How can I understand an AI agent's plan without reading hundreds of lines of markdown?

Coding agents write long plans before they write code: requirements, task lists, data models and design notes, often hundreds of lines of markdown. DiffGuardian's Plans tab reads that plan for you. It gives you a one-screen summary of what the plan will do, draws it as a diagram on your real folders, and lets you ask questions about it, by voice or in text, instead of reading it top to bottom. Plans is an experimental feature: it is off until you turn it on in Settings → Experimental.

#What you get

1. A summary first. One screen answers "what will this change, and what looks wrong?": how many requirements and tasks, which files it will create or edit, and which it names but can't find in your repository. A banner tells you whether to send it back to the agent, give it a closer read, or that nothing was flagged.
2. A diagram instead of prose. The Design tab draws the plan's components on your repository's real folders and colours each one existing, changed, new or missing. Click a box to see its tasks, the requirements they serve, its files, what it connects to and how far along it is.
3. An intent diff. Read the plan the way you read a pull request: what it adds, what it changes, what it names that doesn't exist, and which untouched components sit next to the change and could be affected by it.
4. The data model and contracts as pictures. If the plan has a data model, you get an entity-relationship diagram with every entity, field and relation. API contracts are laid out as readable request and response tables.
5. A conversation. Ask what a requirement means, what's missing, or which tasks depend on each other, and get an answer about this plan. Prefer to talk? Hold the key in the voice dock, ask, and hear the answer. Voice needs Pro.
6. Corrections you can hand back. Flag a finding or a requirement and DiffGuardian builds one correction prompt, citing each flagged item by the id and line the plan uses. You copy it to the agent that wrote the plan, and fix the plan before any code exists.

#Which plans it reads

Plans from Spec Kit (specs/), OpenSpec and Kiro, and loose plain-text PLAN.md files, in your project and in every worktree of it. They are plain files on your machine, and the tab refreshes as the agent edits them. Spec Kit plans bring the most: a data-model file becomes the entity-relationship diagram and a contracts/ folder becomes the API tables. A plain PLAN.md still gets the summary, the findings and the conversation, from whatever structure it has.

Because it reads every worktree, it pairs with watching your agents in their worktrees: read the plan first, then follow the code as it's written.

#What is checked, and what is generated

The summary, the findings and the file statuses come from reading the plan and comparing it with your checkout, with no AI model involved. Findings are structural: a file the plan edits that isn't in your checkout, a requirement no task serves, a task that cites no requirement, a requirement without acceptance criteria, and questions the plan left open.

The Design diagram's boxes and arrows are drawn by the model you choose, and every existing or changed component is checked against a real file or folder in your repository. A component the repository can't back is left out and listed, rather than drawn. Spoken answers are checked against the plan before they're read to you; if one can't be, you're told so instead of hearing it.

#Works alongside your agent and your reviewers

DiffGuardian doesn't write the plan and doesn't judge it for you. Your agent writes it, and DiffGuardian makes it quick to understand, so you can approve or redirect it with confidence before the work starts. It sits next to tools like CodeRabbit, GitHub Copilot and Greptile, which act on the pull request that comes later.

#Limits

• Plans is experimental and off by default. Turn it on in Settings → Experimental.
• It reads a plan's structure. A plan with no requirements or tasks it can parse gets a thin summary, and you're pointed to the Source tab.
• "Nothing flagged" means the analyzer found no structural problems. It does not mean the plan is right, so read what the plan will do.
• The design diagram and chat answers need an AI provider: your own key for Anthropic, OpenAI, Google Gemini or OpenRouter, or a local model.
• Voice answers need Pro.