BudgetLedger interface + InMemoryBudgetLedger reference impl (reserve/settle/
release with refund of unused holds; mirrors the b1dz budget+ledger concept).
withBudget() wraps any SwarmRunner: reserves up front (throws SwarmError 402 when
out of funds, before running), settles actual cost on success, releases on error.
AgentIdentity DID aligns with @logicsrc/agentstack.
8 new tests; 27/27 pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
createSwarm() composes named agents into one SwarmRunner: a router picks the
first agent; any agent hands control to a named peer via `HANDOFF: <name>`,
bounded by maxHandoffs. createLLMRouter() is a cheap-model router. Composes
with the rubric loop and HTTP handler since a swarm is itself a SwarmRunner.
6 new tests (routing, handoff, maxHandoffs cap, unknown-peer/unknown-route
fallbacks, empty-agents guard); 19/19 pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@logicsrc/agentswarm: a self-hosted, framework-agnostic multi-agent runtime
that each app mounts on its own route (e.g. tronbrowser.dev/swarm).
M1 — core:
- createSwarmHandler(): Web (Request)=>Response handler, CORS + validation
- SwarmRunner: injectable engine interface
- createDeepAgentRunner(): deepagents (createDeepAgent) adapter; deepagents +
@langchain/langgraph are optional peers loaded via dynamic import so the core
builds/tests with zero heavy deps
- SwarmError + onRequest gate: seam for x402 metering / auth (402/403)
M3 — rubric self-check (deepagents RubricMiddleware is Python-only, so ported
at the runner layer):
- createRubricRunner(): grades output via an injectable judge and revises until
it passes or maxIterations; passes through untouched when no rubric
- createLLMJudge(): cheap-model grader (lazy langchain initChatModel)
13/13 vitest pass; tsc clean; examples/server.mjs demo verified.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
