Build channel detection returns empty under jj (detached HEAD), silently selecting the wrong profile database

[shuv1337]

Summary

Script.channel falls back to git branch --show-current. In a jj-colocated checkout the working copy is a detached HEAD, so git prints nothing, the channel is baked in as the empty string, and the resulting binary silently uses a different profile database and service file than the one the branch's builds normally use.

https://github.com/Latitudes-Dev/shuvcode/blob/integration-v2/packages/script/src/index.ts#L27-L32

const CHANNEL = await (async () => {
  if (env.OPENCODE_CHANNEL) return env.OPENCODE_CHANNEL
  if (env.OPENCODE_BUMP) return "latest"
  if (env.OPENCODE_VERSION && !env.OPENCODE_VERSION.startsWith("0.0.0-")) return "latest"
  return await $`git branch --show-current`.text().then((x) => x.trim())   // <-- empty under jj
})()

How it presents (this cost an afternoon)

Built integration-v2 from a jj-colocated checkout and installed it. The service then failed to boot:

database migration started  migration=20260804233008_loose_psylocke
background service boot failed  cause="Cause([Die(SQLiteError: index session_pending_session_delivery_seq_idx already exists)])"

The obvious reading — "my profile database is corrupt, the migration ledger and schema disagree" — is wrong, and it sends you straight into ledger forensics on a 1 GB database. What actually happened:

• channel "" → the binary resolved to opencode-.db, a stale 55-migration profile last written 2026-08-12, instead of opencode-integration-v2.db (69 migrations, healthy).
• That stale database happens to have session_pending_session_delivery_seq_idx but not the loose_psylocke ledger row, so the migration re-ran and collided.
• The intended database was fine the whole time. Booting the same binary against it with OPENCODE_DB set: HTTP 200, 547 models, one migration applied (20260824080928_session_dynamic_tools, 4 ms), clean log.

Same root cause for the service file: channel "" writes/reads service-.json rather than service-integration-v2.json, so port and password discovery also diverge from the running deployment.

Workaround: OPENCODE_CHANNEL=integration-v2 bun run --cwd packages/cli build --single.

Why it is worth fixing rather than documenting

The failure is silent and misdirecting. Nothing warns that the channel is empty; you get a database-corruption error message for what is actually a build-metadata problem, against a database that is not the one you were looking at.

Suggested fix

1. When git branch --show-current is empty, fall back before giving up — e.g. jj log -r @ -T 'bookmarks' --no-graph (or git config jj.* / jj bookmark list -r @), then git rev-parse --abbrev-ref HEAD.
2. If the channel still resolves to empty, fail the build loudly (or warn hard) instead of embedding "". An empty channel is never a deliberate choice, and it silently forks a user's data directory.
3. Independently: 20260804233008_loose_psylocke creates an index unconditionally. CREATE INDEX IF NOT EXISTS would make a mixed-history database survivable rather than fatal.

Environment: Arch Linux, integration-v2 @ 4668dfcd, jj-colocated checkout, bun build --single.