v0.9.0

[0.9.0] - 2026-06-03

AgentLane 0.9.0 turns the harness into a more practical host-app surface: generic sub-agents, first-party tool permission policies, approval/run-event streaming, and distributed example docs are now paired with runtime message identity and cancellation improvements.

Added

• Added the generic spawned agent tool with inherited base tools, depth/thread guards, policy modes, and sanitized delegation failures (b3e62ec, fe60246, f55558a, f38eb6d, 8186b08)
• Added first-party tool permission grants, workspace/path-scope policies, approval-required decisions, approval events, and docs for composing local file and command access (8bd7353, 92f9f42, a26092c, 4daa279, fb75688, c6a335d)
• Added high-level harness run events and approval broker lifecycle support so hosts can observe run state, tool calls, approvals, and stream teardown through one event surface (c5a7e91, 1ccc46d)
• Added distributed harness agents documentation and a multi-process clinical inbox copilot example that demonstrates host/worker execution from the harness layer (57b5b31, eeba50f, b3ab07f)

Changed

• Exported AgentFactory from the runtime public API and moved MessageContext to agentlane.runtime for clearer runtime-facing imports (689ac51, d33c59d)
• Let runtime callers provide a message_id to send_message(...) and propagated the resolved id through rejection outcomes and envelopes (a9d0224, e48259a)
• Split first-party harness tool docs into focused pages and refined README/doc examples around current distributed and tool-permission behavior (5d2b14a, 4d9e72d, e862772, 927461b)

Fixed

• Fixed MessageContext cancellation propagation and covered it across in-process and cross-worker runtime paths (43f95ac, 6e50131, ad0221b, 6596ecd)
• Fixed the read base tool's model-facing output and tightened path resolution, allow-all policy boundaries, grep/search docs, and related tool docs consistency (5d69a9c, bf2704b, 35047ac, fd44746, d43400a)