fcx

name

fcx

description

Locate code in a repository — returns compact file:line citations for "where is X", "what calls Y", "where is Z defined" questions, best on large or unfamiliar codebases. Reach for it instead of a multi-file grep/glob/read chain when finding the code is the work. Write a specific query that names the behavior, subsystem, error, or file you are after — vague queries are its main failure mode. It is a locator, not a reasoner: treat every returned citation as a lead and read/verify the cited range before acting on it.

fcx

Fast, autonomous repository explorer (FastContext, RL-trained). Given a natural-language question it searches the codebase with read-only Read/Glob/Grep and returns compact file:line citations as focused evidence — without spending the main agent's context on the exploratory reads. It is a locator: it finds where code lives, it does not judge whether that code is correct.

When to use

• Find code before editing, reviewing, debugging, or explaining it — "where is Y defined", "what calls Z"
• Trace a path across files or layers (request → handler → service → DB) when you know the endpoints
• Map dependencies — what a symbol depends on, or what depends on it
• A repo that is large or unfamiliar, where a manual grep/read chain would burn many turns

When NOT to use

• You already read the exact file this session
• A single obvious grep in one known file (just grep)
• A vague or open-ended ask ("explain everything about auth") — it is tuned for targeted location
• A judgment call ("is this correct", "is this a bug") — that is reasoning, not location

How to query well — this is the main lever

The model is RL-trained to reward precise, targeted citations and to penalize vague, sprawling output. Query quality, not model size, drives accuracy. So:

• Name the target. Reference the behavior, subsystem, symbol, error string, or file. Don't ask "how does this work" — ask "where is the PreToolUse hook registered in settings.json".
• One target per call. Ask for one thing, not a tour. Split compound questions into separate calls.
• Concrete > abstract. Use the repo's own nouns (function names, config keys, error text).

Bad → Good:

• how do hooks get wired and which scripts run → where is the jj guard PreToolUse hook registered in settings.json
• explain the auth flow → where does session validation happen in the auth middleware

If a query comes back vague or wrong, rephrase it more specifically and re-run — that fixes most misses.

Verify before you act

fcx returns leads, not ground truth (a 4B model can cite confidently and be wrong). After a call:

1. Read only the cited ranges, with narrow line windows.
2. Confirm the citation actually answers the question before editing or concluding.
3. Don't re-run broad repo-wide searches for the same thing — that throws away the token savings.

Usage

# Precise answer with file:line citations (run from the repo, or pass --path)
fcx explore "where is request validation handled in the API router" --citation

# Machine-readable output for programmatic use
fcx explore "where are DB migrations defined" --json

Leave --max-turns at its default (8) — that is the model's RL training horizon, so it explores best within it; raising it pushes the model out of distribution and rarely helps. Only nudge it up for a genuinely deep cross-layer trace, and don't expect linear gains.

The first invocation starts the shared local model server (it stays resident across calls); later calls attach in milliseconds. fcx status shows the model server; fcx stop-model tears it down.