The Chewbacca runtime: plan, execute, verify, log

For anything longer than a couple of steps, do not freehand a sequence of commands. Compile the request into one formal plan, check it, run it, and read the trace.

The reason is arithmetic, not style. Twenty improvised steps at 95% reliability each is a 36% success rate (see docs/BENCHMARKS.md). A checked plan that stops at the first failure and gates every irreversible action does not compound errors that way. This is the Genie / executable-semantic-parser idea (Campagna, Xu, Lam; Liang): compile intent into a typed, checkable representation that runs or errors cleanly, rather than letting the model improvise.

The grammar

data/grammar.json defines the command space: stream => query* => action*.

• stream - when it runs: now, every, on_text, on_file, on_calendar
• query - reads, no side effects: texts, screen, web, app_data, file
• action - side effects: click, type, web_click, web_fill, remind, calendar, send_text, send_email, run_shell

Every parameter is typed. Every irreversible action carries confirm: true in its signature, so the gate is part of the type and cannot be forgotten.

Write a plan

{
  "stream": { "name": "now" },
  "query":  [{ "name": "texts", "days": 2, "unanswered": true, "direct": true }],
  "action": [{ "name": "remind", "title": "Reply to Chen about Saturday" }]
}

Run it

chewie plan check plan.json     # type-check only, runs nothing
chewie plan run   plan.json     # execute; confirm-gated actions refuse without --yes
chewie plan run   plan.json --yes

check catches a missing or mistyped parameter before anything touches the machine. run executes queries then actions in order, stops at the first failure, and writes a trace.

Read what happened

chewie log            # the last run, step by step, with timing
chewie log --runs     # recent runs
chewie log --json     # the raw trace

Every step is logged with its status and duration to ~/.chewie/runs/<ts>.jsonl. When a task half-worked, this is how you find the step that failed instead of guessing.

When NOT to use a plan

A single read (chewie see, chewie texts, chewie brief) or a single obvious action does not need the ceremony. The plan runtime earns its keep on multi-step tasks, on anything irreversible, and on anything the user will want to audit afterward.

The end-to-end pattern

The headline example - research leads, draft outreach, stage, confirm - is one plan:

now
  => web (query LinkedIn / a directory for companies under 50 employees)
  => web (extract contact emails)
  => send_text (text the user the list)                    [confirm]
  => send_email (staged drafts, one per lead)              [confirm]

The two outbound steps are confirm: true, so Chewbacca compiles and stages the whole thing, then stops and shows the user exactly what it will send before a single email leaves. That gate is in the grammar, not bolted on, which is why it holds even under "just send them, I trust you."