test(ci): shard cross-platform CLI suite

[marius-kilocode]

The CLI unit suite starts each test file in a separate Bun process to isolate PID-scoped databases, environment state, child processes, and teardown. The full-platform workflow was dominated by process startup and long process-heavy tails, while the Linux HttpApi gate ran only after the unit queue.

This preserves process-per-file isolation, retries, JUnit reporting, and complete eligible coverage on Linux and Windows. Files are ordered largest-first and assigned exactly once through deterministic greedy balancing across two Linux shards and four Windows shards. macOS runs one curated 106-file Darwin-native profile covering every sandbox test plus process, terminal, PTY, watcher, filesystem, permission, Git, worktree, and native-runtime behavior. Platform-neutral coverage remains backed by complete Linux and Windows runs. Non-CLI package tests still run once per OS, and the full HttpApi exerciser runs independently so it overlaps unit jobs.

Windows fixture cleanup forces GC only after a locked-directory failure instead of before every deletion. Persistent-process tests wait for child exit and persisted output before deleting fixtures or simulating reload. Turbo cache lookup excludes installed package manifests, includes shard identity, and shared Linux sandbox setup lives in one composite action. Stable aggregate checks preserve the existing unit (linux) and test (linux) required contexts while shard jobs provide detailed failures.

Measured impact

Cold reference: https://github.com/Kilo-Org/kilocode/actions/runs/28947466351
Cold sharded run: https://github.com/Kilo-Org/kilocode/actions/runs/29031811351
Final profile-trim run: https://github.com/Kilo-Org/kilocode/actions/runs/29082265682

Path	Before	After	Saved	Faster
Complete test workflow	18m44s	11m10s	7m34s	40.4%
Linux critical path	12m24s	6m40s	5m44s	46.2%
Windows critical path	18m02s	7m31s	10m31s	58.3%
macOS curated profile	7m26s / 137 files	5m52s / 106 files	1m34s	21.1%

The macOS profile comparison is measured on consecutive revisions of this branch. The original full-suite macOS reference was 13m20s, but that comparison is not coverage-equivalent because platform-neutral tests are now intentionally covered by the complete Linux and Windows runs.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Since the last review, the branch picked up the test(cli): normalize profile fixture paths commit, which changes a single line in packages/opencode/test/kilocode/test-profile.test.ts: the raw glob-scanned file list (all) now has its path separators normalized before being sorted and used in assertions. Bun.Glob.scan returns OS-native separators, so on Windows the previous all array still contained backslashes, which could break the startsWith("kilocode/sandbox/") filter and the subsequent array-equality assertion on lines 20-21, even though TestProfile.resolve itself already normalizes backslashes internally (covered by the existing "normalizes Windows test paths" test). The fix is minimal, test-only, and contained to a kilocode/-named path, so no kilocode_change markers are required.

The two pre-existing open findings on .github/workflows/test.yml (sandbox-helper setup duplication, and macOS test-profile scope) are unrelated to this commit's file and remain tracked in their existing inline comments.

Files Reviewed (1 file changed since last review)

• packages/opencode/test/kilocode/test-profile.test.ts

Previous Review Summaries (6 snapshots, latest commit 6c96179)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 6c96179)

Status: No Issues Found | Recommendation: Merge

Since the last review, the branch picked up the test(ci): trim platform-neutral macOS coverage commit. It further narrows the Darwin CI profile (packages/opencode/script/kilocode/test-profile.ts) by removing additional platform-neutral test globs (e.g. cli/run/{footer.view,runtime.stdin,scrollback.surface}, several cli/tui/* snapshot tests, tool/external-directory, util/wildcard, several single-file kilocode/* entries, mcp/lifecycle, shell/*, and tightening server/httpapi-pty* to the exact httpapi-pty.test.ts file) that remain fully covered by the complete Linux/Windows unit runs. The companion test (test/kilocode/test-profile.test.ts) adds matching assertions confirming these files are now excluded from the darwin profile while the kilocode/sandbox/* group still resolves to every sandbox test file. The change is self-contained to these two kilocode-named files (no kilocode_change markers required) and consistent with the profile-trimming pattern already in place.

The pre-existing open finding on macOS test-profile scope (.github/workflows/test.yml line 143) is unrelated to this commit's files and remains tracked in its existing inline comment.

Files Reviewed (2 files changed since last review)

• packages/opencode/script/kilocode/test-profile.ts
• packages/opencode/test/kilocode/test-profile.test.ts

Previous review (commit b0623a6)

Status: No Issues Found | Recommendation: Merge

Since the last review, the branch picked up the fix(ci): restore required unit status commit. It adds a new unit-required job whose GitHub check name is pinned to the literal string unit (linux), decoupled from the sharded/dynamic unit matrix job names (which now render as unit (unchanged) when general tests are skipped, instead of colliding with that literal name). This restores a stable, always-present required-status-check name for branch protection while the actual unit matrix job names remain dynamic across shards/OSes. The change is small, self-contained to the kilocode_change CI block, and consistent with the existing required (test (linux)) job's if: always() + needs.<job>.result == 'success' gating pattern.

The pre-existing open finding on macOS test-profile scope (line 143) is unrelated to this commit and remains tracked in its existing inline comment.

Files Reviewed (1 file changed since last review)

• .github/workflows/test.yml

Previous review (commit c340790)

Status: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	1

Issue Details (click to expand)

SUGGESTION

File	Line	Issue
.github/workflows/test.yml	143	macOS CI now runs only the curated darwin test-profile subset (single shard) instead of the complete suite; PR description still describes a third complete macOS shard and full-suite timings

Since the last review, the branch picked up the test(ci): focus macOS on native coverage commit, which drops macOS from three full-suite shards down to one shard scoped to TestProfile's curated Darwin-native file list via the new KILO_TEST_PROFILE=darwin env var, and hardened two flaky persistent-process tests (background-process.test.ts now waits for log flush before reload and for a descendant readiness file before reading its pid, watcher.test.ts now also runs under the macOS profile). These test changes look correct and well-justified, but the PR description's narrative and measured-impact numbers no longer match the shipped macOS scope, which is worth reconciling before merge.

Files Reviewed (5 files changed since last review)

• .github/workflows/test.yml
• packages/opencode/script/kilocode/test-profile.ts
• packages/opencode/test/file/watcher.test.ts
• packages/opencode/test/kilocode/background-process.test.ts
• packages/opencode/test/kilocode/test-profile.test.ts

Fix these issues in Kilo Cloud

Previous review (commit 736850e)

Status: No Issues Found | Recommendation: Merge

The previously flagged duplication of the Zig/bwrap sandbox-helper setup between the unit and httpapi jobs has been resolved by extracting the shared steps into the new composite action .github/actions/setup-linux-sandbox, matching the existing ./.github/actions/setup-bun pattern. The matrix identity fields were renamed from a single name/shard string to structured os/index/total fields (propagated consistently through the job name, Turbo cache key, artifact name, and KILO_TEST_SHARD), a third macOS shard (3/3) was added as described in the PR, and the CLI test-ci Turbo task now also keys its cache on the workflow file itself.

Files Reviewed (3 files)

• .github/actions/setup-linux-sandbox/action.yml
• .github/workflows/test.yml
• turbo.json

Previous review (commit 6339835)

Status: No Issues Found | Recommendation: Merge

The previously flagged duplication of the Zig/bwrap sandbox-helper setup between the unit and httpapi jobs has been resolved by extracting the shared steps into a new composite action, .github/actions/setup-linux-sandbox. Both jobs now reference it via uses: ./.github/actions/setup-linux-sandbox, matching the existing ./.github/actions/setup-bun pattern. This change also adds a third macOS shard (macos-3, 3/3) to further balance the macOS critical path, consistent with the PR description.

Files Reviewed (2 files)

• .github/actions/setup-linux-sandbox/action.yml
• .github/workflows/test.yml

Previous review (commit 443d510)

Status: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	1

Issue Details (click to expand)

SUGGESTION

File	Line	Issue
.github/workflows/test.yml	195	Zig/bwrap sandbox-helper setup steps are duplicated verbatim between the unit and new httpapi jobs; consider a composite action to keep them in sync

Files Reviewed (7 files)

• .github/workflows/test.yml
• packages/opencode/script/kilocode/test-shard.ts
• packages/opencode/script/test-runner.ts
• packages/opencode/test/kilocode/background-process.test.ts
• packages/opencode/test/kilocode/cleanup.ts
• packages/opencode/test/kilocode/test-shard.test.ts
• turbo.json

The sharding matrix, TestShard.split/order load-balancing logic, the httpapi job split-out with its required-job gating, the Windows GC-on-first-lock-failure change, and the child-process exit synchronization fix in background-process.test.ts all check out correctly against the CI/turbo config and the PR's stated intent.

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-sonnet-5-20260630 · Input: 38 · Output: 9.9K · Cached: 904K}

_{Review guidance: REVIEW.md from base branch main}

[marius-kilocode]

Final observed pipeline timing

Compared the original reference workflow with the final green profile-trim workflow using the GitHub workflow createdAt to updatedAt timestamps:

Path	Before	Final observed	Saved	Faster
Complete test workflow	18m44s	6m50s	11m54s	63.5%
Linux critical shard	12m24s	4m22s	8m02s	64.8%
Windows critical shard	18m02s	5m59s	12m03s	66.8%
macOS curated profile	7m26s / 137 files	5m52s / 106 files	1m34s	21.1%

Reference: https://github.com/Kilo-Org/kilocode/actions/runs/28947466351

Final run: https://github.com/Kilo-Org/kilocode/actions/runs/29082265682

Linux and Windows still run the complete eligible CLI suite. macOS runs the curated native-sensitive profile and retains all 11 sandbox test files. The 6m50s figure is the latest observed end-to-end workflow time, not labeled as a controlled cold-cache measurement because cache state may differ between workflow runs.