v0.1.0

[0.1.0] — 2026-05-26

Initial public release. A Model Context Protocol server exposing eight
no-auth astrodynamics tools — TLE lookup, SGP4 propagation, Lambert
solving, ground-station access, time-scale and coordinate-frame
conversion, interplanetary porkchop scans, and B-plane targeting — over
stdio and Streamable HTTP transports, with a regression eval against
GitHub Models on every workflow dispatch.

Added

• tle_lookup — fetch current TLEs from CelesTrak by NORAD catalog
  number, satellite name, or group. Routes through the cache layer with
  per-source TTL and a stale-fallback path when the upstream is
  unreachable (#9).
• sgp4_propagate — propagate a TLE across UTC ISO 8601 epochs and
  return position/velocity in TEME, ICRF, GCRS, ITRS, or CIRS, backed by
  the sgp4 C implementation. Default response size capped to keep the
  MCP envelope small for the LLM consumer (#10).
• lambert_solve — solve Lambert's problem via lamberthub, with
  multi-revolution enumeration and optional two-impulse Δv computation.
  Four solvers exposed; defaults pick the one with the broadest
  convergence basin (#11).
• access_windows — ground-station / observer access intervals over a
  time window, returning AOS / LOS / peak-elevation triples. Named
  station registry plus arbitrary lat/lon/elevation, range filtering,
  and min_elevation_deg gating; backed by skyfield (#12).
• time_convert — convert epochs across UTC, TAI, TT, TDB, UT1, GPS,
  TCB, and TCG in ISO 8601, JD, MJD, J2000-seconds, and Unix
  representations, via astropy.time (#13).
• frame_transform — transform state vectors across ICRF, ITRS, GCRS,
  TEME, CIRS, TIRS, and IAU body-fixed frames, via
  astropy.coordinates (#14).
• porkchop — (depart × arrive) Δv / C3 grid for interplanetary
  transfers, composed from lamberthub and a JPL Horizons ephemeris
  adapter. Returns either a summary, full grid, or ASCII-contour view;
  default response size capped (#15).
• bplane_target — B-plane element calculation and impulsive targeting
  for hyperbolic flybys, fed by the same Horizons adapter (#16).
• FastMCP server primitive with a per-tool description lint that fails
  CI on missing units, undocumented arguments, or empty examples in any
  registered tool (#7).
• astrodynamics-mcp console script with stdio and http
  subcommands. The stdio path is what MCP clients launch as a
  subprocess; the HTTP path exposes the same tool surface over
  Streamable HTTP per the 2025-11-25 spec (#8).
• Typed error hierarchy under astrodynamics_mcp.errors rooted at
  AstroMCPError, with explicit units discipline across every tool
  output — every numeric field carries an explicit unit string and
  every state vector carries an explicit frame string (#3).
• Shared pydantic base schemas with JSON-schema export, so every tool's
  input and output surface is type-checked at the boundary and the same
  schemas drive the docs site's tool reference page (#4).
• XDG-aware on-disk cache layer at ~/.cache/astrodynamics-mcp/, with
  per-source TTLs and a stale-value fallback when the upstream is
  unreachable. Cache key derivation is canonical so a re-run with the
  same arguments hits the same row (#5).
• Data adapters for CelesTrak (gp.php), JPL Horizons, and IERS
  Bulletin A, each cached through the layer above. The Horizons adapter
  speaks the small subset of the API the tools actually need; the IERS
  adapter resolves UT1-UTC for time_convert and frame_transform
  without requiring an astropy data download at runtime (#6).
• Cross-tool validation test suite — TLE round-trip across
  tle_lookup → sgp4_propagate, reference-output regressions for
  Lambert and porkchop, an equivalence test comparing
  lambert_solve's two-impulse Δv against a Hohmann analytic, and
  failure-mode coverage for each typed error code (#17).
• Inspect AI eval suite with a hybrid scorer (trace-match + functional
  tolerance), exercised by eval/tasks.py. Goldens cover every tool at
  the single-tool tier and at least one tool at the sequential and
  planning tiers. The manual GitHub Models gate via eval.yml
  measures whether the LLM picks the right tool and binds the right
  arguments on Claude Sonnet 4.6 and gpt-4.1-mini (#18, #19).
• MkDocs-Material documentation site published at
  astro-tools.github.io/astrodynamics-mcp
  — getting started, per-client setup, tool reference, recipes, data
  sources, eval suite, supported clients, FAQ, and output shaping.
  Deployed to GitHub Pages on every v* tag push (#20).
• Three runnable example sessions — a Hohmann Δv single-tool prompt
  (lambert_solve), Hubble passes from Madrid as a sequential prompt
  (tle_lookup → access_windows), and a Mars 2028 launch-window
  planning prompt (porkchop + follow-up lambert_solve). Each ships
  with a markdown transcript and a Python script that drives the same
  sequence against an in-process MCP server and asserts numerical
  tolerance (#21).
• CI on Ubuntu and Windows × Python 3.10 / 3.11 / 3.12 (6 cells),
  separate lint, typecheck, and minimal install smoke jobs, and a
  release workflow that builds via uv build, publishes to PyPI via
  trusted publishing, and creates the GitHub Release on every v* tag.
  macOS is deferred to v0.2 (#1, #2).