13-git-workflow-commit

13-git-workflow-commit

Fan-out curing workflow: a planner agent inspects all changed files and enqueues one signed commit task per file; executor agents commit each file in parallel via GPG signing. The entire pipeline runs as a single leather workflow run invocation.

stdin (cwd + signing key)
  |
  └─ cli-git-commit-all-in ──► [planner: cli-git-commit-plan]
                                    │  (calls git_enqueue_file_commit N times)
                                    │  (POSTs to LEATHER_INTAKE_URL)
                                    ▼
                            cli-git-commit-file-in ──► [executor × 16: cli-git-commit-file]
                                                            │  (git add + gpg-sign commit)
                                                            ▼
                                                        artifacts/

The planner and executor workers start simultaneously. As the planner enqueues file tasks via the HTTP intake endpoint, executor workers pick them up immediately. workflow run blocks until all queues quiesce, then exits.

Requirements

• leather and shell-mcp binaries on PATH (or built from repo root)
• A local OpenAI-compatible endpoint at $LEATHER_LLM_ENDPOINT
• GPG key on the keyring — set LEATHER_GIT_SIGNING_KEY to the key ID
• curl (used by git_enqueue_file_commit to POST tasks to the intake URL)
• git with GPG commit signing support

Run

From a git repository with uncommitted changes:

export LEATHER_GIT_SIGNING_KEY=<your-gpg-key-id>

echo "Commit all changed files in cwd: $PWD
SIGNING_KEY: $LEATHER_GIT_SIGNING_KEY" | \
  leather workflow run \
    --config /path/to/13-git-workflow-commit/config.yaml \
    --tannery /path/to/13-git-workflow-commit/tannery.yaml \
    --kind cli.git.commit_all \
    --source cli \
    --settle 2s

Or use the make target from examples/:

LEATHER_GIT_SIGNING_KEY=<key-id> make 13

The make target runs against a throwaway scratch repo seeded with sample files so you can try it without touching your real work.

What this shows

• leather workflow run: single command replaces a multi-phase serve loop; blocks until quiescent, then exits with an artifact summary.
• Concurrent fan-out: planner and executor curings run in parallel from the start — executor workers pick up tasks as soon as the planner posts them.
• In-process intake: MCP tool subprocesses POST new hides to the running workflow run via LEATHER_INTAKE_URL instead of writing queue files directly, avoiding cross-process file I/O races.
• GPG-signed commits: each executor commits one file with --gpg-sign, serialized per file via a flock to avoid concurrent index writes.
• shell-mcp integration: all git operations are exposed as MCP tools from a local shell-tools.json tool manifest.

Environment variables

Variable	Default	Description
LEATHER_LLM_ENDPOINT	http://localhost:11434	OpenAI-compatible LLM endpoint
LEATHER_MODEL	llama3	Model name
LEATHER_GIT_SIGNING_KEY	(required)	GPG key ID for commit signing
LEATHER_GIT_DIFF_LINES	12	Lines of diff shown per file to the planner