Skip to content

feat: agent UI kit — conversation primitives, run & governance components, 24-block catalog - #71

Open
forge-ui wants to merge 21 commits into
mainfrom
feat/agent-kit
Open

forge-ui wants to merge 21 commits into
mainfrom
feat/agent-kit

Conversation

@forge-ui

Copy link
Copy Markdown
Owner

Summary

Rebuilds the agent UI around three lines, benchmarked against HeroUI (Pro chat + Agents) and affectUI.

A. Conversation primitives (core) — host-driven, one set instead of overlapping ones

  • Reasoning, ThinkingIndicator (controlled thinking state, live timer, phases)
  • ToolCall / ToolCallGroup (status, timer, args/result, retry, inline approval)
  • Question (multi-question, keyboard answers) and Approval (risk, before→after, scopes once/session/always, deny reason, lapse timer)
  • CitedAnswer / Citation (chips preview the quote and highlight the sentence)
  • RunControls (stop/resume/retry/steer), PromptQueue, AttachmentList, AgentWelcome
  • MessageActions gains version switching, edit and feedback reasons
  • AgentParts now renders reasoning and tool calls with these
  • AgentPanel, ChoiceCard, ConfirmCard, RunProgress move from the agent package into core (agent re-exports them)

B. Generative UI catalog (agent) — 8 → 24 blocks

  • display: record, list, timeline, tags, compare, changes, flow, images, callout
  • data: chart (line/area/bar/hbar/donut), progress, funnel
  • asking/acting: questions, form, confirm, plan, actions
  • ui accepts a component list; prompt lines are generated from the catalog; the site uses a 10-block subset

C. Run & governance (core)

  • AgentPlan (editable before run, retry from failed step), AgentTrail, AgentCheckpoints
  • ChangeLedger (before/after, undo with dependents), AutonomySelector
  • AgentInbox, RunningNow, AwaySummary
  • ContextUsage, TraceWaterfall, AgentTree, AgentLanes

Docs: /components/agent reorganised by layer with demos and full API tables; /cases/agent gains five end-to-end scenarios.

Release

core 0.9.0 and agent 0.5.0. The block registry now requires core ^0.9.0, so publish core (and agent) from this branch before merging.

Test plan

  • core: typecheck, lint, 352 tests (11 new), component audit, package check (budgets raised, API snapshot updated)
  • agent: typecheck, 48 tests
  • site: pnpm check (regenerated inventory, llms, registry, SEO)
  • Chrome screenshots of every new demo and scenario; refund scenario driven end to end

🤖 Generated with Claude Code

forge-ui and others added 6 commits November 17, 2026 18:33
… blocks

- core: Reasoning, ThinkingIndicator, ToolCall(Group), Question, Approval, CitedAnswer,
  PromptQueue, RunControls, AttachmentList, AgentWelcome, AgentPlan, AgentTrail,
  ChangeLedger, AgentCheckpoints, AutonomySelector, AgentInbox, RunningNow, AwaySummary,
  ContextUsage, TraceWaterfall, AgentTree, AgentLanes; MessageActions gains versions,
  edit and feedback reasons; AgentParts renders reasoning and tool calls with them
- AgentPanel, ChoiceCard, ConfirmCard and RunProgress move from the agent package to core
- agent: catalog grows from 8 to 24 blocks (record, list, timeline, compare, flow, progress,
  funnel, form, questions, confirm, plan, actions…); ui accepts a component list and the
  prompt lines are generated from the catalog

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /components/agent: sections for thinking and tools, asking and approval, cited answers,
  composer and panel, runs, governance, operations and observability, with full API tables
- /cases/agent: five end-to-end scenarios (analysis Q&A, refund handling, multi-step plan,
  background task center, multi-agent run)
- icons in the new components take their colour through the color prop

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@vercel

vercel Bot commented Nov 17, 2026 •

Copy link
Copy Markdown

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

Project Deployment Actions Updated
forge Ready Ready Preview Nov 18, 2026 3:29am UTC

forge-ui and others added 2 commits November 17, 2026 20:48
- shared internal pieces: SurfaceCard-style AgentCard (header/footer bands), StatusBadge-soft
  SoftBadge, ApprovalCard choice marks and accent buttons, TextField-look inputs, Collapse motion
- text-sm / text-xs only; StarsLinear thinking header like ThinkingTrace; ToolChips-style rows
- list enter/exit/reorder animations (plan, queue, inbox); FileTypeIcon in attachments;
  ButtonGroup for autonomy, TabBar for inbox filters
- motion and size props across the kit

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
forge-ui and others added 2 commits November 17, 2026 21:11
- /components/agent groups every agent component by layer (conversation and input, thinking
  and tools, asking and approval, answers and citations, runs, governance, operations,
  observability) instead of a separate section for the earlier ones
- ChatInputBar with quick-action cards and its natural motion is the kit's composer; AgentPanel
  passes hasMessages so the cards fold once the chat has messages
- /cases/agent conversation scenarios end with the same composer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
forge-ui and others added 2 commits November 17, 2026 21:20
…nputBar quick actions

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
forge-ui and others added 3 commits November 17, 2026 21:48
…d InlineDiff

- ApprovalCard renders through Question (skippable survey mode, labels, sent status);
  ThinkingTrace through Reasoning; ToolChips through ToolCallGroup/ToolCall; AskAi points to AgentPanel
- Composer, ComposerHeader, ComposerInput, ComposerToolbar, ComposerAction, ComposerModelMenu,
  ComposerSend: ChatInputBar's surface, quick-action fan and send button as parts (SendButton
  moves to a shared internal module)
- ConversationList (grouped by day, search, pin, rename, delete, running/unread) replaces the
  AgentPanel history menu; AgentSession gains updatedAt, pinned, running, unread
- InlineDiff + resolveInlineDiff for reviewing AI edits to prose
- when-to-use / when-not notes for every agent module, a selection table on /components/agent,
  and a document-writing scenario on /cases/agent

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…Diff

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tested all 24 blocks against qwen3.8-flash with 16 business prompts. Fixes from the run:
- inner plain double quotes in strings are escaped before parsing
- lists and strings over a limit are cut to it instead of dropping the block
- compare allows 8 points of 120 characters; confirm fields may carry only the current value
- the prompt says blocks are text, not tool calls, keeps computed totals out of number blocks,
  and asks for forge:confirm before changes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seen in the browser with a real model:
- bar charts render as labelled rows with values instead of unlabelled columns
- donut legend shows full labels, separators, shares and matching colour keys
- form fields use the Forge field look, selects get a chevron instead of the native control

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
forge-ui and others added 2 commits November 17, 2026 23:51
Light and dark, 20 prompts with clicks through questions, choices, confirm and plan:
- line charts: distinct series colours, keyed legend with latest values, middle axis label
- flow: a single line with diamond decisions; only branches and jumps get labelled chips
- tables: compact panel table with separators and no wrapping headers
- blocks with a glued closing fence, a name on the first line or on one line still render
- a call to a tool that does not exist shows as skipped, not as a red failure; the prompt
  says to use only the given tools
- Question keeps its card outline; answered labels drop the colon after a question mark
- confirm fields accept a plain value

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The answer blocks were hand-drawn and did not read like the rest of Forge. They now use the
components the pages use:
- Panel for every block frame; StatusBadge for statuses; Label for tags
- charts as in Forge's chart cases: headline value, BarChart with labels, filled
  SmoothLineChart, interactive DonutChart with value rows, ProgressBar rows
- DataTable for tables, ListGroup + ListItem for lists, DescriptionItem for records,
  HistoryGrouped for timelines, AgentFlowchart for flows, AgentCodeBlock for code, ImageGrid
- forms with TextField, TextArea, SelectOption and Toggle

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Unchanged text was keyed by index and changes by id, so a change with
id "2" collided with the third segment.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch was successfully deployed

1 active deployment
Preview — 762f7a71 Deployed Nov 18, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant