feat: Plan review and the AI assistant in the dashboard by nfebe · Pull Request #67 · flatrun/ui

nfebe · 2026-06-07T22:27:27Z

Front-end for the agent's plan/apply and AI features. Depends on the agent feat/ai-native PR; merge that first (builds against its new endpoints).

Plan review

When a deployment requires plan review, Save/Delete/Setup/service actions transparently switch to a preview modal: summary counts, per-change reasons, expandable before/after diffs, sensitive values masked. Apply/discard the saved plan; stale previews surface as a warning.
Deployment settings moved under Configuration as a reusable sub-tab; split buttons target the whole deployment or a single service.

AI assistant

Global launcher in the header opens a chat against the whole instance; the same chat is seeded from any log view or operation result.
Shows the assistant's investigation inline (each read-only lookup as an expandable step); proposed lookups get CLI-style Allow/Decline buttons per the session's auto-run/approve toggle.
Suggested fixes run only on click through the guarded APIs. Bulky logs go to the model as hidden context, not echoed into the transcript.
New AI Settings tab; new features use lucide icons (PrimeIcons being phased out).

413 tests, lint and build green.

Mutating actions run directly by default. When a deployment has plan review required, the same buttons transparently switch to a preview flow: the change is computed server side and shown in a review modal with summary counts, per-change reasons and expandable before/after diffs, with sensitive values masked behind an explicit reveal. Apply and discard act on the saved plan, and a stale preview surfaces as a warning instead of applying blindly. Deployment settings moved under the Configuration tab, which now uses a reusable attached sub-tab component, and were restyled as a single expandable list with a Require Plan Review switch next to Protected Mode. Start, stop, restart, rebuild and pull are split buttons that target the whole deployment or any single service, and each service row gains its own start, stop, restart and rebuild controls.

Every surface that shows logs now carries an AI action in the log toolbar: deployment logs, operation output in both operation modals, and any future use of the shared log viewer. The assistant opens in one global modal where the user can switch between diagnose, improve, secure and explain, and ask free-text follow-up questions against the same context. When AI is not configured, the entry points stay visible and lead to the settings page instead of silently hiding. Analyses can return suggested actions; each is shown with its command or action and reasoning, and runs only on explicit click through the same APIs a human would use, so permissions, protected mode and plan review apply unchanged, with output shown inline. Failed operations default to diagnosis and successful ones to explanation. The assistant is configured from a new settings tab: any OpenAI-compatible endpoint, model name and a write-only API key, applied to the running agent without a restart.

The AI icon used a glyph that only exists in PrimeIcons 7 while the app pinned version 6, so every assistant button rendered as an empty space; the icon set is upgraded, which is additive and keeps all existing icons. The assistant also kept telling users to enable AI after they had just configured it, because the enabled status was cached from before setup; the status is now re-checked when it reads disabled and refreshed immediately after saving AI settings.

All AI entry points (log toolbar action, assistant modal header and hints, the assist button, the AI settings tab) now render the lucide sparkles icon instead of a PrimeIcons glyph that only exists in a newer icon-font version than the app ships, which left every AI button as an empty space. The icon set version is reverted; new features should use lucide as the project moves away from PrimeIcons, and the settings tab bar can now render lucide icons alongside the legacy class-based ones.

The AI assistant is now a chat, not a one-shot answer. A persistent launcher sits top-center in the dashboard header whenever AI is enabled, opening a general session where you can ask about the instance, its deployments, networks, or anything going wrong, and keep asking follow-ups in the same thread. The chat shows the assistant's investigation inline: each read-only lookup it runs (or proposes) appears as an expandable step with its result, so you see how it reached an answer. A per-session toggle chooses whether those lookups run automatically or pause for your approval, and pending lookups are shown with allow or decline before anything runs. Suggested fixes still execute only on click through the guarded APIs. Every place that shows logs (deployment logs, operation output) and every operation modal opens this same chat, seeded with what you were looking at, so the assistant starts from your context and can dig further with its tools.

… transcript A proposed read-only lookup now shows with Allow and Decline buttons beside it, like a CLI tool prompt: a single pending call resolves on one click, and when several are proposed each gets its own decision (with allow-all and decline-all), showing each outcome as the batch runs. Seeded analyses (from a log view or an operation result) send the logs or output to the model as hidden context while the conversation shows only a short prompt, so the chat no longer echoes a wall of log lines. The header launcher is larger and now reads as running the instance, not just asking, and the empty prompt invites both questions and actions.

sourceant · 2026-06-07T22:27:41Z

Code Review Summary

This PR introduces a robust Plan/Apply flow and an AI assistant integrated into the dashboard. It successfully bridges the gap between raw operations and user-reviewed changes, particularly for destructive actions.

🚀 Key Improvements

Integrated Plan Review modal with diffing and sensitive value masking.
AI Assistant for log analysis and operation troubleshooting with 'Auto-run' lookup capabilities.
Reorganized Deployment Settings into a sub-tab under Configuration.
Introduced SplitActionButtons to support scoped actions (Deployment vs Service).

💡 Minor Suggestions

Consolidate AI context generation logic.
Add console logging for failed AI JSON parsing.
Truncate logs in AI context to prevent token overflow.

sourceant

Review complete. See the overview comment for a summary.

sourceant · 2026-06-07T22:27:41Z

+};
+
+const renderMarkdown = (content: string) =>
+  DOMPurify.sanitize(marked.parse(content, { async: false }) as string);


The default marked.parse behavior can include IDs and other attributes that might bypass basic sanitization or cause CSS collisions. It is safer to explicitly configure the sanitizer to allow only a safe subset of tags, especially when dealing with AI-generated markdown which can be unpredictable.

Suggested change

DOMPurify.sanitize(marked.parse(content, { async: false }) as string);

DOMPurify.sanitize(marked.parse(content, { async: false }) as string, { USE_PROFILES: { html: true } });

sourceant · 2026-06-07T22:27:42Z

+const parts = computed(() => diffLines(props.before ?? "", props.after ?? ""));
+
+const partLines = (part: ChangeObject<string>) => {
+  const lines = part.value.split("\n");


In the partLines function, using lines.pop() modifies the array if the last element is empty. While functional, it's slightly cleaner to filter or slice to handle the trailing newline often returned by diffing libraries.

Suggested change

const lines = part.value.split("\n");

const lines = part.value.split("\n");

return lines.length > 1 && lines[lines.length - 1] === "" ? lines.slice(0, -1) : lines;

cloudflare-workers-and-pages · 2026-06-07T22:31:15Z

Deploying flatrun-ui with Cloudflare Pages

Latest commit:	`7d5ee52`
Status:	✅ Deploy successful!
Preview URL:	https://01abb2d4.flatrun-ui.pages.dev
Branch Preview URL:	https://feat-ai-native.flatrun-ui.pages.dev

View logs

sourceant

Review complete. See the overview comment for a summary.

sourceant · 2026-06-07T22:31:29Z

+
+const parts = computed(() => diffLines(props.before ?? "", props.after ?? ""));


When both 'before' and 'after' are null, diffLines might return an empty array or unexpected result. While types allow null, the UI should ideally handle the empty state explicitly or prevent rendering if both are missing.

Suggested change

const parts = computed(() => diffLines(props.before ?? "", props.after ?? ""));

const parts = computed(() => {

if (props.before === null && props.after === null) return [];

return diffLines(props.before ?? "", props.after ?? "");

});

sourceant · 2026-06-07T22:31:29Z

+  const detail = args.command || args.path || args.service || args.deployment || "";
+  return detail ? `${humanTool(call.name)}: ${detail}` : humanTool(call.name);
+};
+const parseArgs = (raw: string): Record<string, string> => {


JSON.parse can fail if the AI returns malformed arguments. While there is a catch block, logging the error to console might help debug integration issues with different LLM providers.

Suggested change

const parseArgs = (raw: string): Record<string, string> => {

const parseArgs = (raw: string): Record<string, string> => {

try {

return JSON.parse(raw || "{}");

} catch (e) {

console.error("AI Assistant: Failed to parse tool arguments", raw, e);

return {};

}

};

nfebe added 6 commits June 7, 2026 18:53

sourceant Bot suggested changes Jun 7, 2026

View reviewed changes

style: Apply prettier formatting to AI assistant files

7d5ee52

sourceant Bot reviewed Jun 7, 2026

View reviewed changes

nfebe merged commit aa47ce6 into main Jun 7, 2026
5 checks passed

nfebe deleted the feat/ai-native branch June 7, 2026 22:35

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: Plan review and the AI assistant in the dashboard#67

feat: Plan review and the AI assistant in the dashboard#67
nfebe merged 7 commits into
mainfrom
feat/ai-native

nfebe commented Jun 7, 2026

Uh oh!

sourceant Bot commented Jun 7, 2026 •

edited

Loading

Uh oh!

sourceant Bot left a comment

Uh oh!

sourceant Bot Jun 7, 2026

Uh oh!

sourceant Bot Jun 7, 2026

Uh oh!

cloudflare-workers-and-pages Bot commented Jun 7, 2026 •

edited

Loading

Uh oh!

sourceant Bot left a comment

Uh oh!

sourceant Bot Jun 7, 2026

Uh oh!

sourceant Bot Jun 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

	DOMPurify.sanitize(marked.parse(content, { async: false }) as string);
	DOMPurify.sanitize(marked.parse(content, { async: false }) as string, { USE_PROFILES: { html: true } });

	const lines = part.value.split("\n");
	const lines = part.value.split("\n");
	return lines.length > 1 && lines[lines.length - 1] === "" ? lines.slice(0, -1) : lines;


		const parts = computed(() => diffLines(props.before ?? "", props.after ?? ""));

-const parts = computed(() => diffLines(props.before ?? "", props.after ?? ""));
+const parts = computed(() => {
+  if (props.before === null && props.after === null) return [];
+  return diffLines(props.before ?? "", props.after ?? "");
+});

-const parseArgs = (raw: string): Record<string, string> => {
+const parseArgs = (raw: string): Record<string, string> => {
+  try {
+    return JSON.parse(raw || "{}");
+  } catch (e) {
+    console.error("AI Assistant: Failed to parse tool arguments", raw, e);
+    return {};
+  }
+};

Conversation

nfebe commented Jun 7, 2026

Plan review

AI assistant

Uh oh!

sourceant Bot commented Jun 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Code Review Summary

🚀 Key Improvements

💡 Minor Suggestions

Uh oh!

sourceant Bot left a comment

Choose a reason for hiding this comment

Uh oh!

sourceant Bot Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

sourceant Bot Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

cloudflare-workers-and-pages Bot commented Jun 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Deploying flatrun-ui with Cloudflare Pages

Uh oh!

sourceant Bot left a comment

Choose a reason for hiding this comment

Uh oh!

sourceant Bot Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

sourceant Bot Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

sourceant Bot commented Jun 7, 2026 •

edited

Loading

cloudflare-workers-and-pages Bot commented Jun 7, 2026 •

edited

Loading