Repository navigation
v0.2.10
Smith v0.2.10
Smith 0.2.10 makes the path from an empty terminal to a working model easier
to follow. New sessions explain how to start, commands and resource pickers
are easier to operate, /connect can configure any OpenAI-compatible
endpoint, and setup can discover custom-model limits instead of asking for
three numbers. Provider failures that exhaust the retry budget now keep their
actual cause.
Changed: an empty session explains how to start
A configured but empty session now shows a short local guide for typing a
task, selecting a model, adding a connection, and opening help. The guide is
presentation only: it disappears when conversation or active work begins and
never enters model context or session history.
/help opens at the start of its new result, where the same starting-work
guide precedes the complete command and keyboard reference. Existing
conversation remains above it and ordinary non-help results keep their
tail-following behavior.
Fixed: command and picker recovery paths
Pressing Enter now executes the highlighted command completion while
preserving explicitly typed arguments and the existing idle and approval
guards. Command search falls back to descriptions after name-prefix matches,
so intent words such as switch find the model, profile, and provider
commands. The completion menu remains bounded to five rows.
Resource pickers put current and unavailable state before optional
metadata, distinguish an empty inventory from a filter with no matches, and
offer Ctrl+U to clear the filter. Essential choose and cancel keys remain
visible at narrow supported widths. Setup also preserves accepted non-secret
provider, endpoint, model, and context values when moving backward; secret
input is still cleared.
Added: connect any OpenAI-compatible endpoint in-session
The /connect picker now includes openai-compatible. Selecting it, or
typing /connect openai-compatible, enters the existing custom-provider
setup flow at the same session rebuild boundary used by other connection
changes. This covers vLLM, Ollama, LM Studio, Groq, Together, and compatible
gateways without leaving the TUI to run a separate setup command.
The entry follows the same deduplication rules as built-in connection choices.
Unconfigured-provider errors and CLI help now point to the in-session route as
well as smith setup add-provider.
Added: automatic limits for custom models
After a custom model ID is entered, setup makes one bounded best-effort
GET /models request and reads common OpenRouter, Groq, LiteLLM, and vLLM
limit fields. If the endpoint does not publish complete limits, Smith looks
for a same-name model in its reviewed embedded catalog using exact,
case-folded, path-segment, and separator-normalized matches.
Resolved limits skip the numeric questions and identify their source on the
review screen. When neither source resolves the model, setup asks for only the
total context window. Maximum input follows that window and maximum output
uses Smith's existing quarter-window rule, capped at 32,768 tokens. Back opens
the context value pre-filled so it can be reviewed or changed.
The probe has a five-second deadline and a one-MiB response cap. It may send
the selected bearer credential to the configured endpoint, but it makes no
inference request and persists nothing before the normal review and commit.
Runtime preflight remains strict: a model still cannot run without enforceable
limits.
Fixed: retry exhaustion reports the provider cause
A retryable provider error could disappear inside the retry loop, leaving a
turn that exhausted its attempt budget with a silent limit_reached result.
Attempt failures now carry their redaction-safe provider error through the
client protocol. The TUI shows the failure while retrying, and headless output
reports the last attempt error when a failed or limit-stopped turn produces no
answer. Successful retries and causeless output or tool-step limits keep their
existing outcomes.
Upgrading
Existing configuration, credentials, profiles, sessions, and checkpoints
remain compatible. Custom-provider setup may now make one authenticated model
listing request after model entry. The embedded Models.dev snapshot is
refreshed so current aliases can resolve on a fresh installation.
The pinned Agent Runtime revision is 91f5454.
What's Changed
- fix(deps): clear the rustls TLS 1.3 advisory and the yanked chacha20 by @mai1015 in #3
- fix: resume upgraded sessions and make cache/time status readable by @mai1015 in #1
- feat(context): bound inline tool evidence and preserve recoverable output by @mai1015 in #2
New Contributors
Full Changelog: v0.2.9...v0.2.10
