v0.1.44: deep-research skill now dispatches parallel sub-agents (one tab each)

What's New

• deep-research skill refactored to use /dispatching-parallel-agents -- the orchestrating Claude session now dispatches one parallel sub-agent per sub-question, and each sub-agent owns exactly one Chrome tab. This replaces the previous single-coroutine asyncio.gather over multiple tabs.
• True parallelism, independent reasoning per tab, and failure isolation: a sub-agent that hits a bad page no longer poisons the rest of the run.
• Drilldown mode also now uses parallel sub-agents: a second wave of fanout, one sub-agent per needs_depth=true finding (capped at 3 per parent).
• Sub-agents communicate findings to the orchestrator via partial JSON files at local_docs/research/_partial/agent-NN.json, NOT via stdout. Cleanup removes the partial dir.

Migration

No package or CLI behavior changes. Skill ships via the Claude Code plugin marketplace; if you use the deep-research skill, the new version will be picked up automatically after this release syncs to the plugin manifest.

The hard rule for any future skill author: one sub-agent owns one tab, and dispatch must use a single message with multiple Agent tool calls. Sequential agent invocations across messages run serially and lose the parallelism benefit.

Details

• plugin/skills/deep-research/SKILL.md 593 -> 636 lines.
• Step 2a, Step 3, Step 7, and Tips section all rewritten.
• Background on the design tradeoff: a single Python coroutine driving N tabs through one daemon serializes navigation events at the CDP layer and cannot make per-tab decisions about pagination or follow-up clicks. Real sub-agents fix all three.

Full Changelog: v0.1.43...v0.1.44