AI providers skip reading files before editing due to missing code-level enforcement and inconsistent system prompts #30376

Code bugs

After checking all source code related to file modifications, these code-level defects cause or allow AI agents to edit files without reading the current content first.

Bug 1: edit.ts claims "will error if not read first" — but never checks

Code: packages/opencode/src/tool/edit.txt:4

This tool will error if you attempt an edit without reading the file.

Code: packages/opencode/src/tool/edit.ts:69-208 (execute function)

The description tells the AI the tool will error. But the actual execute() function:

• Line 71-77: Only validates filePath and oldString/newString
• Line 88-168: Reads the file internally via Bom.readFile(), performs the edit, writes back
• Never checks whether Read tool was called first

Result: The AI can skip Read entirely and call edit directly. It always succeeds. The description is a lie.

Bug 2: write.ts claims "will fail if not read first" — but never checks

Code: packages/opencode/src/tool/write.txt:5

This tool will fail if you did not read the file first.

Code: packages/opencode/src/tool/write.ts:38-101 (execute function)

Same pattern:

• Line 46-47: Checks if file exists, reads it internally via Bom.readFile() if it does
• Never checks whether Read tool was called first

Bug 3: apply_patch.ts has zero read-before-edit enforcement

Code: packages/opencode/src/tool/apply_patch.ts:30-304 (execute function)

For "update" hunks (line 106-131) and "delete" hunks (line 161-188), the tool reads the file internally and applies the patch. No read-before-edit check exists. Additionally, apply_patch.txt (33 lines) never mentions reading first.

Compounding factor: packages/opencode/src/tool/registry.ts:329-335 replaces edit/write with apply_patch for non-4, non-oss GPT models. These models lose even the textual claim of read-before-edit enforcement.

Bug 4: instruction.ts tracks read files — but edit/write/apply_patch never use it

Code: packages/opencode/src/session/instruction.ts:15-30

function extract(messages: SessionLegacy.WithParts[]) {
  const paths = new Set<string>()
  for (const msg of messages) {
    for (const part of msg.parts) {
      if (part.type === "tool" && part.tool === "read" && part.state.status === "completed") {
        if (part.state.time.compacted) continue
        const loaded = part.state.metadata?.loaded
        if (!loaded || !Array.isArray(loaded)) continue
        for (const p of loaded) {
          if (typeof p === "string") paths.add(p)
        }
      }
    }
  }
  return paths
}

This function already iterates past messages, finds completed Read tool calls, and tracks which files were read. The result is exported via loaded() at line 233-235.

But this data is ONLY used by instruction.ts:resolve() (line 183) for attaching nearby instruction files. It is never passed to or checked by edit.ts, write.ts, or apply_patch.ts.

This is the fix point: the infrastructure to track read files already exists, but the modification tools don't use it.

Bug 5: Compaction silently destroys read content, making reads unreliable

Code: packages/opencode/src/session/compaction.ts:334-338 (prune function)
When compaction prunes old tool outputs, it sets part.state.time.compacted = Date.now().

Code: packages/opencode/src/session/message-v2.ts:304-305

const outputText = part.state.time.compacted
  ? "[Old tool result content cleared]"
  : truncateToolOutput(...)

Code: packages/opencode/src/session/instruction.ts:20

if (part.state.time.compacted) continue

The sequence:

1. Model calls Read on a file, gets its content
2. Context grows, compaction triggers
3. The Read tool output is replaced with "[Old tool result content cleared]"
4. instruction.ts:extract() no longer counts this file as read
5. Model sees the Read call in history but the content is gone
6. Model edits the file based on stale memory
7. edit.ts succeeds (no check) but the edit may be wrong

This is a data integrity bug: compaction invalidates the read guarantee without signaling either the model or the editing tools.

Suggested fixes (for maintainers)

1. Add a Read-before-edit check in edit.ts: before executing, call Instruction.loaded(messages) to verify the target file was read in a recent non-compacted Read call
2. Same for write.ts when writing to existing files
3. Same for apply_patch.ts for "update" and "delete" hunks
4. Add read-before-edit guidance to apply_patch.txt
5. When compaction clears a Read output, ensure the model gets a signal that re-reading is needed

Environment

OpenCode latest, all platforms