- Add kimi-coding provider using Anthropic Messages API
- API endpoint: https://api.kimi.com/coding/v1
- Environment variable: KIMI_API_KEY
- Models: kimi-k2-thinking (text), k2p5 (text + image)
- Add context overflow detection pattern for Kimi errors
- Add tests for all standard test suites
Swapped order of initExtensions() and renderInitialMessages() in init()
so loaded resources (Context, Skills, Prompts, Extensions) appear at the
top of the chat instead of at the bottom when resuming a session.
* fix(tui): keep file suggestions open when typing in Tab-triggered mode
Previously, pressing Tab on an empty prompt or after a space would show
file suggestions, but typing a letter would dismiss them because the
text didn't look like a path pattern. Now the editor tracks whether
autocomplete was triggered via Tab (force mode) or naturally (regular
mode), and uses the appropriate suggestion method when updating.
* fix(tui): hide autocomplete when backspacing slash command to empty
Previously, typing / showed slash command suggestions, but pressing
Backspace to delete it showed file suggestions instead of hiding all
suggestions.
Instead of buffering `\` and waiting for the next key, let it be
inserted immediately. On Enter, check if preceded by `\` and treat as
newline.
Removed backslash handling from the Input component entirely.
- Add huggingface to KnownProvider type
- Add HF_TOKEN env var mapping
- Process huggingface models from models.dev (14 models)
- Use openai-completions API with compat settings
- Add tests for all provider test suites
- Update documentation
fixes#994
Adds support for extended cache retention via PI_CACHE_RETENTION=long:
- Anthropic: 5m -> 1h TTL
- OpenAI: in-memory -> 24h retention
Only applies to direct API calls (api.anthropic.com, api.openai.com).
Proxies and other providers are unaffected.
fixes#967
- Handle pipe-separated IDs from OpenAI Responses API in openai-completions provider
- Strip trailing underscores after truncation in openai-responses-shared (OpenAI Codex rejects them)
- Add regression tests for tool call ID normalization
fixes#1022
Previously, within a single tool-use turn, rate limit retries would
accumulate across separate LLM calls. For example, if each of 3 tool
calls hit a 429 and retried once, the counter would show '3/3' and fail
even though each individual retry succeeded.
Now the counter resets immediately when a successful (non-error)
assistant message arrives, so each LLM call gets a fresh set of retries.
Fixes#1019
429 (Too Many Requests) was incorrectly classified as context overflow,
triggering compaction instead of retry with backoff. The original logic
assumed token-based rate limiting correlates with context overflow, but
these are different concepts:
- Rate limiting (429): requests/tokens per time period (throughput)
- Context overflow: single request exceeds context window (size)
Now 429 errors are handled by the existing retry logic with exponential
backoff, while 400/413 remain as potential context overflow indicators.
fixes#1038
- Add alt+b/alt+f as alternative bindings for cursorWordLeft/cursorWordRight
- Add ctrl+d as alternative binding for deleteCharForward
- Fix ctrl+d in custom editor to perform delete when text is present
(previously it was always consumed, even with non-empty input)
Closes#1043
Co-authored-by: Jason Ish <ish@unx.ca>
Proxies like Portkey omit input_tokens in message_delta events (it's nullable
per the SDK). The previous code unconditionally overwrote usage fields, causing
input token counts to reset to 0.
Now only updates usage fields when they are present (not null), preserving
the correct input_tokens value captured from message_start.
Fixes#1045
* Add get_commands RPC for headless clients
Headless clients like Emacs can now query which commands are available.
Previously they could only discover file-based prompt templates by
scanning the filesystem; extension commands and skills were invisible.
The response includes each command's name, description, and source
(extension, template, or skill). Commands appear in the same order
as the TUI's autocomplete: extension commands first, then templates,
then skills.
Built-in TUI commands (/settings, /fork, etc.) are excluded since
they require the interactive UI. Commands like /compact have dedicated
RPC equivalents instead.
* Add location and path to get_commands response
Clients can show where commands come from (user/project/path) and
display file paths in tooltips. The data is already available on
templates and skills - just exposing it.
When Tab triggers file autocomplete and there's exactly one matching
suggestion, apply it immediately without showing the menu. This makes
path completion faster by eliminating the extra Tab press to confirm.
Add status === 'deprecated' check for OpenCode Zen models, matching
the existing pattern used for GitHub Copilot models. This removes
deprecated models like glm-4.7-free and minimax-m2.1-free from the
generated model catalog.
Move setup callback handling from interactive/rpc modes into AgentSession.newSession().
After setup() runs, sync agent state via replaceMessages() so the LLM has context
and the UI renders the messages properly.
fixes#968
