claude-sdk-rs: Command for Claude Code

Orchestrate — Run an ordered chain of blocks through the SDLC engines, in one session.

Takes an ordered list of work blocks and drives each one end-to-end: spec → (breakdown) → engine → integrate → verify state → next. The engines run as background workflows, so while one block builds, this command prepares the specs for the blocks behind it.

One /orchestrate session drives one repo. Run several repos at once — that is the lane model.

Variables

$ARGUMENTS — one of:

• Inline list: ordered block IDs or spec slugs, space- or comma-separated. /orchestrate OK.3.A OK.3.B
• Lane-file path: a single lane-<name>.json path, authored against .claude/workflows/lane.schema.json (D71). Read its blocks[] array — array order IS chain order — and take each entry's id. The argument stays a single path; there is no second argument for a lane file. /orchestrate planning/bullet-proof-software/lane-okf-core.json
• Flags:
  • --worktree — require worktree isolation for every block in the chain. See step 5.
  • --no-worktree — force plain-branch/in-place for every block, overriding any per-repo default.
  • --engine <task|flow> — force one engine for every block, overriding recommendations.
  • --dry-run — resolve the chain, generate the specs, print the plan. Run no engines.
  • --stop-on-fail (default) / --continue-on-fail.
  • --stop-after <N> — stop the chain cleanly after N blocks have integrated, releasing the repo lease and registry claim exactly as at ordinary lane close (begin-orchestration.md's lane-close release), then reporting the remaining chain the way an early stop already does. Nothing today stops a lane at all — that gap is why the last run stopped for an ad-hoc reason (a stale global-command snapshot, base-template standing rule 10) instead of a decision. This flag and --autonomy below are the two honest, nameable stop conditions: "stop after this many blocks" and "stop when this needs a human." These are flags, not lane-record fields, because they describe this run, not the lane: .claude/workflows/lane.schema.json is validated by mev's LaneRecord, which is deny_unknown_fields — every field added to it is a cross-repo add-then-install before any lane can read it, the exact cost base-template:BT.ticket.lane-schema-has-no-home-for-the-briefing already paid. A per-run knob belongs on the invocation, not on data a sibling lane's tooling must also parse.
  • --autonomy <level> — how far this lane may go without the operator before stopping. Same per-run reasoning as --stop-after: it governs this invocation, not the lane, so it is a flag. This command does not define the level vocabulary; take whatever the operator passes and stop at the next point rule 11's "not yours to decide alone" list would otherwise require a call.