[ling-senpeng13]

Adds the agents feature to the Go SDK as sdk/ai, a port of the Python SDK's agent surface.

What it adds

• Agents — the Agent model and its serializer, Runtime, AgentHandle, and the agent REST client.
• Tools — functions, HTTP, MCP, human input, agent-as-tool, media, and worker tools served from another process, with declared credentials and structured output.
• Composition — guardrails, handoffs and swarm transitions, plan-execute, code execution and CLI runners, routers, terminations, gates, and Agent Skills.
• Operations — deploy and serve, signals and messages, pause and resume, lifecycle callbacks, agent schedules, code executors, and semantic memory.
• Docs and examples — the design docs under docs/ai/, and the Python agent examples ported under examples/agents/.

How it is verified

TestGoldenAgentConfig serializes a Go agent for every fixture under sdk/ai/testdata/agent_config/ and requires the document the Python serializer produces, so the wire format cannot drift from Python's. The Python e2e suites are ported under test/integration_tests/ai/ and replayed against recorded LLM traffic, one directory per test file: 105 tests pass, and the 13 skips are listed with their reasons in that directory's README.

CONDUCTOR_SERVER_URL=http://localhost:8080/api go test -tags integration ./test/integration_tests/ai/

Fixes outside sdk/ai

Six gaps the ports exposed, each a feature that was declared but never wired up: tool-level guardrails now run in the worker; terminations, stop-when predicates and router functions each get the worker the server waits for; Agent.Stateful routes a run to a per-execution worker domain; Runtime.Plan compiles a skill from its own document; an event carries the execution it came from, so a nested agent's request for human input is answered on that execution; and a schedule's input survives being read back.

The lint workflow now compares a pull request against its merge base. The patch API it used refuses a diff this size and then silently lints the whole repository.

Two tests stay skipped until conductor-oss/conductor#1633 lands, which fixes numeric and agent-envelope matching in the LLM recorder.

[NicholasDCole]

All of these files should not be required

[ling-senpeng13]

All of these files should not be required

In addition to porting the Python example tests, I also ported the applicable Python E2E tests with LLM replay to increase test coverage and these json files are the artifacts. Through this process, I found a few SDK bugs as well.

General thoughts:

1. There may be cases where we want to run SDK-specific integration tests with LLM replay. In that case, we would need to keep the corresponding JSON replay files in the SDK repo.
2. It is possible to port the above E2E tests with LLM replay in the same way as the example tests. If we take this approach, I can remove the JSON replay files mentioned above.