Repomix context scoping

<!-- Provenance: re-authored (not copied) from infina-claudekit .claude/skills/repomix/SKILL.md @ 1d662ad (infina-claudekit; MIT-from-upstream yamadashy/repomix, no per-file header). Retargeted for a Multica squad: a generic "pack a repo for an LLM" utility becomes the Architecture reviewer's context-hygiene aid — the one panel seat with the largest window; a plain CLI for whole-repo packing + token budgeting that complements the seat's code-graph tool (code-graph tool does targeted caller-impact, repomix budgets the one-pass read). The value is wired to that seat's existing accuracy guards (the VERIFIED/INFERRED split + the File manifest in its verdict format). Dropped: remote-repo packing, third-party-library security-audit framing, clipboard --copy, the MCP-server mode, multi-LLM token tables, and the references/ companion files (Multica ships only this body). -->

Repomix context scoping

You have the largest usable context in the panel — but "largest" is not "unlimited," and a big window invites the exact failure your accuracy guards warn about: feeling like you read everything when you read part. This skill keeps your single-pass full-repo+diff read honest by sizing the surface before you read it and scoping the pack to what the diff actually touches.

Premise (and when this is inert)

repomix is a plain command-line tool (npx repomix, no global install), not an MCP server — it complements this seat's code-graph tool ({{CODE_GRAPH_TOOL}}): the code-graph tool gives targeted caller-impact, repomix budgets + packs the whole-repo read. It needs shell access. If your seat cannot run shell commands, this skill is inert: read the diff directly and still emit an accurate File manifest, explicitly listing the modules the diff references that you could not load.

Size before you read

npx repomix --token-count-tree [{minTokens}]

Prints a hierarchical token map of the repo (dirs/files with token counts). Omit {minTokens} for the full tree; pass a number like 1000 to hide anything smaller on a large repo. Use it to decide whether the diff's blast radius fits your window, and to spot the token-heavy directories you would otherwise skim and silently under-read.

Scope to the blast radius, not the repo

Pack only the modules the diff touches plus their callers — never the whole monorepo:

npx repomix --include "{globs}" --remove-comments -o {out}.md --style markdown

Scope to the blast radius the {{MODULE_GLOBS}} scan already identified, plus any recurring hot-path flow the retro loop has flagged:

<!-- EARNED:blast-radius-globs --> <!-- ships empty on purpose — the retro/harvest loop fills this from real deliveries --> <!-- /EARNED -->

node_modules/ and dist/ are skipped automatically — repomix both respects .gitignore and applies its own default-ignore patterns (which already exclude node_modules, dist, .git, coverage).

The honesty contract (the real payoff)

repomix's run summary tells you exactly which files and how many tokens you packed. Carry that straight into your verdict's File manifest: read = {what you packed and actually read}; referenced-but-not-loaded = {modules the diff imports that did not fit}. A finding on a module you packed and read is VERIFIED; a finding on a module that didn't fit is INFERRED — never the other way around.

Security

repomix runs Secretlint and respects .gitignore, but never pack .env, secrets, or other credential-shaped files. The packed output is an input to your read that stays local — it is not an artifact you post in a comment or share externally.

Defer

This skill only improves the read and the File manifest. Emit findings and your verdict in your own VERDICT FORMAT (## REVIEW VERDICT … {{{END-REVIEW}}}) — it adds no new finding or verdict format. Strictly read-only: packing source to read it never licenses an edit, push, or merge.