Files
deck_in_a_dash/CLAUDE.md
T
ml 83248561cc
build / build (push) Successful in 33s
Adding gitea action
2026-07-11 20:23:39 +02:00

27 KiB
Raw Blame History

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project goal

A Pebble smartwatch app that produces a daily personalized tarot (Celtic Cross) and astrology (natal chart + transits) reading. This repository holds the core engine: a plain-C, dependency-free library plus a standalone CLI binary (dist/deck-engine) that can be built and run without any Pebble SDK or emulator, deliberately built so it links straight into Pebble C/UI watch code without rework (see "Portability to the watch" below).

Build & run

make          # builds dist/{deck-engine,interpreter-cli}, dist/img/, dist/i18n/,
              # dist/run-engine.sh, dist/run-interpreter.sh, dist/user.properties
make test     # builds and runs engine/tests/smoke_test.c and interpreter/tests/*_test.c
make clean    # removes build/ and the generated binaries/img/i18n (never touches dist/user.properties)
make package  # builds a versioned Linux CLI tarball at dist/ - see "Packaging" below

A single Makefile at the repo root builds both engine/ and interpreter/ — there is no per-module Makefile, and no cd needed before running make. Build output goes to dist/, a sibling of engine/, interpreter/, and res/. There is no make install; dist/ is meant to be run in place.

Two ways to run a reading:

# 1. Direct CLI, all inputs explicit:
dist/deck-engine --seed "2026-07-03" \
  --birth-date 1990-05-14 --birth-time 14:32 --birth-utc-offset 2 \
  --birth-lat 52.5200 --birth-lon 13.4050 \
  [--date YYYY-MM-DDTHH:MM] [--format text|html|json] [--lang en|de] [--i18n-dir <path>]

# 2. dist/run-engine.sh [text|html|json] [lang] — wraps the binary for
#    daily use: seed and --date come from the OS clock (today's date /
#    current UTC time), birth data + default language are read from
#    dist/user.properties next to the script; the optional [lang] arg
#    overrides that language for a single run.

dist/user.properties is seeded once from engine/scripts/user.properties.template with placeholder values (YOUR_BIRTH_DATE_HERE, etc.) — run-engine.sh/ run-interpreter.sh refuse to run (with a clear error) until the placeholders are replaced with real values. Because dist/ itself is disposable (make clean, or just deleting the directory, wipes it), the scripts Makefile target also keeps a durable backup outside dist/: once dist/user.properties looks filled in (no leftover YOUR_...), every build copies it out to engine/scripts/user.properties.local (gitignored); if dist/user.properties is ever missing when a build runs, it's restored from that backup instead of being reseeded from the placeholder template. dist/run-interpreter.sh is the equivalent daily-use wrapper for the interpreter: it runs run-engine.sh's same OS-clock seed/date and user.properties birth data through deck-engine --format json, piped straight into dist/interpreter-cli — see "Interpretation" below.

Run a single smoke test by editing engine/tests/smoke_test.c's main() temporarily, or just read its assertions — there's no test filter flag, the whole suite runs in milliseconds.

Packaging

make package (a Makefile target, not a separate script) builds a fresh all and bundles it into a versioned, self-contained Linux CLI tarball at dist/deck-in-a-dash-<version>-linux-<arch>.tar.gz - deck-engine, interpreter-cli, run-engine.sh/run-interpreter.sh, img/, i18n/, LICENSE, README.md, docs/reading.md (renamed READING.md), and a placeholder user.properties (from engine/scripts/user.properties.template, never dist/user.properties itself - see below). Extracting the tarball anywhere and running the scripts/binaries from inside it works exactly like running them from dist/ in place, since run-engine.sh/ run-interpreter.sh already resolve every path relative to their own location (SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"), not the caller's working directory.

Version: defaults to the current commit's git tag, matching vX.Y.Z (the v is stripped); if HEAD isn't exactly on such a tag, it falls back to a 0.0.0-dev+<short-sha> placeholder with a warning on stderr. Override either with make package VERSION=1.2.3. Push a vX.Y.Z tag to drive a real release version.

Never packages dist/user.properties: the staging step copies files into the package by explicit name only - never a wildcard or recursive copy of dist/ itself - specifically so a filled-in dist/user.properties (real birth data, gitignored, private) can never end up in a distributable package. The package always gets the placeholder template instead, identical to what a first-time make seeds dist/user.properties with.

CI (.gitea/workflows/build.yml): runs make test && make package on every push/PR (plus manual dispatch). The resulting tarball is both kept as a Gitea Actions run artifact and, on push only (not PRs), published over SFTP to dl.ladkau.de under files/deck-in-a-dash/ - requires a DL_SFTP_KEY secret configured on this repo (or inherited from the org/instance level).

Architecture

deck_in_a_dash/
  res/                          Tarot card art (22 RWS Major Arcana JPEGs),
                                 Waite's "Pictorial Key to the Tarot" (PDF,
                                 public domain — source of all card/spread
                                 text), Sepharial's "Transits and Planetary
                                 Periods" (PDF, public domain — source of
                                 interpreter/src/narrative.c's text), logo,
                                 research notes.
  engine/
    third_party/astronomy/      Vendored cosinekitty/astronomy C library
                                 (MIT, pinned commit — see VENDORED.md).
                                 Unmodified from upstream.
    src/
      rng.h/.c                  Deterministic string-seeded PRNG.
      tarot.h/.c                Celtic Cross spread draw logic.
      tarot_data.c               22 cards' names/meanings + 10 position
                                 names/descriptions (Waite's own text).
      astro.h/.c                 Natal chart + daily transits.
      i18n.h/.c                   key=value translation catalog + English fallback.
      reading.h/.c               Ties tarot + astro together; text/HTML output.
      main.c                     CLI entry point.
    i18n/en.lang, i18n/de.lang   Translation source files, one per language.
    tests/smoke_test.c
    scripts/run-engine.sh, scripts/run-interpreter.sh, scripts/user.properties.template
  interpreter/                  Separate module + binary; see "Interpretation" below.
  dist/                         Build output (gitignored-style; see Build & run) and,
                                 after `make package`, versioned release tarballs.
  Makefile                      Single root Makefile, builds both engine/ and
                                 interpreter/, and packages a release (see Packaging).

Data flow / the real API

reading_generate(seed, birth, utc_moment, &DailyReading) in reading.h is the single entry point that matters — the real API surface. It populates a plain struct (NatalChart + DailyTransits + CelticCrossSpread) that a caller walks directly to build a UI. reading_print_text/reading_print_html are desktop-only text/HTML renderers of that same struct, used by the CLI.

Determinism (rng.c)

The Celtic Cross draw must be exactly reproducible from a seed string, on any platform (including the watch's ARM core later), so rng.c is pure integer arithmetic: FNV-1a hashes the seed string, splitmix64 streams from it, and a partial FisherYates shuffle (tarot.c) draws the 10 cards plus a reversal bit per card from that same stream, in one pass.

Known-sharp-edge: rng_next_bounded's rejection-sampling limit must stay a uint64_t. When bound evenly divides 2^32 (true for bound == 2, used by every reversed-card coin flip), the true rejection threshold is 2^32, which silently truncates to 0 in a uint32_t and turns the loop into an infinite one. This exact bug shipped once already — if you touch this function, keep the intermediate type 64-bit.

Astrology (astro.c)

  • Bodies: Sun, Moon, MercuryPluto (Body enum in astro.h, prefixed PLANET_* — deliberately not BODY_*, because Astronomy Engine's own astro_body_t already defines BODY_SUN, BODY_MOON, etc., and both headers are included together in astro.c).
  • Positions must go through Astronomy_GeoVector + Astronomy_Ecliptic, not Astronomy_EclipticLongitude. The latter computes heliocentric longitude and outright rejects BODY_SUN (ASTRO_INVALID_BODY) — wrong for astrology, which needs apparent geocentric position. This also shipped once and was caught by the natal-Sun-sign smoke test.
  • Houses use the Whole Sign system (house 1 = the Ascendant's sign, houses follow zodiacally) — no Placidus/Koch cusp trig. Astronomy Engine has no Ascendant function; it's computed directly in astro.c from sidereal time + a standard low-precision obliquity polynomial + the RAMC/obliquity/latitude identity (Duffett-Smith & Zwart).
  • Every PlanetPosition (astro.h) carries a whole-sign house (1-12), via assign_houses()/whole_sign_house(). DailyTransits.bodies[*].house is always relative to the natal Ascendant, not a fresh "houses for right now" chart — that's the standard, personalized way transits are read (e.g. "transiting Jupiter is in your 5th house"), and it's why astro_compute_daily_transits takes the natal chart as a parameter in the first place (also used for aspect detection).
  • Aspects (conjunction/sextile/square/trine/opposition) use an 8° orb when a luminary (Sun/Moon) is involved, 6° otherwise.

Tarot (tarot.c, tarot_data.c)

Only the 22 Major Arcana are modelled (TAROT_DECK_SIZE) — matches the art actually in res/img/. The 10 Celtic Cross positions (CelticCrossPosition in tarot.h) follow Waite's own original drawing order from The Pictorial Key to the Tarot, Part III §7 ("An Ancient Celtic Method of Divination"): Present → Challenge → Crown → Foundation → Recent Past → Near Future → Attitude → Environment → Hopes/Fears → Outcome. Note this differs from the reordering used by most modern tutorials (which typically put Foundation/Recent Past before Crown) — res/2026_07_03_celtic_cross_spread.txt has an example of that more common (but less original) ordering; don't use it as a reference for the position enum order. Card and position text in tarot_data.c is condensed from Waite's own wording in Part III §3 (public domain).

Translations (i18n.c, engine/i18n/*.lang)

All display text (planet/sign/moon-phase/aspect names, tarot card names and meanings, Celtic Cross position names, and reading.c's section headers) is looked up in a small process-global catalog loaded by i18n_load(path) from a key=value file, via i18n_get(key, fallback). Every call site passes the hardcoded English string as fallback, so a missing file, a missing key, or an untranslated language degrades to English rather than showing a raw key or nothing — this is what lets engine/i18n/de.lang (or any future language) be filled in incrementally.

Keys are namespaced by the slug tables in tarot_data.c/astro.c (card.<slug>.name, card.<slug>.upright/reversed, position.<slug>.name/desc, body.<slug>, sign.<slug>, moonphase.<slug>, aspect.<slug>), plus ui.* keys for reading.c's own labels — deliberately keyed off separate slug strings rather than the C enum names, so renaming an enumerator can never silently break a .lang file. To add a language: copy engine/i18n/en.lang to <code>.lang in the same directory (keys must match exactly) and translate the values — no source changes needed, make picks up any *.lang file there automatically.

i18n_load uses stdio (fopen/fgets), so — like main.c and reading_print_* — it's a desktop-only entry point; i18n_get itself is a pure fixed-size-array lookup with no heap allocation, so it links into the watch as-is given a non-file-based way to populate the catalog, e.g. from a compiled-in resource.

Portability to the watch

rng, tarot, tarot_data, astro, i18n_get, and reading_generate are kept free of stdio/CLI assumptions specifically so they link into the Pebble watchapp unchanged. Only reading_print_text/_html (text formatting), main.c (CLI arg parsing), and i18n_load (reads a file from disk) are desktop-only and don't port as-is.

Interpretation (interpreter/)

A second, separate top-level module and its own binary (dist/interpreter-cli, sibling to dist/deck-engine) that scores how significant a day's transits are, prints a short narrative sentence for each of the top items about what that transit classically means, reads out the full Celtic Cross spread with every card's meaning, and finally prints guidance tying the day's top transit to the spread's Attitude/Outcome cards; see docs/reading.md. Like deck-engine, it supports --format text|html|json and --lang <code>/--i18n-dir <path> for its own output - see main.c's bullets below for both. deck-engine itself is deliberately unchanged beyond gaining --format json; the two binaries compose over that JSON, never by linking reading_generate() or any I/O function together:

make          # -> dist/deck-engine (unchanged; --format json is new) and dist/interpreter-cli
make test     # builds and runs both engine/tests/*_test.c and interpreter/tests/*_test.c

dist/deck-engine ... --format json | dist/interpreter-cli [--format text|html|json] [--lang <code>]
# or: dist/deck-engine ... --format json > reading.json && dist/interpreter-cli reading.json
# or, for daily use (OS clock + user.properties, same as run-engine.sh):
dist/run-interpreter.sh [text|html|json] [lang]

Note --lang means something different on each side of the pipe: deck-engine's own --lang (if passed at all) is irrelevant here, since --format json is deliberately language-independent (see docs/input-output-format.md); dist/interpreter-cli --lang controls its own text/html/json output language, independently.

The root Makefile builds both from one invocation, and keeps them independent compilation units for everything ephemeris/tarot-draw/ random-related — the one exception is engine/src/tarot_data.c and its own i18n.c dependency (pure, deterministic card/position lookup text, see INTERP_TAROT_TEXT_OBJS in the Makefile), whose object files are built once and linked into both binaries.

  • interpret_daily_reading(reading, &out) (significance.h/.c) scores every aspect in reading->transits.aspects[], keeps the top 5 by score in DailyInterpretation.top_items[] (descending, evicting the rest), and classifies the day into a DaySignificance enum (DAY_QUIET ... DAY_MAJOR) from the single highest-scoring item. The enum also carries a trailing DAY_SIGNIFICANCE_COUNT sentinel (not a real level) purely so callers can print the level as a rank, e.g. main.c's "%s (%d/%d)""Notable (2/4)"interp.day_level + 1 out of DAY_SIGNIFICANCE_COUNT. interpretation_deserves_framing() is true only at DAY_MAJORmain.c uses it solely to decide whether to print an extra "worth a deeper Celtic Cross look" line; it's not a gate on guidance.c (below), which runs on every day level.
  • Score = transiting-planet weight (slow/outer planets score higher) × orb tightness (1.0 = exact, ~0 = at the edge) × natal-luminary bonus (transits to natal Sun/Moon score higher than to other planets). The weights and the DAY_QUIET/DAY_NOTABLE/DAY_SIGNIFICANT/DAY_MAJOR thresholds in significance.c are hand-tuned, not derived from anything physical.
  • narrative.c/narrative.h supplies main.c's one-sentence description printed under each top item, condensed from Sepharial's Transits and Planetary Periods (1920, public domain — res/Transits_and_Planetary_Periods.pdf), Chapter VIII "Effects of Transits". Per transiting planet it holds a base description plus a harmonious/discordant addendum, shown only under a trine/sextile or square/opposition respectively (either may be "" if the source gives none); a conjunction is treated as neutral — base only — since Sepharial's own text says a conjunction "takes the nature of what it conjoins," which isn't modelled here. The Moon and Pluto aren't in that chapter (fast lunar transits were out of scope for a year-ahead forecasting book; Pluto wasn't discovered until 1930) — their text is original, written for this project, not drawn from Sepharial. Each narrative is also framed by where it's happening: narrative_print() takes the transiting planet's whole-sign house (1-12, relative to the natal chart, same as everywhere else in this codebase) and prefixes an "In matters of ..." line naming that house's traditional area of life, from a small k_house_area[12] table of standard whole-sign house significations — centuries-old convention, not attributed to any single source (same status as the sign/aspect names already in astro.c). Passing a house outside 1-12 (the same "no data" sentinel reading_io.c uses elsewhere) omits that framing and prints the planet's narrative on its own. Every string here (k_narratives[]'s base/harmonious/ discordant, k_house_area[], and the "In matters of ..." framing template itself) is looked up via i18n_get() under narrative.* keys before falling back to this English text - engine/i18n/de.lang has a full German translation under the same keys.
  • main.c calls i18n_load() once at startup (same pattern as engine/src/main.c, including the same default_i18n_path() helper, duplicated rather than shared since the two binaries are never linked
    • see the compilation-units bullet above) before producing any output, so every i18n_get() call anywhere in the interpreter - not just tarot_data.c's, but narrative.c's/guidance.c's own (below) - respects --lang. --format text's report is printed in a fixed order: significance level → top significant transits (each with its narrative.c sentence) → the full Celtic Cross spread (print_celtic_cross_text(), every position in Waite's own drawing order, each with its card, orientation, position description, and card meaning — via the real tarot_position_name()/tarot_card_name()/ tarot_position_description()/tarot_card_meaning() accessors, not a duplicated table, since tarot_data.c is linked into this binary) → guidance.c's paragraph, deliberately printed last, so it reads as the closing takeaway after the reader has seen the full spread it references. --format html mirrors deck-engine's own HTML styling and reuses the same img/<file> card-art convention (tarot_card_ image_file()); --format json embeds the same localized narrative/ guidance prose as --format text (in whatever --lang was loaded) alongside stable, language-independent slugs (transiting_planet, aspect, card, position, etc., reusing the same small local slug tables reading_io.c needs for parsing) — a deliberate departure from reading_print_json's "slugs only, never prose" policy (next bullet), justified because rendering that prose is this module's job, unlike deck-engine's JSON which is purely a machine interchange format. narrative_print()/guidance_print() only know how to write to a FILE *, so --format json captures their output into a heap string via POSIX open_memstream() (capture_narrative()/capture_guidance()) rather than changing either module's public API just for this one caller.
  • guidance.c/guidance.h is the combined-storytelling piece: it ties interp->top_items[0] (the day's single most significant transit) to the Celtic Cross spread and prints a short "how to meet the day" paragraph on every day — main.c calls it unconditionally after every interpret_daily_reading(), with no day_level gate. The core guidance keys off two things: the top transit's aspect character (harmonious/discordant/neutral, same classification narrative.c uses) crossed with whether the spread's Attitude position — Waite's "Himself: his position or attitude in the circumstances", the position most directly about how the reader is meeting the day — fell upright or reversed (3 × 2 = 6 combinations); this text stays valid regardless of the day's intensity. What does vary with interp->day_level is only the sentence introducing the top transit (k_intro[DAY_SIGNIFICANCE_COUNT], one %s each) — e.g. "With Saturn as today's dominant influence" on DAY_MAJOR vs. "Saturn is only faintly active today, but for what it's worth" on DAY_QUIET. A day with no aspects in orb at all (top_item_count == 0, always DAY_QUIET) has no transiting planet to introduce, so it falls back to k_no_transit_stance, keyed on the Attitude card's orientation alone. All of this guidance text is original, written for this project, since neither Waite nor Sepharial discuss combining astrology and tarot. The paragraph also names the Attitude and Outcome cards and states their actual meaning via tarot_card_meaning() (e.g. what Wheel of Fortune reversed means) — the real Waite text, not a duplicated table, for the same reason as print_celtic_cross_text() above. Every string here (stance/intro/no-transit/label text, plus a guidance.planet.* table used only for the intro sentence's planet name - separate from body.* because "the Sun"/"the Moon" take a definite article the other eight planets don't, in both English and German) is looked up via i18n_get() under guidance.* keys before falling back to this English text - engine/i18n/de.lang has a full German translation. The German guidance.intro.* templates are deliberately phrased so the planet name placeholder is always the sentence's grammatical subject (nominative case) across all four day levels, sidestepping the case-agreement problems a naive word-for-word translation of the English templates would hit (e.g. "with Saturn" needs dative in German, but "Saturn is only faintly active" needs nominative - the four German templates are worded so the placeholder never needs to change case).
  • reading_print_json (engine/src/reading.c) is the wire format: the full DailyReading (natal, transits, spread), using the language-independent astro_*_slug()/tarot_*_slug() accessors (astro.h/tarot.h) rather than astro_*_name()/tarot_*_name() — the JSON must stay identical regardless of --lang, since it's a machine interchange format, not display text.
  • json.c/json.h is a small hand-rolled recursive-descent JSON parser (object/array/string/number/bool/null) - not a general-purpose validating library, just enough to parse deck-engine's own output. Unlike the core engine, it uses malloc; interpreter-cli was never meant to run on the watch itself.
  • reading_io.c walks the parsed JSON down to "transits"."aspects" and fills a DailyReading with those (used for scoring), plus "natal"."bodies"/"transits"."bodies" (each body's sign + whole-sign house, via parse_bodies() — used only for the sign/house context main.c prints alongside each significant event, never for scoring) and "spread"."positions" (each position's card + reversed flag, via parse_spread() — used by main.c's full Celtic Cross readout and by guidance.c's Attitude/Outcome framing, never for scoring). An aspect naming a planet/aspect-type slug it doesn't recognize, or a document missing "transits"."aspects" entirely (as opposed to a present-but-empty array, which is a valid "no aspects today"), makes the whole load fail; by contrast an unrecognized body/card/position slug, a missing house inside "bodies", or a missing "spread" key entirely is silently skipped/zeroed (main.c checks for the body "no data" sentinel before printing sign/house), since all of that is supplementary display context rather than something scoring depends on.
  • Mostly header-only dependency on the engine: significance.h/ reading_io.h/narrative.h/guidance.h #include engine headers (reading.h/astro.h/tarot.h) for the struct/enum definitions, and the root Makefile never compiles or links astro.c, tarot.c, rng.c, reading.c, or the vendored Astronomy Engine into the interpreter's binary or tests — every test in interpreter/tests/ runs against hand-built JSON text or hand-built DailyReading/ DailyInterpretation values, with no real ephemeris/tarot-draw call involved anywhere. Consequently reading_io.c (planet/aspect/sign/ card/position slugs) and main.c (planet/aspect/sign display names) each duplicate a small local table rather than linking astro.c to reuse its own - keep those in sync if the engine's slugs ever change. tarot_data.c/i18n.c are the one exception, actually linked in (see above) - so card/position names and meanings are never duplicated; only the slugs reading_io.c needs for JSON parsing (which tarot_data.c doesn't expose a reverse lookup for) still are.
  • Only Aspects compete for the top 5 (SignificantItemKind is currently just ITEM_ASPECT).

Licensing

  • Engine code (engine/src/, engine/scripts/) and the root Makefile: MIT, per the repo's top-level LICENSE.
  • engine/third_party/astronomy/: vendored MIT code, unmodified — see VENDORED.md for the pinned upstream commit.
  • Card/spread text in tarot_data.c/engine/i18n/en.lang: condensed from A. E. Waite's The Pictorial Key to the Tarot (1911), public domain.
  • engine/i18n/de.lang's card/spread text is an original translation of that same public-domain English text, made for this project — not taken from any specific published German edition.
  • Transit narrative text in interpreter/src/narrative.c: condensed from Sepharial's Transits and Planetary Periods (1920, public domain — res/Transits_and_Planetary_Periods.pdf), except the Moon and Pluto entries, which are original (see narrative.c's own comment for why).
  • Guidance text in interpreter/src/guidance.c: original, written for this project — no source to condense from, since neither Waite nor Sepharial discuss combining astrology and tarot. (The Attitude/Outcome card meanings that same paragraph quotes are Waite's own text via tarot_card_meaning(), covered by the tarot_data.c bullet above, not original text.)
  • engine/i18n/de.lang's narrative.*/guidance.* entries are original German translations of the above two files' text, made for this project — same status as de.lang's card/spread text (not taken from a specific published German edition of Sepharial, since none exists for this condensed, project-specific wording anyway).