@stringsync/splates 0.0.55 → 0.0.57

package/README.md

@@ -47,7 +47,8 @@ removes the worktree. A conflict or a failing check goes back to the agent that
 made the change, which fixes it and merges again. A check that writes a JUnit
 XML report to `$SPLATES_TEST_REPORT` (`bun test --reporter=junit
 --reporter-outfile="$SPLATES_TEST_REPORT"`) queues a task for each test slower
-than 5 seconds.
+than 5 seconds. A check that appends `{"phase": "<name>", "ms": <n>}` lines to
+`$SPLATES_GATE_PHASES` has those phase times kept in the merge's log entry.
 
 **Project notes.** Keep durable conventions, decisions and other project context
 as short Markdown files in `notes/`, one thing per file, edited in place.
@@ -59,11 +60,11 @@ in `splates.config.ts` to move the directory, relative to the checkout root
 or to an absolute path. Notes change through tasks, just as code does.
 Keep sensitive information out of them.
 
-**Heavy jobs.** Hooks (`postCheckout`, `postRewrite`, `preMerge`) run in one
-of the machine's heavy-job slots, two by default (`SPLATES_HEAVY_JOBS` sets
-how many), so several agents testing at once queue rather than run the
-machine out of memory. `splates run -- <command>` runs anything else as
-heavy, such as a test suite an agent runs mid-task, in a slot too.
+**One premerge at a time.** Each merge's gate (`postRewrite` and
+`preMerge`) holds the machine's premerge lock, so merges from several agents
+test one after another rather than run the machine out of memory. A gate
+that runs past 3 minutes is killed, and the agent merges again. A `preLock`
+command, if set, runs before the wait, so checks that need no lock fail first.
 
 **Pushing.** With `remote: "origin"` in `splates.config.ts`, each task's
 branch goes to that remote as `splates/<task>`: after every commit an agent
@@ -251,58 +252,12 @@ traffic, and `--tunnel-name` is refused beside a provider of your own. One
 that is not ready, or fails to start, is logged and skipped: the page keeps
 serving on this machine.
 
-### Worker providers
+### Local worker
 
-To have splates make workers when agents queue for want of a slot, give
-`splates.config.ts` a `workerProvider`, in the repo `splates server` starts in,
-any object that implements `WorkerProvider`:
-
-```ts
-import type { ConfigFile } from "@stringsync/splates-server/config";
-import type { WorkerProvider } from "@stringsync/splates-server/workers";
-
-const myCloud: WorkerProvider = {
-	// the machines it makes, with what each costs you per hour
-	types: async () => [
-		{ name: "small", platform: "linux", slots: 4, dollarsPerHour: 0.05 },
-	],
-	// build an image by running recipe.script as root on a fresh machine
-	snapshot: async (recipe) => ({ id: "img-1", version: recipe.version, builtAt: Date.now() }),
-	// boot spec.snapshot and run spec.boot on it, as the `splates` user
-	create: async (spec) => ({ id: spec.id, type: spec.type }),
-	destroy: async (id) => {},
-	// the machines it made that still exist
-	list: async () => [],
-};
-
-export default { workerProvider: myCloud } satisfies ConfigFile;
-```
-
-A made worker boots from a snapshot splates asks for when there is none, and
-again when splates's version changes or the snapshot is a week old; it holds
-bun, git, the browser and splates, and no credentials. Its boot command spends
-a single-use join token, dialing the page at its tunnel host, so the server
-needs a tunnel. Placement makes one only once agents other agents started are
-queued, and only as far as the worker budget allows; none is kept warm.
-`splates worker add --provider [--type <name>]` makes one by hand, past the
-rate cap, and `splates worker remove` destroys its machine.
-
-The worker budget is in dollars, at each type's list price, so every figure is
-an estimate: a rate cap on what the made workers running at once cost an hour
-($1/h by default), and limits on what they cost in a day ($15) and a month
-($200), from the start of each in local time. Past the rate cap no worker is
-made, and the queue says which limit holds it, as `rate cap $1.00/h reached`.
-Once a day's or month's limit is spent, made workers take nothing new, their
-agents are asked to stop at a checkpoint, and each is destroyed once its last
-agent ends. Each limit warns at 80%, 90% and 100%, once in its window, by Web
-Push and on the page's header. Agents' own API spend is not counted here.
-
-```bash
-bunx @stringsync/splates budget show
-bunx @stringsync/splates budget set day 10        # or `default`
-bunx @stringsync/splates budget raise hour 0.1 --for 1d
-bunx @stringsync/splates budget end 1             # a raise, by its number
-```
+This machine's own worker runs agents over a private Unix socket and
+reconnects when the server restarts. Remote workers cannot join.
+It runs as many agents at once as are started. Use `splates worker list` to
+inspect it, and `splates worker drain local --undo` to resume it.
 
 ### Task ports