intentfile and agent task contracts

intentfile started from a practical annoyance: agents are getting better at doing work, but the handoff format is still usually a prompt, an issue comment, or a loose checklist. Those are easy to improvise and hard to audit. If the work matters, the agent needs a more durable contract than a paragraph of vibes.

The format is intentionally small. An .intent.yaml file names the objective, relevant context, constraints, permissions, acceptance criteria, and proof required before the job can be called done. The point is not to create another heavy planning system. The point is to make done explicit enough that a coding agent, a human reviewer, and a CI job can all reason about the same task.

The CLI turns that file into useful operating surfaces: validate the schema, render a brief for Codex or another agent, generate a Codex /goal prompt, create a starter proof file, and verify the proof against the original contract. That makes the workflow portable across Codex, Claude Code, Cursor, GitHub Actions, MCP, A2A, custom runners, and normal human review.

The proof side is the part I care about most. A good agent workflow should not end with a confident summary. It should return changed files, commands run, acceptance status, and any unresolved questions in a shape that can be checked later. intentfile does not pretend to prove product semantics by itself, but it can make missing proof harder to hide.

The launch demo focuses on Codex /goal because the fit is clean: /goal keeps the agent moving toward a durable objective, while intentfile defines what done means for that objective. That pairing turns a vague instruction into a task contract the agent can keep returning to.