mirror of
https://github.com/stablyai/orca.git
synced 2026-10-21 16:03:26 +00:00
Brennan Benson
3d6a98a834
* fix(native-chat): locate Claude's model row by frame structure
The scraper assumed the model descriptor sits within three rows of the
`Claude Code vX` line. It does not: Claude prints it near the bottom of the
startup frame with the welcome art and release-notes panel in between — eight
rows down at 100 columns. Narrow panes degrade the frame further, dropping the
version from the title row entirely below ~70 columns and wrapping the billing
tail onto its own row. Any one of those made the scrape return null, so the
model picker showed no current selection at all.
Search the frame from its bottom border upward for the row carrying model
metadata, read only the leftmost frame cell so release-notes prose can never
win, accept the frame corner as header proof when the version is gone, and
tolerate the effort suffix being elided to an ellipsis. Catalog families now
match as a leading word, which both survives the resolved-name suffix
("Opus 5 (1M context)") and keeps custom slugs like company/my-haiku-v2 from
being claimed as haiku; an unrecognized name is reported as a custom model.
Fixtures are real: captured from a live claude 2.1.220 by replaying the PTY
bytes through @xterm/headless and serializing exactly as TerminalPane does.
* fix(native-chat): resolve the scraped model against the host's real catalog
The scraper matched the static seed while the picker lists what #12369
discovers from the host CLI, so the two spoke different id spaces. On a current
CLI `list_models` returns `opus[1m]`, `sonnet`, `sonnet[1m]`, `fable` and
`haiku` — no plain `opus` — while the seed only knows families. Reporting
`opus` therefore selected a row the picker had to invent, dropping the host's
own effort and fast-mode descriptors with it. Locating the model row correctly
made this the normal case rather than a rarity, since the scrape now succeeds.
Resolve against the discovered list first, falling back to the seed for aliases
a host no longer lists and to the raw name for genuinely custom models. Matching
requires the family to lead as a whole word and the label's remaining tokens to
appear in order, so `Opus 5 (1M context)` picks `opus[1m]`, plain `Sonnet 5`
keeps `sonnet` instead of being captured by the 1M-context row, and
`opus-internal-v3` stays custom. Most specific label wins.
The hook keeps the screen that parsed so a discovery landing after the first
read re-resolves it, rather than stranding a family id once the frame has
scrolled out of the buffer.
* fix(native-chat): identify option-less models on narrow panes
Live capture at 60 columns: a Haiku session prints a bare `Haiku 4.5` row with
its billing wrapped to the next line. No middot, no effort suffix — nothing
marks it as the model, so it reported nothing. Claude always closes the frame
with the working directory and prints at most the descriptor plus a wrapped
billing line above it, so fall back to walking up from there when no row
carries descriptor metadata. The height bound is what keeps the walk from
climbing into the welcome art.
* test(native-chat): pin re-resolution when discovery lands after the read
Covers the wiring the parser tests cannot reach: the frame is visible at mount
and gone by the time the host's model list arrives, so only the cached screen
can drive the second resolve. Fails against a listener that merely replaces the
models.
* fix(native-chat): prevent stale Claude model reports