feat(cli): search bootstraps itself on first use - #84
Merged
Conversation
`RepositoryParser` discovers the repository the way git does, walking up
from wherever it was invoked. `IndexStorage` did not — it looked for
`./.git` and gave up. So every command worked from the repository root
and reported "not a git repository" one directory down, which is where
most people actually are:
$ cd src && git-semantic search "race condition"
error: not a git repository: .git directory not found
Discovery also resolves the `gitdir:` pointer in a linked worktree's or
submodule's `.git` file. Submodules write that pointer relative to the
submodule directory (`gitdir: ../.git/modules/<name>`); reading the file
by hand resolved it against the process's working directory instead and
put the index outside the repository.
`InvalidGitFile` now marks a `.git` entry that exists but does not
resolve, so a broken worktree keeps its own hint instead of being told it
is not in a repository.
The storage tests built a bare `.git` directory that no git tool would
accept. They now initialize real repositories, which is what the file
headers claimed all along.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`git-semantic completions <shell>` writes a completion script for bash, zsh, fish, elvish, or powershell to stdout. Generated from the same clap definition the parser is built from, so the flags a shell offers cannot drift from the flags that exist. Named for the binary rather than the git alias: `git semantic …` completes through git's own subcommand machinery, which finds `git-semantic` on PATH. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Rust ignores SIGPIPE at startup, so writing to a closed pipe returns EPIPE and `println!` panics on it. `git-semantic completions zsh | head` printed a Rust panic and a backtrace note instead of the first few lines, and `search --json | head` did the same — the two outputs most likely to be piped were the two that broke. Restoring the default disposition on unix makes the process die the way every other tool in the pipeline does. Not covered by a test: whether the child hits EPIPE at all depends on losing a race against the reader's exit, and the completion script is 8 KB, well under a pipe buffer. A test would pass without exercising anything. Verified by hand instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Onboarding was three commands, and the first one a new user reached for
was the wrong one. `search` answered with an error telling them to run
`index`; `index` answered with an error telling them to run `init`. None
of those steps needed a human — the model download location is fixed and
the index mode has a default — so `search` now does them:
$ git-semantic search "race condition in auth"
No index for this repository yet — building one (one-time).
For a faster, message-only index instead: git-semantic index --quick
`init` and `index` still exist for warming a machine or picking a mode.
`init`'s closing message no longer recites a three-step sequence that is
no longer required.
Bootstrap indexes in full mode, matching `git-semantic index`. Quick mode
is faster but its index is not a superset, so bootstrapping into it would
quietly commit the user to the weaker option and a full re-embed to leave
it.
Progress goes to stderr, because a first run can happen underneath
`--json` and a progress line inside a JSON document is not parseable.
Indexing driven by `index` still reports on stdout as before.
Verified by hand end to end — a cold machine with no model, searching
from a nested subdirectory, downloads, indexes, and answers. Not covered
by a test: exercising it needs the 130 MB model, which CI does not have
and should not fetch.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This was referenced Aug 12, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Getting a first result took three commands, and the first one a new user reached for was the wrong one.
searchanswered with an error telling them to runindex;indexanswered with an error telling them to runinit.None of those steps needed a human. The model download location is fixed and the index mode has a default, so
searchdoes them:initandindexstill exist for warming a machine ahead of time or picking a mode.init's closing message no longer recites a three-step sequence that is no longer required.Decisions worth flagging
Full mode, not quick. Quick is faster, but its index is not a superset of full — switching to full later requires re-embedding every commit, and the tool already warns about that. Bootstrapping into quick would quietly commit the user to the weaker option.
Progress on stderr. A first run can happen underneath
--json, and a progress line inside a JSON document is not parseable. Indexing driven byindexstill reports on stdout as before; the newProgresschannel is what keeps those two cases apart.Testing
Verified by hand end to end: a machine with no model, searching from a nested subdirectory of an unindexed repository, downloads the model, indexes, and answers.
Not covered by an automated test — exercising this path needs the 130 MB model, which CI does not have and should not fetch. No existing test invokes
search,index, orinitbeyond--help, so nothing in the suite starts downloading as a side effect of this change.Third of four stacked PRs, on top of #83.