v0.1.21 — analysis-tool visibility

UX-driven release on top of v0.1.20. Drops the hasData-based dynamic filter on the analysis-tool set so the LLM can plan multi-step "load → query → analyse → report" workflows up front.

What changed

Analysis tools are now exposed to the LLM every round, regardless of whether the active session has data loaded. Previously query-sql, describe-data, list-tables, query-preview, suggest-analysis, quick-summary, analyze-data, promote-finding were hidden until a successful load-data, which forced the LLM into round-by-round discovery and broke up-front planning.

The original rationale ("keep tool count low for local LLMs") was for the early gemma2 / gemma3 era; the standard local model (gemma-4-26b-a4b, MoE) and current gpt-oss / qwen3 generation handle 30+ tools without selection-accuracy regression.

Visible result for users:

• Asking "monthly sales totals" without attaching data → the model now proposes load-data (asking for the file path) instead of declining with "I don't have a SQL tool"
• Asking "load /Users/.../sales.csv and tell me the top product" → the model can plan the load + query in one response instead of two round-trips

Escape hatch

New config flag tools.hide_analysis_tools_until_data_loaded (default false) restores the pre-v0.1.21 behaviour for users on weaker local backends. Power-user knob, config-only — not exposed in the Settings UI. To use it, edit ~/Library/Application Support/shell-agent-v2/config.json:

{
  "tools": {
    "hide_analysis_tools_until_data_loaded": true
  }
}

Other touch-ups

• load-data's tool description now advertises the downstream pipeline so the LLM doesn't have to guess what becomes available after a successful load.
• Settings → Tools list now shows all analysis tools by default (consistent with the dispatcher's view) so per-tool MITL toggles are reachable without first loading data.

Documentation

• Design rationale + implementation: docs/en/agent-tool-visibility.md (日本語)
• CHANGELOG