bunqueue v2.9.15
Fixes 4 bugs. Agent-oriented MCP inspection, structured MCP contracts, agent preset and readiness checks, optional bunqueue/agents entrypoint and more.
To install bunqueue v2.9.15
Section titled “To install bunqueue v2.9.15”bun add bunqueue@2.9.15npm install bunqueue@2.9.15TAG=2.9.15docker run -p 6789:6789 -p 6790:6790 ghcr.io/egeominotti/bunqueue:$TAGTo upgrade bunqueue
Section titled “To upgrade bunqueue”bun add bunqueue@latestAgent-oriented MCP inspection
Section titled “Agent-oriented MCP inspection”bunqueue_diagnose_queue accounts for all
queue states and reports observed limits and blockers, bunqueue_observe_job
distinguishes completion, failure and unknown jobs with bounded polling, and
bunqueue_list_jobs_page provides bounded offset pages with explicit projection
and truncation metadata. Scoped queue/job resource templates avoid broad reads.
Structured MCP contracts
Section titled “Structured MCP contracts”Registered tools expose policy annotations,
output schemas and structured outcomes alongside existing text responses.
Dynamic discovery includes complete descriptors, and
bunqueue_get_capabilities reports effective server settings and tools.
Agent preset and readiness checks
Section titled “Agent preset and readiness checks”BUNQUEUE_MCP_PROFILE=agent enables
dynamic discovery and destructive-operation confirmation when those settings
were not explicitly supplied. bunqueue-mcp --doctor --json checks configuration,
paths and TCP readiness without starting consumers or opening workflow SQLite.
Optional bunqueue/agents entrypoint
Section titled “Optional bunqueue/agents entrypoint”Provider-neutral turn consumption, a Claude message adapter and an independent SQLite journal support durable session checkpoints, bounded progress replay and stable external-effect keys. Pending effects require reconciliation before a retry; provider artifacts and cooperative cancellation remain application responsibilities.
Executable integration guidance and offline evaluations
Section titled “Executable integration guidance and offline evaluations”Claude, OpenAI Agents and Mastra examples use caller/CLI model configuration. Five real-MCP scenarios run against embedded and TCP backends with independent state checks. Catalog measurements include schemas and annotations; deterministic task scores are not live-model success rates. Workflow/agent skills and public API references now expose the integration contracts.
Compatibility
Section titled “Compatibility”- Queue scheduling, queue persistence and workflow engine behavior remain unchanged. Existing MCP listing/wait tools and text results remain available. The agents helper is optional and Bun-only, imports no provider SDK and uses its own database.
Bugfixes
Section titled “Bugfixes”- MCP diagnostic prompts omitted prioritized, paused and dependency-blocked work from their accounting and could imply an unobserved cause. They now report complete counts, observed limits and bounded samples.
- MCP write errors after a potentially committed operation could look safely retryable. Structured errors preserve uncertainty and require reconciliation; pre-dispatch refusals remain distinguishable from attempted writes.
- Agent examples could silently accept provider errors or interrupted streams,
share effect state between concurrent runs, or replace an explicit structured
nullresult with text. Typed terminal outcomes, per-run effect identity and output validation now preserve those contracts. - Documentation discovery validation accepts the independently validated
llms-full.txtresource while continuing to reject non-canonical page links and unknown text resources.
Thanks to 1 contributor!
Section titled “Thanks to 1 contributor!”← bunqueue v2.9.14 · Release on GitHub · Commits since v2.9.14 · Full changelog · All posts