AI Chat and Actions

AI Chat & Dry-Run Actions

The chat panel is a retrieval-augmented assistant scoped to your files, with a strict safety model for anything that touches disk: the AI proposes, you decide.

Two modes, one chat

The assistant follows a format rule baked into its system prompt:

• Question / analysis → a normal text answer, citing relevant files by name (grounded in your index).
• Action request (move, rename, delete, tidy, create a folder) → a structured plan only, as JSON:

{
  "summary": "…",
  "operations": [
    { "kind": "move|rename|delete|mkdir", "old_path": "…", "new_path": "…", "reason": "…" }
  ]
}

The model must use exactly the paths present in the folder structure it's given — it can't invent paths. For reorganizations it reasons over the directory tree supplied in-context.

The Dry-Run safety model

When the assistant returns operations, SenseTree renders an Action Plan card: an interactive Before → After diff of every move/rename/delete/mkdir, each with the AI's stated reason. Then:

• Approve → apply_action_plan executes the plan.
• Discard → discard_action_plan throws it away.

Nothing is written to disk until you Approve. The plan is first recorded as a draft in the transaction log; no std::fs change happens at planning time.

Transactional apply with rollback

On Approve, operations run one by one and each success is journaled. If any operation fails (e.g. a file is locked), SenseTree rolls back the ones already done (inverse operations), leaving your disk as it was. After a successful apply it also updates the index — renames update LanceDB paths without re-embedding (Index Sync), and deletes remove vectors — then commits.

Related commands

Command	What it does
chat_with_assistant	Q&A / action chat, retrieval-augmented over your index.
plan_reorganization	Build a reorganization plan for a scope from an instruction (draft only).
apply_action_plan	Execute an approved plan transactionally, with rollback.
discard_action_plan	Drop a pending plan.

Tips

• Be specific about scope. Open the folder you want the assistant to reason about; it plans over that structure.
• Read the reasons. Each operation carries a rationale — a quick sanity check before Approve.
• Reasoning model quality matters. Strict-JSON, path-faithful behavior is easier for solid instruction-following models. If plans are malformed or invent paths, try a better reasoning model or tighten the prompts.