v0.46.0

Release Notes — v0.46.0

Minor: Every model gets a reasoning-effort picker

A reasoning model does not always need to think as hard as it can, and letting it
try costs you time and tokens. Beside the model picker — in the composer, on the
home screen and in the plan composer — there is now an effort picker. Claude goes
from low to max, GPT-5.x from none to xhigh, gpt-oss from low to high, some models
only switch their thinking on or off, and a model with no such control shows no
picker at all.

Each model's levels come from the catalogue, which now records its ladder and its
default, or live from the host where the host publishes them — Ollama's thinking
metadata and OpenRouter's reasoning object. Your pick is remembered per model and
carried on every prompt, and it is stored on the session, so a turn you resume
keeps the effort you chose. Sub-agents a task starts inherit it. From the command
line, ogcode run --effort <level> sets it.

The same catalogue pass brings Claude Sonnet 5.5 in as the Anthropic default,
moves Sonnet 5 to legacy, and adds a Gemini section. GPT-5.4 and later show a note
where the levels would be, because their tools on Chat Completions are reachable
only with reasoning off. Mistral is left out of the picker; reasoning turns its
streamed content into chunk arrays.

Minor: Reasoning goes back to the hosts that need it

Some hosts refuse a tool turn that arrives without the previous reasoning.
DeepSeek V4 answers 400 and the turn stalls; Kimi K3, GLM and Qwen Cloud want
their reasoning returned too. Ogcode now sends it back.

A host that needs it opts in, and for those hosts the previous reasoning is
attached to each assistant message, taken from the stored reasoning so the history
is resent byte for byte. The field is named the way each host names it:
reasoning_content for DeepSeek, Z.ai, Moonshot and Qwen Cloud, reasoning for
Ollama and OGX, and nothing for the hosts that neither want nor accept it. Only
the current model's own reasoning is carried, so the token count reflects what is
really sent.

Minor: The turn summary keeps what a later turn needs

The summary written at the end of a turn is the only record of that turn a future
turn will ever read, and it had been quietly dropping detail. A seven-item
proposal could come back as "proposed 7 definitions", so the next turn knew the
proposal existed but not one word of what it said.

The scribe's instructions now keep every item of a list — its name and its value,
in order, in the original wording — and every identifier, path, command, URL and
error message verbatim; concision applies to the narration around them. Two rules
keep that faithfulness honest. A secret that appears in a tool input is redacted
whatever else the prompt says, and a status is taken only from the agent's own
account of the outcome — a test command shows that a test was run, not that it
passed. Headings must name their subject, so the outline a later turn navigates by
says what each section holds, and a correction or preference from the developer is
kept in their own words under a section of its own.

Other changes

• GitHub Models was retired upstream on 2026-07-30 and no longer answers, so it
  is gone from the provider presets and from the logo belt on the home page. A
  configuration that still names its URL keeps its "GitHub Models" label rather
  than becoming an unnamed slot.
• Renaming a session in the sidebar now focuses the title field on every rename,
  not only the first one, and a row that rebuilds mid-edit no longer saves a
  half-typed title or renames twice on Enter.
• The server, worker and CLI test suites run against a throwaway home now, so a
  go test run no longer migrates the developer's real database, and they send
  nothing to PostHog.
• Dead wiring in the worker that built a second provider registry and model
  catalogue — never called, since each hosted worktree runs a full server — is
  removed, and the web build carries no type errors.