Repository navigation
v0.59.3
Release Notes — v0.59.3
Major: An image is read by a model that can see, even when the chat model cannot
Most coding models still cannot take a picture, so a screenshot, a scanned page
or a chart on disk used to be a dead end for them: read refused a binary file
and view_image was offered only to models the registry says support images.
The new analyze_image tool closes that gap. It runs one fixed vision call of
its own and returns text, so the asking model never has to see the pixels — the
same shape as the deep-search pipeline, applied to an image. Three modes frame
the call: transcribe returns every visible character in reading order without
interpretation, describe gives a thorough factual account of the frame
(diagram components and connections, chart axes and trends, interface elements
and dialog state) and quotes on-image text verbatim, and answer (the default)
answers a pointed question out of what the pixels contain. A question that
is really a debugging inquiry — a crash dialog, a stack trace, a red build, a
not-found page — is answered as evidence, reading and next step, with
observation and inference kept visibly apart. The tool is offered to every agent
that already carries the document tools, rides the first request, and the
binary refusal in read and the blind-model refusal in view_image now both
point at it.
The file side is bounded the way image work has to be: a 20 MB read cap, a
120-second bound on one call, and a passthrough rule that sends a JPEG or PNG
under 4 MB to the provider byte for byte, because re-encoding softens exactly
the text edges OCR depends on. Other formats — GIF, BMP, WebP, and oversize
PNG/JPEG — are decoded and re-encoded. Every result is cached under a SHA-256
of the image bytes plus the mode, question, resolved model and provider, and
the effort, with a seven-day TTL, so a long session never re-sends the same
picture twice; a cached answer says so in its footer. A live answer ends with
_Analyzed via <model> (<provider>) in Xs._ so the caller knows which eyes read
it. effort — low, medium or high — is resolved against the vision model
rather than the router's complexity class, and a level that model does not take
is dropped instead of failing the call. The vision model is mandatory: when no
vision-capable model is available the tool fails loudly and names the setting
to fix, rather than silently answering from the filename.
Minor: The image analysis model is a machine setting, not a per-chat one
Settings → General → Background tasks gains an Image analysis model row,
beside the background model it now shares a picker with. Left unset, image
analysis inherits the session's own model when that model can see images; set,
it applies to every session on the machine. Only vision-capable models are
listed, and the write path refuses a model the registry marks as sightless or
one this machine cannot reach, so the setting cannot be pointed at something
that will fail on first use. Storage is ui_prefs.image_model and
image_provider (migration 078) behind GET and POST /models/image; an empty
model clears the override and restores inheritance.
Minor: Provider concurrency has a ceiling you can set
The process-wide in-flight limit for provider requests was a constant of eight,
chosen to sit just above the indexer's own five. It is now a setting — Settings
→ Provider requests — and it takes effect immediately: the budget reshapes
its index lanes and wakes the waiters already queued instead of leaving them for
the next run, and it never preempts a streaming request that is already open.
Two is the floor, because one would leave the index unable to start beside even
a single interactive request, and zero means "use the default" — eight. The
OGCODE_PROVIDER_MAX_CONCURRENT environment variable still outranks the stored
value; when it is set the screen shows the value read-only and says why. The
ceiling is applied at server boot, at the start of a prompt run and at the start
of an index run, so a CLI invocation and a served session agree. Storage is
ui_prefs.provider_max_concurrent (migration 079) behind GET and
POST /providers/concurrency.
Minor: The model catalog catches up
Claude Fable 5.2 becomes the default: $8/$40 per million tokens with cache
write at $10 and cache read at $0.20, on a full 1M window. Haiku 5.5
replaces Haiku 4.5 at $0.10/$0.50 with the same 1M window, adaptive thinking
and a selectable effort, and Mythos 5.2 follows at $8/$40. Fable 5.1,
Mythos 5.1 and Haiku 4.5 move to legacy rather than disappearing, so existing
sessions keep resolving. OpenAI gains GPT-6.1 Sol at $2/$10 and GPT-6.1
Terra at $1/$5, both on the Responses API only with the 922K input cap.
Google gains Gemini 4 Hydrogen at $1.50/$7.50 and Gemini 4 Carbon at
$2/$10 on a 1M window, Carbon carrying an introductory price through
2026-11-30 and rising to $4/$20 on 2026-12-01. Kimi gains K3.1 Pro at
$2/$10 and K3.1 Pro Highspeed at $4/$20.
Minor: The search ranker gets room to think
The deep-search ranker was capped at 1,024 output tokens, which a reasoning
model can spend before it emits a single ranking — the call then came back
empty and the pipeline fell back to unranked results. The budget is now 4,096,
which is a bound rather than an allocation: nothing is charged for room left
unused.
Minor: A subagent's work shows in a dock above the composer
The subagent indicator leaves the header for a floating dock at the bottom
right, just above the prompt box, where the work it reports happens. Each pill
carries the subagent's title and a spinner while it runs, reveals its workload
and elapsed time on hover, and the region is announced through aria-live so a
screen reader hears a child start and finish. Five pills show at once and the
rest fold into +N more, so a wide fan-out no longer pushes the header around.
Patch: "Project Index" is now "Source"
The screen, the settings group, the tool's own titles (Source, Source / <subdir>) and the empty-index message all speak of the source, matching what
the index actually is. The empty message also says what to do about it: run
Index workspace first to build the source map.
Patch: Prices in a message are no longer typeset as math
The inline-math rule treated a pair of dollar signs anywhere in a paragraph as a
KaTeX span, so a model quoting $3/1M … $0.05/1M got its prices italicised and
its spacing mangled. The pattern now requires a real delimiter — the opening $
hugging its first character, the closing $ hugging its last, a backslash
escaping a literal \$, and no digit directly after the closing $ — so
currency reads as currency and formulas still render.