0180. Swarm owns cross-session coordination

Status: superseded by ADR 0213.

Clankie needs durable peer messages and task ownership across independent coding hosts. Terminal input and pane status describe a running process but cannot establish processing acknowledgment, task ownership or a restart-safe handoff.

Swarm MCP v2 is the coordination owner. Clankie embeds it as a pinned package; its implementation stays in the swarm-mcp repository. The Pi operator runtime mounts the nine tools per conversation and delivers leased envelopes through the existing internal turn queue. Stable conversation identities survive a service restart; each incarnation receives a new fenced session. The native Claude seat uses the selected conversation's tools (global by default) through Clankie's MCP server and receives envelopes through the plugin channel. Worker spawning uses a trusted Herdr route whose token is persisted before creating a pane. Uncertain provisioning retains its task and capacity; it never authorizes a second launch. Herdr's native layout.apply replaces only the newly created tab with a direct argv process, so interactive shell startup cannot consume the worker command. This existing runtime API avoids a Herdr fork or shell-input readiness heuristics. The worker publishes its actual pane ID in a private startup receipt; binding also requires its authenticated runtime observation. A lost response is reconciled against that receipt instead of replaying the launch.

lead holds shared leadership judgment; swarm-lead and herdr-lead select the coordination workflow. Every worker uses the swarm-mcp protocol skill. Herdr continues to own terminals, inspection and process control. Its leadership workflow is an explicit fallback for unenrolled fleets, never a second delivery path for an uncertain Swarm assignment. Native subagents remain appropriate for bounded work inside one host. A configured work tracker holds the durable human-facing deliverable record; it is independent of the coordination transport.

One private coordinator state directory is selected per canonical repository under $CLANKIE_STATE/swarm. Worktrees share their repository's scope and each conversation retains its own actor. The configured profile is clankie. The initial built-in route provisions Claude workers in the requester's worktree, with a capacity of four. Other hosts can participate through their supported Swarm integrations; host availability and delivery support remain explicit. The service and owned stream workers wake from durable inbox events without a model polling loop. A channel-enabled native seat receives the global conversation's wakes; plugin-dir seats have tools but no channel delivery.

The owner outlives the service connection. A service restart reconnects and fences its prior sessions, preserving worker work. Cooperative cancellation waits for the worker's fenced terminal outcome; closing a pane does not prove all descendants stopped. Tool availability remains limited to operator conversations and the authenticated native operator seat; this adds no authority to social Discord lanes.

The repository pins packed dependencies because the public npm package is the legacy release. vendor/README.md records source provenance and regeneration; release packaging includes the dependency graph and bundled skills. The skill source remains in the skills repository, and checkout links resolve through the installed package. No live global MCP configuration is rewritten.

Managed harness amendment — VUH-1407 (2026-09-27)

Local runtime settings now select Claude, Codex or pi independently of worker mode. Codex defaults to gpt-6-astra and uses its app-server protocol. Pi uses RPC with a worker-specific extension that projects its enrolled MCP tools; the operator extension remains conversation-bound. Both reuse the existing managed claim, leased inbox, explicit ack, progress/lease, cancellation and stop protocol. Only an actual worker Swarm tool call establishes readiness. Launch arguments and environment carry configuration and capabilities; global harness settings are untouched. Runtime selection is exposed through the operator API, CLI and TUI, with typed refusals for unsupported harness/mode pairs and no fallback.

The resolved harness is persisted in the intent and launch receipt. Retargeting a route cannot adopt an existing launch. Codex/pi initially support stream; interactive support remains the independent ADR 0194 axis. Remote PC peers keep the shared coordinator and relay. Protocol tests are separate from live canary proof: each installed harness must finish a managed task and release its intent.

The repaired launcher preapproves only swarm_inbox and swarm_task on its own enrolled Swarm server, as explicitly authorized for unattended lifecycle work. Other MCP and shell approval policies remain unchanged. Managed pi selects OpenRouter moonshotai/kimi-k3 explicitly. A disabled retained route still owns stop authority for its verified tokens; disabling new provisioning must not strand old dispatch capacity. Production package tests import the pi extension, and a manifest-to-lockfile check prevents same-version tarball updates from retaining an obsolete dependency snapshot.