fix(agent): follow multiple sessions in one process

[Atchyut Preetham Pulavarthi (theonly1me)]

Overview of Changes

Following 32 sub-sessions with separate CLI processes can exhaust the lead pod's memory. Add one command that follows those sessions in one process.

• Add agent get --sessions <ids...> --follow, which returns the first completion, failure, cancellation or question.
• Read snapshots with waitSeconds: 0, one request at a time, and keep the existing session output and exit codes.
• Keep single-session following and interactive answers on their existing path.
• Add behavioral tests for completion, questions, failure, cancellation, timeout, duplicate ids, invalid arguments and cleanup.
• Add a minor package changeset.

Testing

bun run typecheck
bun run lint --max-warnings 0
bun run format:check
bash scripts/check-naming.sh
bun run knip
bun run test
bun run build

All pass locally. The full suite passed 2762 tests with 0 failures. The agent tests passed 62 tests.

• The new tests failed before implementation because the command could not follow the session list.
• In a disposable ai-task image with a 3 GiB memory limit, 32 separate CLI processes reached the limit and caused 3 kernel OOM kills. Their total JSON response bodies were 24.9 KB.
• The combined command used one CLI process for 32 and 50 sessions. Peak container memory was 191 MiB and 199 MiB, with 0 OOM kills.
• The memory test used a local API server and the arm64 ai-task image. It did not include the lead agent's normal workload.
• The Platform draft requires the CLI release that contains this command. Hard process enforcement remains follow-up work.