[FEATURE]: Optional read-before-write precondition for edit/write tools #33089

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Currently, the model can call edit/write/bash on a file without ever having read that file first. This allows the AI to act on partial context (e.g., a grep summary) and make incorrect edits.

I propose adding an optional read-before-write precondition to the tool runtime. The system would:

1. Track which files have been read in the current session
2. Before allowing edit/write on a file, check if it was read first
3. If not read, require the model to read it first or get explicit user permission

This could be implemented in the existing tool.execute.before plugin hook (session/tools.ts), which already exists but is unused for safety checks. The hook would:

• Check if the target file was read in this session
• If not, either block the edit or append a warning to the tool output

An opt-in configuration in opencode.json would let users enable this:

{
  safety: {
    requireReadBeforeEdit: true
  }
}

About the duplicate flag: This is different from:

• [FEATURE]: optional runtime enforcement of edit tool's read-before-edit precondition #27901 (runtime enforcement of read-before-edit precondition) — that issue requests a configurable flag. This issue additionally proposes the implementation mechanism (track read state in session, enforce in tool.execute.before hook).
• [bug] No pre-flight validation before tool execution despite tool.execute.before hooks #33076 (pre-flight validation) — that issue is about general destructive operation detection. This issue is specifically about read-before-write ordering, which is a simpler, narrowly scoped precondition.
• [FEATURE]: Add isReadOnly and isDestructive flags to tool interface #20296 (isReadOnly/isDestructive flags) — that issue is about tool classification. This issue is about file-level access ordering, not tool-level safety classification.