v0.1.45: deep-research skill is -c-only, no LLM API key required

What's New

• The deep-research plugin skill no longer invokes openbrowser-ai -p anywhere. No LLM API key required to run the skill.
• Previously the skill had an optional Step 2b fallback that shelled out to subprocess.run(["openbrowser-ai", "-p", ...]) for sub-questions returning <2 findings. That path required OPENAI_API_KEY (or Anthropic / Google) because -p runs the full Browser Agent LLM loop. Removed entirely.
• Step 2b retry path rewritten to dispatch another /dispatching-parallel-agents round with broader search strategy: alternative engines (Bing, DuckDuckGo, Wikipedia), query reformulation, lower sentence-length thresholds. Still pure -c.

Why

The skill's architecture (orchestrator + parallel sub-agents driven via -c heredocs) already covers what the -p fallback was doing. The -p path also coupled the skill to the user's LLM provider configuration and cold-started a fresh browser per fallback invocation. Pure subtraction.

Migration

If you previously set OPENAI_API_KEY / ANTHROPIC_API_KEY / GOOGLE_API_KEY only because the deep-research skill needed it, you can unset them. The skill no longer touches any model.

Details

• Skill drives the daemon via openbrowser-ai -c only. Daemon executes raw CDP / JS through its Python namespace, no model loaded. See cli.py:434-490 get_llm() for the -p LLM coupling that the skill now avoids.
• plugin/skills/deep-research/SKILL.md 636 -> 645 lines.
• Sections rewritten: intro, Setup, Step 2b (entire section), Tips.
• Other 6 sibling skills audited; none use -p. No changes there.

Full Changelog: v0.1.44...v0.1.45