* feat(a2a): durable A2A protocol — client, server, observability
Add Agent2Agent (A2A) protocol support to the AI module, in both directions.
Client: CALL_AGENT / GET_AGENT_CARD / CANCEL_AGENT_TASK system tasks call remote
A2A agents over JSON-RPC with poll, streaming (SSE), and push modes. Durable by
design — deterministic messageId, state in the execution (not a thread), liveness
guards, and a push backstop. SSRF-guarded (IPv6 ULA + cloud-metadata; redirects off).
Server: expose any workflow as an A2A agent (one agent per workflow), idempotent
message/send -> startWorkflow (RETURN_EXISTING), tasks/get / tasks/cancel, and
multi-turn resume (a follow-up message/send completes the paused HUMAN/WAIT task
instead of starting a duplicate).
Observability: Micrometer counters via the shared Monitors registry and MDC
correlation keys across the A2A code paths.
Gated by conductor.integrations.ai.enabled (client) and conductor.a2a.server.enabled
(server).
* test(a2a): unit, real-agent interop, and real-engine durability tests
- Unit/wire tests (MockWebServer), embedded-agent e2e, push callback, server
JSON-RPC dispatch + multi-turn resume, mapper/worker, and observability.
- A2ASdkInteropTest: drives the client against the official a2a-sdk reference
agent launched as a subprocess (discovery, send, poll, streaming, message-mode);
self-skips when no Python with a2a-sdk is available.
- A2ADurableEngineEndToEndTest (test-harness): CALL_AGENT through the real decider
+ AsyncSystemTaskExecutor + Redis, proving crash/restart resume from persistence.
* docs(a2a): integration guide, examples, UI task types, design notes
- docs/devguide/ai/a2a-integration.md (AI Cookbook nav) with worked examples:
call/expose, multi-turn resume request/response, and push end-to-end.
- ai/examples: call / get-card / server + streaming / push / multi-turn / cancel,
indexed in the examples README; runnable interop + durable demos under test
resources.
- ui-next: register CALL_AGENT / GET_AGENT_CARD / CANCEL_AGENT_TASK task types.
- design/a2a: protocol study, durability proposal, and server design notes.
Note: a CI step to install a2a-sdk (so A2ASdkInteropTest runs in the build job)
is left as a follow-up — it needs a token with the `workflow` scope to land.
* refactor(a2a): move inbound server auth to enterprise (OSS open by default)
The A2A server's optional shared-secret api-key (conductor.a2a.server.api-key)
is removed from OSS: the server is now open by default, matching OSS Conductor
REST. Inbound authentication (API keys, OAuth/OIDC, mTLS, per-skill scopes,
signed Agent Cards) belongs to the enterprise build; front OSS with a
gateway/firewall.
Unchanged and kept in OSS as safe-by-default guards: client SSRF protection,
push-callback token auth, cross-agent execution isolation, and client→remote
per-call auth headers.
Removes authorized()/apiKey + the two api-key tests; docs and design notes
updated.
* docs(a2a): error-handling/troubleshooting, multi-agent + client multi-turn + LLM-pick examples
Docs: add an error-handling & retries section (FAILED vs FAILED_WITH_TERMINAL_ERROR
mapping for HTTP/JSON-RPC/SSRF/liveness) + a troubleshooting table; flesh out the
client multi-turn section with a worked SWITCH-on-input-required snippet; add an
"orchestrating multiple agents" subsection.
Examples (validated by ExampleWorkflowValidationTest):
- 27-a2a-multi-agent: FORK_JOIN calling agents in parallel → JOIN.
- 28-a2a-llm-pick-skill: GET_AGENT_CARD → LLM_CHAT_COMPLETE → CALL_AGENT.
- 29-a2a-client-multi-turn: SWITCH on input-required, re-call with same context/taskId.
* refactor(a2a)!: rename CALL_AGENT→AGENT, CANCEL_AGENT_TASK→CANCEL_AGENT + add agentType
Generalize the agent task types ahead of multi-runtime support:
- CALL_AGENT → AGENT (class CallAgentTask → AgentTask), CANCEL_AGENT_TASK →
CANCEL_AGENT. GET_AGENT_CARD kept (discovery/"Agent Card" is A2A-specific).
- New input field `agentType` (default "a2a") on all three tasks — the extension
point for native runtimes (langgraph, openai, …). Unknown values are rejected
with a clear error; only "a2a" is implemented today.
Updated across enum, handlers, request models, mapper, callback, tests,
test-harness, examples, docs, design notes, and the ui-next task registration.
BREAKING CHANGE: workflows using type "CALL_AGENT"/"CANCEL_AGENT_TASK" must switch
to "AGENT"/"CANCEL_AGENT".
* docs(a2a): clean up post-rename residue (grammar, example names, phrasing)
Fix 'A AGENT'→'An AGENT' grammar (A2AMetrics, design), reword example descriptions and a doc sentence left awkward by the mechanical rename, and update legacy example workflow names (a2a_call_agent_* → a2a_agent_*).
* feat(a2a): server-side message/stream (SSE)
Exposed workflow-agents now support A2A message/stream. The endpoint returns an
SSE stream driven off a dedicated daemon pool: the initial Task, status-update
events as the workflow's A2A state changes, artifact-update events as output is
produced, and a final status-update at a terminal/input-required state (or when
the stream window elapses). The agent advertises capabilities.streaming=true.
A web-agnostic A2AStreamSink keeps SseEmitter out of A2AWorkflowAgent so the
stream driver stays unit-testable. Config: conductor.a2a.server.stream-poll-
interval-millis (500), stream-max-duration-seconds (300).
Tests: a mocked unit test asserting the task → artifact-update → final
status-update sequence, and a real-HTTP loopback (our client message/stream
against the server's SSE) aggregating to a completed task with artifacts.
Docs + design notes updated (streaming moves from follow-up to shipped).
* Update A2ADurableEngineEndToEndTest.java
* Update A2ADurableEngineEndToEndTest.java
* refactor(a2a): simplify client/server task code (no behavior change)
- AgentTask: replace hand-rolled isBlank/firstNonBlank/stripTrailingSlash/
asInt/asLong helpers with commons-lang3 StringUtils + Number-only reads;
reuse parseRequest().getPollIntervalSeconds() for pollInterval; dedup
stateOf via A2AResults; fix the stale {@link #isAsyncComplete} javadoc to
describe the real backstop-poll mechanism
- A2ACallbackResource: drop the deprecated query-param token path (untested,
no legacy clients on a feature branch); StringUtils for isBlank
- A2AServerResource/A2AWorkflowAgent: delete the duplicate basePath()/URL
building (delegate to the agent's agentUrl/new agentCardUrl); drop the
no-op `id == null ? null : id` ternaries
- remove dead fields A2AMessage.referenceTaskIds and PushNotificationConfig.id
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(a2a): add mermaid use-case diagrams (client & server directions)
Diagrams rendering both A2A directions — Conductor as client (workflow calls
remote agents via the AGENT task) and as server (a workflow exposed as an A2A
agent):
- docs/devguide/ai/a2a-integration.md: both-directions overview flowchart,
AGENT lifecycle sequence (poll/push/stream), FORK_JOIN→JOIN multi-agent
fan-out, and the server message/send→workflow→tasks/get sequence
- design/a2a/08: overview flowchart + a sequence per direction (A/B)
- design/a2a/10: server sequence emphasizing the streaming and multi-turn
resume paths
Rendered via pymdownx.superfences (mkdocs) and natively on GitHub.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(ai)!: rename package ai.models -> ai.model (singular)
Align the AI module with conductor-oss package conventions: every model
package in the codebase (com.netflix.conductor.common.model, core.model,
grpc/proto/model, ai.a2a.model) is singular `model` — `ai.models` was the
only plural `models` package in the entire repo, and sat one letter away
from the singular ai.a2a.model, which read as an inconsistency.
Pure mechanical rename: git mv of the package dir (history preserved) plus
package-decl/import/FQN updates across the module (main + test) and one doc
reference in ai/CONTRIBUTING.md. No type or behavior changes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix /a2a endpoint not being reachable
* style: spotlessApply
* fix and upgrade version
* refactor for agentspan
* feat: add SQLite (sqlite-vec) vector database backend
Add an embedded, zero-infrastructure VectorDB backend implemented with the
sqlite-vec extension, alongside the existing pgvector/MongoDB/Pinecone backends.
- SqliteVectorDB stores embeddings in a vec0 virtual table and runs KNN via the
sqlite-vec MATCH operator (l2/cosine/l1).
- The native vec0 binaries (linux/macos/windows) are downloaded + checksum-verified
at build time (downloadSqliteVec) and bundled as jar resources; SqliteVecExtensions
extracts the right one at runtime.
- SqliteVectorDBAutoConfiguration registers a "default" instance automatically when
conductor.db.type=sqlite and conductor.integrations.ai.enabled=true, so the
all-in-one SQLite server gets a working vector store with no external dependency.
- Adds unit tests, a real end-to-end round-trip test, a RAG example
(30-rag-sqlite-vec.json), and documentation.
Note: a dedicated Linux e2e CI job (.github/workflows/ci.yml) is intentionally not
included here because the push token lacks the GitHub `workflow` scope; it must be
committed separately with a workflow-scoped token.
* test: enable workflow execution lock in integration profile to fix flaky specs
The integration-test profile disabled conductor.app.workflow-execution-lock-enabled
(since 2021), so acquireLock() was a no-op and the background WorkflowSweeper could
decide a workflow concurrently with specs' manual sweep()/asyncSystemTaskExecutor
calls. Two simultaneous decides could schedule duplicate tasks, intermittently
failing exact task-count assertions (FailureWorkflowSpec, and
HierarchicalForkJoinSubworkflowRetrySpec).
Enable the lock (matching production defaults; local_only lock is already configured)
so decides are serialized. Full :conductor-test-harness:test passes (252 tests).
* fix(build): wire downloadSqliteVec as a resource source so all consumers depend on it
The generated sqlite-vec resource directory is consumed by sourcesJar (and other
resource consumers), but only processResources depended on downloadSqliteVec. Gradle's
task-dependency validation (run as part of `build`) failed sourcesJar with an implicit
dependency error. Register the task provider itself as the resource srcDir so every
consumer depends on it automatically.
* fix(ai): harden sqlite-vec backend after review
Follow-up fixes from a deep review of the SQLite/sqlite-vec backend:
- SqliteVectorDB.upsertEmbeddings: restore autoCommit(true) in a finally block
so connections are never returned to the Hikari pool in manual-commit mode.
- SqliteVectorDB.getSqliteClient: validate the extension path against an
allowlist before interpolating it into load_extension(...).
- SqliteVectorDBAutoConfiguration.resolveDbPath: strip JDBC URL query params
(e.g. ?busy_timeout=15000&journal_mode=WAL) so the derived *_vectordb.db path
is a valid filename.
- SqliteVecExtensions: re-extract the bundled binary if the cached temp file was
deleted between calls (stale-cache guard).
- VectorDBProvider: warn when an auto-registered default is skipped because an
explicit instance with the same name already exists.
Adds SqliteVectorDBAutoConfigurationTest (resolveDbPath cases incl. query
params), SqliteVecExtensionsTest (platform detection), and a rollback/
autocommit-restore test in SqliteVectorDBTest.
* feat(ui-next): config forms for all AI / agentic task types
Wire up the workflow-editor task form dispatcher, Add-Task menu, icons, task
generators and JSON schemas for every AI task type exposed by the `ai/` module,
so each gets a real config form instead of the generic JSON editor.
- New bespoke forms: AGENT, GET_AGENT_CARD, CANCEL_AGENT, LIST_MCP_TOOLS,
CALL_MCP_TOOL, GENERATE_IMAGE, GENERATE_AUDIO, GENERATE_VIDEO, GENERATE_PDF,
LLM_SEARCH_EMBEDDINGS.
- Wire the existing-but-unwired LLM forms (chat/text complete, embeddings,
index text/document, search index, get document) into TaskFormContent.
- LLM_CHAT_COMPLETE: OSS plain "Instructions" textarea (enterprise prompt-name
picker remains pluggable via the plugin registry).
- A2A forms: agentType as a fixed A2A/Conductor radio; multi-row headers via the
shared HTTP headers editor; pushNotification is a boolean toggle.
- CALL_MCP_TOOL arguments accept inline JSON or a ${variable} reference.
- Media-gen + search-embeddings reuse the field-driven LLMFormFields for
provider/model/vectorDB/embedding selection (server-backed dropdowns).
- Full field coverage vs the backend input models; numeric fields use coerceTo
so the workflow JSON stores real numbers while still accepting ${variables}.
- A2A tasks added to the Agentic Orchestration quick-add section.
* fix(ui-next): address PR review on AI task forms
- Schemas.ts: add the new AI/agentic task types to genericSchema's type enum so
the code (JSON) editor no longer warns "type is not one of ..." for AGENT,
GENERATE_*, *_MCP_*, LLM_SEARCH_EMBEDDINGS, CHUNK_TEXT, LIST_FILES, PARSE_DOCUMENT.
- AgentTaskForm: fix spacing so the "Push backstop poll" field's floating label
no longer collides with the push-notification toggle (group switches, separate
the input row).
- LLMSearchEmbeddingsTaskForm: add spacing so the helper text no longer collides
with the Query field's floating label.
- QuickAddMenu: fill the two empty Agentic Orchestration slots (Call MCP Tool,
Generate Image) for a clean two-row grid.
* fix(ui-next): snug node height for AI/agentic task cards
AI task cards render header-only content (no custom body), but fell through to
the DEFAULT node height (100px), leaving empty space below the label on the
canvas. Size the AI/agentic header-only task family to 80px (matching the
header-only FORK_JOIN node) so the cards fit their content.
---------
Co-authored-by: Nicholas Cole <68611647+NicholasDCole@users.noreply.github.com>
Co-authored-by: nicholascole <nicholas.colesd@gmail.com>
* Restore Redis Sentinel auth configuration
* Support multiple sentinel addresses in redis-lock configuration
Split CONDUCTOR_REDIS_LOCK_SERVER_ADDRESS on semicolons so that
multiple sentinel endpoints can be provided, matching how the cluster
mode already splits on commas. This improves sentinel HA by allowing
the lock client to discover the master even if one sentinel is down.
Example: redis://sentinel-0:26379;redis://sentinel-1:26379
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Document semicolon-separated sentinel addresses for redis-lock
Add a note to the deployment guide explaining that multiple sentinel
endpoints can be provided using semicolons when serverType is SENTINEL,
improving high availability for the lock client.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Fixes failing int test
* reverts test-harness fixes
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* Overhaul documentation site with modern design and new content
- Redesign homepage with hero section, scrolling logo wall, value strip, feature cards, architecture section, FAQ, and CTA
- Add new architecture pages: Durable Execution, Agents & AI, JSON + Code Native
- Add Quickstart guide with polyglot worker examples (Java, Python, JS, Go, C#, Ruby, Rust)
- Modern CSS theme: Instrument Serif + DM Sans + IBM Plex Mono, solarized code blocks, dark mode support
- Replace curl examples with Conductor CLI commands across all pages
- Add SEO meta descriptions to 30+ pages targeting durable code execution, workflow engine, saga pattern keywords
- Restructure mkdocs.yml navigation with tabs, search, and content features
- Add MkDocs Material theme overrides and serve-docs.sh helper script
Viren Baraiya