tanuki-slice

Chunk GitLab merge request discussion threads into deterministic, budget-aware batches for LLM and agentic workflows.

Reviewing a 400-comment MR with an LLM blows past every context window. tanuki-slice scrapes the MR's discussions via the GitLab API, groups comments and replies by file, and splits them into token-budgeted chunks that preserve file locality and thread context — so each chunk fits your model and can be processed in isolation.

Features

• Stdlib-only HTTP client (no requests / httpx dependency).
• Greedy bin-packing that keeps same-file threads together.
• Oversized threads get their own chunk instead of being silently dropped.
• Resolved threads filtered out by default; opt in with --include-resolved.
• Deterministic ordering (file path, then discussion id) for stable diffs between runs.
• Strict typing, py.typed shipped for downstream consumers.

Installation

uv sync

Or from PyPI once published:

pip install tanuki-slice

Usage

chunk — split MR discussions

# Set your GitLab token
export GITLAB_TOKEN="glpat-xxx"

# Chunk an MR's discussions into JSON
tanuki-slice chunk --project-id 123 --mr-iid 45

# Custom token budget and output file
tanuki-slice chunk --project-id 123 --mr-iid 45 --budget 8000 -o chunks.json

# Quick summary view
tanuki-slice chunk --project-id 123 --mr-iid 45 --summary

# Include resolved threads
tanuki-slice chunk --project-id 123 --mr-iid 45 --include-resolved

# Self-hosted GitLab
tanuki-slice chunk --project-id 123 --mr-iid 45 --gitlab-url https://gitlab.example.com

review — LLM code review

# Required env
export GITLAB_TOKEN="glpat-xxx"
export ANTHROPIC_API_KEY="sk-ant-xxx"

# Review a MR (post inline + summary comments)
tanuki-slice review --project-id 123 --mr-iid 45

# Dry-run: print findings only
tanuki-slice review --project-id 123 --mr-iid 45 --dry-run

# Multi-focus + raise cap
tanuki-slice review --project-id 123 --mr-iid 45 --focus security --focus style --max-findings 20

# CI-friendly: skip confirm
tanuki-slice review --project-id 123 --mr-iid 45 --yes

Re-runs are safe: every posted comment embeds a fingerprint marker, and the next run reads existing notes to skip already-raised findings. Project-level defaults can be placed in a tanuki.toml file with a [review] table; CLI flags override it.

Library

from tanuki_slice import GitLabMRChunker

chunker = GitLabMRChunker(token="glpat-xxx")
chunks = chunker.scrape_and_chunk(project_id=123, mr_iid=45, token_budget=4000)

for chunk in chunks:
    print(chunk.to_dict())

Output shape

Each chunk serializes to:

{
  "chunk_id": "1/3",
  "mr_metadata": { "project_id": 123, "mr_iid": 45, "title": "...", "...": "..." },
  "files": [
    {
      "path": "src/foo.py",
      "threads": [
        {
          "discussion_id": "abc",
          "line": 42,
          "resolved": false,
          "notes": [{ "id": 1, "author": "alice", "body": "...", "created_at": "..." }]
        }
      ]
    }
  ],
  "stats": { "thread_count": 7, "estimated_tokens": 3820 }
}

Token counts are rough (~4 chars/token) — good enough for budgeting, not for billing.

Configuration

Environment Variable	Description	Default
GITLAB_TOKEN	Personal access token with api scope	required
GITLAB_URL	GitLab instance URL	https://gitlab.com
ANTHROPIC_API_KEY	Anthropic API key (required for review)	required for review

tanuki.toml in the project root can set review defaults; keys under the [review] table map 1:1 to CLI flags (focus, model, max_findings, max_diff_tokens, gitlab_url).

Status

The chunk and review commands are available. Future work: webhook daemon, multi-provider LLMs, and cross-chunk review for mega-MRs.

Development

uv sync
uv run pytest
uv run ruff check src/ tests/
uv run mypy src/

License

MIT — see LICENSE.