Setup Gitignore

Setup gitignore

Contract

Inputs

No required user input. Operate on the current Git repository.

Optional input is a set of empirical ignore patterns explicitly confirmed by the user after reviewing current untracked-path clusters. Never infer confirmation and never add an empirical pattern silently.

Procedure

1. Locate the repository. Run git rev-parse --show-toplevel, change the working directory used by this procedure to that exact root, and stop with not-a-repository if it fails. Do not inspect global Git ignore configuration. Done when: the repository root is confirmed or not-a-repository is reported.

2. Detect template keys at depth two. Enumerate regular files at repository-relative depth 0, 1, or 2, excluding .git/. Apply every matching row, deduplicate keys, and sort keys lexicographically before joining them with commas: Done when: the sorted key list is produced (possibly empty).

   Evidence filename	gitignore.io key(s)
   tsconfig.json, tsconfig.*.json	node,typescript
   package.json	node
   pyproject.toml, requirements.txt, setup.py, setup.cfg, Pipfile	python
   Cargo.toml	rust
   go.mod	go
   dune-project, dune, *.opam	ocaml
   pom.xml, build.gradle, build.gradle.kts	java
   settings.gradle.kts with Kotlin source files	java,kotlin
   CMakeLists.txt	cmake,c++
   Gemfile, *.gemspec	ruby
   composer.json	php

   A filename may contribute several keys; an empty result is valid and means no network template is requested.

3. Capture current state. Count and retain the exact list from git status --short --untracked-files=all entries beginning ?? , excluding .gitignore. If .gitignore exists, read its exact bytes and copy it to /tmp/gitignore-snapshot-<UTC-YYYYMMDDTHHMMSSZ>.bak before modifying it. If snapshot creation fails, stop without writing. Done when: the untracked list is captured and the snapshot exists or snapshot-failed is reported.

4. Confirm empirical patterns. Group untracked paths first by top-level directory; group root files by lowercase extension, with extensionless root files as (no extension). Show every path in each group. Ask the user which exact directory (name/) or extension (*.ext) patterns to add. Normalize confirmed patterns to repository-relative slash form, reject .., absolute paths, and patterns matching .gitignore, then sort and deduplicate them. Declining empirical additions means an empty empirical block, not cancellation of the language and local-baseline composition. Done when: empirical patterns are confirmed, normalized, and deduplicated.

5. Fetch the language block. If the sorted key list is non-empty, run scripts/compose-gitignore.sh <comma-separated-keys>. Use its stdout verbatim after normalizing line endings to LF and removing trailing blank lines. If the script fails, stop with template-fetch-failed and leave .gitignore unchanged; there is no silent fallback. If the key list is empty, the language block is empty and the script is not run. Done when: the language block is fetched and normalized or template-fetch-failed is reported.

6. Use these exact local baselines. They are inline data, not support-file placeholders:

   # === AI TOOLING ===
   .claude/settings.local.json
   .cursor/
   .windsurf/
   .aider.chat.history.md
   .aider.input.history
   
   # === IDE / EDITOR ===
   .idea/
   .vscode/*
   !.vscode/extensions.json
   !.vscode/settings.json
   !.vscode/tasks.json
   !.vscode/launch.json
   *.swp
   *.swo
   *~
   .DS_Store
   Thumbs.db
   

   Done when: the exact local baselines are available unchanged for section assembly.

7. Build managed sections. Produce these four anchors in this exact order, each followed by its normalized body and one blank line except the final block:

   # === LANGUAGE TEMPLATES ===
   <successful API body, or empty>
   
   # === AI TOOLING ===
   <the five AI patterns above>
   
   # === IDE / EDITOR ===
   <the ten IDE/editor patterns above>
   
   # === EMPIRICAL ===
   <confirmed patterns, one per line>
   

   Treat a non-comment, non-blank pattern line as a duplicate when its exact trimmed text has already appeared earlier in unmanaged user content or an earlier managed section; first occurrence wins. Preserve comment lines supplied by the API. Do not introduce angle-bracket text into the actual file. Done when: the four managed sections are built in order with duplicates removed.

8. Merge deterministically. If no .gitignore exists, the candidate is the four managed sections. If one exists, require each managed anchor to occur zero or one time. Duplicate or out-of-order managed anchors are invalid-managed-sections. Preserve all bytes before the first managed anchor and after/between managed regions that are not part of a managed block. Replace each existing managed block from its anchor through the line before the next managed anchor; append missing managed blocks in canonical order. Normalize only generated managed blocks to LF; do not rewrite preserved user content. Ensure exactly one terminal newline. Done when: the candidate is merged with preserved user content and exactly one terminal newline, or invalid-managed-sections is reported.

9. Obtain write approval. Show the complete before/candidate diff using any available local diff renderer; do not require difft. If the user declines, leave .gitignore unchanged. If approved, write the candidate to .gitignore atomically in the repository root. Done when: the user approves and the file is written, or the user declines and the file is unchanged.

10. Prove idempotence and report. Run steps 2, 5, 7, and 8 in memory against the newly written bytes with the same confirmed empirical inputs. Require the second candidate to be byte-identical; otherwise restore the snapshot (or remove a newly created .gitignore) and report idempotence-failed. Recount untracked entries with the same command as step 3 and report both counts plus every remaining untracked path. Done when: the second candidate is byte-identical and before/after untracked counts are reported, or idempotence-failed is reported.

Failure and recovery

Output

The repository-root .gitignore, created or updated with LANGUAGE TEMPLATES, AI TOOLING, IDE / EDITOR, and EMPIRICAL anchors. The response reports before/after untracked counts, remaining paths, and the snapshot path when an existing file was merged.