Repomix Analysis

Repomix Code Analysis

Use Repomix to pack a repository into a single AI-friendly file (usually XML), then ask the user to run the analysis in a large-context model (Gemini) and paste back the result.

Default behavior: broad strokes first. Do a quick whole-repo measurement pass, then re-pack with obvious non-code/unrelated noise excluded. If that filtered whole-repo pack fits within ~1,000,000 tokens, do not over-optimize—hand it off to Gemini.

Non-negotiables

• Treat Repomix output files as read-only artifacts. Make changes only in the original repo files.
• Keep Repomix security checks enabled by default. Use --no-security-check only if the user explicitly requests it and confirms the output is safe to share externally.
• If packing a remote repo is required but network/git access is blocked by the sandbox, ask the user to run outside the sandbox.
• Do not open/search the Repomix output (XML) before Gemini handoff. Decide what to exclude using only token stats + folder/filename paths.
• Exception: only do a targeted search when the user explicitly asks to find a specific known identifier.

Workflow

1. Define the analysis question.
2. First pass (measurement): generate a whole-repo pack (default to XML) and measure tokens (--top-files-len, --token-count-tree). Treat this pack as measurement-only; you usually won’t upload it to Gemini. Delete it after you’ve captured token stats (keep only final Gemini pack(s)).
3. Define an initial “obvious noise” filter (file types + folders) from the token stats (no content inspection; use only paths).
4. Second pass (filtered whole repo): re-pack the whole repo with that initial filter applied. If the filtered pack is less than 1,000,000 tokens, stop here and handoff to Gemini (upload the filtered pack). Do not spend time “curating” code files (unless there are clearly some auxiliary non-code artifacts that must be excluded anyway).
5. If still over budget, widen exclusions in this order (re-pack and re-check after each): other non-essential → tests → examples → drop either docs or code based on the question.
6. If still over budget after that, reduce scope at folder granularity based on the question + token tree (avoid file-by-file selection). Prefer docs for usage questions; prefer code for implementation questions. Use --include-full-directory-structure so Gemini still sees the full tree.
7. If needed, create a small number of semantic packs by subsystem/folder (e.g. 2–5 packs), not a pile of single-file packs (user will run the prompt once per pack, sequentially).
8. Handoff pack(s) to Gemini. If Gemini’s answer is vague/weird but points at relevant areas, do a second pass: re-pack focusing on those folders and ask again with a narrower question.

Command templates (copy/paste)

Prefer the system-installed repomix binary when available:

bashCopy code
repomix --version

If repomix is not on PATH, replace repomix below with pnpm dlx repomix@latest (or npx repomix@latest).

Local repo (first pass / measurement) → XML pack + token stats

bashCopy code
repomix source_directory --style xml -o repomix-output.your-file-name.xml --top-files-len 100 --token-count-tree

Local repo (second pass / filtered whole repo) → XML pack + token stats

Start with a small “obvious noise” ignore list and adjust it based on the first pass token stats.

bashCopy code
repomix source_directory --style xml -o repomix-output.filtered.xml --top-files-len 100 --token-count-tree \
  --ignore "**/dist/**,**/build/**,**/target/**,**/coverage/**,**/node_modules/**,**/*.svg,**/*.png,**/*.jpg,**/*.jpeg,**/*.gif,**/*.pdf,**/*.zip" \
  --remove-empty-lines --truncate-base64

Focused pack (semantic split) with full tree context

bashCopy code
repomix source_directory --style xml -o engine.xml --top-files-len 100 --token-count-tree \
  --include "src/engine/**,src/shared/**,README.md" \
  --ignore "**/*.svg,**/*.png,**/*.jpg,**/*.min.js,**/*.map,**/dist/**,**/build/**,**/node_modules/**" \
  --include-full-directory-structure

Reduce tokens without losing code (much) semantics

bashCopy code
repomix source_directory --style xml -o repomix-output.your-file-name.xml --remove-empty-lines --truncate-base64

Architecture-first (structure/signatures over bodies)

bashCopy code
repomix source_directory --style xml -o repomix-output.your-file-name.xml --compress

Split output by size (for tool file-size limits)

bashCopy code
repomix source_directory --style xml -o repomix-output.your-file-name.xml --split-output 1mb

Note: --split-output groups by top-level directory; a single file/directory will never be split across multiple output files. Prefer semantic splitting for analysis quality.

Remote repository pack

bashCopy code
repomix --remote user/repo --remote-branch main --style xml -o repomix-output.your-file-name.xml --top-files-len 100 --token-count-tree

Add git context (only when it helps the question)

bashCopy code
repomix source_directory --style xml -o repomix-output.your-file-name.xml --include-diffs --include-logs --include-logs-count 25

Scope / filtering heuristics (practical)

• Default: pack the whole repo first to get token stats. Apply an initial “obvious noise” filter (non-code/unrelated folders + file types), then re-pack. If still over budget, widen exclusions (other non-essential → tests → examples → drop either docs or code based on the question). Only then exclude code/docs by relevance (prefer folders/modules).
• Keep: entrypoints, core source (src/), configuration, schemas/migrations, key docs (README, SPEC, ADRs), and tests that define behavior.
• Exclude first: build outputs (dist/, build/, target/), caches, coverage, vendored deps, large assets, huge fixtures/dumps, generated code blobs, and large repo-meta (e.g. changelogs/release notes) unless directly relevant.
• Use .repomixignore for repeated iterations; use --ignore for one-offs.
• Use --include-full-directory-structure when you pack only a subset but still want the model to see the full repo tree.
• Prefer folder-level --include/--ignore patterns (and a few key docs/config files) over curating individual files.

Include vs ignore (priority rules)

• --include is a positive filter (pick candidates). --ignore is a negative filter (remove candidates). If a file matches both, it is excluded (ignore wins).
• Ignore sources apply in this order (highest priority first): custom ignore patterns (--ignore / ignore.customPatterns) → ignore files (.repomixignore, .ignore, .gitignore, .git/info/exclude) → default patterns (e.g. node_modules/**, dist/**). If --include “doesn’t work”, it’s usually because one of these ignore sources is still excluding the file.
• To include files that are being excluded by default ignore sources, disable the relevant layer: --no-gitignore, --no-dot-ignore, and/or --no-default-patterns.
• --include-full-directory-structure only affects the Directory Structure section; it does not override ignore filtering.

Gemini handoff (what to ask the user to do)

When the pack(s) are ready, ask the user to upload the file(s) to Gemini and run a prompt like this:

textCopy code
You are analyzing a repository packed by Repomix (attached).
Use the attached Repomix file(s): <filenames>.

Task:
<the exact question and objective>

Constraints:
- Cite file paths for every non-trivial claim.
- When reasoning about behavior, trace cross-file call paths and data flow.
- Avoid speculation. If information is missing or ambiguous, state what’s missing and what additional pack/file would resolve it.

Output format:
- Summary (3–7 bullets)
- Key files/modules (path → responsibility)
- Detailed analysis
- Actionable next steps / follow-up questions (only if needed)

If you created multiple semantic packs (e.g., engine.xml then plugins.xml), provide one prompt per pack, specify the order, so user can carry forward a short summary between runs.