kochetkov-ma__claude-brewcode/.sembleignore

# brewcode-meta: version=5.2.0 generated_by=brewcode:semble-setup
# brewcode:semble — managed file. Regenerate with
#   semble-guidance.sh install --part ignore --force
# Edit it freely: any change makes it `user_modified`, and the installer then
# leaves it alone (a backup is taken before --force overwrites).
#
# WHY THIS FILE EXISTS
# semble 0.5.4 builds its ignore set in index/file_walker.py:_load_ignore_for_dir
# from exactly two files per directory — ./.gitignore and ./.sembleignore. It
# never reads the user's global excludes file (core.excludesFile / ~/.gitignore*)
# and never asks git. So a directory that is invisible to `git status` only
# because of a GLOBAL ignore rule is still fully indexed and still comes back as
# search evidence. `.claude/` is the common case.
#
# ORDER MATTERS, AND IT WORKS IN OUR FAVOUR
# _load_ignore_for_dir concatenates ./.gitignore lines FIRST and ./.sembleignore
# lines SECOND into one GitIgnoreSpec, and _is_ignored keeps the LAST pattern
# that matched. A rule here therefore overrides a conflicting rule in the
# sibling .gitignore — including a `!` un-ignore. That is the only lever for the
# bypass described next.
#
# THE NEGATION BYPASS (file_walker.py:_is_ignored, the `found` flag)
# A `!` un-ignore pattern whose text ends in a file extension — `!keep.png`,
# `!web/docs/package-lock.json`, `!*.json` — sets `found = True`, and `_walk`
# then yields the file **even though its suffix belongs to no content type**.
# So a .gitignore negation can drag binaries and lockfiles into the corpus that
# `--content` alone can never reach, and no change to the content set removes
# them. Measured on this workspace: one negated `package-lock.json` was 552
# chunks (5.9% of the whole index) and two negated `.png` files added 143 chunks
# of decoded binary garbage. The two blocks below exist to re-ignore exactly
# that class of file.
#
# Only paths that are never project source belong here. Leaving noise indexed is
# cheaper than hiding something you wanted to find.

# --- Claude Code working directories ---------------------------------------
# Scratch, vendored upstream copies, generated reports and machine state.
# NOT excluded, because they are project-authored: .claude/skills/,
# .claude/agents/, .claude/rules/, .claude/commands/, .claude/hooks/,
# .claude/scripts/, .claude/tasks/.
.claude/tmp/
.claude/reports/
.claude/backups/
.claude/logs/
.claude/semble/
.claude/projects/
.claude/history/
.claude/file-history/
.claude/shell-snapshots/
.claude/statsig/
.claude/todos/
.claude/ide/

# --- Build output and caches semble does not skip by default ----------------
# Its built-in list already covers .git .hg .svn __pycache__ node_modules
# .venv venv .tox .mypy_cache .pytest_cache .ruff_cache .cache .semble .next
# dist build .eggs — these are the ones it misses.
target/
coverage/
htmlcov/
.gradle/
.astro/
.turbo/
.parcel-cache/
.nuxt/
.svelte-kit/
.output/
.docusaurus/
.terraform/
.dart_tool/
_site/

# --- Vendored dependency trees ----------------------------------------------
# Conventional names for "someone else's source, copied in". Every one of these
# is upstream code the question is never about.
vendor/
third_party/
bower_components/
.yarn/
Godeps/

# --- Generated bundles ------------------------------------------------------
*.min.js
*.min.css
*.bundle.js
*.map

# --- Binary and non-text assets ---------------------------------------------
# GENERIC AND ZERO-RISK. None of these suffixes maps to a language, so with a
# plain .gitignore these lines are a no-op. They earn their place only against
# the negation bypass above: when a `!logo.png` slips one through, semble reads
# it with errors="replace" and indexes the mojibake. There is no repo in which
# decoded binary is the answer to a question.
*.png
*.jpg
*.jpeg
*.gif
*.bmp
*.tiff
*.webp
*.avif
*.ico
*.icns
*.svgz
*.pdf
*.woff
*.woff2
*.ttf
*.otf
*.eot
*.mp3
*.mp4
*.wav
*.mov
*.webm
*.zip
*.gz
*.tgz
*.bz2
*.xz
*.7z
*.rar
*.jar
*.war
*.class
*.so
*.dylib
*.dll
*.exe
*.bin
*.o
*.a
*.pyc
*.pyo
*.wasm
*.db
*.sqlite
*.sqlite3
*.parquet
*.avro
*.pack
*.idx

# --- Dependency lockfiles ---------------------------------------------------
# Machine-written dependency resolution. `pnpm-lock.yaml` is a .yaml and so is
# in the config bucket outright; the rest reach the corpus only through the
# negation bypass. No lockfile has ever been the answer to a "how does this
# work" question, and one of them was 5.9% of this workspace's index.
package-lock.json
npm-shrinkwrap.json
yarn.lock
pnpm-lock.yaml
bun.lockb
composer.lock
Gemfile.lock
Cargo.lock
poetry.lock
uv.lock
Pipfile.lock
pdm.lock
go.sum
gradle.lockfile
packages.lock.json

# --- Per-repo exclusions ----------------------------------------------------
# NOTHING BELOW THIS LINE SHIPS PRE-FILLED. The rules above hold in any repo;
# the two biggest sources of wasted result slots do not, because they are
# layout-specific and no static pattern can recognise them:
#
#   1. DUPLICATE TREES — the same file committed at two or three paths (a
#      mirror for another agent runtime, a vendored copy of your own plugin, a
#      generated port). Semble has no dedup: N copies means N chances to fill a
#      result slot with the same text. On this workspace three mirrors of one
#      plugin tree were 2202 chunks and took 15 of 80 result slots across 16
#      queries.
#   2. LONG CHANGELOGS — RELEASE-NOTES.md / CHANGELOG.md. Genuinely useful for
#      "when did X land", genuinely ruinous for "how does X work": a 24k-line
#      changelog here was 503 chunks and took 9 of 80 slots. Exclude it only if
#      you do not ask semble history questions.
#
# `semble-guidance.sh install --part ignore` MEASURES this repo and appends what
# it found below, in a delimited "measured candidates" block:
# duplicate trees (byte-identical copies of files that live somewhere else) and
# paths carrying a disproportionate share of the corpus, with exact chunk counts
# when an index exists and byte share before that. Every proposal is written
# COMMENTED OUT and excludes nothing until you uncomment it - a wrong exclusion
# fails silently, so the scan proposes and you decide. Re-running only ever adds
# paths it has never proposed; your edits inside that block survive.
#
# To see the same measurement without installing:
#
#   scripts/semble-project.sh candidates
#
# Then add them here, one per line, with a comment saying why. Root-anchor a
# path that must not match deeper copies of the same name (`/skills/` hits only
# the top-level directory; `skills/` would hit every `*/skills/` in the repo).

# --- brewcode:semble measured candidates ---
# Measured in THIS repo by `semble-project.sh candidates` (exact chunk counts, 3207 files scanned).
# PROPOSALS ONLY - every line below is commented out and excludes nothing.
# Uncomment what you agree with; delete what you do not. A re-run only ever
# adds paths it has never proposed, so your edits here survive.
# /.codex/   duplicate-tree  13.6%  102 of 107 files are byte-identical copies of files under brewtools
# /RELEASE-NOTES.md   heavy-file  5.6%  one file is 6% of the corpus chunks
# --- end brewcode:semble measured candidates ---