spawn

First run /relay:list — if a live idle session already owns this worktree/branch/topic, use /relay:send on that session instead of spawning fresh (cheaper, keeps context). Only spawn when nothing relevant is idle, the relevant session is dead/stalled, or you're upgrading to a stronger model (a session's model is fixed at launch).

Choosing each executor's model — decide per packet, by judgment demanded and cost of a wrong-but- plausible result, never by file count:

• Haiku-class — mechanical, fully-specified, verifiable-by-command: rename sweeps, boilerplate tests from a shown template, config plumbing, doc formatting. The packet must be tight (exact files, exact acceptance commands). If the packet needs a "why" explained, it's not haiku work.
• Sonnet-class (the workhorse, default) — bounded features, bugfixes with a repro, test suites, refactors within one module. Most packets land here.
• Opus-class — unknown-root-cause debugging, cross-cutting changes, anything touching core logic/ledgers/parity tests, and any task where a wrong-but-plausible result would likely survive your review. Pay for judgment exactly where review can't catch its absence.
• Upgrade signals: two fix-list rounds haven't landed it (respawn stronger + --supersede); ambiguity you can't spec away in the packet. Downgrade signal: your acceptance criteria could be checked by a script.
• Effort rides the same judgment (packet EFFORT: line or --effort): mechanical/ script-checkable → low/medium; leave unset for ordinary work; xhigh for unknown-root-cause or core-logic packets (max when correctness beats cost). Fixed per process, like the model.
• Version hygiene: pass the TIER alias (haiku/sonnet/opus) and let relay resolve the concrete id through this machine's CLI — never type version ids (…-4-6) from memory; a stale id silently pins an old model. relay doctor shows what each alias resolves to here.

Write the task-specific packet content (ROLE / REQUIRED READING / WORK PACKET only — the standing GATES and REPORT FORMAT are in the executor agent's system prompt, and relay appends the per-packet report path / self-diff / closing line) to a file, then run:

${CLAUDE_PLUGIN_ROOT}/bin/relay spawn $worktree $topic $packet --model $model --lead "$CLAUDE_CODE_SESSION_ID"

$worktree is the ABSOLUTE path of the project directory the executor works in — the shared project dir (e.g. /tmp/calc), NOT a per-task or per-module name. Parallel executors on the same project all pass the same worktree; they just touch different files in it. (relay resolves it to an absolute path and refuses to spawn if it isn't an existing directory.)

Pass --lead "$CLAUDE_CODE_SESSION_ID" (bash expands it to your own lead session id) so the executor inherits your ownership — it's stamped with your lead id and project.

(Call relay via ${CLAUDE_PLUGIN_ROOT}/bin/relay — Claude Code substitutes the plugin's absolute path — not bare relay, which often isn't on the Bash tool's non-interactive PATH.)

Relay lints every outgoing packet (advisory ⚠/ℹ lines at spawn and send — e.g. the text mentions Linear but no MCP: line; big reading pinned to 200K; the packet tells the executor to commit or to ask someone). Fix what it flags before the executor burns a turn on it; relay lint <packet> --worktree <dir> runs the same checks without sending.

Succeeding a retired session? Pass --seed <retired_session_id> (or a path to its successor-seed.md) to inherit what that session did on this territory — its packet index and each report's outcome/risk/UNVERIFIED lines get appended to this packet as context, ahead of the GATES. See /relay:retire.

(--model optional — omit for the default. Add --keep if you already know a follow-up packet is coming soon and don't want auto-close to park it in between (it can always be resumed; relay keep <sid> --off unpins). Add --name <label> for a custom session name, or --scope <tag> for the short area tag shown in /relay:list. Add --pane to open this executor as a split pane in your own window instead of a tab (--tab to force a tab), overriding the executor_layout config default; iTerm only, falls back to a tab if your session can't be found. Executors launch with NO MCP servers by default — saves tokens every turn and removes a side-effect surface. If the packet genuinely needs one, declare it IN the packet with a line MCP: linear (strict allowlist of servers configured in ~/.claude.json / the worktree's .mcp.json; comma-separate several) or MCP: inherit (everything, incl. plugin/connector MCPs like Chrome or Gmail). The packet is the source of truth — relay reads the line at spawn and, on a later /relay:send, relaunches the executor's conversation with the wider set if it lacks what the new packet declares. --mcp SPEC on the command line overrides the line. Decide per packet; don't declare inherit by habit. Context window: executors default to the 1M window (executor_default_context, shipped 1m — the window is a ceiling, not consumption, so a bounded packet costs the same at either). To pin one executor down, put CONTEXT: 200k in its packet; haiku always runs 200K. The window can't change on resume — a heavy session is widened by /relay:retire + respawn.)