Start the app and the server from bun run dev, and leave the routines worker to start.sh

[zopeVaibhav]

Fixes #467

What this changes

bun run dev fanned out with --filter '*', which selects all three workspaces, and the third of them is the routines worker. That worker is handed DATABASE_URL, SERVER_INTERNAL_URL and WORKER_SHARED_SECRET by scripts/start.sh and by nothing else, so the copy this command started read none of them and threw at boot on every run. It has never started successfully on any machine — worker/package.json has one commit in its history, and .env.example ships the secret empty by design, so even adding --env-file=../.env would leave a fresh clone throwing.

The command now names its two packages, which is what README.md:339 and docs/development.md:34 already say it does. build and typecheck still use --filter '*', so the worker is compiled and typechecked exactly as before.

Making the worker start instead would have been the worse fix. scripts/start.sh:329 records that its pgrep guard matches bun worker/src/index.ts specifically, because the cd worker && bun src/index.ts form also matches three Dockerfiles. A worker started by bun run dev runs under that second form, so start.sh would not see it and would start a second worker beside it.

One line in the repository does read the other way, and it is worth naming rather than leaving to be
found. CHANGELOG.md:113 says "scripts/start.sh looks for a process it does not start the way
bun run dev does", so somebody was aware a worker process comes out of this command. It is written
as a hazard rather than as an endorsement, and the hazard is hypothetical, because that worker throws
before it can claim anything. Nothing else in the repository asks for a worker here: git log -S
shows the --filter '*' in this script has never been edited since the initial commit, and both
places that describe the command scope it to two processes. If the arrangement was deliberate, say so
and I will close this — the alternative fix, wiring --env-file=../.env into worker/package.json,
is one line, though .env.example shipping the secret empty means a fresh clone would still throw.

Where it runs

• New state that outlives a request? None. The change is a package script.
• What happens on the second replica? Nothing differs. No deployment runs bun run dev; Kubernetes fires routines from the CronJob and a laptop from scripts/start.sh.
• Anything serialised? None added. One fewer process contends for the routine lease locally.
• Anything fanned out to a browser? No.
• New listener, port, or schedule? None. This removes a process that bound no port and never ran.

Boundary and audit

• Unchanged. No acting call, refusal, or audit path is touched.

Changelog

• Added under Unreleased.

Proof

Reproduced on origin/main 1c7bd92. bun run --filter worker dev — the exact command
--filter '*' issues for that workspace:

worker dev: error: WORKER_SHARED_SECRET is not set, so this worker cannot authenticate itself
worker dev:   to /internal/routines/run and no routine could be fired.
worker dev:       at loadWorkerEnv (worker/src/env.ts:32:15)
worker dev:       at worker/src/index.ts:52:3

The new test fails against main's script and passes against this one. Same test file, only
package.json swapped for the one on main:

(fail) OpenBot workspace > dev starts the app and the server, and not the routines worker
  @@ -3,3 +3,3 @@
     "server",
  +   "worker",

End to end. bun run dev on this branch, on spare ports so it did not collide with a running
stack: VITE ready, app on 3910, OpenBot server listening on http://localhost:3901, and zero lines
in the whole run mentioning either the worker or WORKER_SHARED_SECRET.

The worker still starts by its supported path. Launched the way scripts/start.sh:335 launches
it, from the repository root with the three settings it expects:

OpenBot worker status: idle
{"type":"routine-sweep","offered":[],"considered":0,"fired":[],"skipped":[]}
{"type":"routine-sweep-purge","purged":0}

Gates. Typecheck clean across all three workspaces — the worker is still compiled, because
build and typecheck keep --filter '*'. Lint 604 files, format 600 files, no diff.

Suite, run in one session against a same-session control, with the two
agent-handoff-*.integration files excluded because they fail most of their tests on bare main and
make any comparison meaningless:

	pass	skip	fail	tests
origin/main @ 1c7bd92	2609	25	8	2642
this branch	2612	25	8	2645

The failing test names are byte-identical on both sides, as is the @langchain/core/messages
resolution error, so none of them belong to this change. The difference is exactly the three tests
added here.