Review

Verified the diff against the codebase. The mechanics of the change are clean — only the experiment bundle widens, exposure capture and --beta/--fast precedence are unchanged, version bump is correct. But I'd hold this PR until the sandbox volume-mount lands, ideally on the same branch.

Blocker: sandbox ships broken for the primary use case

You flagged this in the PR description, but I want to underscore the consequence: once 1.0.39 is released, every user bucketed into fast_provision=test for spawn claude local (or codex/cursor/kilo/opencode) gets an agent in a container that cannot see their working directory.

packages/cli/src/local/local.ts:503-542 — pullAndStartContainer runs:

docker run -d --name <name> <image>

No -v, no -w. dockerInteractiveSession (lines 544-557) also docker execs without a workdir. Coding agents are the headline use case for spawn, so the test variant would regress them silently for ~50% of unopted users.

There is no in-code gate that says "don't include sandbox until PostHog is flipped on" — once this is merged + auto-updated, the bundle is live as soon as the flag is enabled, and that decision lives outside this repo.

Recommendation: land Option 1 (the -v \"\$(pwd)\":/workspace -w /workspace one-liner) on this branch before merging. That's small enough to keep this as a single coherent PR.

Other concerns (non-blocking)

1. Inconsistency with --fast (packages/cli/src/index.ts:966-968). --fast currently expands to tarball, images, parallel, docker — explicitly not sandbox. The test variant now ships a strictly different bundle than the user-facing fast flag. Either align them (add sandbox to --fast, or drop it from the experiment) or document why they intentionally diverge. The current state means a user who reads the docs and types --fast gets a different experience than a user silently bucketed to test.

2. No unit test for the bundle. feature-flags.test.ts only covers getFeatureFlag's return values. There's no test that asserts variant === \"test\" results in SPAWN_BETA containing images,docker,sandbox. A small spy on betaFeatures after main() would prevent regressions when someone tweaks this list later.

3. docker + sandbox interaction is untested. docker is a cloud-side flag (Hetzner/GCP host image); sandbox is local-only. They shouldn't conflict in practice (different cloud paths), but the experiment now bundles both and we have no test that exercises the combination on either path. Low risk, but worth a sentence in the PR body confirming you've reasoned through it.

Approving criteria

• Add -v + -w to pullAndStartContainer (and consider dockerInteractiveSession) OR drop sandbox from the bundle for now
• One unit test asserting the test-variant bundle composition
• Brief note on --fast vs experiment-bundle divergence

Once the sandbox path actually works for coding agents, this is a clean ship.

Filed from Slack by SPA

Follow-up commit pushed: 8f87330

Acknowledged the cwd-isolation-is-intentional clarification — dropped that concern. Pushed a follow-up commit addressing the three non-blocking items from my review.

What changed

1. Aligned `--fast` with the experiment bundle (`packages/cli/src/index.ts`)

Before: `--fast` → `[tarball, images, parallel, docker]`, experiment test variant → `[images, docker, sandbox]`. Sandbox was reachable only via the silent A/B or via `--beta sandbox` directly, not via the explicit `--fast` opt-in.

After: `--fast` → `[tarball, images, parallel, docker, sandbox]`. The explicit opt-in now exercises the same surface as the experiment plus the speed-ups outside the experiment's scope. Comment updated to note the alignment is intentional and to keep both bundles in sync.

2. Extracted bundle into a testable helper (`packages/cli/src/shared/feature-flags.ts`)

New `expandFastProvisionVariant(variant: string): readonly string[]` lives next to `getFeatureFlag`. `index.ts` now does:

const variant = getFeatureFlag(\"fast_provision\", \"control\");
const variantStr = isString(variant) ? variant : \"control\";
betaFeatures.push(...expandFastProvisionVariant(variantStr));

The `isString` narrowing is required because flags can be string OR boolean. Pure function, drop-in replacement, behavior unchanged for the test variant.

3. Added unit tests pinning the bundle (`packages/cli/src/tests/feature-flags.test.ts`)

expandFastProvisionVariant(\"test\")    → [\"images\", \"docker\", \"sandbox\"]
expandFastProvisionVariant(\"control\") → []
expandFastProvisionVariant(\"\")        → []
expandFastProvisionVariant(\"rollout\") → []

Guards against silent drift if someone tweaks the list. Was 11 tests in the suite, now 14.

4. Docker + sandbox interaction (verified, no code change needed)

Confirmed by reading the cloud entry points: `hetzner/main.ts` and `gcp/main.ts` only check `betaFeatures.includes("docker")`; `local/main.ts` only checks `betaFeatures.includes("sandbox")`; `digitalocean/main.ts` only checks `"images"`. A given `spawn ` run picks exactly one cloud, so the flags never combine in the same execution path. The "untested combination" concern is moot — they're orthogonal.

5. Version bump: 1.0.39 → 1.0.40, per `.claude/rules/cli-version.md`.

Verification

• `bunx @biomejs/biome check src/` → clean
• `bun test` → 2188/2188 pass (was 2185 + 3 new bundle tests)
• `bun test src/tests/feature-flags.test.ts` → 14/14 pass

Diff is +59/-7 across 4 files.

Filed from Slack by SPA

Reverted --fast change in 12f7063

Misread your scope on the prior commit — pushed a revert.

`--fast` is back to `[tarball, images, parallel, docker]` (unchanged from before). The experiment bundle in `expandFastProvisionVariant` stays at `[images, docker, sandbox]` and is reachable only via the silent `fast_provision` flag bucketing or via explicit `--beta sandbox` / `--beta docker` / `--beta images`.

What survives from the prior commit:

• `expandFastProvisionVariant(variant)` helper in `shared/feature-flags.ts`
• 3 unit tests pinning the bundle composition (test → 3 features, control/unknown → empty)
• `index.ts` calling the helper instead of inline-ing the list

Updated the helper's docstring to say it's experiment-only, no relation to `--fast`.

Bumped 1.0.40 → 1.0.41. Biome clean, 2188/2188 tests pass.

Filed from Slack by SPA