feat(ai-agent): question, plan and result cards in the AI sidebar by Suharshit · Pull Request #25 · Suharshit/draftly

Suharshit · 2026-09-15T18:51:23Z

Stacked on #23 (#21 → #22 → #23 → this). The base is feat/ai-design-turn-engine, so this diff shows step 4 only. After the PR below it merges, change this PR's base to main.

Why

Step 3 made the design agent conversational, but the sidebar still showed its questions and plans as plain text, so users had to type answers and approvals. This is step 4 of the AI sessions plan: dedicated cards with real controls.

What

QuestionsCard:
- numbered questions, each with its short "why"
- option chips (aria-pressed) plus a free-text answer per question; a typed answer wins over a chip
- Send answers, disabled until something is answered, and Skip, plan with assumptions
- once sent, the card stays read-only with the chosen answers
PlanCard:
- summary
- key decisions highlighted (choice, rationale, alternatives considered)
- components with role labels, folded after 6 behind "Show all N"
- key flows and assumptions
- Draw this plan, with an "or describe changes below" hint
ResultCard: what was added to the canvas, with a collapsible list of the decisions.
Which card is active: only the newest non-failed assistant reply is interactive, and only while no turn is pending, which mirrors the server's rules for open questions and the plan to draw. The composer placeholder changes while questions or a plan are open.
useAiSession.sendTurn: sends message, answers, generate and skip. The optimistic user message matches what the server stores (answers paired with their questions, the fixed generate and skip texts). Messages now expose kind and payload.
Shared code: TurnInput, the generate and skip texts and the payload readers moved into the client-safe lib/ai/agent-schema.ts, and session-turns.ts uses them instead of its own copies. Question and plan payloads also store the agent's short reply, shown as the card intro; older messages show none.

Verification

pnpm typecheck and pnpm lint pass.
29 static render checks (react-dom/server):
- open vs answered questions: chips, inputs, disabled send, skip, labelled groups, pressed chip, typed answer, no controls once answered
- plan sections: highlighted decisions, the alternatives line only when there are alternatives, 6 of 8 components with show-all, flows, assumptions
- the draw button only on the interactive plan, and empty sections omitted
- the result card
- the payload readers, including step 2 results without decisions
15 database turn checks pass after the shared-reader refactor.
Live flow through the Trigger worker (gemini-3.5-flash-lite): message → 3 questions (15.8s), skip → plan with 9 components and 3 decisions (12.3s), generate → "Added 9 components and 8 connections to the canvas." (12.6s).
Not checked in a browser: clicking chips, typing answers, send, skip and draw with a signed-in user.

🤖 Generated with Claude Code

The AI sidebar kept its chat in React state only, so a reload lost the transcript and any in-flight run. This adds the storage layer the multi-turn design agent will build on (not wired to the UI yet). - AiSession / AiMessage Prisma models + additive migration; sessions cascade with their project, messages cascade with their session - lib/ai/session-store.ts: 7-day sliding expiry, 10 sessions per user per project, 60 messages per session; reads ignore expired rows - GET/POST /api/projects/[projectId]/ai-sessions and GET/DELETE /api/projects/[projectId]/ai-sessions/[sessionId], private to the session's creator - Daily Vercel Cron cleanup at /api/cron/ai-sessions/cleanup, gated by CRON_SECRET (public in proxy.ts) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

A reload no longer loses the chat or an in-flight run. Generation is still one-shot; this changes persistence and resume only. - POST /api/projects/[projectId]/ai-sessions/[sessionId]/turns stores the prompt and a PENDING reply carrying the run id (replaces POST /api/ai/design, which is removed) - Replies are settled from runs.retrieve whenever a session is read, so a turn is saved even if the tab closed mid-run; concurrent reads settle once - useAiSession replaces useDesignAgent: reopens the last chat, resumes pending runs via Realtime with a 4s polling fallback, stores new chats lazily on the first message - Sidebar session bar with saved-chat history (delete, expiry) and New chat - Limits moved to lib/ai/session-limits.ts for client use; formatRelativeTime extracted to lib/relative-time.ts Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Each turn is still one design-agent run, but the agent now keeps a requirements brief, asks up to 3 rounds of clarifying questions when an unknown would change the architecture, proposes a plan (components, flows, decisions with rationale and alternatives, assumptions), and only draws the diagram once the plan is approved. - lib/ai/agent-schema.ts: brief, question, plan, turn-analysis and run-result schemas; list limits trimmed in code rather than enforced as maxItems, which models don't reliably honour - lib/ai/prompts.ts, lib/ai/design-agent-engine.ts: analyze/plan/ generate on generateText + Output.object; code-enforced decision rules; per-step thinking level (Gemini 3) or budget (Gemini 2.x); one retry on schema mismatch; per-call timeouts - design-agent task: single attempt, maxDuration 300s - Turns route accepts message, answers, generate and skip; stored replies are validated run output (QUESTIONS / PLAN / RESULT) and run failures are shown as friendly messages Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The AI Architect chat now renders the design agent's turns as cards instead of plain text, so users answer, skip and approve with buttons. - QuestionsCard: option chips and a free-text answer per question, Send answers and Skip; read-only with the chosen answers once sent - PlanCard: summary, highlighted key decisions with rationale and alternatives, components (folded after 6), flows, assumptions, and Draw this plan - ResultCard: canvas summary with collapsible decisions - Only the newest non-failed reply is interactive, and only while no turn is pending, mirroring the server's rules - useAiSession.sendTurn sends message, answers, generate and skip, with optimistic user text matching what the server stores - TurnInput, the generate/skip texts and payload readers now live in the client-safe agent-schema and are shared with the server; question and plan payloads also store the agent's reply Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

vercel · 2026-09-15T18:51:29Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
draftly	Ready	Preview	Sep 15, 2026 6:58pm UTC

coderabbitai · 2026-09-15T18:51:29Z

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 0d22e3f0-a81e-4769-ba18-1bea3940025c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

GitHub main has steps 1-3 only; steps 4-6 were merged into stacked base branches. Content matches local exactly; a stacked merge into main conflicts in 11 files, so main needs a sync commit whose tree is the stack head. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* feat(ai-agent): multi-turn design agent - clarify, plan, then generate Each turn is still one design-agent run, but the agent now keeps a requirements brief, asks up to 3 rounds of clarifying questions when an unknown would change the architecture, proposes a plan (components, flows, decisions with rationale and alternatives, assumptions), and only draws the diagram once the plan is approved. - lib/ai/agent-schema.ts: brief, question, plan, turn-analysis and run-result schemas; list limits trimmed in code rather than enforced as maxItems, which models don't reliably honour - lib/ai/prompts.ts, lib/ai/design-agent-engine.ts: analyze/plan/ generate on generateText + Output.object; code-enforced decision rules; per-step thinking level (Gemini 3) or budget (Gemini 2.x); one retry on schema mismatch; per-call timeouts - design-agent task: single attempt, maxDuration 300s - Turns route accepts message, answers, generate and skip; stored replies are validated run output (QUESTIONS / PLAN / RESULT) and run failures are shown as friendly messages Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-agent): question, plan and result cards in the AI sidebar (#25) * feat(ai-sessions): add persistent storage for AI design sessions The AI sidebar kept its chat in React state only, so a reload lost the transcript and any in-flight run. This adds the storage layer the multi-turn design agent will build on (not wired to the UI yet). - AiSession / AiMessage Prisma models + additive migration; sessions cascade with their project, messages cascade with their session - lib/ai/session-store.ts: 7-day sliding expiry, 10 sessions per user per project, 60 messages per session; reads ignore expired rows - GET/POST /api/projects/[projectId]/ai-sessions and GET/DELETE /api/projects/[projectId]/ai-sessions/[sessionId], private to the session's creator - Daily Vercel Cron cleanup at /api/cron/ai-sessions/cleanup, gated by CRON_SECRET (public in proxy.ts) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-sessions): run the AI Architect chat on stored sessions A reload no longer loses the chat or an in-flight run. Generation is still one-shot; this changes persistence and resume only. - POST /api/projects/[projectId]/ai-sessions/[sessionId]/turns stores the prompt and a PENDING reply carrying the run id (replaces POST /api/ai/design, which is removed) - Replies are settled from runs.retrieve whenever a session is read, so a turn is saved even if the tab closed mid-run; concurrent reads settle once - useAiSession replaces useDesignAgent: reopens the last chat, resumes pending runs via Realtime with a 4s polling fallback, stores new chats lazily on the first message - Sidebar session bar with saved-chat history (delete, expiry) and New chat - Limits moved to lib/ai/session-limits.ts for client use; formatRelativeTime extracted to lib/relative-time.ts Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-agent): multi-turn design agent - clarify, plan, then generate Each turn is still one design-agent run, but the agent now keeps a requirements brief, asks up to 3 rounds of clarifying questions when an unknown would change the architecture, proposes a plan (components, flows, decisions with rationale and alternatives, assumptions), and only draws the diagram once the plan is approved. - lib/ai/agent-schema.ts: brief, question, plan, turn-analysis and run-result schemas; list limits trimmed in code rather than enforced as maxItems, which models don't reliably honour - lib/ai/prompts.ts, lib/ai/design-agent-engine.ts: analyze/plan/ generate on generateText + Output.object; code-enforced decision rules; per-step thinking level (Gemini 3) or budget (Gemini 2.x); one retry on schema mismatch; per-call timeouts - design-agent task: single attempt, maxDuration 300s - Turns route accepts message, answers, generate and skip; stored replies are validated run output (QUESTIONS / PLAN / RESULT) and run failures are shown as friendly messages Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-agent): question, plan and result cards in the AI sidebar The AI Architect chat now renders the design agent's turns as cards instead of plain text, so users answer, skip and approve with buttons. - QuestionsCard: option chips and a free-text answer per question, Send answers and Skip; read-only with the chosen answers once sent - PlanCard: summary, highlighted key decisions with rationale and alternatives, components (folded after 6), flows, assumptions, and Draw this plan - ResultCard: canvas summary with collapsible decisions - Only the newest non-failed reply is interactive, and only while no turn is pending, mirroring the server's rules - useAiSession.sendTurn sends message, answers, generate and skip, with optimistic user text matching what the server stores - TurnInput, the generate/skip texts and payload readers now live in the client-safe agent-schema and are shared with the server; question and plan payloads also store the agent's reply Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

* feat(ai-sessions): add persistent storage for AI design sessions The AI sidebar kept its chat in React state only, so a reload lost the transcript and any in-flight run. This adds the storage layer the multi-turn design agent will build on (not wired to the UI yet). - AiSession / AiMessage Prisma models + additive migration; sessions cascade with their project, messages cascade with their session - lib/ai/session-store.ts: 7-day sliding expiry, 10 sessions per user per project, 60 messages per session; reads ignore expired rows - GET/POST /api/projects/[projectId]/ai-sessions and GET/DELETE /api/projects/[projectId]/ai-sessions/[sessionId], private to the session's creator - Daily Vercel Cron cleanup at /api/cron/ai-sessions/cleanup, gated by CRON_SECRET (public in proxy.ts) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-sessions): run the AI Architect chat on stored sessions A reload no longer loses the chat or an in-flight run. Generation is still one-shot; this changes persistence and resume only. - POST /api/projects/[projectId]/ai-sessions/[sessionId]/turns stores the prompt and a PENDING reply carrying the run id (replaces POST /api/ai/design, which is removed) - Replies are settled from runs.retrieve whenever a session is read, so a turn is saved even if the tab closed mid-run; concurrent reads settle once - useAiSession replaces useDesignAgent: reopens the last chat, resumes pending runs via Realtime with a 4s polling fallback, stores new chats lazily on the first message - Sidebar session bar with saved-chat history (delete, expiry) and New chat - Limits moved to lib/ai/session-limits.ts for client use; formatRelativeTime extracted to lib/relative-time.ts Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-agent): multi-turn design agent - clarify, plan, then generate Each turn is still one design-agent run, but the agent now keeps a requirements brief, asks up to 3 rounds of clarifying questions when an unknown would change the architecture, proposes a plan (components, flows, decisions with rationale and alternatives, assumptions), and only draws the diagram once the plan is approved. - lib/ai/agent-schema.ts: brief, question, plan, turn-analysis and run-result schemas; list limits trimmed in code rather than enforced as maxItems, which models don't reliably honour - lib/ai/prompts.ts, lib/ai/design-agent-engine.ts: analyze/plan/ generate on generateText + Output.object; code-enforced decision rules; per-step thinking level (Gemini 3) or budget (Gemini 2.x); one retry on schema mismatch; per-call timeouts - design-agent task: single attempt, maxDuration 300s - Turns route accepts message, answers, generate and skip; stored replies are validated run output (QUESTIONS / PLAN / RESULT) and run failures are shown as friendly messages Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-agent): question, plan and result cards in the AI sidebar The AI Architect chat now renders the design agent's turns as cards instead of plain text, so users answer, skip and approve with buttons. - QuestionsCard: option chips and a free-text answer per question, Send answers and Skip; read-only with the chosen answers once sent - PlanCard: summary, highlighted key decisions with rationale and alternatives, components (folded after 6), flows, assumptions, and Draw this plan - ResultCard: canvas summary with collapsible decisions - Only the newest non-failed reply is interactive, and only while no turn is pending, mirroring the server's rules - useAiSession.sendTurn sends message, answers, generate and skip, with optimistic user text matching what the server stores - TurnInput, the generate/skip texts and payload readers now live in the client-safe agent-schema and are shared with the server; question and plan payloads also store the agent's reply Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(ai-agent): canvas-aware generation with roles, async edges and repair The design agent now sees what is already on the canvas, so follow-up requests extend the existing design instead of redrawing it, and generated diagrams match the starter templates' visual conventions. - lib/ai/canvas-summary.ts: compact canvas listing with refs (ex-N), kinds, used connection points, connections and notes, passed to the analyze, plan and generate steps - Generation schema: role + kicker instead of legacy dark colorId, sync/async delivery (async edges dashed); edges may reference existing components by ref - buildCanvasGraph: resolves refs, folds redrawn or repeated components into the existing/first node, drops unknown and duplicate links, and assigns connection points counting existing edges - lib/ai/graph-validation.ts + one repair call: duplicate/redrawn components, unknown refs, self loops, repeated links, >4 connections, unconnected components, missing plan components - Shared NODE_ROLES / getRoleFill and SHAPE_KICKERS in types/canvas.ts (starter templates use them) - Plan text no longer shows internal canvas refs Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(ai-sessions): document AI design sessions in context files and README Brings the tracked docs in line with AI sessions steps 1-5. Docs only. docs.md is gitignored, so its matching refresh stays local. - context/architecture-context.md: AI Design Agent section (turn = run, turn steps, code-enforced rules, settle on read, reliability, model config) and invariants 5-9 - context/project-overview.md: canvas snapshots in Vercel Blob (was "filesystem"); chat cards and canvas-aware generation - context/ui-context.md: AI sidebar session bar, saved chats and cards follow the paper conventions - context/progress-tracker.md: step 6 entry and current goal - README.md: local AI design agent setup Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(spec): generate a Markdown technical spec from the canvas The Specs tab now turns the current canvas into a Markdown technical spec that downloads automatically: an overview, key decisions highlighted as callouts, one Mermaid diagram section per group of connected components, a component reference, connections, and risks. - spec-agent Trigger task: reads the room, builds the spec graph in code, makes one structured model call for the prose (sanitised against the graph, internal refs stripped), renders the Markdown in code; aborts on an empty or oversized canvas - Stored on read like AI turns: GET /api/projects/[id]/spec uploads a finished run to private Vercel Blob and records it on Project (guarded on the run id); POST starts a run (409 while one is pending); GET .../spec/markdown downloads the file - Decisions recorded in the user's AI design sessions are included and tagged as recorded - Migration: TaskRun.kind (DESIGN or SPEC) and Project spec fields - useSpecGenerator and SpecPanel replace the placeholder Specs tab - Shared lib/ai/model.ts, lib/ai/run-failure.ts and lib/canvas-room.ts used by both agents Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(tracker): record stacked merges of #25-#27 and the main sync need GitHub main has steps 1-3 only; steps 4-6 were merged into stacked base branches. Content matches local exactly; a stacked merge into main conflicts in 11 files, so main needs a sync commit whose tree is the stack head. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(tracker): record merging the AI stack into main Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

Suharshit and others added 4 commits September 15, 2026 22:16

vercel Bot deployed to Preview September 15, 2026 18:52

View deployment

Merge branch 'feat/ai-design-turn-engine' into feat/ai-agent-cards

2d2d109

vercel Bot deployed to Preview September 15, 2026 18:58

View deployment

Suharshit mentioned this pull request Sep 15, 2026

feat(ai-agent): canvas-aware generation with roles, async edges and repair #26

Merged

Suharshit merged commit ceab34e into feat/ai-design-turn-engine Sep 15, 2026
5 checks passed

This was referenced Sep 15, 2026

docs(ai-sessions): document AI design sessions in context files and README #27

Merged

feat(spec): generate a Markdown technical spec from the canvas #28

Merged

Navigation Menu

Repository navigation

feat(ai-agent): question, plan and result cards in the AI sidebar - #25

feat(ai-agent): question, plan and result cards in the AI sidebar#25
Suharshit merged 5 commits into
feat/ai-design-turn-enginefrom
feat/ai-agent-cards

Suharshit commented Sep 15, 2026

Uh oh!

vercel Bot commented Sep 15, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Sep 15, 2026 •

edited

Loading

Review skipped

Uh oh!

Uh oh!

This branch was successfully deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Suharshit commented Sep 15, 2026

Why

What

Verification

Uh oh!

vercel Bot commented Sep 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Sep 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

Uh oh!

This branch was successfully deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

vercel Bot commented Sep 15, 2026 •

edited

Loading

coderabbitai Bot commented Sep 15, 2026 •

edited

Loading