v3.19.0

Six Tier-1 quick-win features focused on long-task reliability, interop, and reach — plus drag-and-drop and a bigger iteration budget.

Added

• Live task list (todo_write tool). The model maintains a checkable plan during long multi-step tasks and re-grounds itself each turn; rendered as a panel (✓ done / ▸ in-progress / ◻ pending). The biggest quality lever for long agentic runs, especially on smaller local models. Works on native + Ollama providers; resets on /clear.
• AGENTS.md interop. Reads the cross-tool AGENTS.md project-instructions standard (walks to the repo root; nearest wins), with CLAUDE.md as a fallback — so Kodiqa respects instruction files already present in repos set up for other agents.
• OpenRouter provider. /key openrouter unlocks hundreds of models by their vendor/model id (auto-discovered; openrouter/auto routes for you). Its /models reports context_length, so the context window is auto-detected.
• /effort reasoning dial. Set low/medium/high — applied to OpenAI reasoning models (o-series/gpt-5) and OpenRouter; other providers use their default (no 400 risk).
• Per-persona model binding. /persona <name> <model> binds a model to a persona so switching personas auto-switches the model (strong model for architect, cheap one for code). Persisted.
• Per-category auto-approve. /approve write|command|delete|clipboard [on|off] auto-approves a whole category without dropping to full auto mode.
• Drag-and-drop images & files. Dropping a file into the terminal now attaches it — images are sent to the model to view — instead of erroring as an "unknown command." Handles the backslash-escaped / quoted paths terminals paste, plus multiple files.

Changed

• /recommend refreshed with Devstral (Mistral × OpenHands agentic coder) and the expert-pruned Cerebras Qwen3-Coder-REAP-25B-A3B (~coding-lossless, ~20% lighter), each fit-annotated; both pull via the HuggingFace GGUF fallback.
• Default max_iterations raised 15 → 40 so long multi-step tasks don't stop early. Still overridable in ~/.kodiqa/config.json; /budget remains the cost backstop, and continue resumes past the cap.

84 commands, 27 tools, 7 cloud providers. 637 tests passing.