docs(architecture): define agent-managed compute plan

[drewstone]

Summary

• stress-test the current coordination MCP, supervisor, provider adapter, journals, and public run APIs against distributed agent-managed compute
• centralize the active audit, architecture, failure behavior, roadmap, and validation plan in docs/agent-managed-compute/
• correct claims that completed-tree replay already provides coordinator restart recovery
• document the acyclic knowledge/runtime package direction and required dependency alignment

Key findings

• coordination works locally, but its event state is process-local and the remote HTTP endpoint is unauthenticated
• supervised trees restart from the root instead of recovering live work
• delegation idempotency is local to one queue instance
• safe failover also requires provider command fencing, not only a database ownership claim

Checks

• pnpm run build
• pnpm run typecheck
• pnpm run lint
• pnpm run docs:check
• pnpm exec vitest run tests/loops/coordination-mcp.test.ts tests/loops/supervisor-agent.test.ts tests/loops/coordination.test.ts tests/loops/event-bus.test.ts tests/mcp/task-queue-durable.test.ts src/runtime/environment-provider.test.ts (48/48)
• local link validation for all six new documents
• no em dash or en dash in docs/agent-managed-compute/

[tangletools]

✅ Auto-approved drewstone PR — 2f2ca0cf

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-15T18:16:41Z}