The moon's eye over your codebase.
A GitHub-native agent that turns a plain-English issue into a high-quality, test-first pull request — with you in the loop at every gate.
You describe the change in plain English. Tsukinome does the rest — and stops for you at the moments that matter.
Drafts a functional spec with testable acceptance criteria, committed to the repo.
Asks one batched set of questions — only when something genuinely can't be inferred.
Commits an implementation plan and waits. You reply /approve before any code is written.
Implements test-first — red → green → refactor, one commit per task, in an isolated sandbox.
Self-reviews the diff, then opens a pull request with the spec, plan, and tests attached.
Actionable PR review comments trigger a bounded, test-first fix loop — never runaway.
Most code agents are fire-and-forget scripts that dump your whole repo into a context window. Tsukinome is built on three deliberate choices.
Runs on a persistent backend, so a run can suspend and resume across human gates — clarify, approve, review — for days. A stateless Action can't hold that state. You approve the plan before a single line is written.
CocoIndex + pgvector give the agents semantic retrieval of just the relevant code, instead of stuffing the whole repo into the prompt. Model tiering — Haiku, Sonnet, Opus — plus prompt caching keep every run cheap and metered.
No per-repo config files, no workflow YAML, no CI wiring. Install the GitHub App once and open an issue — it just works, on any supported repo.
Every external dependency sits behind an interface with a real implementation and a fake — the whole pipeline is unit-tested without a single key.
No workflow files, no CI wiring. Install the App, open an issue, and stay in control of every gate.
One click from the GitHub Marketplace. Pick the repositories you want Tsukinome to work on — it needs read/write on issues, pull requests, and contents. No files added to your repo.
Describe the change in plain English — a bug, a small feature, a refactor. Tsukinome acknowledges it, then drafts a spec with testable acceptance criteria.
If something genuinely can't be inferred, it asks one batched set of clarifying questions right in the issue thread. Reply, and it refines the spec.
It commits an implementation plan and waits. Nothing gets built until you say so.
A self-reviewed, test-first pull request lands — spec, plan, and tests attached, with a per-run cost summary. Leave review comments for a bounded fix loop, or merge.
An experiment in agentic engineering that keeps humans in command.