Recipes

Everyday work starts with ordinary requests.

“Fix it”#

Name the broken behavior. The agent proposes a mode, quick for a settled one-commit fix, and waits for your yes; "fix it in quick" chooses the mode yourself. The change gets one record, and quick runs frame, plan, build, verify and close in one line each, with no review; the record's review section says none at quick. Name regular or careful when a fix deserves a fresh reviewer. A hidden consequential choice becomes a concrete question, and a known failure starts with diagnose, which builds a reproducer before the mode is chosen.

“Stop here” and “carry on”#

Say stop at any point. The agent sets the change's record to stopped and rewrites its "next" line: the step reached and the action that resumes it. A request for another change stops the open one the same way before it is proposed. Later, in the same session or a fresh one, greenline status lists each open and stopped change with its "next" line; ask "what were we doing?" or "carry on with that change", and the agent resumes at that step on the change's branch, under the record's mode, with no new proposal.

“Prove it”#

Ask “Did this fix it?” or “Prove it is faster.” Verify needs a falsifiable claim, an appropriate baseline, and a check that could fail, run through greenline check so its facts are recorded with the change. A performance claim needs a relevant workload and a measured result; a missing baseline or noise makes the result inconclusive, with its limit named. What a result relies on is copied into the change's folder, .greenline/ledger/<id>/, so it can be checked again.

“What should we build next?”#

The agent can explore ideas against current product evidence and recorded decisions. Discussion alone creates no change and no record. Choosing an idea hands it to a mode: the agent proposes one for the idea's first change and waits.

“Find out which design works”#

Frame answers such a question by work rather than by asking. A bounded prototype answers it through execution, in the change's working files or on a throwaway branch, never in the shipping code. Research reads primary sources and records what remains uncertain. Both return their answer to the change's frame section. A read-only request keeps repository bytes unchanged.

Changing how the agent works#

State your preference directly. It applies within your request; you do not need to repeat it merely because a skill suggests a different default. A substantive conflict should be explained with its consequence.

Two settings in .greenline/manifest.json change the loop itself, and greenline inspect edits both. proceed lets the agent go on without waiting: at quick it goes on for quick changes and still waits for the others, at regular for quick and regular ones; at none, the default, it waits for every mode. trials enables a trial skill, a start or a habit not yet a default, such as idea-refine, which otherwise does not install. Run greenline sync after changing trials.

Persistent engineering choices belong in .greenline/DECISIONS.md with scope and reasons. greenline status lists dash bullets under ## House rulings outside AGENTS.md's managed markers. It does not interpret other prose as structured rulings; those instructions still govern the agent. The installed WORK.md shows the optional stanza and where to write it. Longer governing documents can be referenced from the decisions book.

Knowledge and settled choices#

The agent works within your decisions. Advice it reads, from garden when the connector is enabled, does not silently replace a framework or change an exclusion. It brings back a consequential conflict with evidence. Nothing runs while the agent is idle.

An inspector save leaves the latest change's hashes and time for the next invocation. The agent checks relevant consequences while preserving your intent. Direct edits remain visible through the files and Git history.

Reading another option#

Ask “Compare Go options for this service without changing files.” The agent answers from the installed methods and the repository's evidence. With the garden connector enabled it may also consult garden, when the comparison is an open choice, and it names what it consulted in the reply. Reading another language does not add it to your repository's choices.

Editing work by hand#

Change records and plan documents are yours, and you can edit them. A record is checked against its mode whenever greenline status reads it: it reports a change record or plan file that does not parse, an open or stopped record that does not match its mode, and a cited file that is missing; see the reference. A done record is judged while its latest check ran under the current installation; after an update, or with no check at all, it is history and is read for its form alone.