fix(ci): stop mise run ci racing bootstrap, and itself, on a fresh tree

[yashau]

Two ways mise run ci was unreliable. Both are tasks in one run reaching the same files at the same time, so they are fixed the same way — with wait_for, which this repo already uses for exactly this.

1. It raced bootstrap on a fresh tree

depends runs in parallel, so listing bootstrap first in [tasks.ci] did not make it finish first. openapi:check and build:mcp need the pnpm workspace installed and did not declare it, so on a fresh clone or worktree they started before bootstrap:js and failed with:

Error [ERR_MODULE_NOT_FOUND]: Cannot find package '@scalar/hono-api-reference'
  imported from packages/app/src/lib/server/http/app.ts

The gate then aborted the whole DAG seconds in — bootstrap never even appeared in the output, and lint, typecheck, test:js and e2e never ran. So the first mise run ci in any new worktree failed on something unrelated to the change under test, and named a missing npm package rather than a missing bootstrap step.

That also quietly undercut the guarantee AGENTS.md makes about this command being a superset of ci.yml: the real workflow works because it installs dependencies as an explicit earlier step.

Four tasks now declare what they need: openapi, openapi:check, test:mcp, build:mcp.

openapi carries a comment because it is the counterintuitive one — the script imports nothing from node_modules itself, it reaches them by loading the Hono router.

Deliberately left alone: lint:loc, test:scripts, test:action and version:check also shell out to node, but their scripts import only Node builtins and relative paths, so they do not need the workspace. version:check passing during the failing run corroborates that.

2. It raced itself

lint:docs and lint:svelte each re-run a check that typecheck also reaches through pnpm -r. That overlap is intentional and both tasks' comments explain why — this PR keeps it. But it has to be sequenced, not merely tolerated, because both halves write generated state:

• astro check syncs its content layer through a temporary file renamed over packages/docs/node_modules/.astro/data-store.json. Two of them racing that rename leaves the loser reporting an ENOENT from MutableDataStore for a .tmp path the other process already moved — which reads as a broken docs build rather than as two copies of one check. Observed intermittently, and it passes when run on its own, which is what makes it look like a docs problem.

[UnknownFilesystemError]
  Caused by: ENOENT: no such file or directory, rename
    'packages\docs\node_modules\.astro\data-store.json.tmp'
    -> 'packages\docs\node_modules\.astro\data-store.json'

• lint:svelte is ordered for the same structural reason: its first step is svelte-kit sync, which generates into .svelte-kit/. Two copies of one generator writing one directory is a race whether or not it has surfaced. To be explicit: this one has not been observed failing — it is fixed because it is the same shape, not because it broke.

Both now wait_for = ["bootstrap:js", "typecheck"].

Verification

Wiped every node_modules in the tree — root, all five packages/*, e2e, action — and ran the gate from that state, which is the exact condition that used to fail within seconds.

• openapi:check now waits for bootstrap:js and passes.
• Sequencing is visible in the log: [typecheck] Finished in 31.14s at line 2341, then [lint:svelte] $ … and [lint:docs] $ … at 2342-2343. No overlapping astro check.
• Full gate green, all 32 leaves: test:rust 663, miri 127, test:js 1152, e2e 101, plus the whole lint/typecheck/audit/deny set.

Note on wait_for semantics: it orders tasks within a run that schedules both, so it does not turn mise run openapi:check on a never-bootstrapped tree into a self-installing command — mise run bootstrap is still the documented setup step. That is the same contract every existing wait_for in this file has, and it keeps each leaf individually runnable as AGENTS.md promises.

🤖 Generated with Claude Code