# Track 2: Auto, hand delivery to the agent ## Who it is for and what you will have done For someone who has seen the stages once (track 1) and now wants to hand the work to the agent, watch it move through the gates and reports on its own, and learn how to check in on it and step back in later. After this track, one small intent will have been delivered by the agent end to end, and pausing and resuming that delivery will feel familiar. ## Before you start Run `/plastic-update` first (`$plastic-update` on Codex), so the commands below match what is actually installed. Work in a sandbox: a throwaway git repository, or a global-store intent. Nothing in this track touches a real project. ## Stations ### 1. Board a small intent and choose auto Start from an active intent (create one first with `/plastic-intent-creating` if none exists, the same way as track 1 station 1). Type `/plastic-auto`. Artifact: the delivery lock arms, and the agent announces it is taking over the intent for autonomous delivery. Checkpoint: name the one precondition auto needs before it will start: an active intent must already exist for the intent you name. (Say "auto" with nothing named, and it can pick a queued intent from the dashboard's queue itself.) ### 2. What auto does, and what stays with the user No new command at this station. Watch how the work splits. Auto owns How (the plan, the checklist, the action files) and Exec (the code, the tests, the mechanical close) from here on. Inside Exec it follows a few fixed habits: it syncs its working copy with the main line before touching anything, ticks each task the moment it lands rather than batching several into one later edit, and independently verifies its own work (running the test suite, or checking the changed file) before presenting anything back to you. For a task shaped like an audit or a sweep, checking many files rather than building one artifact, it also drops a short methods report into `resources/` before the gate, so you can review how it checked, not just what it found. The user keeps two things: the rulings made along the way, and the review points, moments auto is built to pause for, such as confirming a project path or stopping before a destructive action with no safe way back. When auto tells you to run a command yourself ("run /plastic-intent-speccing"), that is an instruction for you to type; it is a different thing from the prompts auto hands to its own dispatched subagents, and the two are never mixed up in what it tells you. Checkpoint: name one thing auto will always stop and ask about, rather than decide alone. ### 3. Walking the gates No new command. Auto still honors every hard gate a guided session would hit: the delivery lock (one owner at a time), the code gate (shut until `plan.md` and `checklist.md` exist), the create gate (intents only come from the tool that makes them, never hand-authored), the links gate (a `## Links` section changes only to match the file's own frontmatter), and the bash gate (a shell command cannot write project code the code gate would have blocked). Checkpoint: read one gate message from the run so far and name which of the five gates it came from. ### 4. Reading the per-stage reports No new command. At each stage boundary (What, Why, How, Exec, Done) the agent briefs in a fixed three-line shape: State (what happened and why it matters), Risk (the one thing that could bite, or "nothing flagged"), and Call (the decision left to the user, or the call the agent is taking on its own). That is the depth for a medium or large intent. A small intent gets one briefing, at How, folding in what the earlier stages would have said. Checkpoint: in the most recent report, point at the State line, the Risk line, and the Call line. ### 5. Continue and where-was-I after time away Type `/plastic-continuing`. Artifact: the current state, presented and then the session stops. If a specific intent is named, the agent reads its stage and savepoint and resumes exactly there, rather than starting over. Checkpoint: after stepping away and running this command, name the stage the intent resumed at and how that matched what was actually on disk. ## Wrap and where to go next Auto keeps the same stages and the same gates as guided; the only difference is who steers. Read [`pick-your-mode.md`](https://github.com/zalom/plastic/blob/main/docs/guides/pick-your-mode.md) for the honest trade-off between guided and auto, and [`using-plastic-with-claude-code.md`](https://github.com/zalom/plastic/blob/main/docs/guides/using-plastic-with-claude-code.md) for how that choice feels day to day and how it connects to roadmap-driven delivery. For denial messages met along the way, read [`what-the-gates-are-telling-you.md`](https://github.com/zalom/plastic/blob/main/docs/guides/what-the-gates-are-telling-you.md). Note on terms: "guided" means the user starts each stage with a command and the agent narrows the thinking inside it, the same shape track 1 walked. "Manual", editing project files outside the lifecycle entirely, is not a third mode; it is the anti-pattern both guided and auto exist to avoid.