Agent Mode

Agent Mode runs an autonomous AI agent that can read files, write code, run shell commands, and iterate until the task is done. Unlike Inline Editing which replaces selected code, Agent Mode operates on your entire project with full tool access.

How It Works

1. Describe the task — pass a natural-language prompt or visually select text and run :DwightAgent.
2. Planning — by default the agent creates a plan first. Use --jump to skip planning, or --plan to force it.
3. Execution — the agent reads files, writes code, runs tests, and iterates. Each tool call appears in the status buffer.
4. Lessons — after completion, Dwight extracts lessons (patterns that worked, errors and fixes) for future sessions.
5. Review — check the diff with :DwightDiffReview or step through the session with :DwightReplay latest.

:DwightAgent Add input validation to the user creation endpoint
:DwightAgent --plan Refactor the auth module to use JWT
:DwightAgent --jump Fix the failing test in test_users.py

Planning

By default, the agent creates a brief plan before writing code. The plan appears in the status buffer so you can see the approach before execution begins.

• --plan — always create a plan first (useful for complex tasks)
• --jump — skip planning and start executing immediately (useful for simple fixes)
• No flag — agent decides based on task complexity

Visual Range

Select text in any buffer and use it as the prompt:

:'<,'>DwightAgent

The selected text becomes the task description. This works well with the Whiteboard — write your thoughts, select the relevant paragraph, and fire it to the agent.

Agent Status Buffer

While the agent runs, :DwightAgentStatus toggles a live status buffer showing:

• Current tool call (file read, write, bash command)
• Agent reasoning and decisions
• Progress through the plan
• Errors and retries

The buffer updates in real-time. Close it with q or toggle it off.

Diff Review

After the agent finishes, review all changes in one view:

:DwightDiffReview

Opens a split buffer with the full diff of every file the agent modified. This is the same diff format as git diff — familiar and easy to scan.

Lessons

Dwight extracts lessons from every agent session — what worked, what failed, what to do differently. Lessons are stored in .dwight/lessons.json and automatically included in future agent prompts.

:DwightLessons                " View all lessons with metadata
:DwightLessons stats          " Summary: total, stale, usage counts
:DwightLessons consolidate    " Merge similar lessons
:DwightLessons clear          " Remove all lessons

Each lesson tracks:

• Keywords for matching to future tasks
• Reinforcement count — how many times the same lesson was learned
• Use count — how many times it was injected into a prompt
• Staleness — lessons older than 30 days are flagged in :checkhealth dwight

Session Logs

Every agent session is logged to .dwight/agentic-logs/ as JSONL. Browse past sessions:

:DwightAgentLog              " Telescope picker of past sessions
:DwightSessionLog            " View the persistent session log

See Session Replay for stepping through sessions event by event.

Tips

• Start with --plan for unfamiliar tasks. The plan shows you what the agent intends to do before it writes code. Cancel early if the approach is wrong.
• Use visual range from the Whiteboard. Write detailed requirements on the whiteboard, select the relevant section, and fire with :'<,'>DwightAgent.
• Review lessons periodically. Run :DwightLessons consolidate to merge duplicates and keep the lesson set focused.
• Check the diff before committing. :DwightDiffReview after every agent run catches unexpected changes.

Commands

See Also

• Auto Mode -- for multi-step tasks that run multiple agent sessions sequentially
• Inline Editing -- for scoped, in-buffer edits instead of autonomous agents
• Session Replay -- to step through agent sessions event by event
• Whiteboard -- to draft prompts and fire them to the agent