Inline comments on plans and .md files → attach to composer → send to agent

[Daniel-tabz]

This was generated by AI during triage.

Summary

Add the ability to leave inline comments on plans (and on .md files in the codebase) by selecting/hovering text, typing a comment in a small popup, and attaching those comments to the chat composer. The collected comments can then be sent to an agent as actionable feedback to revise the plan/spec.

This mirrors the "review comments" UX people know from Google Docs / GitHub PR review, but wired directly into Helmor's agent loop: comment on the plan → send → agent fixes the plan.

Motivation

Today, when an agent produces a plan (plan-review card) or when a .md spec/plan lives in the repo, the only way to give feedback is to write a free-form prose message describing what to change and where. This is imprecise and slow:

• You have to re-describe which part of the plan you're talking about ("in the third bullet under Migration…").
• Multiple small notes get squashed into one paragraph.
• The agent loses the anchor between your comment and the exact text it refers to.

Spec-driven development is increasingly common: teams keep plans and specs as .md files and iterate on them. A precise, inline, comment-to-agent flow makes reviewing and revising those documents dramatically faster.

Proposed UX

1. Select text in a rendered plan or markdown document. On selection (or on hovering the selection), a small inline affordance appears — an icon/button (e.g. a comment bubble) near the selection.
2. Click the icon → a small popup with a text input appears, anchored to the selection. The user types their comment ("make this step optional", "this is wrong, we use X not Y", etc.).
3. Submit the comment → it becomes an attached comment "chip" on the chat composer, showing the quoted/anchored text + the user's note. The original selected text stays highlighted to show it has a pending comment.
4. Repeat for multiple selections — several comments can be queued, each anchored to a different span.
5. Send to agent → all queued comments are bundled into the prompt sent to the agent, with enough anchoring context (the quoted span + comment + source: which plan or which file/path) for the agent to locate and revise the right text.

Where it should work

• Agent plan content rendered in the chat thread (the plan-review card / assistant markdown).
• .md / .markdown / .mdx files opened in the codebase editor — both the rendered preview mode and ideally source mode. This is the spec-driven use case.

Relevant code (pointers for implementation)

Markdown rendering is centralized, which makes a single shared selection/comment layer feasible:

• Markdown renderer (shared by chat + editor preview): src/components/streamdown-loader.tsx, src/components/streamdown-components.tsx
• Plan rendering: PlanReviewCard in src/features/panel/message-components/content-parts.tsx, type guards in src/features/panel/message-components/shared.ts, assistant markdown in src/features/panel/message-components/assistant-message.tsx. Plan parts already carry metadata (plan markdown, planFilePath, allowedPrompts).
• Markdown editor surface: src/features/editor/index.tsx; EditorViewMode ("source" | "preview") and isMarkdownPath() in src/lib/editor-session.ts. Preview mode reuses LazyStreamdown.
• Composer + attached chips/badges (the model for comment chips): src/features/composer/index.tsx, the context-bar chip strip src/features/composer/context-bar/index.tsx, and inline decorator/badge nodes (file-badge-node.tsx, custom-tag-badge-node.tsx) + src/components/inline-badge/index.tsx.
• Sending to agent: AgentSendRequest in src/lib/api.ts (already supports files, images, pastedTexts ranges) and startAgentMessageStream(). A new comments? payload (quoted span + note + source anchor) would slot in here, plus the matching Rust AgentSendRequest in src-tauri.
• Popover/selection UI primitives available: src/components/ui/popover.tsx, tooltip.tsx, hover-card.tsx, context-menu.tsx. No global text-selection toolbar exists yet — this feature introduces one.

Open questions / scope to refine

• Anchoring strategy: how robustly do we map a selected span back to source text the agent can find? For plans, quoting the exact text may be enough. For .md files, include the file path + quoted span (and maybe line numbers in source mode).
• Comment persistence: are queued comments ephemeral (cleared after send) or do they persist on the message/file until resolved? MVP can be ephemeral, attached only to the composer.
• Source vs preview in the editor: Monaco has its own selection model — supporting source mode may need a Monaco-specific affordance vs. the DOM-selection affordance used in rendered markdown.
• How comments are serialized into the prompt so the agent reliably edits the right plan/spec.

Suggested MVP

1. DOM text-selection → floating comment icon → popup input, scoped to rendered plan content in chat first.
2. Submitted comments become composer chips (reuse the context-bar/inline-badge pattern), each showing quoted text + note.
3. On send, serialize comments into the agent prompt with quoted-span anchoring.
4. Follow-up: extend the same affordance to .md preview mode in the editor, then source mode.

Labels

enhancement