feat(aircoding): AirCoding V2 baseline — deterministic multi-agent architecture

Forked from OpenCode v1.17.4 with multi-agent system: - 5 agents: aircoding, scheduler, worker, architect, reviewer - Deterministic DAG scheduling engine (coordinator_tick) - Tool whitelists as hard enforcement - AirCoding validation plugin - V1 requirements: C4 docs, ADR, AGENTS.md, debug-log.md - Design documents in docs/
2026-06-13 21:53:43 +08:00
commit deaf7eb806
5757 changed files with 1169969 additions and 0 deletions
--- a/packages/opencode/test/lib/cli-process.ts
+++ b/packages/opencode/test/lib/cli-process.ts
@@ -0,0 +1,459 @@
+// Subprocess test harness for the opencode CLI. Spawns the real binary against
+// a TestLLMServer running in-process at a random port, with full env isolation.
+//
+// This is the missing test tier: in-process tests can't catch bugs that span
+// argv parsing → server boot → SDK call → event consumption → exit code (like
+// the original /event race or #27371's invalid-model hang).
+//
+// Configuration flows through opencode's built-in test affordances:
+//   - OPENCODE_CONFIG_CONTENT      : provider config inline, no files to find
+//   - OPENCODE_TEST_HOME           : pins os.homedir() → tmpdir
+//   - OPENCODE_DISABLE_PROJECT_CONFIG : skip walking up for opencode.json
+//   - OPENCODE_PURE                : skip external plugin discovery + install
+//   - OPENCODE_DISABLE_AUTOUPDATE / AUTOCOMPACT / MODELS_FETCH : no background work
+// Plus HOME / XDG_* pointing at the tmpdir for belt-and-suspenders isolation.
+//
+// Today only `opencode.run` is fully wired. The shape supports adding more
+// builders (`opencode.serve(opts)`, `opencode.acp(opts)`, `opencode.auth(...)`)
+// without changing the fixture. Long-lived commands like `serve` will need a
+// different return shape — see the TODO at the bottom of OpencodeCli.
+import { test, type TestOptions } from "bun:test"
+import { FSUtil } from "@opencode-ai/core/fs-util"
+import { AppProcess } from "@opencode-ai/core/process"
+import { Deferred, Duration, Effect, Layer, Queue, Scope, Stream } from "effect"
+import { FetchHttpClient, HttpClient } from "effect/unstable/http"
+import { ChildProcess } from "effect/unstable/process"
+import path from "node:path"
+import { TestLLMServer } from "./llm-server"
+import { testProviderConfig } from "./test-provider"
+import { it } from "./effect"
+
+const opencodeRoot = path.resolve(import.meta.dir, "../../")
+const cliEntry = path.join(opencodeRoot, "src/index.ts")
+
+export const testModelID = "test/test-model"
+
+// Wrap a Bun subprocess pipe (or any ReadableStream<Uint8Array>) as a Stream.
+// Centralizes the `evaluate` + `onError` boilerplate and tags errors with the
+// stream name so a stderr/stdout failure is greppable in logs.
+function fromBunStream(name: string, get: () => ReadableStream<Uint8Array>) {
+  return Stream.fromReadableStream({
+    evaluate: get,
+    onError: (cause) => new Error(`${name} stream error: ${String(cause)}`),
+  })
+}
+
+// Long-lived processes (serve, acp) all want the same stderr drain: read every
+// chunk, push to a tail buffer, swallow stream errors (the child closing the
+// pipe is normal). `log: true` surfaces a real protocol error to logs so a
+// regression doesn't silently disappear.
+function forkStderrDrain(stream: ReadableStream<Uint8Array>, into: string[]) {
+  return Effect.forkScoped(
+    fromBunStream("stderr", () => stream).pipe(
+      Stream.decodeText(),
+      Stream.runForEach((chunk) => Effect.sync(() => into.push(chunk))),
+      Effect.ignore({ log: true }),
+    ),
+  )
+}
+
+function isolatedEnv(home: string, configJson: string): Record<string, string> {
+  return {
+    OPENCODE_TEST_HOME: home,
+    HOME: home,
+    XDG_CONFIG_HOME: path.join(home, ".config"),
+    XDG_DATA_HOME: path.join(home, ".local/share"),
+    XDG_STATE_HOME: path.join(home, ".local/state"),
+    XDG_CACHE_HOME: path.join(home, ".cache"),
+    OPENCODE_CONFIG_CONTENT: configJson,
+    OPENCODE_DISABLE_PROJECT_CONFIG: "1",
+    OPENCODE_PURE: "1",
+    OPENCODE_DISABLE_AUTOUPDATE: "1",
+    OPENCODE_DISABLE_AUTOCOMPACT: "1",
+    OPENCODE_DISABLE_MODELS_FETCH: "1",
+    OPENCODE_AUTH_CONTENT: "{}",
+  }
+}
+
+export type RunResult = {
+  readonly exitCode: number
+  readonly stdout: string
+  readonly stderr: string
+  readonly durationMs: number
+}
+
+export type SpawnOpts = { readonly timeoutMs?: number; readonly env?: Record<string, string> }
+
+// Typed equivalent of constructing argv for `opencode run`. New flags should
+// land here so tests stay grep-able and refactor-safe.
+export type RunOpts = SpawnOpts & {
+  readonly model?: string
+  readonly agent?: string
+  readonly format?: "default" | "json"
+  readonly command?: string
+  readonly printLogs?: boolean
+  readonly extraArgs?: string[]
+}
+
+// `opencode serve` is a long-lived process — it never exits on its own.
+// `serve(opts)` therefore returns a handle inside the caller's Scope: the
+// subprocess is killed when the scope closes (test end), and the URL the
+// server actually bound to (port 0 means OS-assigned) is parsed off stdout.
+export type ServeOpts = SpawnOpts & {
+  readonly port?: number
+  readonly hostname?: string
+  readonly extraArgs?: string[]
+  // How long to wait for the "listening on http://..." line before failing.
+  // Default 15s — startup is dominated by bun's transpile + plugin init, not
+  // the actual listen() call.
+  readonly readyTimeoutMs?: number
+}
+
+export type ServeHandle = {
+  // Full URL the server is bound to, e.g. "http://127.0.0.1:54321". Use this
+  // as the base for HTTP requests in tests — never assume the port.
+  readonly url: string
+  readonly hostname: string
+  readonly port: number
+  // Sends SIGTERM. The scope finalizer also calls this, so tests rarely need
+  // to invoke it directly — useful for tests that assert exit behavior.
+  readonly kill: () => void
+  // Resolves with the exit code once the process exits. Bun returns a number.
+  readonly exited: Promise<number>
+}
+
+// `opencode acp` speaks newline-delimited JSON-RPC over stdin/stdout. It is
+// long-lived and exits cleanly when stdin is closed. The handle exposes the
+// duplex stream as send/receive rather than raw pipes so tests don't have to
+// reimplement framing on every call site.
+export type AcpOpts = SpawnOpts & {
+  readonly cwd?: string
+  readonly extraArgs?: string[]
+}
+
+export type AcpHandle = {
+  // Writes a single JSON-RPC message to the child's stdin as one ndjson line.
+  readonly send: (msg: object) => Effect.Effect<void>
+  // Resolves with the next parsed JSON-RPC line from the child's stdout.
+  // Lines are buffered in a queue so multiple receives in a row won't drop
+  // anything. Pair with `Effect.timeout` if a test wants a deadline.
+  readonly receive: Effect.Effect<unknown>
+  // Closes stdin. ACP exits cleanly on stdin EOF; the scope finalizer also
+  // calls this, so tests only need it when asserting exit behavior.
+  readonly close: () => void
+  readonly exited: Promise<number>
+}
+
+export type OpencodeCli = {
+  // High-level: run a single prompt against the test model. Short-lived.
+  readonly run: (message: string, opts?: RunOpts) => Effect.Effect<RunResult>
+  // Spawn `opencode serve` and wait until it's listening. Long-lived: the
+  // returned handle is killed when the caller's Scope closes. Fails if the
+  // listening line doesn't appear within `readyTimeoutMs`.
+  readonly serve: (opts?: ServeOpts) => Effect.Effect<ServeHandle, Error, Scope.Scope>
+  // Spawn `opencode acp` and return a duplex JSON-RPC handle. Long-lived:
+  // the subprocess exits on stdin close, which the scope finalizer triggers.
+  readonly acp: (opts?: AcpOpts) => Effect.Effect<AcpHandle, Error, Scope.Scope>
+  // Escape hatch: any CLI invocation with full control over argv. Used to test
+  // commands that don't yet have a typed builder.
+  readonly spawn: (args: string[], opts?: SpawnOpts) => Effect.Effect<RunResult>
+  // Convenience assertion. Dumps captured stderr/stdout on mismatch so CI
+  // failures are debuggable without re-running locally.
+  readonly expectExit: (result: RunResult, expected: number, label?: string) => void
+  // Parse `--format json` stdout into one event object per non-empty line.
+  // The CLI writes `JSON.stringify({ type, sessionID, ... }) + EOL` for each
+  // event (see src/cli/cmd/run.ts `emit`). Throws on a malformed line so
+  // tests fail loudly rather than silently skipping data.
+  readonly parseJsonEvents: (stdout: string) => Array<Record<string, unknown>>
+}
+
+export type CliFixture = {
+  readonly llm: TestLLMServer["Service"]
+  readonly home: string
+  readonly opencode: OpencodeCli
+}
+
+// Provisions a TestLLMServer + tmpdir + spawn helper and invokes fn. Cleans
+// up the tmpdir on scope exit. TestLLMServer.layer is provided internally so
+// the caller doesn't need to wire it up — the fixture's lifetime is tied to
+// the surrounding Scope.
+export function withCliFixture<A, E>(
+  fn: (input: CliFixture) => Effect.Effect<A, E, Scope.Scope | HttpClient.HttpClient>,
+): Effect.Effect<A, E | unknown, Scope.Scope> {
+  return Effect.gen(function* () {
+    const llm = yield* TestLLMServer
+    const fs = yield* FSUtil.Service
+    const appProc = yield* AppProcess.Service
+
+    // FileSystem.makeTempDirectoryScoped handles both creation and scope-tied
+    // cleanup — replaces the old mkdir + addFinalizer pair.
+    const home = yield* fs.makeTempDirectoryScoped({ prefix: "oc-cli-" })
+
+    const configJson = JSON.stringify(testProviderConfig(llm.url))
+    const env = isolatedEnv(home, configJson)
+
+    const spawn = Effect.fn("opencode.spawn")(function* (args: string[], opts?: SpawnOpts) {
+      const start = Date.now()
+      const timeoutMs = opts?.timeoutMs ?? 30_000
+      // stdin: "ignore" so the child doesn't see a piped stdin and block
+      // on `Bun.stdin.text()` (see src/cli/cmd/run.ts — non-TTY stdin is
+      // consumed as the prompt). The old Process.run wrapper defaulted to
+      // ignore; ChildProcess.make defaults to pipe, so we set it explicitly.
+      const command = ChildProcess.make("bun", ["run", "--conditions=browser", cliEntry, ...args], {
+        cwd: home,
+        env: { ...env, ...opts?.env },
+        extendEnv: true,
+        stdin: "ignore",
+      })
+      // Pass timeout to appProc.run rather than wrapping with
+      // Effect.timeoutOrElse externally: AppProcess.run is itself scoped, so
+      // its built-in timeout triggers the acquireRelease kill finalizer
+      // inside cross-spawn-spawner *before* surfacing the AppProcessError —
+      // guaranteeing the child is dead by the time the test continues.
+      // External timeoutOrElse interrupts the run fiber but races the
+      // scope close, which can leak the child past the test boundary.
+      //
+      // Catch AppProcessError (timeout OR spawn failure) and synthesize a
+      // non-zero result so the test sees it via the usual `expectExit`
+      // path rather than as an unhandled Effect failure.
+      const result = yield* appProc.run(command, { timeout: Duration.millis(timeoutMs) }).pipe(
+        Effect.catchTag("AppProcessError", (err) =>
+          Effect.succeed({
+            command: err.command,
+            exitCode: err.exitCode ?? -1,
+            stdout: Buffer.alloc(0),
+            stderr: Buffer.from((err.stderr ?? String(err.cause ?? err.message)) + "\n"),
+            stdoutTruncated: false,
+            stderrTruncated: false,
+          } satisfies AppProcess.RunResult),
+        ),
+      )
+      return {
+        exitCode: result.exitCode,
+        stdout: result.stdout.toString(),
+        stderr: result.stderr.toString(),
+        durationMs: Date.now() - start,
+      }
+    })
+
+    const run = (message: string, opts?: RunOpts): Effect.Effect<RunResult> => {
+      const argv: string[] = ["run"]
+      if (opts?.printLogs) argv.push("--print-logs")
+      argv.push("--model", opts?.model ?? testModelID)
+      if (opts?.agent) argv.push("--agent", opts.agent)
+      if (opts?.format) argv.push("--format", opts.format)
+      if (opts?.command) argv.push("--command", opts.command)
+      if (opts?.extraArgs) argv.push(...opts.extraArgs)
+      argv.push(message)
+      return spawn(argv, opts)
+    }
+
+    const serve = Effect.fn("opencode.serve")(function* (opts?: ServeOpts) {
+      const argv = ["serve"]
+      // Default port 0 — let the OS pick a free port, parse the actual one
+      // off stdout. Hard-coded ports flake under parallel tests.
+      argv.push("--port", String(opts?.port ?? 0))
+      if (opts?.hostname) argv.push("--hostname", opts.hostname)
+      if (opts?.extraArgs) argv.push(...opts.extraArgs)
+
+      // Acquire the subprocess; release sends SIGTERM and awaits exit on
+      // scope close. Wrapped in Effect.ignore so a flaky kill doesn't surface
+      // as a finalizer error during test teardown.
+      const proc = yield* Effect.acquireRelease(
+        Effect.sync(() =>
+          Bun.spawn(["bun", "run", "--conditions=browser", cliEntry, ...argv], {
+            cwd: home,
+            env: { ...process.env, ...env, ...opts?.env },
+            stdout: "pipe",
+            stderr: "pipe",
+          }),
+        ),
+        (p) =>
+          Effect.promise(() => {
+            p.kill()
+            return p.exited
+          }).pipe(Effect.ignore),
+      )
+
+      // Tail buffer so timeout failures can include stderr context. The fork
+      // also keeps the OS pipe buffer from filling and wedging the child.
+      const stderrChunks: string[] = []
+      yield* forkStderrDrain(proc.stderr, stderrChunks)
+
+      // Watch stdout line-by-line for the listening sentinel. Format
+      // (see src/cli/cmd/serve.ts):
+      //   "opencode server listening on http://<host>:<port>"
+      const readyRe = /listening on (http:\/\/([^\s:]+):(\d+))/
+      const readyDeferred = yield* Deferred.make<{ url: string; hostname: string; port: number }>()
+      yield* Effect.forkScoped(
+        fromBunStream("stdout", () => proc.stdout).pipe(
+          Stream.decodeText(),
+          Stream.splitLines,
+          Stream.runForEach((line) => {
+            const m = line.match(readyRe)
+            return m ? Deferred.succeed(readyDeferred, { url: m[1], hostname: m[2], port: Number(m[3]) }) : Effect.void
+          }),
+          Effect.ignore({ log: true }),
+        ),
+      )
+
+      const readyTimeoutMs = opts?.readyTimeoutMs ?? 15_000
+      const match = yield* Deferred.await(readyDeferred).pipe(
+        Effect.timeoutOrElse({
+          duration: Duration.millis(readyTimeoutMs),
+          orElse: () =>
+            Effect.fail(
+              new Error(
+                `opencode serve did not become ready within ${readyTimeoutMs}ms\n` +
+                  `stderr (last 2000):\n${stderrChunks.join("").slice(-2000)}`,
+              ),
+            ),
+        }),
+      )
+
+      return {
+        url: match.url,
+        hostname: match.hostname,
+        port: match.port,
+        kill: () => {
+          proc.kill()
+        },
+        exited: proc.exited as Promise<number>,
+      } satisfies ServeHandle
+    })
+
+    const acp = Effect.fn("opencode.acp")(function* (opts?: AcpOpts) {
+      const argv = ["acp"]
+      if (opts?.cwd) argv.push("--cwd", opts.cwd)
+      if (opts?.extraArgs) argv.push(...opts.extraArgs)
+
+      // Acquire the subprocess. Release ends stdin (clean shutdown — ACP exits
+      // on stdin EOF) and falls back to SIGTERM if it doesn't exit promptly.
+      // 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], {
+            cwd: opts?.cwd ?? home,
+            env: { ...process.env, ...env, ...opts?.env },
+            stdin: "pipe",
+            stdout: "pipe",
+            stderr: "pipe",
+          }),
+        ),
+        (p) =>
+          // Graceful shutdown: close stdin (ACP exits on EOF), give it a
+          // window to exit, then SIGTERM. The Effect.timeoutOrElse expresses
+          // exactly that race without raw setTimeout or Promise.race.
+          Effect.gen(function* () {
+            yield* Effect.sync(() => p.stdin.end())
+            yield* Effect.promise(() => p.exited).pipe(
+              Effect.timeoutOrElse({
+                duration: Duration.seconds(2),
+                orElse: () =>
+                  Effect.sync(() => {
+                    p.kill()
+                  }),
+              }),
+            )
+            yield* Effect.promise(() => p.exited)
+          }).pipe(Effect.ignore),
+      )
+
+      const stderrChunks: string[] = []
+      yield* forkStderrDrain(proc.stderr, stderrChunks)
+
+      // Each ndjson line becomes one queue entry. JSON.parse failures are
+      // surfaced as the raw string so a malformed protocol message doesn't
+      // silently wedge the test in `receive`.
+      const responses = yield* Queue.unbounded<unknown>()
+      yield* Effect.forkScoped(
+        fromBunStream("stdout", () => proc.stdout).pipe(
+          Stream.decodeText(),
+          Stream.splitLines,
+          Stream.runForEach((line) => {
+            if (line.length === 0) return Effect.void
+            let parsed: unknown
+            try {
+              parsed = JSON.parse(line)
+            } catch {
+              parsed = { _rawLine: line }
+            }
+            return Queue.offer(responses, parsed)
+          }),
+          Effect.ignore({ log: true }),
+        ),
+      )
+
+      return {
+        // `proc.stdin.write` returns `number | Promise<number>`. The promise
+        // form is the backpressure signal — if we don't await it, rapid
+        // successive sends can interleave under pipe-buffer-full conditions
+        // and corrupt the ndjson framing.
+        send: (msg: object) =>
+          Effect.promise(async () => {
+            const ret = proc.stdin.write(JSON.stringify(msg) + "\n")
+            if (typeof ret !== "number") await ret
+          }),
+        receive: Queue.take(responses),
+        // proc.stdin.end() is idempotent in Bun; no try/catch needed.
+        close: () => proc.stdin.end(),
+        exited: proc.exited as Promise<number>,
+      } satisfies AcpHandle
+    })
+
+    const opencode: OpencodeCli = { run, serve, acp, spawn, expectExit, parseJsonEvents }
+
+    return yield* fn({ llm, home, opencode })
+    // FetchHttpClient is provided so test bodies can `yield* HttpClient.HttpClient`
+    // and hit endpoints on `opencode.serve()` without rolling their own fetch.
+  }).pipe(
+    Effect.provide(
+      Layer.mergeAll(TestLLMServer.layer, FetchHttpClient.layer, FSUtil.defaultLayer, AppProcess.defaultLayer),
+    ),
+  )
+}
+
+function parseJsonEvents(stdout: string): Array<Record<string, unknown>> {
+  return stdout
+    .split("\n")
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0)
+    .map((line) => JSON.parse(line) as Record<string, unknown>)
+}
+
+// Convenience for the common assertion pattern. Dumps stderr/stdout when
+// the exit code doesn't match — saves debugging time on CI failures.
+function expectExit(result: RunResult, expected: number, label = "opencode") {
+  if (result.exitCode === expected) return
+  const tail = (s: string, n: number) => (s.length > n ? "..." + s.slice(-n) : s)
+  // eslint-disable-next-line no-console
+  console.error(`[${label}] expected exit ${expected}, got ${result.exitCode} after ${result.durationMs}ms`)
+  // eslint-disable-next-line no-console
+  console.error(`[${label}] stderr (last 2000):\n${tail(result.stderr, 2000)}`)
+  // eslint-disable-next-line no-console
+  console.error(`[${label}] stdout (last 500):\n${tail(result.stdout, 500)}`)
+  throw new Error(`${label}: expected exit ${expected}, got ${result.exitCode}`)
+}
+
+// `cliIt.live(name, fixture => effect)` is the same as
+// `it.live(name, () => withCliFixture(fixture))` — one fewer nesting level at
+// every call site. Use this for any test that needs the opencode CLI fixture.
+//
+// Subprocess tests must run against the real clock — a TestClock-paused
+// environment can't drive a child process. If you need `.only` or `.skip`, fall
+// back to `it.live` + `withCliFixture` directly.
+// Body's R is `Scope.Scope | never` so tests can yield* scope-requiring
+// resources (e.g. `opencode.serve`) without an extra `Effect.scoped` wrapper —
+// `withCliFixture`'s outer scope is the natural lifetime.
+export const cliIt = {
+  live: <A, E>(
+    name: string,
+    body: (input: CliFixture) => Effect.Effect<A, E, Scope.Scope | HttpClient.HttpClient>,
+    opts?: number | TestOptions,
+  ) => it.live(name, () => withCliFixture(body), opts),
+  concurrent: <A, E>(
+    name: string,
+    body: (input: CliFixture) => Effect.Effect<A, E, Scope.Scope | HttpClient.HttpClient>,
+    opts?: number | TestOptions,
+  ) => test.concurrent(name, () => Effect.runPromise(Effect.scoped(withCliFixture(body))), opts),
+}
--- a/packages/opencode/test/lib/effect.ts
+++ b/packages/opencode/test/lib/effect.ts
@@ -0,0 +1,177 @@
+import { test, type TestOptions } from "bun:test"
+import { ConfigV1 } from "@opencode-ai/core/v1/config/config"
+import { Cause, Duration, Effect, Exit, Layer } from "effect"
+import * as Scope from "effect/Scope"
+import * as TestClock from "effect/testing/TestClock"
+import * as TestConsole from "effect/testing/TestConsole"
+import { memoMap } from "@opencode-ai/core/effect/memo-map"
+import type { Config } from "@/config/config"
+import { TestInstance, withTmpdirInstance } from "../fixture/fixture"
+import { InstanceStore } from "@/project/instance-store"
+
+type Body<A, E, R> = Effect.Effect<A, E, R> | (() => Effect.Effect<A, E, R>)
+type InstanceOptions<E, R> = {
+  git?: boolean
+  config?: Partial<ConfigV1.Info> | (() => Partial<ConfigV1.Info>)
+  init?: (directory: string) => Effect.Effect<void, E, R>
+}
+
+function isInstanceOptions<E, R>(
+  options: InstanceOptions<E, R> | number | TestOptions | undefined,
+): options is InstanceOptions<E, R> {
+  return !!options && typeof options === "object" && ("git" in options || "config" in options || "init" in options)
+}
+
+function instanceArgs<E, R>(
+  options?: InstanceOptions<E, R> | number | TestOptions,
+  testOptions?: number | TestOptions,
+): { instanceOptions: InstanceOptions<E, R> | undefined; testOptions: number | TestOptions | undefined } {
+  if (typeof options === "number") return { instanceOptions: undefined, testOptions: options }
+  if (isInstanceOptions(options)) return { instanceOptions: options, testOptions }
+  return { instanceOptions: undefined, testOptions: options }
+}
+
+const body = <A, E, R>(value: Body<A, E, R>) => Effect.suspend(() => (typeof value === "function" ? value() : value))
+
+type Runner = <A, E, R, E2>(value: Body<A, E, R | Scope.Scope>, layer: Layer.Layer<R, E2>) => Promise<A>
+
+const isolatedRun: Runner = (value, layer) =>
+  Effect.gen(function* () {
+    const exit = yield* body(value).pipe(Effect.scoped, Effect.provide(layer), Effect.exit)
+    if (Exit.isFailure(exit)) {
+      for (const err of Cause.prettyErrors(exit.cause)) {
+        yield* Effect.logError(err)
+      }
+    }
+    return yield* exit
+  }).pipe(Effect.runPromise)
+
+// Builds the test layer through the shared process-wide memoMap so cached
+// services (Bus, Session, …) match Server.Default's instances. Use for tests
+// that publish to an in-process HTTP server and need pub/sub identity with
+// the server's handlers.
+const sharedRun: Runner = (value, layer) =>
+  Effect.gen(function* () {
+    const scope = yield* Scope.make()
+    const ctx = yield* Layer.buildWithMemoMap(layer, memoMap, scope)
+    const exit = yield* body(value).pipe(Effect.scoped, Effect.provide(ctx), Effect.exit)
+    yield* Scope.close(scope, Exit.void)
+    if (Exit.isFailure(exit)) {
+      for (const err of Cause.prettyErrors(exit.cause)) {
+        yield* Effect.logError(err)
+      }
+    }
+    return yield* exit
+  }).pipe(Effect.runPromise)
+
+const make = <R, E>(testLayer: Layer.Layer<R, E>, liveLayer: Layer.Layer<R, E>, run: Runner = isolatedRun) => {
+  const effect = <A, E2>(name: string, value: Body<A, E2, R | Scope.Scope>, opts?: number | TestOptions) =>
+    test(name, () => run(value, testLayer), opts)
+
+  effect.only = <A, E2>(name: string, value: Body<A, E2, R | Scope.Scope>, opts?: number | TestOptions) =>
+    test.only(name, () => run(value, testLayer), opts)
+
+  effect.skip = <A, E2>(name: string, value: Body<A, E2, R | Scope.Scope>, opts?: number | TestOptions) =>
+    test.skip(name, () => run(value, testLayer), opts)
+
+  const live = <A, E2>(name: string, value: Body<A, E2, R | Scope.Scope>, opts?: number | TestOptions) =>
+    test(name, () => run(value, liveLayer), opts)
+
+  live.only = <A, E2>(name: string, value: Body<A, E2, R | Scope.Scope>, opts?: number | TestOptions) =>
+    test.only(name, () => run(value, liveLayer), opts)
+
+  live.skip = <A, E2>(name: string, value: Body<A, E2, R | Scope.Scope>, opts?: number | TestOptions) =>
+    test.skip(name, () => run(value, liveLayer), opts)
+
+  const instance = <A, E2, E3 = never>(
+    name: string,
+    value: Body<A, E2, R | InstanceStore.Service | TestInstance | Scope.Scope>,
+    options?: InstanceOptions<E3, R | Scope.Scope> | number | TestOptions,
+    opts?: number | TestOptions,
+  ) => {
+    const args = instanceArgs(options, opts)
+    return test(
+      name,
+      () => run(body(value).pipe(withTmpdirInstance(args.instanceOptions)), liveLayer),
+      args.testOptions,
+    )
+  }
+
+  instance.only = <A, E2, E3 = never>(
+    name: string,
+    value: Body<A, E2, R | InstanceStore.Service | TestInstance | Scope.Scope>,
+    options?: InstanceOptions<E3, R | Scope.Scope> | number | TestOptions,
+    opts?: number | TestOptions,
+  ) => {
+    const args = instanceArgs(options, opts)
+    return test.only(
+      name,
+      () => run(body(value).pipe(withTmpdirInstance(args.instanceOptions)), liveLayer),
+      args.testOptions,
+    )
+  }
+
+  instance.skip = <A, E2, E3 = never>(
+    name: string,
+    value: Body<A, E2, R | InstanceStore.Service | TestInstance | Scope.Scope>,
+    options?: InstanceOptions<E3, R | Scope.Scope> | number | TestOptions,
+    opts?: number | TestOptions,
+  ) => {
+    const args = instanceArgs(options, opts)
+    return test.skip(
+      name,
+      () => run(body(value).pipe(withTmpdirInstance(args.instanceOptions)), liveLayer),
+      args.testOptions,
+    )
+  }
+
+  return { effect, live, instance }
+}
+
+// Test environment with TestClock and TestConsole
+const testEnv = Layer.mergeAll(TestConsole.layer, TestClock.layer())
+
+// Live environment - uses real clock, but keeps TestConsole for output capture
+const liveEnv = TestConsole.layer
+
+export const it = make<never, never>(testEnv, liveEnv)
+
+export const testEffect = <R, E>(layer: Layer.Layer<R, E>) =>
+  make<R, E>(Layer.provideMerge(layer, testEnv), Layer.provideMerge(layer, liveEnv))
+
+// Variant of `testEffect` that builds the test layer through the shared
+// process-wide memoMap so services like Bus/Session resolve to the same
+// instances Server.Default uses. Use when a test needs pub/sub identity with
+// an in-process HTTP server — most tests should stick with `testEffect`.
+export const testEffectShared = <R, E>(layer: Layer.Layer<R, E>) =>
+  make<R, E>(Layer.provideMerge(layer, testEnv), Layer.provideMerge(layer, liveEnv), sharedRun)
+
+export const awaitWithTimeout = <A, E, R>(
+  self: Effect.Effect<A, E, R>,
+  message: string,
+  duration: Duration.Input = "2 seconds",
+) =>
+  self.pipe(
+    Effect.timeoutOrElse({
+      duration,
+      orElse: () => Effect.fail(new Error(message)),
+    }),
+  )
+
+export const pollWithTimeout = <A, E, R>(
+  self: Effect.Effect<A | undefined, E, R>,
+  message: string,
+  duration: Duration.Input = "5 seconds",
+) =>
+  Effect.gen(function* () {
+    while (true) {
+      const result = yield* self
+      if (result !== undefined) return result
+      yield* Effect.sleep("20 millis")
+    }
+  }).pipe(
+    Effect.timeoutOrElse({
+      duration,
+      orElse: () => Effect.fail(new Error(message)),
+    }),
+  )
--- a/packages/opencode/test/lib/filesystem.ts
+++ b/packages/opencode/test/lib/filesystem.ts
@@ -0,0 +1,10 @@
+import path from "path"
+import { Effect, FileSystem } from "effect"
+
+export const writeFileStringScoped = Effect.fn("test.writeFileStringScoped")(function* (file: string, text: string) {
+  const fs = yield* FileSystem.FileSystem
+  yield* fs.makeDirectory(path.dirname(file), { recursive: true })
+  yield* fs.writeFileString(file, text)
+  yield* Effect.addFinalizer(() => fs.remove(file, { force: true }).pipe(Effect.orDie))
+  return file
+})
--- a/packages/opencode/test/lib/llm-server.ts
+++ b/packages/opencode/test/lib/llm-server.ts
@@ -0,0 +1,779 @@
+import { NodeHttpServer, NodeHttpServerRequest } from "@effect/platform-node"
+import * as Http from "node:http"
+import { Deferred, Effect, Layer, Context, Stream } from "effect"
+import * as HttpServer from "effect/unstable/http/HttpServer"
+import { HttpRouter, HttpServerRequest, HttpServerResponse } from "effect/unstable/http"
+
+export type Usage = { input: number; output: number }
+
+type Line = Record<string, unknown>
+
+type Flow =
+  | { type: "text"; text: string }
+  | { type: "reason"; text: string }
+  | { type: "tool-start"; id: string; name: string }
+  | { type: "tool-args"; text: string }
+  | { type: "usage"; usage: Usage }
+
+type Hit = {
+  url: URL
+  body: Record<string, unknown>
+}
+
+type Match = (hit: Hit) => boolean
+
+type Queue = {
+  item: Item
+  match?: Match
+}
+
+type Wait = {
+  count: number
+  ready: Deferred.Deferred<void>
+}
+
+type Sse = {
+  type: "sse"
+  head: unknown[]
+  tail: unknown[]
+  wait?: PromiseLike<unknown>
+  hang?: boolean
+  error?: unknown
+  reset?: boolean
+}
+
+type HttpError = {
+  type: "http-error"
+  status: number
+  body: unknown
+}
+
+export type Item = Sse | HttpError
+
+const done = Symbol("done")
+
+function line(input: unknown) {
+  if (input === done) return "data: [DONE]\n\n"
+  return `data: ${JSON.stringify(input)}\n\n`
+}
+
+function tokens(input?: Usage) {
+  if (!input) return
+  return {
+    prompt_tokens: input.input,
+    completion_tokens: input.output,
+    total_tokens: input.input + input.output,
+  }
+}
+
+function chunk(input: { delta?: Record<string, unknown>; finish?: string; usage?: Usage }) {
+  return {
+    id: "chatcmpl-test",
+    object: "chat.completion.chunk",
+    choices: [
+      {
+        delta: input.delta ?? {},
+        ...(input.finish ? { finish_reason: input.finish } : {}),
+      },
+    ],
+    ...(input.usage ? { usage: tokens(input.usage) } : {}),
+  } satisfies Line
+}
+
+function role() {
+  return chunk({ delta: { role: "assistant" } })
+}
+
+function textLine(value: string) {
+  return chunk({ delta: { content: value } })
+}
+
+function reasonLine(value: string) {
+  return chunk({ delta: { reasoning_content: value } })
+}
+
+function finishLine(reason: string, usage?: Usage) {
+  return chunk({ finish: reason, usage })
+}
+
+function toolStartLine(id: string, name: string) {
+  return chunk({
+    delta: {
+      tool_calls: [
+        {
+          index: 0,
+          id,
+          type: "function",
+          function: {
+            name,
+            arguments: "",
+          },
+        },
+      ],
+    },
+  })
+}
+
+function toolArgsLine(value: string) {
+  return chunk({
+    delta: {
+      tool_calls: [
+        {
+          index: 0,
+          function: {
+            arguments: value,
+          },
+        },
+      ],
+    },
+  })
+}
+
+function bytes(input: Iterable<unknown>) {
+  return Stream.fromIterable([...input].map(line)).pipe(Stream.encodeText)
+}
+
+function responseCreated(model: string) {
+  return {
+    type: "response.created",
+    sequence_number: 1,
+    response: {
+      id: "resp_test",
+      created_at: Math.floor(Date.now() / 1000),
+      model,
+      service_tier: null,
+    },
+  }
+}
+
+function responseCompleted(input: { seq: number; usage?: Usage }) {
+  return {
+    type: "response.completed",
+    sequence_number: input.seq,
+    response: {
+      incomplete_details: null,
+      service_tier: null,
+      usage: {
+        input_tokens: input.usage?.input ?? 0,
+        input_tokens_details: { cached_tokens: null },
+        output_tokens: input.usage?.output ?? 0,
+        output_tokens_details: { reasoning_tokens: null },
+      },
+    },
+  }
+}
+
+function responseMessage(id: string, seq: number) {
+  return {
+    type: "response.output_item.added",
+    sequence_number: seq,
+    output_index: 0,
+    item: { type: "message", id },
+  }
+}
+
+function responseText(id: string, text: string, seq: number) {
+  return {
+    type: "response.output_text.delta",
+    sequence_number: seq,
+    item_id: id,
+    delta: text,
+    logprobs: null,
+  }
+}
+
+function responseMessageDone(id: string, seq: number) {
+  return {
+    type: "response.output_item.done",
+    sequence_number: seq,
+    output_index: 0,
+    item: { type: "message", id },
+  }
+}
+
+function responseReason(id: string, seq: number) {
+  return {
+    type: "response.output_item.added",
+    sequence_number: seq,
+    output_index: 0,
+    item: { type: "reasoning", id, encrypted_content: null },
+  }
+}
+
+function responseReasonPart(id: string, seq: number) {
+  return {
+    type: "response.reasoning_summary_part.added",
+    sequence_number: seq,
+    item_id: id,
+    summary_index: 0,
+  }
+}
+
+function responseReasonText(id: string, text: string, seq: number) {
+  return {
+    type: "response.reasoning_summary_text.delta",
+    sequence_number: seq,
+    item_id: id,
+    summary_index: 0,
+    delta: text,
+  }
+}
+
+function responseReasonDone(id: string, seq: number) {
+  return {
+    type: "response.output_item.done",
+    sequence_number: seq,
+    output_index: 0,
+    item: { type: "reasoning", id, encrypted_content: null },
+  }
+}
+
+function responseTool(id: string, item: string, name: string, seq: number) {
+  return {
+    type: "response.output_item.added",
+    sequence_number: seq,
+    output_index: 0,
+    item: {
+      type: "function_call",
+      id: item,
+      call_id: id,
+      name,
+      arguments: "",
+      status: "in_progress",
+    },
+  }
+}
+
+function responseToolArgs(id: string, text: string, seq: number) {
+  return {
+    type: "response.function_call_arguments.delta",
+    sequence_number: seq,
+    output_index: 0,
+    item_id: id,
+    delta: text,
+  }
+}
+
+function responseToolArgsDone(id: string, args: string, seq: number) {
+  return {
+    type: "response.function_call_arguments.done",
+    sequence_number: seq,
+    output_index: 0,
+    item_id: id,
+    arguments: args,
+  }
+}
+
+function responseToolDone(tool: { id: string; item: string; name: string; args: string }, seq: number) {
+  return {
+    type: "response.output_item.done",
+    sequence_number: seq,
+    output_index: 0,
+    item: {
+      type: "function_call",
+      id: tool.item,
+      call_id: tool.id,
+      name: tool.name,
+      arguments: tool.args,
+      status: "completed",
+    },
+  }
+}
+
+function choices(part: unknown) {
+  if (!part || typeof part !== "object") return
+  if (!("choices" in part) || !Array.isArray(part.choices)) return
+  const choice = part.choices[0]
+  if (!choice || typeof choice !== "object") return
+  return choice
+}
+
+function flow(item: Sse) {
+  const out: Flow[] = []
+  for (const part of [...item.head, ...item.tail]) {
+    const choice = choices(part)
+    const delta =
+      choice && "delta" in choice && choice.delta && typeof choice.delta === "object" ? choice.delta : undefined
+
+    if (delta && "content" in delta && typeof delta.content === "string") {
+      out.push({ type: "text", text: delta.content })
+    }
+
+    if (delta && "reasoning_content" in delta && typeof delta.reasoning_content === "string") {
+      out.push({ type: "reason", text: delta.reasoning_content })
+    }
+
+    if (delta && "tool_calls" in delta && Array.isArray(delta.tool_calls)) {
+      for (const tool of delta.tool_calls) {
+        if (!tool || typeof tool !== "object") continue
+        const fn = "function" in tool && tool.function && typeof tool.function === "object" ? tool.function : undefined
+        if ("id" in tool && typeof tool.id === "string" && fn && "name" in fn && typeof fn.name === "string") {
+          out.push({ type: "tool-start", id: tool.id, name: fn.name })
+        }
+        if (fn && "arguments" in fn && typeof fn.arguments === "string" && fn.arguments) {
+          out.push({ type: "tool-args", text: fn.arguments })
+        }
+      }
+    }
+
+    if (part && typeof part === "object" && "usage" in part && part.usage && typeof part.usage === "object") {
+      const raw = part.usage as Record<string, unknown>
+      if (typeof raw.prompt_tokens === "number" && typeof raw.completion_tokens === "number") {
+        out.push({
+          type: "usage",
+          usage: { input: raw.prompt_tokens, output: raw.completion_tokens },
+        })
+      }
+    }
+  }
+  return out
+}
+
+function responses(item: Sse, model: string) {
+  let seq = 1
+  let msg: string | undefined
+  let reason: string | undefined
+  let hasMsg = false
+  let hasReason = false
+  let call:
+    | {
+        id: string
+        item: string
+        name: string
+        args: string
+      }
+    | undefined
+  let usage: Usage | undefined
+  const lines: unknown[] = [responseCreated(model)]
+
+  for (const part of flow(item)) {
+    if (part.type === "text") {
+      msg ??= "msg_1"
+      if (!hasMsg) {
+        hasMsg = true
+        seq += 1
+        lines.push(responseMessage(msg, seq))
+      }
+      seq += 1
+      lines.push(responseText(msg, part.text, seq))
+      continue
+    }
+
+    if (part.type === "reason") {
+      reason ||= "rs_1"
+      if (!hasReason) {
+        hasReason = true
+        seq += 1
+        lines.push(responseReason(reason, seq))
+        seq += 1
+        lines.push(responseReasonPart(reason, seq))
+      }
+      seq += 1
+      lines.push(responseReasonText(reason, part.text, seq))
+      continue
+    }
+
+    if (part.type === "tool-start") {
+      call ||= { id: part.id, item: "fc_1", name: part.name, args: "" }
+      seq += 1
+      lines.push(responseTool(call.id, call.item, call.name, seq))
+      continue
+    }
+
+    if (part.type === "tool-args") {
+      if (!call) continue
+      call.args += part.text
+      seq += 1
+      lines.push(responseToolArgs(call.item, part.text, seq))
+      continue
+    }
+
+    usage = part.usage
+  }
+
+  if (msg) {
+    seq += 1
+    lines.push(responseMessageDone(msg, seq))
+  }
+  if (reason) {
+    seq += 1
+    lines.push(responseReasonDone(reason, seq))
+  }
+  if (call && !item.hang && !item.error) {
+    seq += 1
+    lines.push(responseToolArgsDone(call.item, call.args, seq))
+    seq += 1
+    lines.push(responseToolDone(call, seq))
+  }
+  if (!item.hang && !item.error) lines.push(responseCompleted({ seq: seq + 1, usage }))
+  return { ...item, head: lines, tail: [] } satisfies Sse
+}
+
+function modelFrom(body: unknown) {
+  if (!body || typeof body !== "object") return "test-model"
+  if (!("model" in body) || typeof body.model !== "string") return "test-model"
+  return body.model
+}
+
+function send(item: Sse) {
+  const head = bytes(item.head)
+  const tail = bytes([...item.tail, ...(item.hang || item.error ? [] : [done])])
+  const empty = Stream.fromIterable<Uint8Array>([])
+  const wait = item.wait
+  const body: Stream.Stream<Uint8Array, unknown> = wait
+    ? Stream.concat(head, Stream.fromEffect(Effect.promise(() => wait)).pipe(Stream.flatMap(() => tail)))
+    : Stream.concat(head, tail)
+  let end: Stream.Stream<Uint8Array, unknown> = empty
+  if (item.error) end = Stream.concat(empty, Stream.fail(item.error))
+  else if (item.hang) end = Stream.concat(empty, Stream.never)
+
+  return HttpServerResponse.stream(Stream.concat(body, end), { contentType: "text/event-stream" })
+}
+
+const reset = Effect.fn("TestLLMServer.reset")(function* (item: Sse) {
+  const req = yield* HttpServerRequest.HttpServerRequest
+  const res = NodeHttpServerRequest.toServerResponse(req)
+  yield* Effect.sync(() => {
+    res.writeHead(200, { "content-type": "text/event-stream" })
+    for (const part of item.head) res.write(line(part))
+    for (const part of item.tail) res.write(line(part))
+    res.destroy(new Error("connection reset"))
+  })
+  return yield* Effect.never
+})
+
+function fail(item: HttpError) {
+  return HttpServerResponse.text(JSON.stringify(item.body), {
+    status: item.status,
+    contentType: "application/json",
+  })
+}
+
+export class Reply {
+  #head: unknown[] = [role()]
+  #tail: unknown[] = []
+  #usage: Usage | undefined
+  #finish: string | undefined
+  #wait: PromiseLike<unknown> | undefined
+  #hang = false
+  #error: unknown
+  #reset = false
+  #seq = 0
+
+  #id() {
+    this.#seq += 1
+    return `call_${this.#seq}`
+  }
+
+  text(value: string) {
+    this.#tail = [...this.#tail, textLine(value)]
+    return this
+  }
+
+  reason(value: string) {
+    this.#tail = [...this.#tail, reasonLine(value)]
+    return this
+  }
+
+  usage(value: Usage) {
+    this.#usage = value
+    return this
+  }
+
+  wait(value: PromiseLike<unknown>) {
+    this.#wait = value
+    return this
+  }
+
+  stop() {
+    this.#finish = "stop"
+    this.#hang = false
+    this.#error = undefined
+    this.#reset = false
+    return this
+  }
+
+  contentFilter() {
+    this.#finish = "content_filter"
+    this.#hang = false
+    this.#error = undefined
+    this.#reset = false
+    return this
+  }
+
+  toolCalls() {
+    this.#finish = "tool_calls"
+    this.#hang = false
+    this.#error = undefined
+    this.#reset = false
+    return this
+  }
+
+  tool(name: string, input: unknown) {
+    const id = this.#id()
+    const args = JSON.stringify(input)
+    this.#tail = [...this.#tail, toolStartLine(id, name), toolArgsLine(args)]
+    return this.toolCalls()
+  }
+
+  pendingTool(name: string, input: unknown) {
+    const id = this.#id()
+    const args = JSON.stringify(input)
+    const size = Math.max(1, Math.floor(args.length / 2))
+    this.#tail = [...this.#tail, toolStartLine(id, name), toolArgsLine(args.slice(0, size))]
+    return this
+  }
+
+  hang() {
+    this.#finish = undefined
+    this.#hang = true
+    this.#error = undefined
+    this.#reset = false
+    return this
+  }
+
+  streamError(error: unknown = "boom") {
+    this.#finish = undefined
+    this.#hang = false
+    this.#error = error
+    this.#reset = false
+    return this
+  }
+
+  reset() {
+    this.#finish = undefined
+    this.#hang = false
+    this.#error = undefined
+    this.#reset = true
+    return this
+  }
+
+  item(): Item {
+    return {
+      type: "sse",
+      head: this.#head,
+      tail: this.#finish ? [...this.#tail, finishLine(this.#finish, this.#usage)] : this.#tail,
+      wait: this.#wait,
+      hang: this.#hang,
+      error: this.#error,
+      reset: this.#reset,
+    }
+  }
+}
+
+export function reply() {
+  return new Reply()
+}
+
+export function httpError(status: number, body: unknown): Item {
+  return {
+    type: "http-error",
+    status,
+    body,
+  }
+}
+
+export function raw(input: {
+  chunks?: unknown[]
+  head?: unknown[]
+  tail?: unknown[]
+  wait?: PromiseLike<unknown>
+  hang?: boolean
+  error?: unknown
+  reset?: boolean
+}): Item {
+  return {
+    type: "sse",
+    head: input.head ?? input.chunks ?? [],
+    tail: input.tail ?? [],
+    wait: input.wait,
+    hang: input.hang,
+    error: input.error,
+    reset: input.reset,
+  }
+}
+
+function item(input: Item | Reply) {
+  return input instanceof Reply ? input.item() : input
+}
+
+function hit(url: string, body: unknown) {
+  return {
+    url: new URL(url, "http://localhost"),
+    body: body && typeof body === "object" ? (body as Record<string, unknown>) : {},
+  } satisfies Hit
+}
+
+function isTitleRequest(body: unknown): boolean {
+  if (!body || typeof body !== "object") return false
+  return JSON.stringify(body).includes("Generate a title for this conversation")
+}
+
+namespace TestLLMServer {
+  export interface Service {
+    readonly url: string
+    readonly push: (...input: (Item | Reply)[]) => Effect.Effect<void>
+    readonly pushMatch: (match: Match, ...input: (Item | Reply)[]) => Effect.Effect<void>
+    readonly textMatch: (match: Match, value: string, opts?: { usage?: Usage }) => Effect.Effect<void>
+    readonly toolMatch: (match: Match, name: string, input: unknown) => Effect.Effect<void>
+    readonly text: (value: string, opts?: { usage?: Usage }) => Effect.Effect<void>
+    readonly tool: (name: string, input: unknown) => Effect.Effect<void>
+    readonly toolHang: (name: string, input: unknown) => Effect.Effect<void>
+    readonly reason: (value: string, opts?: { text?: string; usage?: Usage }) => Effect.Effect<void>
+    readonly fail: (message?: unknown) => Effect.Effect<void>
+    readonly error: (status: number, body: unknown) => Effect.Effect<void>
+    readonly hang: Effect.Effect<void>
+    readonly hold: (value: string, wait: PromiseLike<unknown>) => Effect.Effect<void>
+    readonly reset: Effect.Effect<void>
+    readonly hits: Effect.Effect<Hit[]>
+    readonly calls: Effect.Effect<number>
+    readonly wait: (count: number) => Effect.Effect<void>
+    readonly inputs: Effect.Effect<Record<string, unknown>[]>
+    readonly pending: Effect.Effect<number>
+    readonly misses: Effect.Effect<Hit[]>
+  }
+}
+
+export class TestLLMServer extends Context.Service<TestLLMServer, TestLLMServer.Service>()("@test/LLMServer") {
+  static readonly layer = Layer.effect(
+    TestLLMServer,
+    Effect.gen(function* () {
+      const server = yield* HttpServer.HttpServer
+      const router = yield* HttpRouter.HttpRouter
+
+      let hits: Hit[] = []
+      let list: Queue[] = []
+      let waits: Wait[] = []
+      let misses: Hit[] = []
+
+      const queue = (...input: (Item | Reply)[]) => {
+        list = [...list, ...input.map((value) => ({ item: item(value) }))]
+      }
+
+      const queueMatch = (match: Match, ...input: (Item | Reply)[]) => {
+        list = [...list, ...input.map((value) => ({ item: item(value), match }))]
+      }
+
+      const notify = Effect.fnUntraced(function* () {
+        const ready = waits.filter((item) => hits.length >= item.count)
+        if (!ready.length) return
+        waits = waits.filter((item) => hits.length < item.count)
+        yield* Effect.forEach(ready, (item) => Deferred.succeed(item.ready, void 0))
+      })
+
+      const pull = (hit: Hit) => {
+        const index = list.findIndex((entry) => !entry.match || entry.match(hit))
+        if (index === -1) return
+        const first = list[index]
+        list = [...list.slice(0, index), ...list.slice(index + 1)]
+        return first.item
+      }
+
+      const handle = Effect.fn("TestLLMServer.handle")(function* (mode: "chat" | "responses") {
+        const req = yield* HttpServerRequest.HttpServerRequest
+        const body = yield* req.json.pipe(Effect.orElseSucceed(() => ({})))
+        const current = hit(req.originalUrl, body)
+        if (isTitleRequest(body)) {
+          hits = [...hits, current]
+          yield* notify()
+          const auto: Sse = { type: "sse", head: [role()], tail: [textLine("E2E Title"), finishLine("stop")] }
+          if (mode === "responses") return send(responses(auto, modelFrom(body)))
+          return send(auto)
+        }
+        const next = pull(current)
+        if (!next) {
+          hits = [...hits, current]
+          yield* notify()
+          const auto: Sse = { type: "sse", head: [role()], tail: [textLine("ok"), finishLine("stop")] }
+          if (mode === "responses") return send(responses(auto, modelFrom(body)))
+          return send(auto)
+        }
+        hits = [...hits, current]
+        yield* notify()
+        if (next.type !== "sse") return fail(next)
+        if (mode === "responses") return send(responses(next, modelFrom(body)))
+        if (next.reset) {
+          yield* reset(next)
+          return HttpServerResponse.empty()
+        }
+        return send(next)
+      })
+
+      yield* router.add("POST", "/v1/chat/completions", handle("chat"))
+      yield* router.add("POST", "/v1/responses", handle("responses"))
+
+      yield* server.serve(router.asHttpEffect())
+
+      return TestLLMServer.of({
+        url:
+          server.address._tag === "TcpAddress"
+            ? `http://127.0.0.1:${server.address.port}/v1`
+            : `unix://${server.address.path}/v1`,
+        push: Effect.fn("TestLLMServer.push")(function* (...input: (Item | Reply)[]) {
+          queue(...input)
+        }),
+        pushMatch: Effect.fn("TestLLMServer.pushMatch")(function* (match: Match, ...input: (Item | Reply)[]) {
+          queueMatch(match, ...input)
+        }),
+        textMatch: Effect.fn("TestLLMServer.textMatch")(function* (
+          match: Match,
+          value: string,
+          opts?: { usage?: Usage },
+        ) {
+          const out = reply().text(value)
+          if (opts?.usage) out.usage(opts.usage)
+          queueMatch(match, out.stop().item())
+        }),
+        toolMatch: Effect.fn("TestLLMServer.toolMatch")(function* (match: Match, name: string, input: unknown) {
+          queueMatch(match, reply().tool(name, input).item())
+        }),
+        text: Effect.fn("TestLLMServer.text")(function* (value: string, opts?: { usage?: Usage }) {
+          const out = reply().text(value)
+          if (opts?.usage) out.usage(opts.usage)
+          queue(out.stop().item())
+        }),
+        tool: Effect.fn("TestLLMServer.tool")(function* (name: string, input: unknown) {
+          queue(reply().tool(name, input).item())
+        }),
+        toolHang: Effect.fn("TestLLMServer.toolHang")(function* (name: string, input: unknown) {
+          queue(reply().pendingTool(name, input).hang().item())
+        }),
+        reason: Effect.fn("TestLLMServer.reason")(function* (value: string, opts?: { text?: string; usage?: Usage }) {
+          const out = reply().reason(value)
+          if (opts?.text) out.text(opts.text)
+          if (opts?.usage) out.usage(opts.usage)
+          queue(out.stop().item())
+        }),
+        fail: Effect.fn("TestLLMServer.fail")(function* (message: unknown = "boom") {
+          queue(reply().streamError(message).item())
+        }),
+        error: Effect.fn("TestLLMServer.error")(function* (status: number, body: unknown) {
+          queue(httpError(status, body))
+        }),
+        hang: Effect.gen(function* () {
+          queue(reply().hang().item())
+        }).pipe(Effect.withSpan("TestLLMServer.hang")),
+        hold: Effect.fn("TestLLMServer.hold")(function* (value: string, wait: PromiseLike<unknown>) {
+          queue(reply().wait(wait).text(value).stop().item())
+        }),
+        reset: Effect.sync(() => {
+          hits = []
+          list = []
+          waits = []
+          misses = []
+        }),
+        hits: Effect.sync(() => [...hits]),
+        calls: Effect.sync(() => hits.length),
+        wait: Effect.fn("TestLLMServer.wait")(function* (count: number) {
+          if (hits.length >= count) return
+          const ready = yield* Deferred.make<void>()
+          waits = [...waits, { count, ready }]
+          yield* Deferred.await(ready)
+        }),
+        inputs: Effect.sync(() => hits.map((hit) => hit.body)),
+        pending: Effect.sync(() => list.length),
+        misses: Effect.sync(() => [...misses]),
+      })
+    }),
+  ).pipe(Layer.provide(HttpRouter.layer), Layer.provide(NodeHttpServer.layer(() => Http.createServer(), { port: 0 })))
+}
--- a/packages/opencode/test/lib/snapshot.ts
+++ b/packages/opencode/test/lib/snapshot.ts
@@ -0,0 +1,73 @@
+// Shared normalization helpers for cross-OS-stable snapshot tests.
+//
+// Every snapshot test that captures subprocess output, file paths, or other
+// OS-flavored strings hits the same two issues:
+//   1. Bun emits CRLF line endings on Windows stderr; LF elsewhere.
+//   2. Path separators differ (\ on Windows, / on POSIX), and macOS's
+//      /var/folders symlink resolves to /private/var/folders.
+//
+// These helpers exist so each test doesn't reinvent the same regexes.
+//
+// Use individually for fine-grained control, or compose them via
+// `normalizeForSnapshot` for the common "snapshot subprocess output" path.
+import fs from "node:fs"
+import os from "node:os"
+
+const TMP = os.tmpdir()
+const REAL_TMP = fs.realpathSync(TMP)
+
+/**
+ * Collapses CRLF to LF. Bun's subprocess pipes emit native line endings —
+ * snapshots captured on macOS/Linux contain LF, so a Windows run without
+ * this step always diffs.
+ */
+export function stripCrlf(text: string): string {
+  return text.replaceAll("\r\n", "\n")
+}
+
+/**
+ * Converts Windows-style `\` separators to POSIX `/` so paths render
+ * identically across OSes. Use for path strings you want stable in a
+ * snapshot, not for filesystem operations.
+ */
+export function toPosixPath(p: string): string {
+  return p.replaceAll("\\", "/")
+}
+
+/**
+ * Strips both the OS-level `os.tmpdir()` and its realpath form (macOS
+ * `/var/folders` → `/private/var/folders`) from text, replacing each
+ * occurrence with `marker` (default `<TMPDIR>`).
+ */
+export function withTmpdirStripped(text: string, marker = "<TMPDIR>"): string {
+  return text.replaceAll(REAL_TMP, marker).replaceAll(TMP, marker)
+}
+
+/**
+ * Separator-agnostic match class for path-style strings. Use inside a
+ * larger regex when you want to match both `/` (POSIX) and `\` (Windows)
+ * boundaries — e.g. `<TMPDIR>${PATH_SEP}oc-cli-[a-z0-9]+`.
+ */
+export const PATH_SEP = "[/\\\\]"
+
+/**
+ * One-shot normalization for the common case: strip CRLF, strip tmpdir,
+ * then apply any caller-supplied path regex substitutions. Does NOT
+ * blanket-replace `\` with `/` — that would mangle non-path backslash
+ * content (regex literals in help text, etc.). Use `toPosixPath` or
+ * `PATH_SEP` in your own regex when you need separator agnosticism.
+ */
+export function normalizeForSnapshot(
+  text: string,
+  options?: {
+    readonly tmpdirMarker?: string
+    readonly pathReplacements?: ReadonlyArray<readonly [RegExp, string]>
+  },
+): string {
+  let out = stripCrlf(text)
+  out = withTmpdirStripped(out, options?.tmpdirMarker)
+  for (const [pattern, replacement] of options?.pathReplacements ?? []) {
+    out = out.replace(pattern, replacement)
+  }
+  return out
+}
--- a/packages/opencode/test/lib/test-provider.ts
+++ b/packages/opencode/test/lib/test-provider.ts
@@ -0,0 +1,37 @@
+// Shared provider config for tests that need opencode to talk to a fake LLM
+// over a real HTTP endpoint. Registers a single provider `test` with a single
+// model `test-model` (i.e. `--model test/test-model`), pointed at the URL the
+// caller supplies (typically a TestLLMServer instance).
+//
+// Used by:
+//   - test/lib/run-process.ts          (subprocess CLI tests)
+//   - test/server/httpapi-sdk.test.ts  (in-process SDK tests)
+export function testProviderConfig(llmUrl: string) {
+  return {
+    formatter: false,
+    lsp: false,
+    provider: {
+      test: {
+        name: "Test",
+        id: "test",
+        env: [],
+        npm: "@ai-sdk/openai-compatible",
+        models: {
+          "test-model": {
+            id: "test-model",
+            name: "Test Model",
+            attachment: false,
+            reasoning: false,
+            temperature: false,
+            tool_call: true,
+            release_date: "2025-01-01",
+            limit: { context: 100_000, output: 10_000 },
+            cost: { input: 0, output: 0 },
+            options: {},
+          },
+        },
+        options: { apiKey: "test-key", baseURL: llmUrl },
+      },
+    },
+  }
+}
--- a/packages/opencode/test/lib/websocket.ts
+++ b/packages/opencode/test/lib/websocket.ts
@@ -0,0 +1,46 @@
+export class FakeWebSocket {
+  static CONNECTING = 0
+  static OPEN = 1
+  static CLOSING = 2
+  static CLOSED = 3
+
+  readyState = FakeWebSocket.CONNECTING
+  closed = false
+  sent: string[] = []
+  listeners = new Map<string, Set<(event: { data?: unknown }) => void>>()
+
+  constructor(
+    readonly url: string,
+    readonly options?: { headers?: Record<string, string> },
+  ) {}
+
+  addEventListener(type: string, listener: (event: { data?: unknown }) => void) {
+    const current = this.listeners.get(type) ?? new Set<(event: { data?: unknown }) => void>()
+    current.add(listener)
+    this.listeners.set(type, current)
+  }
+
+  send(data: string) {
+    this.sent.push(data)
+  }
+
+  close() {
+    if (this.readyState === FakeWebSocket.CLOSED) return
+    this.closed = true
+    this.readyState = FakeWebSocket.CLOSED
+    this.emit("close", {})
+  }
+
+  open() {
+    this.readyState = FakeWebSocket.OPEN
+    this.emit("open", {})
+  }
+
+  message(data: unknown) {
+    this.emit("message", { data })
+  }
+
+  emit(type: string, event: { data?: unknown }) {
+    this.listeners.get(type)?.forEach((listener) => listener(event))
+  }
+}