fix: preserve parent RuntimeContext in harness subagent delegation

[guslegend0510]

Summary

• add RuntimeContext.fork() to carry string extras, typed attributes, and ToolExecutionContext into child calls
• pass the parent RuntimeContext through harness subagent invocation paths
• add regression tests for invokeAgent and invokeAgentStream

Validation

• mvn -pl agentscope-harness -am "-Dtest=RuntimeContextTest,DefaultAgentManagerRuntimeContextTest" test

[codecov]

Codecov Report

❌ Patch coverage is 82.75862% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
.../java/io/agentscope/core/agent/RuntimeContext.java	80.00%	1 Missing and 3 partials ⚠️
.../agentscope/harness/agent/tool/AgentSpawnTool.java	50.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[AgentScopeJavaBot]

🤖 AI Review

This PR adds RuntimeContext.fork() to preserve parent context (string extras, typed attributes, and ToolExecutionContext) when delegating to harness subagents, and threads the parent RuntimeContext through DefaultAgentManager and AgentSpawnTool invocation paths. The design is clean: fork() creates isolated snapshot copies of attribute maps (copy-on-fork semantics), preventing parent-child interference, while sharing the ToolExecutionContext by reference (appropriate for a DI container). The old overloads delegate to new ones with null parentContext, maintaining backwards compatibility. Test coverage is solid, verifying both sync and stream paths with typed attributes, string extras, ToolExecutionContext identity, and agentState isolation. One pre-existing gap worth noting: the plain Agent fallback in invokeAgent/invokeAgentStream still calls agent.call(prompt) without the computed ctx — not introduced by this PR, but now more visible.

[AgentScopeJavaBot]

🤖 AI Review

This PR adds RuntimeContext.fork() to preserve parent context (string extras, typed attributes, and ToolExecutionContext) when delegating to harness subagents, and threads the parent RuntimeContext through DefaultAgentManager and AgentSpawnTool invocation paths. The design is clean: fork() creates isolated snapshot copies of attribute maps (copy-on-fork semantics), preventing parent-child interference, while sharing the ToolExecutionContext by reference (appropriate for a DI container). The old overloads delegate to new ones with null parentContext, maintaining backwards compatibility. Test coverage is solid, verifying both sync and stream paths with typed attributes, string extras, ToolExecutionContext identity, and agentState isolation. One pre-existing gap worth noting: the plain Agent fallback in invokeAgent/invokeAgentStream still calls agent.call(prompt) without the computed ctx — not introduced by this PR, but now more visible.

[iuin8]

We're building an enterprise agent-governance platform on AgentScope-Java RC3, and this PR fixes a blocker for us: harness subagent delegation drops the parent RuntimeContext, so our typed governance identity (tenant + granted capabilities) never reaches spawned subagents — the whole permission/audit arc silently no-ops at the subagent layer.

We reviewed the patch against RC3 and it looks complete on the delegation surface: RuntimeContext.fork(sessionId, userId) deep-copies string + typed attributes and shares ToolExecutionContext, and parentCtx is now threaded into all AgentSpawnTool call sites (spawn/send + the three execLocalSync paths + execSpawnTask). 👍

Two notes that might help it land:

• CI is currently red (windows build failure, ubuntu cancelled) — looks infra/unrelated to the diff; happy to help test if useful.
• HarnessGateway.runSubagentStream (the expose_to_user direct-dialog path, tracked in [Bug]:HarnessGateway.runSubagentStream 未向暴露的子 Agent 传递 RuntimeContext / MsgContext #1700) isn't touched here, so that path still loses context — worth a follow-up rather than expanding this PR.

Looking forward to seeing this in a release.

[guslegend0510]

@chickenlj 回复你了，麻烦看一下，谢谢您