Commit Graph
5 Commits
Author SHA1 Message Date
Dylan M. Taylor f180fc0d35 Use chrono instead of hand-rolled UTC conversion (#219)
* Use chrono for UTC/calendar conversion instead of hand-rolled copies

Replaces the hand-rolled civil_from_days and Day::to_days in
crates/ai-core/src/history/report.rs with chrono, added as a workspace
dependency (std-only, so ai-core still compiles for wasm). The day/month
grouping, ISO and CSV timestamps and the Day/Month display are byte-identical
and pinned by the existing tests. AGENTS.md now says time goes through chrono.

Follow-up to storytold/cadcraft#1785

* history: preserve out-of-range conversion failure in civil_from_days

civil_from_days used `i32::try_from(..).unwrap_or(0)` before calling
NaiveDate::from_num_days_from_ce_opt, so a CE day count past i32 (e.g.
from Day::from_days(i64::MAX)) passed a valid day 0 to chrono and read
as 0000-12-31 instead of the 0000-01-01 fallback the None arm gives for
every other out-of-range date. Convert with `.ok().and_then(..)` so both
the integer-conversion and NaiveDate-construction failures share the one
fallback. Regression test covers i64::MAX and i64::MIN.

Doc comments on the chrono-touched to_days/from_days/civil_from_days now
state the out-of-range behaviour (0 / 0000-01-01) for the docstring gate.

* Convert every i32 year, not only chrono's range

Day holds any i32 year, but NaiveDate stops near ±262143, so years like
262144 read as 0 / 0000-01-01. The Gregorian calendar repeats every 400
years: chrono now converts the day within its cycle and whole cycles are
added back, in both directions. Values match the pre-chrono formula
(tested for 262144, -262145, i32::MAX and i32::MIN). Nonexistent days
still read as 0, and day counts whose year doesn't fit i32 still read as 0000-01-01.

* Name the 400-year range extension as the one exception to the time rule

The rule said never to hand-roll calendar math, but the range fix
(shifting by whole 400-year Gregorian cycles so dates past chrono's year
range keep their old output) is cycle arithmetic, so the rule and the code
disagreed. The rule now names that exception: the library still computes
month, day and time, and only whole cycles are carried in the year.

* Say the library computes month and day; time-of-day is plain arithmetic

Narrows the range-extension wording: the date library computes the
calendar fields, and splitting seconds into hours, minutes and seconds
is ordinary arithmetic, which is what the code does.
2026-10-23 11:17:51 -04:00
Dylan M. Taylor a719521c13 security: bound ai-core redaction, template rendering and model checks; consent for params-only requests
- redact matches secrets in the original text only, so markers are never searched again
  (CodeRabbit CWE-400, secret.rs:47)
- render_text/render_value share one MAX_RENDERED_BYTES budget across repeated slots
  (CodeRabbit CWE-400, spec/template.rs:122)
- Connection::validate checks model-ID uniqueness with a set instead of a pairwise scan
  (CodeRabbit CWE-407, connection.rs:205)
- Disclosure::kinds counts parameters as text, so a request carrying only params still needs
  an agreement (CodeRabbit CWE-863, consent.rs:123)
2026-10-22 21:37:34 -04:00
skylar taylor-barrick 559992953f craft-ai-core: record which address a test's models came from
A test is not instant and the address field stays editable while it runs,
so its answer can arrive after the address has changed. `set_models` then
recorded the models against whatever the field held by the time it
returned, filing one endpoint's models under another — the mislabelling
`models_fit` exists to prevent, reached the asynchronous way rather than
the synchronous ones already closed.

`set_models` now takes the address the test was started against and
returns whether the models were taken, so an answer that outlived its
address is dropped and the caller can say the result is out of date. The
parameter replaces the old signature rather than joining it: there was no
second legitimate caller, and leaving an address-free version beside it
would have left the racy call the shorter, easier one to reach — against
this module's rule that the rules live once so two screens cannot
disagree.

Also fixes a pre-existing mismatch found while testing this:
`from_connection` recorded the loaded address untrimmed while `prepare`
compares it trimmed, so a stored endpoint with stray whitespace dropped
its models on the first save.
2026-10-16 23:07:53 -06:00
skylar taylor-barrick a40245bd36 craft-ai-core: track model provenance by address; require the inputs a model needs 2026-10-16 20:26:06 -06:00
Jeremy bd175be54f craft-ai-core: rename from craft-ai-model; lapse each kind of consent on its own; keep models found at a new address
Consent expiry is now per kind of data: renewing text no longer renews an expired image permission. A connection edit keeps models found by a test at the new address, and drops only the ones saved for the old one.
2026-10-15 20:06:57 +01:00