test(cli): opt-in pre-built binary for faster subprocess spawns

packages/opencode/script/prebuild-test-cli.ts

@@ -0,0 +1,95 @@
+#!/usr/bin/env bun
+// Build a pre-compiled `opencode` binary for subprocess tests, then expose
+// it at `dist/test-cli/bin/opencode` for the harness to consume.
+//
+// Why: each `bun run --conditions=browser src/index.ts <cmd>` spawn pays
+// ~15s of JIT + plugin init + DB migration in isolation mode. The
+// pre-compiled binary cuts that to ~5s — a 3x improvement on subprocess
+// tests that touch the DB (mcp, providers list, etc.).
+//
+// Usage:
+//   bun script/prebuild-test-cli.ts
+//   export OPENCODE_TEST_CLI_PATH="$PWD/dist/test-cli/bin/opencode"
+//   bun test test/cli/
+//
+// The harness (see test/lib/cli-process.ts) reads OPENCODE_TEST_CLI_PATH; if
+// set, it spawns the binary directly instead of `bun run src/index.ts`. If
+// unset, it falls back to dev mode — so this script is strictly opt-in.
+//
+// Build cost amortizes after ~1 spawn that touches the DB. Recommended for
+// CI, manual `bun test test/cli/` runs, and any local iteration where the
+// CLI surface itself isn't under change. Skip for normal src/* editing — the
+// dev path picks up source changes without rebuild.
+import { $ } from "bun"
+import fs from "node:fs/promises"
+import path from "node:path"
+
+const dir = path.resolve(import.meta.dirname, "..")
+process.chdir(dir)
+
+const platform = process.platform === "win32" ? "win32" : process.platform === "darwin" ? "darwin" : "linux"
+const arch = process.arch === "x64" ? "x64" : process.arch === "arm64" ? "arm64" : "x64"
+const targetDir = path.join(dir, "dist", `opencode-${platform}-${arch}`)
+const binaryName = process.platform === "win32" ? "opencode.exe" : "opencode"
+const builtBinary = path.join(targetDir, "bin", binaryName)
+const stableBinary = path.join(dir, "dist", "test-cli", "bin", binaryName)
+
+const force = process.argv.includes("--force")
+
+// Walk src/ and return the newest mtime seen. Faster than `git status` for
+// the freshness check and works for uncommitted edits. Returns 0 on error
+// so a missing src/ tree forces a rebuild via the comparison below.
+async function newestMtimeMs(root: string): Promise<number> {
+  let max = 0
+  async function walk(p: string) {
+    let entries: { name: string; isDirectory: () => boolean; isFile: () => boolean }[]
+    try {
+      entries = await fs.readdir(p, { withFileTypes: true })
+    } catch {
+      return
+    }
+    for (const entry of entries) {
+      const full = path.join(p, entry.name)
+      if (entry.isDirectory()) {
+        if (entry.name === "node_modules" || entry.name === "dist") continue
+        await walk(full)
+      } else if (entry.isFile()) {
+        const stat = await fs.stat(full).catch(() => null)
+        if (stat && stat.mtimeMs > max) max = stat.mtimeMs
+      }
+    }
+  }
+  await walk(root)
+  return max
+}
+
+async function fresh(): Promise<boolean> {
+  const binStat = await fs.stat(builtBinary).catch(() => null)
+  if (!binStat) return false
+  const srcMs = await newestMtimeMs(path.join(dir, "src"))
+  return binStat.mtimeMs > srcMs
+}
+
+if (!force && (await fresh())) {
+  console.log(`Test CLI binary is up to date: ${builtBinary}`)
+} else {
+  console.log(`Building test CLI binary for ${platform}-${arch}...`)
+  const start = Date.now()
+  await $`bun script/build.ts --single --skip-embed-web-ui --skip-install`
+  console.log(`Build complete in ${Date.now() - start}ms: ${builtBinary}`)
+}
+
+// Verify the binary exists and is executable before symlinking — catches
+// a silently-failed build that left a stale or partial output behind.
+await fs.access(builtBinary, fs.constants.X_OK).catch(() => {
+  throw new Error(`Built binary missing or not executable: ${builtBinary}`)
+})
+
+await fs.mkdir(path.dirname(stableBinary), { recursive: true })
+await fs.rm(stableBinary, { force: true })
+await fs.symlink(builtBinary, stableBinary)
+console.log(`Symlinked stable path: ${stableBinary}`)
+console.log(``)
+console.log(`To use in tests:`)
+console.log(`  export OPENCODE_TEST_CLI_PATH="${stableBinary}"`)
+console.log(`  bun test test/cli/`)

packages/opencode/test/cli/help/help-snapshots.test.ts

@@ -131,7 +131,9 @@ describe("opencode CLI help-text snapshots", () => {
           expect(normalize(result.stderr)).toMatchSnapshot(`opencode ${argv.join(" ")} --help`)
         }
         if (failures.length > 0) {
-          throw new Error(`Help text failed for:\n  ${failures.join("\n  ")}`)
+          // Keep the failure in the Effect channel — symmetric with the
+          // Effect.fail inside the partition above, not a defect.
+          yield* Effect.fail(new Error(`Help text failed for:\n  ${failures.join("\n  ")}`))
         }
       }),
     180_000,

packages/opencode/test/cli/smokes/read-only.test.ts

@@ -0,0 +1,113 @@
+// Tier-A smoke tests for read-only commands. Each test asserts only that the
+// command exits 0 and produces *some* output in the isolated harness env.
+//
+// These are not behavioral tests — they're the cheapest possible signal that
+// the dependency-layer wiring (config load, DB init, server boot, provider
+// resolution) doesn't crash for the broad class of "no inputs, no side
+// effects" commands. A regression in any shared layer (an Effect.fail that
+// propagates out of a service constructor, a renamed env var, a broken DB
+// migration) will fail one or more of these tests.
+//
+// If a future change should make one of these commands intentionally fail in
+// an empty env, update the assertion + add a note explaining the new contract.
+//
+// Speed: each test pays ~1.5s for bun startup. 7 tests serialize within this
+// file. Tracked as task #16 — investigate bun pre-warm if the suite grows.
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { cliIt } from "../../lib/cli-process"
+
+describe("opencode read-only commands (smoke)", () => {
+  // `mcp list` reads MCP server config and pings each one. With the empty
+  // OPENCODE_CONFIG_CONTENT={} we provide, no servers should be configured
+  // and the command should report that cleanly.
+  cliIt.live(
+    "mcp list: exits 0",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["mcp", "list"])
+        opencode.expectExit(r, 0, "mcp list")
+      }),
+    60_000,
+  )
+
+  // `providers list` enumerates credentials + env-resolved providers.
+  // (Not config-injected ones — those don't appear here by design.) We just
+  // assert it produced the two section headers, which proves the resolver
+  // walked both sources without crashing.
+  cliIt.live(
+    "providers list: exits 0 and prints credential/environment sections",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["providers", "list"])
+        opencode.expectExit(r, 0, "providers list")
+        expect(r.stdout).toContain("Credentials")
+        expect(r.stdout).toContain("Environment")
+      }),
+    60_000,
+  )
+
+  // `models` lists models from configured providers. Our test/test-model
+  // should appear because it's wired into the test provider config.
+  cliIt.live(
+    "models: exits 0 and lists the test model",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["models"])
+        opencode.expectExit(r, 0, "models")
+        expect(r.stdout).toContain("test/test-model")
+      }),
+    60_000,
+  )
+
+  // `agent list` walks the agent config. Empty config means no agents
+  // configured; the command should still exit 0 with a "no agents" line or
+  // similar. We don't pin the message — just exit cleanly.
+  cliIt.live(
+    "agent list: exits 0",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["agent", "list"])
+        opencode.expectExit(r, 0, "agent list")
+      }),
+    60_000,
+  )
+
+  // `session list` reads the session DB. Fresh OPENCODE_TEST_HOME means
+  // empty DB. Exit 0 with no sessions.
+  cliIt.live(
+    "session list: exits 0",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["session", "list"])
+        opencode.expectExit(r, 0, "session list")
+      }),
+    60_000,
+  )
+
+  // `stats` aggregates token usage from the session DB. Empty DB → all zeros.
+  cliIt.live(
+    "stats: exits 0",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["stats"])
+        opencode.expectExit(r, 0, "stats")
+      }),
+    60_000,
+  )
+
+  // `db path` prints the DB file location. Under harness isolation the DB
+  // resolves to SQLite's `:memory:` (no on-disk pollution between tests);
+  // in production it'd be a path under OPENCODE_TEST_HOME / XDG_DATA_HOME.
+  // Accept either form — both prove the resolver ran without crashing.
+  cliIt.live(
+    "db path: exits 0 and prints a path or :memory:",
+    ({ opencode }) =>
+      Effect.gen(function* () {
+        const r = yield* opencode.spawn(["db", "path"])
+        opencode.expectExit(r, 0, "db path")
+        expect(r.stdout.trim()).toMatch(/^(:memory:|[/\\].+\.(db|sqlite|sqlite3))$/i)
+      }),
+    60_000,
+  )
+})

packages/opencode/test/lib/cli-process.ts

@@ -31,6 +31,16 @@ import { it } from "./effect"
 const opencodeRoot = path.resolve(import.meta.dir, "../../")
 const cliEntry = path.join(opencodeRoot, "src/index.ts")
 
+// Argv prefix for spawning the CLI. If OPENCODE_TEST_CLI_PATH is set,
+// subprocess tests spawn the pre-built binary directly (~3x speedup on
+// isolation-env spawns that hit DB migration; produced by
+// `bun script/prebuild-test-cli.ts`). Otherwise falls back to dev mode —
+// strictly opt-in, default behavior unchanged.
+const prebuiltCli = process.env["OPENCODE_TEST_CLI_PATH"]
+const cliArgv: readonly string[] = prebuiltCli
+  ? [prebuiltCli]
+  : ["bun", "run", "--conditions=browser", cliEntry]
+
 export const testModelID = "test/test-model"
 
 // Wrap a Bun subprocess pipe (or any ReadableStream<Uint8Array>) as a Stream.
@@ -196,7 +206,7 @@ export function withCliFixture<A, E>(
       Effect.promise(async () => {
         const start = Date.now()
         // Process.run pipes stdout/stderr by default and returns them as Buffers.
-        const result = await Process.run(["bun", "run", "--conditions=browser", cliEntry, ...args], {
+        const result = await Process.run([...cliArgv, ...args], {
           cwd: home,
           timeout: opts?.timeoutMs ?? 30_000,
           env: { ...process.env, ...env, ...opts?.env },
@@ -235,7 +245,7 @@ export function withCliFixture<A, E>(
       // as a finalizer error during test teardown.
       const proc = yield* Effect.acquireRelease(
         Effect.sync(() =>
-          Bun.spawn(["bun", "run", "--conditions=browser", cliEntry, ...argv], {
+          Bun.spawn([...cliArgv, ...argv], {
             cwd: home,
             env: { ...process.env, ...env, ...opts?.env },
             stdout: "pipe",
@@ -306,7 +316,7 @@ export function withCliFixture<A, E>(
       // Either way we await proc.exited so the test scope doesn't leak.
       const proc = yield* Effect.acquireRelease(
         Effect.sync(() =>
-          Bun.spawn(["bun", "run", "--conditions=browser", cliEntry, ...argv], {
+          Bun.spawn([...cliArgv, ...argv], {
             cwd: opts?.cwd ?? home,
             env: { ...process.env, ...env, ...opts?.env },
             stdin: "pipe",