Repository navigation
v0.59.0
Release Notes — v0.59.0
Major: One background model for everything that is not the conversation
Titles, turn-memory summaries, compaction, command risk assessment, intent tool
matching, auto-indexing, deep search, memory recall and the task sub-agents all
run on a model you never picked. They now share one machine-wide choice under
Settings → General → Background tasks, stored on the server and reachable at
GET/POST /api/models/background. The default, Inherit from active session,
changes nothing: each call keeps using the model of the session doing the work.
Pin a model and it serves every background call; a model named explicitly by a
caller still wins, a pinned model that is later removed falls back to inheriting
rather than failing, and a model that is not being served is refused when you
save it instead of at the first summary. Because a verdict belongs to the model
that gave it, the command-risk cache is keyed by model and command, and compaction
now sizes its request to the smaller of the session window and the background
model's window instead of overrunning a smaller summarizer. A new guide,
Background model, covers what counts as background work and where the line
sits.
Major: Word documents and PDFs are read page by page
A .docx is now parsed from its own XML parts rather than flattened: paragraph
styles resolve through the basedOn chain, text boxes and footnotes stay in
document order, and a page boundary comes from an explicit break, a style's
pageBreakBefore or a sectPr page start, so pagination matches what Word shows.
Tracked changes no longer leak a reverted style or break into the live document,
and a break inside a table cell no longer swallows the row. Encoding is handled
by charset labels and UTF-16 BOMs, every part is bounded against zip bombs, and
unspaced scripts such as Chinese or Japanese paginate by volume instead of piling
into one enormous page. A corrupt or empty document.xml is an error, not an
empty document.
Reading follows the same page model. doc_index replaces the listing-only
docx_index and pdf_index and returns the page catalogue a step can choose
from; read_docx_page and read_pdf_page accept an end_page range and mark each
page with --- page N ---, so a long contract costs only the pages actually read.
A listing whose file changed, vanished or now carries an older timestamp than the
index is labelled stale and points at the index tool, and the indexer itself
reports per-document progress, publishes a terminal phase even when a run fails,
sweeps only the documents whose chunks failed, and makes ogcode index exit
non-zero instead of printing a success it did not have.
Major: Every model call says what it is for
Each provider request now carries X-Client-Task, naming one of fourteen call
kinds: the turn, a sub-agent, indexing, a note, search, recall, intent matching,
turn memory, compaction, a title, the risk gate, a rewrite, a probe or credential
validation. A delegated child adds what its spawner predicted about the work —
X-Client-Complexity as easy, medium or hard, and X-Client-Workload as
read-heavy or write-heavy — so a router can serve a child that only reads from a
cheaper model than one that edits. Both predictions persist on the session and are
read once per turn, so a resumed or restarted child still sends what it was
briefed with. task and teammate accept both as arguments and bounce a
misspelling rather than silently sending nothing. The fields never enter the
request body, so prompts and their cached prefixes are unchanged.
Minor: A child's memory stays under its parent
Session recall used to index a sub-agent's turn summaries against the child, so
what a teammate learned disappeared from the conversation that asked for it. A
child's memory now rolls up to the root session that started the work, resolved
through the board's spawn edges, and the command palette hides the child's own
row instead of listing it beside the parent's. The board also publishes a new
session only after its spawn edge exists, and the sidebar merges a refreshed
roster into a new array, so a hard reload nests running teammates instead of
painting them flat.
Minor: The client survives the state it was opened under
An audit pass over the web client fixed eleven high-severity issues and a long
tail of smaller ones. Escape now closes only the topmost overlay and returns
focus where it came from; a failed configuration fetch no longer wipes a stored
API key; a note poll stops after three misses instead of running forever; a
device iframe survives a status blip; and switching sessions can no longer land an
older page of messages on the newer session. Diffs keep --- and +++ lines that
are content rather than headers, the HTML preview loses allow-same-origin, the
sidebar mounts once per page, and the model picker on first run reloads instead of
reading a stale list. Smaller hardening covers pointer capture on the resize grip,
a debounced colour picker, a windowed project tree for very large repositories, and
client-side pageview tracking.
Patch: Deep search falls back instead of ranking a chain of thought
The ranker's output budget rises from 256 to 1024 tokens. Probing showed a
reasoning model spending the whole 256 on its thinking and returning no numbers
with finish_reason: length, so the cap now fits an answer; it remains a bound,
not an allocation. When an empty answer does end because the budget ran out, the
search falls back to the pages it found instead of parsing the unfinished
reasoning for digits, which had been ranking sources by line numbers quoted in
the model's own thoughts. A model that answers as reasoning and stops normally is
still accepted.