WHY
The bet

The scarce resource isn't code generation. It's converged context.

Agents write code faster than anyone can read it — generation is cheap; knowing that what was built is what was asked for is the expensive part. Supervision doesn't scale. shipd takes a different position: capture intent as a specification, gate everything on it, verify against it mechanically. Five commitments follow — design constraints, not aspirations, every one enforced by the pipeline today.

01

Intent is a file in the repository.

Every change starts as three artifacts — plan.md, testable requirement deltas, tasks.md — versioned next to the code they produce. A chat transcript enforces nothing; a checked-in spec is reviewable, diffable, and permanent. If it isn't written down, it doesn't get built.

02

No code before the gate passes.

A deterministic, LLM-free check compares the plan against the actual codebase before implementation begins. Insufficient context parks the plan as rejected, with the findings written into it. The system never builds on a guess — it stops and says what's missing.

03

Nothing merges until someone tried to break it.

An independent validator attacks every scenario in the spec against the real, running code. Then CI must pass, and the semantic review must be explicitly dispositioned. Only then does the PR merge. No diff-skimming, no rubber stamps.

04

Every question is asked exactly once.

Answers are recorded — in the spec, the workspace wiki, the memory store, the oracle's standing positions — and consulted before anything reaches you. A question you answered in March does not come back in August.

05

shipd ships shipd.

175 changes, 36 verified capabilities, 11 epics — every one planned, built, refuted, reviewed, and merged by this pipeline, including an autopilot that delivers approved epics to merged PRs unattended. The delivery board is our own status page, not a demo.

Compile context into specs, and you can ship software you didn't watch being written.

GET STARTED