121 releases 129 tags

RSS feed

• v0.58.0 40cf04f3a8

  Compare

  v0.58.0 Stable

  joeyr released this 2026-09-22 00:31:06 -05:00 | 41 commits to main since this release

  The weather tenant is complete at nine tools. W2, W3 and W4 all shipped together:
  list_stations, get_model_run, get_taf, search_wind_shifts, get_observations,
  get_alerts, get_forecast_discussion and get_hourly_forecast, alongside W1's
  get_forecast_grid.

  One release rather than three, deliberately, and the reason is the release rule's own
  argument. That rule exists so nothing running in production lacks a version to name it
  by. These three phases were built and verified in one session and deploy together behind
  one connector reconnect, so three tags would have named three moments that never existed
  apart — fictional history rather than a finer-grained record. The entry below is organised
  by phase instead.

  The build overturned three of its own plan's premises, each by measurement. The
  Open-Meteo forecast response carries no issuance time at all, so the run time for a
  tool named get_model_run had to be found in a second document. The configured global
  model is three-hourly, not hourly, so its hourly series is interpolated — the grid's
  native-versus-filled problem arriving through a different door. And the per-request model
  choice could not be built on nulls-outside-the-domain because there are none: a
  coordinate outside a model's domain is an explicit HTTP 400.

  Both of W0's open questions are answered, and one of them without waiting a day. The
  plan said the gridpoint reissue cadence needed samples spread across a day; fifteen
  offices sampled at one instant gave the same distribution, because offices issue on
  staggered schedules. Median issuance age 5.48 hours, maximum 10.36, minimum 0.43 —
  so the 3-hour warn line would have fired on 10 of 15 normal gridpoints and a 12-hour
  line on none. HRRR's horizon is anchored to its run: the extended 00/06/12/18Z cycles run
  48 hours and the hourly cycles between them run 18, so a request gets 42 to 48 hours and
  hours=48 stands as a default.

  1805 tests green, 155 of them this tenant's. 40 mutants were run against the new
  guards and all 40 were caught — two survived the first pass, and both were the failure
  CLAUDE.md warns about rather than a gap in the code: one mutant was ineffective, and one
  test asserted a true thing the broken code also satisfied. The second is written up under
  Fixed.

  Added

  • list_stations — every reporting station near a coordinate, nearest first, with
    what can be read from each. The phase that proves "anywhere" is real, since bbox
    is the mechanism that works outside the United States and /points is the one that does
    not. The widening ladder is 0.5 → 1 → 2 → 4 → 8 degrees, per product, stopping each as
    soon as it answers.
  • get_model_run — an hourly model forecast from whichever configured slug covers the
    coordinate, with the run's initialisation time, its native step, and how much of the
    window actually carried data.
  • get_taf — the terminal forecast for the nearest airfield, with its change groups
    read for what they mean rather than as a list of wind samples. One station, never
    several; station is the one argument that overrides the coordinate.
  • search_wind_shifts — a scan over the model run and the TAF, with the raw grid as
    corroboration reported alongside and never as the scan surface.
  • get_observations — what the weather is actually doing, from NWS inside the United
    States and from the aviation network everywhere else.
  • get_alerts, get_forecast_discussion, get_hourly_forecast — the context
    tools. The discussion comes back verbatim and nothing is parsed out of it.
  • Seven network-free reader modules, each fixture-driven like calendar/expand.py and
    for the same reason: stations.py, taf.py, model.py, observations.py, alerts.py,
    hourly.py, shifts.py.
  • weather/units.py — the conversion layer, moved out of grid.py now that it serves
    four endpoint shapes rather than one.
  • A preflight check per configured model slug, advisory and never fatal, modelled on
    _check_espnff_season. The plan specified validating each slug against Open-Meteo's
    model list; there is no list — /v1/models 404s and the invalid-slug rejection names
    the bad value without enumerating the good ones. What replaces it is one real forecast
    request per slug at a point inside that slug's own coverage, which validates the string
    and that the model answers. Its detail line reports the run state, which is where the
    three-hour native step surfaces without anybody going looking for it.
  • envelope.from_epoch and envelope.from_iso — one conversion point for the three
    timestamp shapes these upstreams send: an ISO string with an offset, an ISO string with
    none, and epoch integers.
  • envelope(deterministic=True) — for a station list or a grid resolution, which
    carry no issuance time because they are mappings rather than products. The fields stay
    identical so a caller reading the booleans never has to know which kind it holds; only
    the sentence differs, because "do not describe it as current" warns about the wrong thing
    on a mapping and teaches a reader to discount the note that matters.

  Changed

  • client._classify no longer calls every 4xx "an argument problem rather than an
    outage", and it carries the body out on the exception. Open-Meteo answers both a
    coverage boundary and a misspelled slug with a 400, and the reason string is the only
    thing that distinguishes them — so a fallback that merely caught a 400 would have
    routed every CONUS request to the global model forever on a typo, and looked like working
    software. WeatherBadRequest carries up to 2 KiB of the body, read while it arrives and
    redacted by coordinate shape, because NWS's own 400 on /alerts/active echoes the
    rejected value and that string reaches a tool response. An unrecognised 400 is never
    fallen back from.
  • The gridpoint staleness line is 12 hours, revised from 3 by use — the first
    threshold on this tenant to be revised, which is what the table's own comment invited.
    See the measurement above. The 24-hour absurdity ceiling did not move.
  • The model's staleness thresholds are derived per model from its own document rather
    than read from THRESHOLDS. Measured availability lags: 1.53 h for the hourly
    high-resolution model, 5.51 h for ncep_gfs013, 7.12 h for ecmwf_ifs025. So a
    normal freshest run is ~2.5 hours old on one and ~13 on another, and a single row would
    have marked every global-model answer stale.
  • CLAUDE.md corrected: the raw gridpoint does not carry generatedAt. That section
    claimed both /gridpoints/{wfo}/{x},{y} and /forecast/hourly carry it. Checked across
    sixteen offices and then directly on one cell: the raw gridpoint's scalar properties are
    @id, @type, updateTime, validTimes, forecastOffice, gridId, gridX and
    gridY, and generatedAt is not among them. The original measurement was real and came
    from the hourly endpoint; the claim about the raw one was inferred and never checked.
    The rule is stronger for the correction: on the endpoint that does carry both, the gap
    measured ten hours, and generatedAt changed between two requests fourteen minutes
    apart on the same unchanged grid.
  • docs/weather-plan.md is down to four items from a full build plan: the recorded
    frontal passage the wind-shift thresholds must be tuned against, the live pass, W5, and
    one operator decision. A plan that shrinks to its outstanding items rather than emptying
    is that rule working — Canvas's deletion is why, since three live checks were sitting
    only in that plan when it went.
  • .env.example says the global model's native step is three hours, since that is an
    operator decision the boot log now prints.

  Fixed

  • A wind shift that went out and came back reported the return as a remnant.
    shifts._collapse treated a candidate window that began exactly where the previous one
    ended as overlapping it, so a wind veering 120 degrees and then backing 120 reported the
    veer and a 60-degree fragment of the back. Two rotations that meet at a point are two
    events. Found by writing the test, which is the discipline in CLAUDE.md working as
    advertised.
  • get_observations reported a mistyped station as an upstream failure. The fallback
    to the aviation network was unguarded, so a 204 — which is how that host says "no such
    station" — escaped as a transport error and lost the sentence that names the typo. It now
    reaches the empty-versus-invalid branch, which asks /stations/{id} and says which of the
    two facts it is.
  • A cap test that could not reach the branch it named. It asserted the error detail was
    short, which the final slice guarantees on its own — so it passed against a client with
    the read cap removed. It now counts the chunks the upstream was actually asked for: a
    body noticed after it has arrived is a body already in memory. This is the a test that
    cannot reach a branch is not coverage of it failure, caught by the mutation sweep rather
    than by review.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads