has no carrier fields); the goal is to prevent leakage.

  • #5 — lexer: en/em-dash normalised, typographic punctuation silently skipped, N-M as range:

    • En-dash (U+2013) and em-dash (U+2014) → ASCII hyphen via pre-scan normalisation.
    • ;, &, ~, @, ASCII apostrophe ('), smart quotes (' ' " "), ellipsis (…), ≤, ≥, middle-dot, bullet → dropped silently.
    • N-M between two numbers emits a :range token (equivalent to ..) when the prior emitted token was a :number. parse_tempo updated to accept :range as a separator alongside :minus between tempo segments.
    • Elixir structural note: implemented as a normalize_typographic_chars/1 pre-scan pass (source string normalisation) rather than the TS in-place byte mutation approach — idiomatic for Elixir's immutable binary model.

  • #6 — trailing-dot number typos (12., 7.): consume_number_like stops before consuming a . that has no digit after it; the clean integer is emitted and tokenize_dots skips the dangling dot silently.

  • #7 — stray top-level ALL-CAPS sections (NUTRITION:, SUMMARY:, NOTES:): parse_sections detects an ALL-CAPS keyword followed by : and skips the entire indented body using skip_until_matching_dedent/2. No error is emitted; compile reports ok: true and the PHASES section parses normally.

  • #8 — two-tier exercise-ref resolution:

    • Unknown TYPE values (e.g. TYPE summary) silently fall back to :workout instead of String.to_atom-ing arbitrary values.
    • Cardio modalities (running, walking, cycling, rowing, elliptical, swimming, jump_rope, hiking) accepted as exercise refs in sets×reps form.
    • resolve_exercise_ref/1 replaces ad-hoc validation: tier 1 auto-corrects high-confidence typos (Jaro-Winkler ≥ 0.85, e.g. pushup → push_up); tier 2 accepts unknown refs as-is so compile succeeds with the model's literal name preserved in exercise_ref.

Added

  • %AST.Exercise{} gains fields: rpe_min, rpe_max, rir_min, rir_max.
  • WplAi.Parser (private): resolve_exercise_ref/1, maybe_consume_reps_unit_suffix/1, consume_simple_activity_modifiers/1, expect_minus_or_range/1, @reps_modifier_follow, @simple_activity_modifier_keywords, @cardio_modality_set.
  • WplAi.Lexer (private): normalize_typographic_chars/1, @dash_replacements, @silent_skip_replacements, last_emitted_token_is_number?/1.
  • 49 new regression tests in test/wpl_ai/v1_12_fixes_test.exs.

[1.6.7] — 2026-05-04

Fixed — parser bug: dash-prefixed typed MeasurementSpec (TS parity with @gymbile/wpl-ai@1.10.5)

  • #G — dash-prefixed typed MeasurementSpec parsing: parse_typed_measurement_list/2 now correctly handles - questionnaire_score questionnaire psqi note "text" as a single %AST.MeasurementSpec{} node (with metric, questionnaire, and note fields) rather than emitting a bare spec with only metric set and discarding the remaining qualifiers.

[1.6.6] — 2026-05-05

Fixed — 7 silent-failure parser/lexer bugs (TS parity with @gymbile/wpl-ai@1.10.4)

  • Bug 1 — digit-leading TAGS value (531): TAGS 531, strength now produces tags: ["531", "strength"] instead of tags: []; parse_tag_list accepts number tokens.
  • Bug 2 — digit-leading identifier in TAGS list (1rm_estimate): TAGS strength_test, 1rm_estimate, powerlifting no longer truncates after the digit-leading item; the number + bare_word token sequence is glued into a single tag string.
  • Bug 3 — colon-qualified contraindication name (acsm:cardiac_rehab_phase_2): the parser now accepts prefix:suffix form in the contraindication-name slot; glues the colon token into a single qualified identifier string.
  • Bug 4 — unknown REQUIRES directive silently terminates block: any unrecognised keyword or bare_word inside a REQUIRES block now produces a ParseError with type invalid_structure: "Unknown REQUIRES directive: '...'. Recognized: contraindication, fitness, equipment, age, time_commitment."
  • Bug 5 — trigger completion (no-arg) swallows subsequent sections: parse_trigger now emits a ParseError with type invalid_structure: "Unsupported checkpoint trigger 'completion' — use 'at N weeks' or 'at N days'."
  • Bug 6 — unknown phase type silently drops the phase: parse_phase detects non-recognized type keywords and emits a ParseError with type invalid_structure: "Unknown phase type '<x>'. Allowed: accumulation, intensification, realization, deload, base, build, peak, recovery, transition."
  • Bug 7 — jogging 10m in cooldown produces malformed recovery_exercise + phantom m orphan: the cooldown block parser now routes <bare_word> <number> <time_unit> EOL to an inline CardioActivity with modality, cardio_type: :continuous, and total_duration populated correctly.

Added

  • Invalid-parser conformance runner: conformance_test.exs now also iterates wpl/conformance/invalid/parser/ fixtures, verifying that WplAi.Parser.parse/1 returns {:error, ...} with errors matching the expected type and message fields.

[1.6.5] — 2026-05-05

Fixed

  • compiler: emit progress.points_system instead of progress.points — matches schema field name. Previous "points" key produced SCHEMA_VIOLATION when validated.
  • compiler: day-scoped activity ID counter — auto-IDs were previously per-block; same activity kind in two blocks of the same day collided (DUPLICATE_ID). Now monotonic across all blocks in a day.

[1.6.4] — 2026-05-04

Added

Fixed

  • validator: refresh MeasurementMetric + Questionnaire vocabulary to schema 1.6.0: WplAi.Validator knows the canonical 24-value MeasurementMetric enum and the 8-value Questionnaire enum; legacy string items are checked against the combined (legacy + enum) set; typed MeasurementSpec items have metric checked against the enum set, and questionnaire (when metric == "questionnaire_score") checked against the questionnaire set.

[1.6.3] — 2026-05-04

Changed

  • test: per-module unit tests (lexer, parser, vocabularies, exercise_matcher) — 93 new tests.

[1.6.2] — 2026-05-05

Fixed

  • Bug 1 — Structured tempo emit — tempo 3 - 1 - 1 - 0 (and 3-0-X-1 forms) now emits the structured object {eccentric, pause_bottom, concentric, pause_top} instead of a raw string. X in the concentric position sets explosive_concentric: true (TS parity).
  • Bug 2 — weight N% bw unit field — percentage_bodyweight weight spec now emits unit: "bw" instead of unit: "%" (TS parity).
  • Bug 3 — Calories kcal unit omitted — when calorie unit is "kcal" (the schema default), the unit field is now omitted from the compiled output. Only kcal_per_kg and multiplier_of_tdee are emitted (TS parity).
  • Bug 4 — Nutrition timing at_time maps to type: "absolute" — timing at 07:30 now emits {type: "absolute", time: "07:30"} (with no seconds, no offset field). Previously emitted {type: "at_time", time: "07:30:00"} (TS parity).
  • Bug 5 — Nutrition timing before_workout/after_workout drops body content — before_workout and after_workout were not in the keywords list, causing them to be tokenized as bare words and break the timing parser. Adding them to the lexer keyword list restores full body parsing after a timing directive (TS parity).
  • Bug 6 — Habit prescription nesting — habit activities now nest target, frequency, and reminder_times under a prescription key instead of emitting them flat on the activity. Also fixed a parser bug where parsing frequency in a habit body called parse_plan_habit_body instead of parse_habit_body, silently dropping reminders (TS parity).
  • Bug 7 — bodyweight keyword as exercise modifier — bare bodyweight after reps/sets (e.g., pull_up 3x8 bodyweight) now attaches as weight: {type: "bodyweight"} on the exercise prescription instead of generating a phantom SimpleActivity (TS parity).

[1.6.1] — 2026-05-04

Fixed

  • metadata.language default — compiler now always emits metadata.language: "en" when the DSL does not specify a language, matching TS compiler behaviour (TS parity).
  • Auto-derived activity display name — compiler now derives name from the exercise_ref / modality / category token for Exercise, Cardio, Nutrition, Meditation, Recovery, and Habit activities, matching the TS humanise() helper exactly. Acronyms (HIIT, AMRAP, EMOM, RPE, RIR, 1RM) are uppercased; all other words are title-cased. Explicit name set in the DSL is preserved as-is (TS parity).

[1.6.0] — 2026-05-05

Added

  • Contraindication.severity + require_clearance (schema v1.6.0) — extends the contraindication DSL to contraindication <name> [severity <low|moderate|high>] [action <action>] where action now includes require_clearance. The old arrow form (contraindication <name> -> <action>) is preserved for back-compat. Compiler emits severity only when present.
  • Reps.amrap (schema v1.6.0) — DSL: <exercise> NxAMRAP (compact) or Nx amrap (space-separated, case-insensitive). Compiler emits prescription.reps: { amrap: true }. Sets count is preserved from N.
  • ExercisePrescription.to_failure (schema v1.6.0) — optional modifier to_failure in the exercise modifier chain. Compiler emits prescription.to_failure: true when present; field omitted otherwise.
  • Weight.metric qualifier (schema v1.6.0) — optional metric <1rm|e1rm|training_max|daily_max> after a weight N% rm spec. Compiler emits weight.metric: "<canonical>" (e.g. "1RM", "e1RM", "training_max", "daily_max"). Omitted when not specified (back-compat).
  • RecoveryExercise extensions (schema v1.6.0) — optional modifiers on recovery exercise lines: modality <enum> (7 values: static_stretch | dynamic_stretch | pnf | smr_foam_roll | smr_ball | breathwork | mobility_drill), intensity <1-10> → emits intensity_rpe, body <token> → emits body_part. Optional indented pnf <Ns> contract <Ns> relax <int> contractions continuation line emits { contraction_seconds, relax_seconds, contractions }. Recovery exercises now compiled under prescription.exercises to match TS schema.
  • Checkpoint typed MeasurementSpec (schema v1.6.0) — measure: lists now accept bare metric tokens (emitting { metric: "<value>" }) and <metric> questionnaire <enum> [note "text"] (emitting full typed spec). Quoted strings preserved as plain strings (back-compat). Added support for TS-style inline CHECKPOINT "Name": blocks with at N weeks trigger form.
  • Cardio intensity.target.min_bpm/max_bpm (schema v1.6.0) — intensity bpm N..M now compiles to prescription.intensity: { type: "bpm", target: { min_bpm: N, max_bpm: M } }.
  • Emitted version bumped to "1.6.0" — compiler now emits "version": "1.6.0" in all compiled plans.

[1.5.0] — 2026-05-04

Added

  • Phase.type enum (schema v1.5.0) — DSL: PHASE "Name" <type> (N weeks): where <type> is one of accumulation | intensification | realization | deload | base | build | peak | recovery | transition. Compiler emits phase.type: "<value>". When omitted, no type key is emitted on the phase object.
  • Week.is_deload (schema v1.5.0) — DSL: WEEK N deload (optional token immediately after the week number, before the optional name string). Compiler emits week.is_deload: true. When absent, the field is omitted (not emitted as false).
  • SubPlanActivity (schema v1.5.0) — new activity variant subplan <plan-id> (optionally followed by a quoted name string) inside any block. Compiler emits { type: "sub_plan", id: "sub_plan_N", sub_plan_ref: "<plan-id>", name?: "<optional>" }.

[1.4.0] — 2026-05-04

Added

  • Per-kg macros + per-kg cals + TDEE multiplier (schema v1.4.0) — DSL accepts unit suffixes g_per_kg on protein, carbs, fat lines and kcal_per_kg / multiplier_of_tdee on calories lines. Compiler emits the unit verbatim into MacroRange.unit / Calories.unit. Default units remain "g" and "kcal" when no suffix is given.
  • Weight.percentage_bodyweight (schema v1.4.0) — DSL: weight N% bw or weight N% bodyweight. Compiler emits Weight with type: "percentage_bodyweight", value: N, unit: "%". Existing weight N kg (absolute) and weight N% rm (percentage_1rm) forms are unchanged.

[1.3.0] — 2026-05-04

Fixed

  • Compiler now emits the canonical $schema URL https://wpl.dev/schemas/wpl/v1.schema.json (previously emitted https://gymbile.com/schemas/wpl/v1).

Added

  • MuscleGroup + MovementPattern enums — DSL: <exercise> NxR muscles primary <m1>, <m2> secondary <m3> pattern <p>. Compiler emits primary_muscles, secondary_muscles, and movement_pattern on ExerciseActivity. Supports all 22 MuscleGroup values and all 13 MovementPattern values from schema v1.3.0.
  • Cardio zone_model — DSL: zone N model <zone_model> qualifier inside a cardio block. Compiler emits intensity.zone_model (7 values: hr_3_zone_seiler, hr_5_zone, hr_7_zone, power_coggan_7_zone, pace_critical_speed, rpe_borg_10, rpe_borg_20). New intensity type intensity power N emits intensity.type: "power".
  • Plan-level ATHLETE_THRESHOLDS block — top-level DSL section (parallel to PHASES). Accepts hr_max N bpm, lthr N bpm, resting_hr N bpm, ftp N watts, vo2max N, critical_pace N, body_weight N kg, one_rm <exercise> N kg. Compiler emits plan.athlete_thresholds with field-name suffixes matching schema v1.3.0 (hr_max_bpm, lthr_bpm, resting_hr_bpm, ftp_watts, vo2max_ml_kg_min, critical_pace_seconds_per_km, body_weight_kg, one_rm: [{ exercise_ref, value, unit }]).

[1.0.0] — 2026-05-04

Added

  • Initial extract from gymbile_backend. Compiler emits WPL schema 1.0.0.
  • WplAi.parse/1 — WPL-AI DSL text → WplAi.AST.Document struct.
  • WplAi.compile/1 — AST → WPL JSON map (string keys, "version": "1.0.0").
  • WplAi.to_wpl/1 — parse + compile in one step.
  • WplAi.decompile/1 — WPL JSON → WPL-AI text (round-trip).
  • WplAi.tokenize/1 — exposes the lexer for tooling / syntax highlighting.
  • WplAi.validate/1 — fast validity check without full compilation.
  • WplAi.ExerciseMatcher — Jaro-Winkler fuzzy matching for exercise references.
  • WplAi.Errors — structured error types (LexerError, ParseError, CompileError) with LLM-optimised formatting helpers.
  • Significant-indentation lexer (Python-style INDENT/DEDENT).
  • Recursive-descent parser covering: header, goals, requirements, personalization, phases/weeks/days/blocks, exercise / cardio / nutrition / meditation / recovery / habit / simple activities, progress checkpoints, top-level HABITS section.

Notes

Phase 2 will update the emitted schema version and bring the compiler to parity with @gymbile/wpl-ai v1.6.0. Every plan valid under schema 1.0.0 will continue to compile correctly.