fix(test): deflake tool.shell metadata streaming test (LAC-2722)

[lacymorrow]

Summary

Fixes the flaky tool.shell abort > streams metadata updates progressively test (LAC-2722) seen on GitHub-hosted linux runners, e.g. run 29081382435.

Problem

The test asserted updates.length > 1 for a command emitting two output chunks separated by sleep 0.1. Under full-suite load, both chunks arrived within one flush window and coalesced into a single metadata update — the assertion was racing the scheduler, not testing the streaming contract.

Fix

Make the producer deterministic instead of raising thresholds: the fixture command now emits the second chunk only after the test observes the first metadata update, signaled via a tmpdir sentinel file written from the metadata callback. Two distinct updates are guaranteed by construction, independent of runner load.

Also adds an explicit 10s command timeout and 15s test timeout so a genuine streaming regression fails cleanly instead of hanging on the gate.

Testing

• bun test test/tool/shell.test.ts -t 'streams metadata updates progressively' — 10 consecutive sequential passes
• Same test 5/5 passes under artificial 6-core busy-loop CPU load (reproducing the coalescing condition)
• Full test/tool/shell.test.ts — 23 pass / 0 fail (3 consecutive runs)
• bun run typecheck in packages/opencode — clean

Acceptance criteria

• Deterministic pass on linux runners under load: the second chunk cannot be emitted until the first update is observed, so coalescing is impossible by construction. CI on this PR + one subsequent run will confirm across two consecutive CI runs.

[gemini-code-assist]

Code Review

This pull request updates the progressive metadata streaming test in shell.test.ts to use a file-based gating mechanism, ensuring the test is robust under runner load. Feedback suggests optimizing the gate-writing logic to avoid redundant synchronous disk I/O since the output is cumulative, and addressing potential cross-platform failures on Windows due to the POSIX-specific shell syntax used in the test command.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Since output is cumulative, output?.includes("first") will evaluate to true for every metadata update after the first one. This causes fs.writeFileSync(gate, "") to be called repeatedly on every subsequent chunk, performing redundant synchronous disk I/O.

We can optimize this by introducing a local boolean flag (gateWritten) to ensure the gate file is only written once.

          let gateWritten = false
          const updates: string[] = []
          const result = yield* run(
            {
              command: "echo first; until [ -f \"" + gate + "\" ]; do sleep 0.05; done; echo second",
              timeout: 10_000,
            },
            {
              ...ctx,
              metadata: (input) =>
                Effect.sync(() => {
                  const output = (input.metadata as { output?: string })?.output
                  if (output) updates.push(output)
                  if (output?.includes("first") && !gateWritten) {
                    gateWritten = true
                    fs.writeFileSync(gate, "")
                  }
                }),
            },
          ).pipe(Effect.ensuring(Effect.sync(() => fs.rmSync(gate, { force: true }))))

[lacymorrow]

Done in 47867e0 — added a gateWritten flag so the gate file is written exactly once.

[gemini-code-assist]

The command uses POSIX-specific shell syntax (until [ -f ... ]; do sleep 0.05; done). This will fail with syntax errors on Windows if the default shell is PowerShell or CMD (which is common on Windows development environments where the SHELL environment variable is not set).

Consider wrapping this test with withShell({ label: 'bash', shell: bash }, ...) if bash is defined, or using a cross-platform node/bun one-liner to ensure compatibility across all platforms.

[lacymorrow]

Good catch — on win32 with no SHELL set, Shell.acceptable() falls back to pwsh/powershell, where until [ -f ... ] doesn't parse. Fixed in 47867e0 by replacing the shell loop with a bun -e one-liner built the same way as the existing fill helper (charcode-encoded, & -prefixed for PowerShell, evalarg quoting), so the fixture runs under bash, pwsh, and cmd.