Sidekick orchestration

When delegation is useful, first run sidekick doctor ENGINE and sidekick capabilities ENGINE --model MODEL --json. Then use sidekick spawn ENGINE NAME --dir PATH -- "PROMPT", passing separate controls only when the evidence matrix marks them native: --model, --provider, --transport, --effort, --permission, --sandbox, and --workspace-trust. Treat --mode as legacy compatibility input, never as a universal permission/effort/sandbox field. Start one background sidekick wait NAME... --all --json for the batch and do useful independent work. Harvest through sidekick result NAME --json; require terminal status and inspect exitCode. Use sidekick send for follow-ups, adopt for existing sessions, tail for live logs, and cancel to stop the process tree without discarding the resumable session. send does not accept control overrides because the persisted control snapshot is immutable. Never busy-poll.

Controls are persisted as separate requested, applied, and effective observations. An argv flag does not prove provider-reported effective state. Unsupported, malformed, model-dependent-unverified, or runtime tool-version-mismatched explicit requests fail closed with exit 2.

Codex effort is value-gated for any model: --effort minimal|low|medium|high|xhigh|max becomes -c model_reasoning_effort=.... Sidekick does not maintain a curated Codex model allowlist. The selected model/provider may still ignore a level it does not support. Prefer medium as a default balance; use max only for the hardest quality-first reviews. Devin family slug glm-5.2 and variant UID glm-5-2 must not be silently aliased. Do not use --mode high as effort.

Do not concurrently spawn workers with the same --dir; harness session discovery and workspace state can be directory-scoped, causing contention, serialization, or overlapping edits. Use a separate git worktree and --dir for each parallel worker. Otherwise serialize the work and continue the existing run with sidekick send.