Skip to content

AGENT.md

Latest commit

 

History

History
111 lines (87 loc) · 5.48 KB

AGENT.md

File metadata and controls

111 lines (87 loc) · 5.48 KB

Note++ — AI Agent Mode Spec

Status: design locked, implementation in progress.

Summary

A toggle in the AI panel that switches the chat from suggest-and-apply mode (current) to agent mode. In agent mode, the AI's response is treated as the new content for the file/selection, and the user sees a Monaco diff editor modal with Apply / Reject buttons before any change lands.

Multi-turn: each follow-up message rebuilds the system prompt from the current editor content, so iterative refinement works ("actually only do public functions" → AI starts from the current file state, not the original).

Locked decisions

Question Answer
Tier B — Diff preview (Monaco diff editor modal)
Multi-turn Yes. Each turn rebuilds the system prompt from current editor content
Scope Selection if present, else whole file (same as current chat)
Dangerous edits No special confirmation. Diff preview already provides review; Ctrl+Z handles rollback if applied in error
What shows in chat thread A summary line ("📝 Proposed change — review diff"), not the raw file content (which is huge)
Default mode on panel open Chat mode (current behaviour). User must explicitly toggle to Agent
Toggle persistence Per-session only — does not persist across app restarts (safer default)

Architecture

Agent mode ON:
  user types prompt
    → build NEW system prompt with current file content
    → REPLACE aiMessages[0] (system) with the new one
    → append user message
    → call /api/chat as before (keeps multi-turn history)
    → on token: stream into a hidden buffer, NOT the chat thread
    → on done:
        - chat thread shows only: "📝 Proposed change · review diff →"
        - open diff modal: original (left) vs AI output (right)
        - Apply → editor.executeEdits(...) replaces selection or full file
        - Reject → close modal, no editor change

UI

Agent toggle (AI panel header)

  • Small badge button next to "+ New chat": 🤖 Chat ↔ ⚡ Agent
  • Visually distinct when in agent mode (orange/red accent)
  • Tooltip: "Toggle Agent mode — AI's reply replaces editor content via diff preview"

Diff modal (#ai-diff-dialog)

┌─ Review AI change ────────────────────────────── × ┐
│  [Apply ✓]  [Reject ✕]                            │
├────────────────────────────────────────────────────┤
│  Original                  │  AI proposal          │
│  ─────────────────────────────────────────────────│
│  function foo(x) {         │  /** Adds 1 to x */  │
│    return x + 1;           │  function foo(x) {   │
│  }                         │    return x + 1;     │
│                            │  }                    │
│  ─────────────────────────────────────────────────│
│  ↑ red = removed        ↓ green = added           │
└────────────────────────────────────────────────────┘
  • Monaco's built-in monaco.editor.createDiffEditor does the heavy lifting (gutter markers, hunk navigation, syntax highlighting carries over from the file's language)
  • Modal closes on Apply, Reject, Esc, or × button
  • Reject preserves the chat history → user can ask "actually only public functions" and try again

Chat-thread render in agent mode

Instead of dumping the (huge) file content as the assistant's message, we show:

🤖 Assistant
📝 Proposed change — 24 lines added, 3 removed.   [Review diff →]

The "Review diff →" link reopens the diff modal even after closing it.

System prompt (agent mode)

You are an expert coding assistant in agent mode.

Output ONLY the complete updated content of the file/selection. No explanations,
no markdown fences, no preamble. Your response IS the new content — it will
directly replace what's currently there.

Preserve everything you weren't asked to change. Match indentation, style,
language conventions exactly.

Current scope: <selection or whole file>
Current content (this is what you must output an updated version of):
\`\`\`
<content>
\`\`\`

Implementation Phases

  1. HTML/CSS — agent toggle button + diff modal shell
  2. Toggle state — aiAgentMode boolean + toggleAiAgentMode() + visual state
  3. Capture mode — when aiAgentMode is on, route streaming tokens to a hidden buffer instead of the chat thread; replace assistant chat message with the summary line
  4. System-prompt rebuild — on each turn in agent mode, rebuild aiMessages[0] from the current editor content
  5. Diff modal — Monaco diff editor inside the modal, Apply/Reject wiring, Esc dismiss, "Review diff →" link to reopen
  6. Apply — editor.executeEdits on selection or whole-file range; close modal; set tab dirty
  7. Reject — just close modal; chat history retained for follow-up

Out of Scope (v1)

  • Hunk-by-hunk Apply / Reject (just whole-diff for now)
  • Multi-file edits (one file per turn — future)
  • Tool use / terminal access (future "Composer" tier)
  • Auto-commit after Apply (defer to manual save)