Every release, and what it changed.

Durable approval and recovery

• Durable approval gates can suspend plan, round, task-dispatch, and tool-call decisions, persist content-bound pending requests and first-wins reviewer decisions in a checkpoint MemoryStore, and resume only the approved plan, task, or validated tool invocation. Missing state, stale hashes, schema drift, and tampering fail closed, while derived execution receipts carry review evidence without making telemetry authoritative. New checkpoint writes use schema version 4 to preserve approval continuation state.
• Mid-turn checkpoint recovery now preserves built-in runner messages, turn and token accounting, pending tool calls, and independently committed tool results. Restore replays committed results without re-executing their tools, reruns only missing calls, and exposes the stable model-issued toolCallId to tools and per-call gates for external idempotency.

Structured input and conversation history

• Agent.run(), Agent.stream(), and OpenMultiAgent.runAgent() accept complete structured message histories.
• Persistent Agent.prompt() conversations can accept one structured user turn, including image input, with runtime validation, defensive copies, and full-message access in beforeRun.
• AgentConfig.history restores persisted conversation messages into a new Agent so later prompt() calls can continue an earlier conversation.

Rich tool results

• Rich tool results separate application-owned ToolResult.data from validated, model-visible text, image, and file modelOutput.
• Built-in adapters and MCP map supported content explicitly, while runner context, checkpoints, recovery, progress, evaluation, and privacy-safe trace summaries preserve the richer result.

Compatibility

• Existing string inputs, string/error tool results, and compression markers keep their previous behavior. Structured input and rich tool output are additive; adapter- and model-specific media limits still apply.
• Checkpoint readers remain compatible with version 1, version 2, and version 3 snapshots. OMA cannot atomically commit an arbitrary external side effect together with its checkpoint write, so consequential integrations still need external idempotency for that crash window.
• runTeam() and runTasks() remain text-only. Process and ACP backends remain task-grained and reject structured input rather than silently dropping it.
• Existing approval callbacks can continue returning allow or deny. Suspension is additive; task gates with live verification wiring, standalone runAgent(), the simple-goal shortcut, process backends, and ACP backends do not expose every resumable private tool-loop path.

Packages

• @open-multi-agent/core: 1.15.0
• create-oma-app: 0.8.0; generated starters pin core 1.15.0
• @open-multi-agent/otel: remains at 0.1.1 and is not republished; its @open-multi-agent/core@^1.11.0 dependency remains compatible

Install

npm install @open-multi-agent/[email protected]
npm create oma-app@latest my-oma

Breaking changes

This release shipped as a minor version, so a caller on a ^1.x range receives it without an explicit upgrade step. npm reports an engines mismatch as an EBADENGINE warning rather than an install failure, so a Node 18 project installs 1.14.0 successfully and then fails at run time.

• Node.js 20 or newer is now required across @open-multi-agent/core, @open-multi-agent/otel, and create-oma-app. Node 18 reached end of life on 2025-04-30.
• The bundled openai dependency moved from v4 to v6. A project that also depends on openai directly resolves a nested second copy until it moves to v6 as well.
• Plans that were previously accepted can now fail at validation time. Invalid task dependency graphs, unsatisfiable task requirements, and invalid coordinator plans are rejected before execution instead of running partially. Each of these surfaces a defect that was previously silent, so correct graphs and rosters are unaffected. The individual entries are under Correctness and safety below.

Adaptive plan recovery

A run can now revise the not-yet-executed part of its task graph instead of carrying a plan that stopped fitting the work.

• Opt in with recovery.mode: 'repairable', then supply a Replanner or an onTaskOutcome callback that proposes an append-only PlanPatch.
• Patches are validated for agent eligibility, limits, task states, references, and the resulting DAG, then gated through the optional onPlanPatch approval.
• Application happens atomically at a task-outcome barrier, before downstream dispatch or failure cascade, so no original dependent starts against a stale plan.
• Revision history is exposed in results, progress events, and observability spans, and persisted through checkpoints with rollback when a checkpoint write fails.

See adaptive recovery.

Execution routing

• Added opt-in hybrid semantic routing through executionRouting: { strategy: 'hybrid' }. Deterministic Team decisions stay authoritative; deterministic Single decisions receive one structured semantic assessment.
• Added a provider-neutral TaskProfiler interface, a built-in LLMTaskProfiler, a strict task-profile schema, typed routing failures with timeout and fallback metadata, and semantic-routing observability.
• Automatic runTeam() routing remains deterministic, so existing runs are unchanged.

Provider support

• DeepSeek V4 Flash reasoning controls: AgentConfig.thinking.enabled maps to DeepSeek’s native thinking.type, and thinking.effort accepts the DeepSeek-only value 'max' without forwarding it to OpenAI, Azure OpenAI, or GitHub Copilot.
• Refreshed DeepSeek model guidance, retired aliases, examples, CLI reference, and scaffold defaults.

Correctness and safety

• Task requirements are enforced as global hard constraints across every scheduling strategy. A task whose requiredTools, requiredCapabilities, requiredBackend, or requiredProvider no agent satisfies is rejected instead of being dispatched to an ineligible agent.
• Invalid task dependency graphs are rejected up front instead of executing a partially valid plan.
• The coordinator fails closed on an invalid plan rather than continuing with a plan it could not validate.
• OpenAI user aborts are classified as cancellation rather than as a retryable failure, and unsupported custom tool calls are rejected explicitly.

Compatibility

• Every public export from 1.13.0 is still exported, and new result and configuration fields remain optional.
• Adaptive recovery adds a version 2 task-queue snapshot. fromSnapshot() still accepts version 1, so checkpoints written by earlier releases remain restorable.
• @open-multi-agent/[email protected] is republished to carry the Node 20 engines floor. Its @open-multi-agent/core@^1.11.0 range is unchanged.
• [email protected] pins generated starters to core 1.14.0.

Thanks

• HarperZ9 contributed the mid-task recovery conformance fixture in #452, which pins the idempotency outcomes of four recovery cases against the current task-level checkpoint granularity.

Execution routing and governance

• Added pluggable execution routing for runTeam() through explicit mode, custom ExecutionRouter implementations, and the built-in DeterministicRouter.
• Added structured governance declarations for required or preferred roles, ordered review paths, budget-aware degradation, and post-execution governanceConclusion checks.
• Added privacy-preserving execution receipts and routing-decision trace linkage.
• Consequential tools can now be declared with ToolDefinition.consequential. Undeclared runs expose a machine-readable disclosure flag and can require confirmation through onToolCall.
• Automatic routing now recognizes structured Chinese, Japanese, and Korean goals using script-aware information length.

Scheduling and task execution

• Task DAG execution is event-driven by default: downstream work starts as soon as its dependencies complete.
• Added configurable dependency-first, round-robin, least-busy, capability-match, and weighted composite scheduling strategies.
• Agents and tasks can declare structured capabilities and hard requirements, with optional strict assignee validation.
• TeamRunResult.taskResults preserves task-scoped results by task ID.
• Explicit tasks can pass raw, structured, or combined dependency payloads with bounded role and provenance metadata.
• Added OrchestratorConfig.onTaskDispatch(task) for native per-task pipeline approval.

Model routing and reliability

• Model routes can declare ordered fallbacks for retryable worker-provider failures.
• Fixed CJK keyword extraction and zero-score fallback so eligible agents are selected consistently.
• Governed planOnly runs now return the declared role DAG without executing agents.
• Execution modes, governance floors, and per-run budget conflicts now follow a documented precedence order and disclose overrides or degradation.

Compatibility

• Existing onApproval integrations retain legacy round scheduling and callback semantics.
• Progress events from independent DAG branches may now interleave. Consumers should correlate events by task ID instead of relying on round adjacency.
• Raw dependency output remains the default. Structured handoffs, governance declarations, consequential confirmation, and custom execution routing are opt-in.
• New result fields remain optional for compatibility with older serialized results and caller-authored fixtures.

Packages

• @open-multi-agent/core: 1.13.0
• create-oma-app: 0.6.0; generated starters pin core 1.13.0
• @open-multi-agent/otel: remains at 0.1.0 and is not republished

Install

npm install @open-multi-agent/[email protected]
npm create oma-app@latest my-oma

Thanks to @LambIessz for the retryable model-route fallback in #424.

Evaluation V1 (#403–#409)

• New @open-multi-agent/core/eval and /eval/file entry points.
• Define reusable Scorer implementations and versioned EvalSet fixtures.
• Run deterministic offline evaluations with reports, stores, gates, CLI commands, and reference scorers.
• Sample and score production runs online without changing the business result.

Offline Run Viewer (#394, #411)

• Open saved results and traces locally without a running service.
• Inspect task-level model, provider, token, and cost information through the viewer and CLI.

Per-run metadata (#396)

• Top-level run APIs accept bounded metadata that is propagated into results, traces, and checkpoint restores.

Zero-key onboarding (#414)

• [email protected] can install and run a deterministic Demo without API keys or model requests.
• Demo output clearly discloses simulated model responses and produces Markdown, JSON, and HTML reports.
• New --no-install and --no-run flags support controlled scaffolding workflows.

Examples catalog (#412)

• Added a machine-readable examples catalog, schema, and coverage validation.

Fixes (#390, #417)

• The process backend cleans up descendant processes after the parent exits.
• v1.12.1 fixes installed oma binaries so the CLI launches correctly through npm-created symlinks. This was found during the v1.12.0 post-publish verification.

Compatibility

• Existing run modes remain compatible.
• Evaluation is disabled by default, and online evaluation does not change the business result.
• Scorer failures are recorded as scorer_error and excluded from score aggregates instead of being counted as zero.
• @open-multi-agent/[email protected] is not republished and remains compatible with core 1.12.1.
• #397, #400, #401, and #402 are behavior-preserving internal extractions, not new public APIs.

Install

npm install @open-multi-agent/[email protected]
npm create oma-app@latest my-oma

Features

• Observability v2: trace spans and run identity (#371, #373 by @JackChen-me). Every run now carries a stable runId, attempt, traceId, and rootSpanId, and emits the new TraceRecord v2 schema: a proper span tree covering run, coordinator, task, agent, LLM, tool, retry, delegation, consensus, and checkpoint, with DAG, synthesis, and restore relationships expressed as links. The existing seven-field onTrace callback keeps working unchanged. See docs/observability.md.

• Observability v2: sink and exporter lifecycle (#374 by @JackChen-me). The new @open-multi-agent/core/observability subpath ships the public TraceSink / TraceExporter contract plus BatchingTraceSink (bounded queue, batched export, backoff retries, priority-aware drop), CompositeSink, FilteringSink, SensitiveDataProcessor, and LegacyCallbackTraceSink. Tracing stays metadata-only by default: prompts, completions, tool payloads, credentials, and reasoning content are never captured.

• Observability v2: trace stores (#375, #384 by @JackChen-me). A storage-independent TraceStore contract (append, get, query, delete, retention) with two references: the zero-dependency InMemoryTraceStore, and FileTraceStore behind @open-multi-agent/core/observability/file, an append-only NDJSON store with crash-safe compaction and explicit flush, close, and diagnostics. FileTraceStore targets local development, tests, CLIs, and modest single-process services.

• New package: @open-multi-agent/otel 0.1.0 (#376, #385 by @JackChen-me). A separately installable adapter that maps TraceRecord v2 spans onto OpenTelemetry. Its only peer dependency is @opentelemetry/api ^1.9.0 (npm 7+ installs it automatically), it never initializes or replaces the global TracerProvider, and it versions independently of core. It requires core 1.11.0; core 1.10.0 does not contain the v2 APIs. Core itself remains installable and runnable without any OpenTelemetry packages. See docs/observability-migration.md.

• Per-call tool gate (#377 by @scarab-systems). Optional onToolCall on AgentConfig / OrchestratorConfig runs after Zod validation and before execution, returning allow or deny per call. A deny becomes a normal error ToolResult (never a throw), and a throwing or invalid gate fails closed. Documented with a runnable risk-gated bash example in #383. See docs/tool-configuration.md.

• Generic process backend (#378 by @scarab-systems). The new @open-multi-agent/core/process subpath runs a local command as an agent: protocol-neutral, with stdin or argument prompt delivery, cancellation, and redacted stderr on failure. It complements the existing protocol-aware ACP backend. See docs/external-agents.md.

• Bash command risk classifier (#382 by @JackChen-me). classifyBashCommand behind the @open-multi-agent/core/classifiers subpath returns a safe / review / high level with a reason, splitting compound commands and taking the highest risk. It is a dependency-free heuristic meant to pair with the tool gate, explicitly not a security boundary.

• create-oma-app 0.4.0: production starters (#370 by @JackChen-me). npm create oma-app@latest now offers production-ready pr-review and security starters alongside demo, with cloud or Ollama provider selection. Starters generate Markdown, JSON, and OMA DAG dashboard reports, analyze GitHub strictly read-only, and redact secrets before model analysis.

• Adaptive customer support recipe (#386 by @JackChen-me). A new cookbook example where runTeam engages only the specialists relevant to each ticket instead of the full roster.

Fixes

• Cancelled and timed-out runs no longer report success (#371 by @JackChen-me). Cancellation and whole-run timeout paths could previously surface as successful results; they now report their real outcome through the normalized status.

Compatibility notes

• onTrace remains fully supported and is not deprecated. LegacyCallbackTraceSink can now be configured directly as a sink, preserving the seven legacy event shapes without also wiring onTrace.
• Results now always include identity and status. The new fields stay optional in TypeScript for this release, and success is still present, derived from status.code === 'ok'.
• Dependency ownership is now expressed per package: OpenTelemetry packages belong exclusively to @open-multi-agent/otel, and core must stay importable without them.

Install

npm install @open-multi-agent/[email protected]
npm install @open-multi-agent/[email protected]   # optional OpenTelemetry adapter, requires core 1.11.0

New project: npm create oma-app@latest (scaffold updated to [email protected], which pins core 1.11.0).

Thanks to @scarab-systems and @JackChen-me.

Features

• External coding agents over ACP (#360 by @JackChen-me). OMA can now orchestrate an external coding agent over the Agent Client Protocol as a first-class team member, alongside LLM agents in the same task DAG. Declare an agent with backend: { kind: 'acp', ... } and it runs a local coding CLI (for example Claude Code via the official @agentclientprotocol/claude-agent-acp adapter) instead of an LLM, sharing the same shared memory, cascade-on-failure, and token budget as any other agent. See docs/external-agents.md.

• Orchestrator cost budget (#358 by @LambIessz). New maxCostBudget and estimateCost on orchestrator config let you set a user-defined spend cap. The estimator receives the effective model, provider, phase, and task id so you can price per model, and the cap is enforced across runAgent, runTasks, runTeam, consensus, and synthesis, mirroring the existing token-budget boundaries.

• Per-agent scoped tool credentials (#362 by @JackChen-me). New AgentConfig.credentials is a per-agent secret bag threaded into ToolUseContext.credentials. Tool code reads context.credentials?.SEARCH_API_KEY instead of closing a single shared secret over every tool, so each agent holds only the credentials it was assigned. A compromised or misbehaving subagent no longer inherits the coordinator’s full access.

• Secret redaction on the persistence path (#359 by @JackChen-me). Redaction previously stopped at the observability layer, so a secret an agent emitted into its answer was scrubbed in traces but written in the clear to shared memory and any on-disk checkpoint. The new RedactingStore, a MemoryStore decorator, redacts values at the single choke point every shared-memory and checkpoint write passes through, and is structure-aware for JSON so checkpoint snapshots stay valid with only secrets masked. Closes #339.

• MessageBus persisted in checkpoints (#363 by @LambIessz). Checkpoint and restore now carry MessageBus state (messages and per-agent read position), so a resumed run keeps its inter-agent message history. Live callback subscribers stay process-local and are not serialized. Closes #343.

• Trace span parent linkage (#354 by @tlysanhuo). Every trace event now carries a spanId and optional parentId, with stable agent span ids so LLM and tool spans point to their owning agent span, worker agents link under their task span, and plan_ready links under the coordinator decomposition. A team run is now reconstructable as a proper span tree. Closes #340.

Fixes

• Windows: kill the full process tree on bash timeout (#357 by @Bobuyoucrypto). A timed-out bash command on Windows now kills its full process tree with taskkill /T /F and returns promptly, instead of leaving background children alive and hanging until they exit on their own. Exit codes 124 (timeout) and 130 (abort) are reported authoritatively, and the suite now passes on Windows.

Install

npm install @open-multi-agent/[email protected]

New project: npm create oma-app@latest (scaffold updated to [email protected], which pins core 1.10.0).

Thanks to @LambIessz, @tlysanhuo, @Bobuyoucrypto, and @JackChen-me.

Features

• Durable memory: FileStore (#347 by @JackChen-me). A zero-dependency, filesystem-backed MemoryStore, so checkpoint/resume survives a process restart out of the box. Writes are atomic (temp file, fsync, rename) and reads come from an in-memory mirror; a corrupt state file fails loud instead of silently. Until now the only bundled store was InMemoryStore, so durability meant writing your own.
• Error-aware task retry (#346 by @JackChen-me). Retry (opt-in via maxRetries > 0) now classifies failures: isRetryableError() skips provably-terminal errors (most 4xx, token-budget, aborted calls) instead of burning attempts, while 429, 5xx, network blips, and per-call timeouts still retry. Adds equal jitter and honors abort signals, on both streaming and non-streaming paths.
• Per-call LLM timeout (#344 by @JackChen-me). New opt-in AgentConfig.callTimeoutMs (and on CoordinatorConfig) bounds a single adapter.chat() call, so one stalled request no longer hangs the whole run. Uniform across every adapter and surfaced as LLMCallTimeoutError, distinct from a deliberate abort or a real API error. Unset preserves current behavior exactly.
• Run-level metrics on TeamRunResult (#345 by @lesbass). TeamRunResult now carries a metrics rollup (total tokens, retries, error/failure/completed counts, and task-latency aggregates) computed from per-task data, and the dashboard renders a matching summary bar.
• Vercel AI SDK 7 support (#348 by @JackChen-me). The optional ai peer range now spans ^5 || ^6 || ^7, so you can use the AI SDK bridge on the latest release without peer-dependency warnings. No adapter code changes. The AI SDK 7 bridge needs Node >= 22 (core stays >= 18).

Docs

• Plan preview and replay guide, surfaced from both READMEs (#349 by @JackChen-me).

Install

npm install @open-multi-agent/[email protected]

New project: npm create oma-app@latest

Thanks to @lesbass and @JackChen-me.

Patch release with two install/runtime fixes.

Fixes

• defaultModel now reaches every agent, not just the coordinator. Previously OrchestratorConfig.defaultModel only applied to the coordinator pass. Workers, runAgent, delegated, and consensus agents never inherited it, so it was effectively dead for executing agents. They now inherit defaultModel the same way they already inherited defaultProvider, defaultBaseURL, and defaultApiKey. AgentConfig.model becomes optional in an orchestrated run; a standalone new Agent() still requires an explicit model and throws a clear error if it is missing. (#323)
• Vercel AI SDK v6 is now supported. The optional ai peer range widened from ^5.0.0 to ^5.0.0 || ^6.0.0, so installs that pair the framework with AI SDK v6 (for example via @ai-sdk/react@3, which pulls ai@6) no longer fail with ERESOLVE. The adapter works against both majors. (#324)

Docs

README and examples cleanups: clearer out-of-the-box vs peer-install provider guidance, a one-click Vercel deploy starter link, corrected example run paths after the reorg, and Ecosystem additions. (#320–#322, #325–#331)

Also republishes [email protected] (template pin bumped to core 1.8.1).

Features

• Checkpoint and resume (#294 by @mvanhorn, #314 by @JackChen-me). A long run can now survive a crash or restart. Persist progress to any MemoryStore with checkpoint: true on runTeam / runTasks / runFromPlan, then resume with orchestrator.restore(...). On resume, runTeam re-runs coordinator synthesis so you still get a final answer, and checkpointing into the team’s own shared-memory store no longer re-serializes the whole store per task.
• Consensus verification reaches runTeam (#301 by @nuthalapativarun). The per-task judge loop from 1.7.0 now applies to coordinator-generated tasks. Pass the judges with RunTeamOptions.verifyJudges, and the coordinator opts a task in with verify: true or a partial config.
• Native reasoning round-trip on Bedrock (#302 by @nuthalapativarun). Extended-thinking blocks round-trip through Bedrock’s signature protocol with full fidelity, moving echoesReasoning from never to own-issued.

Fixes

• Dashboard: the run details panel is back to a right-hand sidebar on desktop instead of stacking under the canvas (#308 by @JackChen-me).

Onboarding and examples

• npm create oma-app scaffolds a runnable multi-agent demo in one command (#305, #306 by @JackChen-me).
• The full-app examples run end to end on a single DeepSeek key (#311 by @JackChen-me).
• New consensus pattern example (#297 by @nuthalapativarun).

Install

npm install @open-multi-agent/[email protected]

New project: npm create oma-app@latest

Thanks to @mvanhorn, @nuthalapativarun, and @JackChen-me.

⚠️ Breaking: built-in tools are now opt-in / default-deny (#289 by @JackChen-me)

A no-tools agent used to receive every built-in implicitly, including an unsandboxed bash. Tool output flows back to the model, so under prompt injection that was a remotely triggerable exec + exfiltration path. Built-in tools (bash, file_*, grep, glob, delegate_to_agent) now need a positive grant: with neither tools nor toolPreset set, an agent resolves to zero.

Migrating:

• One-line restore of the old allow-all: defaultToolPreset: 'full' on OrchestratorConfig.
• Or grant per agent via tools / toolPreset.
• Custom tools (customTools / addTool) are unaffected: registration is the grant, and disallowedTools is still honored.

Features

• Consensus / adversarial verification (#280 by @CodingBangboo). runConsensus() runs proposers against judges, plus an optional per-task verify hook that puts a task’s own result through the same judge loop.
• Deterministic model routing (#286 by @cat0825). Opt-in modelRouting sends different orchestration calls (coordinator, synthesis, workers, delegated) to different models by match rules, without mutating your team config. Also fixes routing being bypassed on pooled agents.
• MiniMax-M3 is the new default (#292 by @octo-patch). Up to a 1M-token context (512K guaranteed) with image input. M2.7 and MiniMax-M2.7-highspeed stay available if you pin them explicitly.

Docs

• Model routing guide: docs/model-routing.md, linked from both READMEs (#293 by @JackChen-me).
• TencentDB-Agent-Memory integration cookbook under examples/ (#295 by @JackChen-me).
• LiteLLM added to the OpenAI-compatible providers table (#283 by @RheagalFire).

Install

npm install @open-multi-agent/[email protected]

Thanks to @CodingBangboo, @cat0825, @octo-patch, and @RheagalFire. Extra thanks to @CodingBangboo, who landed the consensus primitive (#280) and has been a steady reviewer this cycle.

Features

• Replay persisted team plans, so a previously decomposed plan can be re-run without a new coordinator pass (#285 by @cat0825)
• Add a Tencent Hunyuan provider adapter (#281 by @KaitlynFeng)
• Add a robust regex fallback for malformed single-string JSON tool calls from local models (#269 by @apollo-mg)

Fixes

• Validate message content at every adapter entry, surfacing a clear InvalidMessageError instead of a deep content.some is not a function crash (#288 by @JackChen-me, implementing the fail-fast approach @apollo-mg diagnosed in #268)
• Preserve memoryScope and retry config when a plan is replayed (#287 by @JackChen-me)
• Support structured shared-memory handoff between agents (#284 by @cat0825)

Internal

• Extract a shared repairToolArgs helper across the OpenAI-compatible adapters (#282 by @JackChen-me)
• Add unit tests for the error types (#290 by @Oxygen56)

Docs

• Move the Ecosystem section above Examples in the README (#277 by @JackChen-me)

Install

npm install @open-multi-agent/[email protected]

Thanks to @KaitlynFeng, @apollo-mg, @cat0825, and @Oxygen56.

open-multi-agent v1.5.0

Features

• Per-agent filesystem sandbox (#264 by @JackChen-me). file_read/file_write/file_edit/grep/glob now resolve every path, symlinks included, inside each agent’s cwd, defaulting to <cwd>/.agent-workspace. bash stays unsandboxed. Behavior change below.
• Cross-provider reasoning text fallback (#260 by @MyPrototypeWhat). preserveReasoningAsText carries reasoning across providers that can’t echo native reasoning blocks, as inline <thinking> text instead of dropping it. Phase 2 of #223.
• MiMo provider (#265 by @kidoom).
• Doubao (Volcengine) provider shortcut (#261 by @kidoom, #236 by @janelawrence). Target Doubao models without custom baseURL wiring.
• DeepSeek V4 default model names (#250 by @JackChen-me). Behavior change below.

Fixes

• Secrets redacted from traces, bash output, and dashboard payloads (#263 by @JackChen-me). API keys and tokens no longer leak into observability surfaces.
• MCP: dedupe normalized tool names, clean up client/transport on connect failure (#256 by @JackChen-me).
• Orchestrator honors abortSignal before synthesis and rejects ambiguous task deps (#255 by @JackChen-me).
• Loop detector replays history and warns on text-only loops (#254 by @JackChen-me).
• DeepSeek echoes reasoning_content on V4 tool-calling (#251 by @JackChen-me).

Behavior changes

• Filesystem tools sandbox to <cwd>/.agent-workspace by default (#264). Agents that read or wrote outside process.cwd(), or used relative paths, now get a sandbox error. Set OrchestratorConfig.defaultCwd (or AgentConfig.cwd) to your root, process.cwd() for the old wide default, or null to disable. bash is not sandboxed.
• DeepSeek CLI default is now deepseek-v4-flash (#250). Affects the oma CLI run without --model; library users are unaffected (no hardcoded default in the adapter). deepseek-chat and deepseek-reasoner keep routing until DeepSeek retires them on 2026-07-24.

Package metadata

• docs/ no longer ships in the npm tarball. README links now point to the GitHub copy, so they always track the latest version. If you read docs from node_modules/@open-multi-agent/core/docs/ (uncommon), switch to the GitHub URLs.

Examples and docs

• Provider examples: Moonshot (#259), Qwen (#257), and an index of existing ones (#258), all by @goodneamtakenbydogs.
• Bilig WorkPaper MCP integration example (#247 by @gregkonush).
• README restructured with built-in capabilities surfaced and zh synced (#273 by @JackChen-me), CLAUDE.md slimmed to a code map (#274 by @JackChen-me), integrations governance README replacing the stale DECISIONS.md (#248 by @JackChen-me).
• Tests and CI: DeepSeek tool-calling via chat() and stream() (#252 by @btroops), OpenAI adapter fallback assertions (#253), CI on npm ci (#262), and a CI check asserting the npm tarball ships only dist/ plus metadata.

Install

npm install @open-multi-agent/[email protected]

Thanks to @kidoom, @janelawrence, @MyPrototypeWhat, @goodneamtakenbydogs, @btroops, and @gregkonush.

v1.4.2

Drop-in safe patch: three opt-in additions and three new examples. All new APIs default to off.

New (opt-in)

• RunTeamOptions.revealCoordinator (#245 by @JackChen-me). When true, prepends a team-context block (goal, full roster, this worker’s assignee identity) to every worker prompt under runTeam. Default false keeps existing prompts byte-identical. runTasks and the short-circuit single-agent path ignore it. Closes #244.
• enableReasoningTextReplay (#234 by @matthewYang08). Opt-in replay of framework reasoning blocks as inline <thinking> text on OpenAI-family requests, capped by maxReasoningReplayChars (default 1200). Also closes a latent 400 against OpenAI when a reasoning-only assistant message emitted {content: null} with no tool_calls. Refs #223.
• ReasoningBlock.provenance + LLMAdapter.capabilities (#243 by @MyPrototypeWhat). Phase 1 of #223: additive IR fields, per-adapter echoesReasoning: 'never' | 'own-issued' declarations, and a shared reasoningBlockToInlineText helper. No behavior change yet; Phase 2 wires capabilities into the outbound fallback.

Examples

• Doubao (ByteDance) provider example via Volcengine Ark (#232 by @nuthalapativarun).
• Zhipu GLM provider example (#231 by @nuthalapativarun).
• Narrative puzzle hint arbitration cookbook: 5-agent multi-source conflict resolution with external safety veto (#235 by @suans4746-del). Closes #212.

Documentation

• compressToolResults and maxToolOutputChars documented in the context management guide (#233 by @nuthalapativarun).

Install

npm install @open-multi-agent/[email protected]

Thanks to @MyPrototypeWhat, @nuthalapativarun, @matthewYang08, and @suans4746-del.

Full Changelog: https://github.com/open-multi-agent/open-multi-agent/compare/v1.4.1…v1.4.2

v1.4.1

Patch release: two bug fixes for OpenAI-family adapters and the runner, plus Vercel AI SDK as an optional adapter. The previous npm package (@jackchen_me/open-multi-agent) is now formally deprecated.

Bug fixes

• OpenAI-family adapters: guard choices[0] with optional chaining (#220). Prevents crashes when providers return empty choices arrays. Thanks @dvirarad.
• Runner: defer maxTokenBudget break until after tool_result is appended (#221). Fixes orphaned tool_use blocks when the token budget is hit mid-turn. Thanks @CodingBangboo.

New (optional)

• Vercel AI SDK adapter via the dedicated @open-multi-agent/core/ai-sdk subpath (#229). Bridges LLMAdapter to AI SDK’s generateText / streamText. Optional peer dep ai; main import is unaffected if you don’t use it. Thanks @ibrahimkzmv.

Documentation

• MiniMax provider community offer and setup guide (#219).
• Label hero GIF as post-run replay (#222).

Deprecation

• The previous npm package, @jackchen_me/open-multi-agent, is now formally deprecated. Existing installs continue to work; new installs should use npm install @open-multi-agent/core. The npm CLI shows a deprecation message automatically.

Full Changelog: https://github.com/open-multi-agent/open-multi-agent/compare/v1.4.0…v1.4.1

Highlights

Official org package

Open Multi-Agent now has an official organization package:

npm install @open-multi-agent/core

New projects should use @open-multi-agent/core.

Plan-only orchestration

Adds PlanOnly mode so teams can inspect the coordinator’s task DAG before running agent work. (#203 by @CodingBangboo)

LLM adapter improvements

• Preserve reasoning blocks across Anthropic and Gemini turns. (#205 by @MyPrototypeWhat)
• Forward reasoning_effort and backfill sampling-parameter parity across OpenAI-compatible, Copilot, and Azure paths. (#209 by @MyPrototypeWhat)
• Add a Mistral provider example and README entry. (#206 by @mvanhorn)

Shared memory TTL

SharedMemory entries can now expire by turn count. (#213 by @MyPrototypeWhat)

Fixes

• Keep text-tool extraction depth non-negative when a stray closing brace appears. (#217 by @voidborne-d)
• Fix truncation behavior and tighten coordinator dependency guidance. (#215 by @CodingBangboo)

Examples and Docs

• Add paper replication triage cookbook example. (#202 by @DaiMao-UT)
• Add rare disease information triage example. (#211 by @oooooowoooooo)
• Refresh README, hero animation, badges, docs, and repository links for the new GitHub organization. (#214 and #218 by @JackChen-me)

Compatibility

No intentional runtime API breaks were introduced. The package identity changed to @open-multi-agent/core.

The previous package path, @jackchen_me/open-multi-agent, remains supported during the migration window and is also published at 1.4.0.

Install

npm install @open-multi-agent/[email protected]

Legacy path during the migration window:

npm install @jackchen_me/[email protected]

Features

Streaming reasoning events

StreamEvent now supports a reasoning type that carries the model’s thinking tokens in real time. ReasoningBlock is also added to the ContentBlock union for non-streaming paths. Supported on Anthropic and OpenAI providers. (#174 by @SiMinus)

onAgentStream and onPlanReady hooks

Two new orchestrator hooks (runTeam only): onAgentStream delivers real-time per-token streaming events during agent runs, and onPlanReady fires after the coordinator decomposes the goal into a task DAG — return false to abort before any agent work starts. (#182, #181 by @tizerluo; #184, #183 by @JackChen-me)

Agent Observation Pipeline: new trace events

plan_ready and agent_stream trace events join the trace pipeline, enabling downstream observers to react to plan generation and streaming agent output. (#188 by @ibrahimkzmv)

AWS Bedrock adapter

New LLM adapter for Amazon Bedrock, supporting the full adapter contract (chat + stream). (#194 by @CodingBangboo)

ToolCallTrace includes input/output

ToolCallTrace now carries the tool’s input and output payloads, making it useful for debugging and audit without inspecting the raw conversation. (#124 by @MyPrototypeWhat)

Fixes

• Strip image blocks before summarize compression to avoid ballooning token cost. (#196 by @MyPrototypeWhat)
• Preserve tool_use/tool_result pairing during sliding-window truncation, fixing orphaned tool blocks. (#193 by @MyPrototypeWhat)
• onAgentStream path now forwards the full RunOptions into the streaming runner so onTrace, delegation, and run metadata work during streaming. (#184 by @JackChen-me)
• onPlanReady abort path now reports the real coordinator token cost instead of zero, and catches thrown callbacks. (#183 by @JackChen-me)

Examples

• Express customer support pipeline: multi-agent triage, routing, and resolution. (#191 by @CodingBangboo)
• Personalized interview simulator: dynamic question generation with structured output. (#189 by @mmjwxbc)
• Incident postmortem DAG: reconstruct timeline from logs and deploys. (#187 by @binghuaren96)

Docs

• README capability tagline, per-area contributor attribution, and a JSDoc fix. (#190, #186, #180 by @JackChen-me)

Install

npm install @jackchen_me/[email protected]

Thanks to @SiMinus, @tizerluo, @ibrahimkzmv, @CodingBangboo, @MyPrototypeWhat, @mmjwxbc, and @binghuaren96 for the external contributions.

New capabilities

Agent delegation

Agents in an orchestrated run can now hand a sub-prompt to another agent on the team and receive its final output as a tool result. Opt-in via registerBuiltInTools(registry, { includeDelegateTool: true }). Five guards: self-delegation, unknown agent, cycle detection, configurable depth cap (maxDelegationDepth, default 3), and pool deadlock. Delegated runs’ token usage rolls into the parent’s maxTokenBudget so sub-agents cannot silently bypass it. (#123 by @JackChen-me)

runTeam DAG dashboard CLI

oma runTeam ... --dashboard writes a static HTML view of the resolved task graph after a run, including dependencies and per-task status. (#122 by @ibrahimkzmv, follow-up docs in #141 by @JackChen-me)

outputSchema enforcement and defineTool passthrough

The previously advisory outputSchema on AgentConfig is now enforced: results are parsed and validated, with one retry on validation failure. defineTool schemas pass through to the LLM provider. (#149 by @Xin-Mai)

Pluggable shared memory

TeamConfig.sharedMemoryStore accepts any MemoryStore implementation (Redis, SQLite, your own). sharedMemory: true keeps the existing in-process default. (#157 by @JackChen-me)

Advanced LLM sampling

top_p, top_k, repetition_penalty, min_p, and extraBody are now first-class on agent and coordinator configs. Payload spread order is fixed so extraBody overrides sampling parameters but never transport. (#163 by @apollo-mg)

parallelToolCalls exposed for OpenAI

Was previously hardcoded; now configurable per agent. (#173 by @JackChen-me)

Two new providers

• Azure OpenAI adapter, closing the long-standing #24. (#143 by @Klarline)
• Qiniu provider for users on Chinese infrastructure. (#154 by @JackChiang233, follow-up README/CLI docs in #165 by @JackChen-me)

Fixes

• Context compaction persistence and turn dropping. compact strategy was losing turns and not persisting compressed history. (#161 by @apollo-mg)
• OpenAI mixed-content message ordering. Tool messages must precede user messages in mixed content; previously emitted in the wrong order. (#178 by @voidborne-d)
• Provider type widening on configs. AgentConfig, CoordinatorConfig, and OrchestratorConfig were not using the full SupportedProvider union. (#158 by @JackChen-me)

Behavior changes

#163 removed two implicit defaults that some users may have relied on:

• parallel_tool_calls: false is no longer forced. If you need the old behavior, set parallelToolCalls: false explicitly (now exposed via #173).
• The default frequency_penalty override has been removed.

These are behavior changes, not API breaks, but worth checking if you depended on the old defaults.

The same PR also moved the local <think> tag parsing out of the agent layer into tool/text-tool-extractor.ts. This is internal cleanup with no user-visible impact.

Examples and cookbook

Nine new examples and a category reorganization (#125 by @JackChen-me):

• Meeting summarizer pattern. (#139 by @mvanhorn, moved into cookbook/ in #140 by @JackChen-me)
• Translation / back-translation cookbook. (#145 by @zouhh22333-beep)
• Competitive monitoring. (#146 by @pei-pei45)
• Multi-perspective code review, upgraded to structured output and free providers. (#150 by @Kinoo0)
• Contract review DAG with step-level retry. (#155 by @fault-segment)
• Research aggregation with schema. (#159 by @Optimisttt)
• Engram integration: memory store, toolkit, two demos. (#160 by @Agentscreator, Ecosystem section refresh in #151 by @JackChen-me)
• @agentsonar/oma integration: sidecar from the agentsonar team detecting cross-run delegation cycles, repetition, and rate bursts. (e7aecf3 by @JackChen-me)
• Cost-tiered pipeline comparing flagship vs mixed model tiers. (#164 by @HuXiangyu123)
• OpenRouter provider example. (#167 by @kenrogers)
• local-quantized.ts showing tuned sampling on vLLM and llama-server. (ff987cf by @JackChen-me)

Docs and infrastructure

• README refresh: positioning, branding, hero block, integrations, examples section. (#126, #176, #177, #179 by @JackChen-me)
• CLAUDE.md architecture map synced with the current src/ layout. (#171 by @JackChen-me)
• CLI dashboard documented and added to the flag table. (#141 by @JackChen-me)
• Real badges and Codecov integration. (#127, #128, #129, #130 by @JackChen-me)
• npm registry pinned to npmjs.org via repo-level .npmrc (#170 by @JackChen-me).
• Extended LLM adapter coverage for issue #54. (#144 by @jadegold55)

Install

npm install @jackchen_me/[email protected]

Thanks to @ibrahimkzmv, @mvanhorn, @Klarline, @jadegold55, @zouhh22333-beep, @Kinoo0, @apollo-mg, @Optimisttt, @Agentscreator, @pei-pei45, @fault-segment, @Xin-Mai, @HuXiangyu123, @JackChiang233, @kenrogers, and @voidborne-d for the external contributions that make this release.

Full changelog: https://github.com/JackChen-me/open-multi-agent/compare/v1.2.0…v1.3.0

First minor release since 1.1.0. MCP integration, three new LLM providers, context management strategies, a CLI, tool output cost controls, and fixes for abort and error propagation.

Features

• MCP integration. New connectMCPTools() wires any MCP server (stdio) directly into agent tool use. @modelcontextprotocol/sdk is an optional peer dependency. Runnable example at examples/16-mcp-github.ts. (#89, by @ibrahimkzmv)

• Three new LLM providers. First-class provider: 'deepseek' (deepseek-chat, deepseek-reasoner), provider: 'minimax' (global and China endpoints via MINIMAX_BASE_URL), and verified Groq via OpenAI-compatible baseURL in examples/19-groq.ts. (#113 and #114 by @hkalex; #121 by @mvanhorn)

• Context management strategies. New AgentConfig.contextStrategy keeps long runs under token ceilings with four strategies: sliding-window, summarize, compact (rule-based, no extra LLM call), and custom. (#88 by @ibrahimkzmv; #111, #119 by @JackChen-me)

• Tool output cost controls. New AgentConfig.maxToolOutputChars and per-tool ToolDefinition.maxOutputChars truncate large outputs (head + tail with a marker). New AgentConfig.compressToolResults compresses older tool results once the agent has moved on; errors are never compressed. (#110, #115, #116, #117, #118 by @JackChen-me)

• CLI (oma). New binary for shell and CI with oma run, oma task, oma provider, JSON-first output, and stable exit codes. Docs at docs/cli.md. (#107 by @ibrahimkzmv)

• AgentConfig.customTools. Inject tool definitions at config time from the orchestrator. Bypasses preset/allowlist filtering but still respects disallowedTools. (#109, #112 by @JackChen-me)

• glob built-in tool. Find files by glob pattern, sorted by modification time. (#102 by @ibrahimkzmv)

Fixes

• AbortSignal propagation. Abort now reaches tool execution, the Gemini adapter, and the abort queue path. (#104 fixes #99, #100, #101, by @JackChen-me)

• Error event propagation. AgentRunner.run() now surfaces error events to callers. (#103 fixes #98, by @JackChen-me)

Examples

• examples/16-mcp-github.ts: full MCP wiring
• examples/17-minimax.ts, examples/18-deepseek.ts, examples/19-groq.ts: provider quickstarts
• examples/with-vercel-ai-sdk/: Next.js + OMA runTeam() + AI SDK useChat

Docs

• READMEs (EN/ZH) expanded: CLI, MCP, context strategies, tool output control, customTools. ZH caught up with EN on items that shipped in 1.1.

Install

npm install @jackchen_me/[email protected]

Thanks to @hkalex, @ibrahimkzmv, and @mvanhorn for the external contributions that make this release.

Full changelog: https://github.com/JackChen-me/open-multi-agent/compare/v1.1.0…v1.2.0

First minor release since 1.0.1. Six new features, two fixes, two new examples, and one behavior change you should read before upgrading.

⚠️ Behavior change (read this before upgrading)

Agents now run with default-deny, dependency-scoped context (#87). An agent only sees results from tasks it explicitly dependsOn, instead of every prior task in the run. This prevents context leakage between unrelated agents and keeps token usage predictable in larger teams.

If your existing teams relied on agents implicitly seeing all prior task output, add explicit dependsOn edges in your task graph. No API change is required for runTeam() users whose coordinator already produces a sensible DAG.

This change was prompted by a combination of competitive analysis (XCLI scopes sub-agent context to a minimum file set + tool allowlist by default) and a public post on X by guk2472 flagging inter-agent context pollution as the real production killer in multi-agent systems. Thanks for the signal.

Features

• AbortSignal support for runTeam() and runTasks() (#69). Cancel a run mid-flight from the caller.
• Skip coordinator for simple goals in runTeam() (#70). Single-agent goals no longer pay the coordinator round-trip.
• Token budget management at agent and orchestrator level (#71). Stops runs that exceed a configured budget instead of silently burning tokens.
• Tool allowlist / denylist / preset (#83). Restrict which tools an agent can call without rebuilding the registry.
• Customizable coordinator (#85). Override the coordinator’s model, system prompt, tools, toolPreset, and disallowedTools via CoordinatorConfig.
• Dependency-scoped agent context (#87). See behavior change above.

Fixes

• Per-agent mutex prevents concurrent runs on the same Agent instance from corrupting state (#77).
• Duplicate progress events in the short-circuit path for runTeam() are gone, and completedTaskCount is no longer double-incremented (#82).

Examples

• Multi-source research aggregation (#79)
• Multi-perspective code review (#80)

Docs

• README top fold rewritten and Examples section trimmed (#95)
• Coverage badge updated to 88% (#57)
• DECISIONS.md restructured to signal openness on MCP and A2A

Install

npm install @jackchen_me/[email protected]

What’s new since 0.2.0

Features

• Structured output — optional outputSchema (Zod) on any agent, with auto-retry on validation failure (#36, #38)
• Task retry with exponential backoff — maxRetries, retryDelayMs, retryBackoff per task (#37)
• Observability — onTrace callback emits structured spans for LLM calls, tool calls, tasks, and agent runs (#40)
• Lifecycle hooks — beforeRun / afterRun on AgentConfig for prompt rewriting and result post-processing (#45)
• Human-in-the-loop — onApproval callback between task execution rounds to gate the next batch (#46)
• Loop detection — detects stuck agents repeating the same tool calls or text, with configurable warn / terminate / custom handler (#49)
• Grok (xAI) adapter — first-class support with dedicated GrokAdapter (#44)
• Fallback tool-call extraction — local models that emit tool calls as plain text are now handled automatically (#47)

Testing & quality

• 340 tests, 71% line coverage across src/ (#53)
• Coverage badge added to README (#55)

Full changelog

https://github.com/JackChen-me/open-multi-agent/compare/v0.2.0…v1.0.0