- Replace ZSH_AUTOSUGGEST_ALLOW_EMPTY_BUFFER with AI_MIN_INPUT
- Add ZSH_AUTOSUGGEST_AI_DEBUG environment variable
- Add debug logging function to diagnose failures
- Update history lines default from 20 to 5
- Update pwd history preference default to no
Min input provides clearer semantics: set to 0 for empty-buffer
suggestions or higher to require minimum input. Debug logging helps
diagnose missing suggestions by showing API request flow.
Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
Change endpoint configuration to use base URL pattern,
automatically appending /chat/completions path.
- Update default endpoint to base URL format
- Add automatic path construction in strategy
- Update README examples to use base URLs
- Update endpoint description to "base URL"
Follows OpenAI SDK standard pattern where users configure
base URL and library appends specific paths.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Enable AI suggestions on empty prompts with enhanced
environmental context.
- Update AI_MIN_INPUT default from 3 to 0
- Add ALLOW_EMPTY_BUFFER opt-in config variable
- Remove empty-buffer guards in modify, suggest, enable
- Add zle-line-init hook for prompt-time suggestions
- Enhance history gathering with PWD-aware priority
- Add env context for dir listing, git branch, status
- Implement dual prompts: predict vs complete modes
- Add prompt artifact stripping for $ and > prefixes
- Update README with empty buffer configuration
- Add tests for empty buffer and artifact stripping
Empty buffer suggestions require zsh 5.3+ and work best with
AI strategy, leveraging directory context, git state, and
command history to predict likely next actions.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Add new AI strategy that queries OpenAI-compatible LLM APIs
to generate intelligent command completions based on partial
input, working directory, and recent shell history.
- Add AI strategy implementation with JSON escaping
- Add context gathering with PWD prioritization
- Add response normalization for clean suggestions
- Add configuration defaults for AI settings
- Add comprehensive test suite with mocked responses
- Update README with setup guide and examples
Enables LLM-powered completions via ZSH_AUTOSUGGEST_AI_API_KEY
with silent failure and fallback to next strategy. Supports
OpenAI, Ollama, and custom endpoints. Requires curl and jq.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
To avoid wrapping the built-in widgets (e.g. `autosuggest-fetch`,
`autosuggest-toggle`), we were ignoring all widgets whose names start
with `autosuggest-`. This had the downside of preventing wrapping of
user-defined widgets whose names happened to also start with that
prefix.
By being more specific about the exact built-in widgets we want to avoid
wrapping, we can allow users to define widgets whose names start with
`autosuggest-`.
See GitHub issue #496.
Set ZSH_AUTOSUGGEST_COMPLETION_IGNORE to a glob pattern to have the
completion suggestion strategy never make suggestions when the buffer
matches the pattern.
This can be helpful when some completion routines you have are
particularly expensive and you want to prevent them from running
automatically on every keystroke.
See GitHub issue #463.
Set ZSH_AUTOSUGGEST_HISTORY_IGNORE to a glob pattern to have the history
and match_prev_cmd suggestion strategies never make suggestions matching
that pattern.
For example, set to "cd *" to never suggest any `cd` commands from
history (see issues #340 and #425). Or set to "?(#c50,)" to never
suggest anything 50 characters or longer (see issue #429).
$max_cursor_pos in this case was not the correct value to use. It was
calculated based on the old length of the $BUFFER. After the suggestion
is accepted, we need to recalculate the new max cursor length and use it
to set the $CURSOR.
Fixes issue #452. Follow-up to issue #302 (PR #450).
Typing `d` and then `l` runs `vi-delete` and then `vi-forward-char`. However,
by default, `vi-forward-char` is configured to accept the suggestion. So in
that case, the suggestion was being accepted and the cursor set to the end of
the buffer before the deletion was run.
The reason the user doesn't see the suggestion accepted is that `vi-delete`
doesn't finish until the movement widget is run, so we're already inside of a
`modify` when `accept` is called. `modify` unsets `POSTDISPLAY` before calling
the original widget so when we get to the accept function, `POSTDISPLAY` is
empty and thus accepting the suggestion is a no-op.
The fix is to make sure we reset the cursor to the correct place before running
the original widget.
We skip the test for versions of zsh below 5.0.8 since there was a bug in
earlier versions where deleting the last char did not work.
See http://www.zsh.org/mla/workers/2014/msg01316.html
When using async mode, stale background processes will not be cancelled
when a new one starts. This shouldn't cause any real issues since the
processes should eventually finish and be cleaned up anyway, and
removing the handler with `zle -F` means that stale suggestions should
never be shown.
Command substitution via $() trims trailing newlines so the old approach
to reading everything from the fd was preventing suggestions from ending
with newlines.
Found the read solution here: https://stackoverflow.com/a/15184414/154703
Just remove up to and including the first null byte and after and
including the last null byte.
I also looked into using `${${(0)line}[2]}`, but it fails when `$line`
starts with a null byte, since the first split string will be empty and
thus not included in the resulting array.
