AI edits

AI edits let you request code changes in plain language and then choose whether to apply them.

What you can do

Ask for new code, refactors, fixes, or tests.
Review proposed diffs before touching your files.
Accept to apply or Reject to discard.
Switch models directly in the AI panel with the model picker.
Keep separate model selections per conversation slot.
Use agent modes to steer the AI toward a specific task type.

Agent modes

Type / in the prompt input to open the agent mode picker. Each mode tells the AI what kind of task you need:

Mode	What it does
build	Full code edits — the default editing workflow.
plan	Analysis only — the agent reasons about your code and proposes an approach but does not edit files.
general	Research and multi-step tasks — broader reasoning across several files.

Start typing after / to filter the list (for example, /plan). The selected mode appears as a highlighted tag in your prompt so you can see which mode is active before sending. If you don’t pick a mode, the agent infers build or general from your prompt.

Webflow and Figma are not modes

There is no /webflow or /figma mode. When you connect Webflow or Figma, the agent automatically gains the matching MCP tools and uses them when relevant. Keep using /build (or /general) — the agent works across your code, the Webflow Designer, and your Figma designs as needed.

Conversations

Ssscript supports multiple parallel conversations and a history drawer.

Conversation slots

The AI panel uses a carousel of conversation slots. Each slot keeps its own prompt history, model selection, and pending edits. You can have up to 10 persisted slots and switch between them using the dot indicators at the top of the panel.

New conversation — press ⌘N (macOS) or Ctrl+N (Windows/Linux), or use File > New Conversation from the menu bar.
Close conversation — press ⌘W / Ctrl+W when in Agent view. The conversation is archived automatically.

Past conversations

Press ⌘B / Ctrl+B in Agent view to open the Past Conversations drawer. It shows up to 20 archived conversations. From the drawer you can:

Restore a previous conversation back into an active slot.
Remove conversations you no longer need.

On app startup, the last 3 conversation slots are restored automatically.

While the agent is running

You don’t have to wait for a response to finish:

Queue a follow-up — press Enter while a run is in progress to queue your next prompt. Queued prompts appear under Queued after current reply and send automatically when the current reply completes.
Merge & resend — press ⌘Enter / Ctrl+Enter while requesting to stop the current run and immediately resend your original prompt merged with what you just typed.
Stop — use the Stop control to halt the run. Any pending (unwritten) diffs from that run are discarded.

Auto-apply

By default, successful AI diffs are applied immediately — you can use Revert to undo. If you prefer to review every diff before it touches your files, disable auto-apply from the AI panel. When auto-apply is off, each diff shows Accept and Reject buttons and nothing is written to disk until you choose.

Revert and compare

When edits have been written to disk, the AI panel gives you a safety net:

Revert last AI edit — a banner (Last AI changes are on disk. Revert if needed.) lets you undo the most recent applied change.
Compare — open a side-by-side diff of the current file against the version before the AI edit.
History steps — for files with multiple accepted edits, each step in the conversation history offers its own Revert and Compare actions, so you can roll back to any earlier version.

Models

Ssscript supports multiple model families through the built-in AI proxy provider. Models are grouped by tier:

Fast tier (lower latency, lighter tasks)

Provider	Model
OpenAI	GPT-5 Mini
OpenAI	GPT-5.1 Instant
Anthropic	Claude Haiku 4.5
xAI	Grok 4.1 Fast
xAI	Grok Code Fast 1
Google	Gemini 3 Flash

Smart tier (deeper reasoning, complex edits)

Provider	Model
OpenAI	GPT-5.1
Anthropic	Claude Sonnet 4.6
xAI	Grok 4.1 Reasoning
Google	Gemini 3 Pro
Moonshot	Kimi K2.5
Alibaba	Qwen3 Next
Alibaba	Qwen3 Coder Next

For runtime wiring (structured plan-mode output, runtime auth, parallel conversation slots, diff preview limits), see Agent runtime integration. Internally the runtime refers to these as agent tags; in the app you select them through the / picker (/plan, /build, /general).

Model selection is available from the AI panel header:

Auto uses the default project configuration.
Explicit model choices are persisted for each conversation slot.

The in-app picker shows 13 explicit models plus Auto. For lightweight agent paths (plan mode, exploration), the runtime may use a smaller model automatically.

Webflow paste

When the AI produces HTML and CSS (either in fenced code blocks or as file diffs), the app can convert them into a Webflow-compatible clipboard format (XSCP). A Webflow Paste card appears in the chat with:

Copy for Webflow — copies the converted XSCP JSON to your clipboard so you can paste it directly into the Webflow Designer.
Copy HTML / Copy CSS — fallback buttons if conversion fails or you want the raw code instead.

The conversion pipeline extracts HTML/CSS from the assistant’s response, optionally refines it for Webflow compatibility via a lightweight LLM pass, then runs the XSCP converter. If the HTML cannot be converted to XSCP, you still get the raw HTML and CSS to paste manually.

Access requirements Paid

You must be signed in to send prompts when AI is account-gated.
If your session is stale, the app can ask you to sign in again.
If your plan does not include AI access, you will see the upgrade path in-app.

What the session looks like

You will see activity rows as the assistant works, including connection state.
The connection message appears as:

[connecting] Ssscript
Endpoint targets are available from the top status icon tooltips in the AI panel (Cloud and AI icons).

When no edits are produced, the app shows a neutral outlined notice instead of a warning-style alert.

Usage tracking and cost

AI usage tracking records token breakdown (prompt, completion, reasoning) per user.

If the AI gateway returns a cost value, that value is used directly.
If cost is not returned, spend is estimated using model-specific pricing.

When something goes wrong

Soft outcomes — If you see “No code changes were proposed” or similar, that is often expected (e.g. analysis-only prompts). Try a more concrete follow-up.
Real errors — If the panel shows a red error (not a soft warning), you can use Report this error to send a diagnostic bundle to support. You must be signed in. The button does not appear for neutral “no changes” notices or warning-style messages.

For details on what is included in that bundle (prompt tail, model, activity, terminal snippet, paths, runtime/MCP state), see Bug reports & AI diagnostics.

Automatic crash recovery (unhandled errors elsewhere in the app) may still show a toast asking whether to submit a report — that path is separate from the AI panel button but uses the same reporting pipeline.

Prompt quality tips

Name files or features directly in your prompt.
Use follow-up prompts to refine the previous result.
If a build fails, ask the AI to fix the exact error shown in the terminal.
Use /plan first to get the AI’s analysis, then follow up with /build to implement.
Mention file paths with @ to give the agent focused context.