Fastcontext

fastcontext

!command -v fastcontext >/dev/null 2>&1 || echo '> ⚠️ **fastcontext CLI not found on PATH.** Install and configure it before using this skill: https://github.com/T0mSIlver/fastcontext#installation'

A read-only repository-exploration subagent. It searches and reads files in a separate process, then returns a compact <final_answer> block of path:line citations. Delegating to it keeps broad exploration out of your context window — you get the evidence, not the file dumps. It never edits; you act on what it finds.

When to use

Run it before you answer/edit/review/debug code you're not already sure about:

• Understand or explain how something works
• Locate where a symbol or behavior is defined or used ("where is X?", "what calls Z?")
• Trace logic across files or layers (request → handler → service → DB)
• Assess blast radius ("what breaks if I change X?")

Prefer it over manual grep/glob/read chains whenever the answer spans more than one file.

When NOT to use

• You already read the exact file this session
• A single obvious grep in one known file
• A pure write/generate task needing no exploration

Usage

# Machine-readable: prints ONLY the <final_answer> citation block to stdout
fastcontext -q "<specific, detailed question>" --citation

# Large exploration (architecture, multi-module trace): give it more turns
fastcontext -q "<complex question>" --citation --max-turns 16

# Drop --citation to also get a prose explanation (more context, some noise)
fastcontext -q "<question>"

Output on stdout:

<final_answer>
src/app/router.py:42-58 (request validation)
src/app/service.py:10-33 (handler that calls it)
</final_answer>

Parse the path:line-range entries and read only those spans — that's the point.

Notes

• stdout = answer, stderr = diagnostics. Read stdout; ignore stderr.
• Tuning belongs in your fastcontext config, not in each call. The only flag worth reaching for per-run is --max-turns 16 on a large exploration.
• Treat cited ranges as candidate evidence, not ground truth. Ranges the model never opened are dropped, but it can still cite a real file at the wrong lines — open each span and confirm it actually answers your question before relying on it.
• Negative claims are unreliable. Positive citations are safe to open directly, but any "X does not exist / is not referenced" conclusion needs a direct grep before you act on it — fastcontext has reported a pattern absent from a file whose literal first entry was that pattern.
• An empty <final_answer> block with exit 0 is a failed run, not a clean "nothing found". It happens when citation validation drops every citation (e.g. the model cited paths with a bogus prefix) or the model answered in prose only. Re-ask with a rephrased or narrower question, or drop --citation to see the prose; treat "nothing found" as unproven until a direct grep agrees.
• The answer holds at most ~25 citations (a safety cap); ask a narrower question if you need more.
• A nonzero exit code means the run failed (e.g. the endpoint was unreachable); the error is on stderr and stdout stays empty. Exit 0 with a non-empty <final_answer> block is a good run — retry or fall back to manual exploration on a nonzero exit or an empty block.
• Ask specific questions; run several queries for a multi-part investigation.
• Cap concurrency at 2. Run at most 2 fastcontext sessions at once; queue the rest.