Skip to content

docs: Phase 21 spec — point-and-tell feedback#120

Merged
vicmaster merged 2 commits into
masterfrom
phase-21-spec
Jul 2, 2026
Merged

docs: Phase 21 spec — point-and-tell feedback#120
vicmaster merged 2 commits into
masterfrom
phase-21-spec

Conversation

@vicmaster

Copy link
Copy Markdown
Owner

What

Planning artifact only (no code): spec for Phase 21 — point-and-tell feedback, plus the VISION.md phase entry.

The feature: the human clicks any element in the viewer, types a note, and it lands as structured feedback ({ nodeId, comment } on canvas.metadata.feedback[]) that the agent reads via a new get_feedback tool and closes via resolve_feedback. The inverse of canvas_evaluate — human critique becomes machine-readable, ending the "the third card, no the OTHER one" prose loop.

Why it's cheap for what it buys

Verified against current src/ — every hard piece already exists:

  • renderer stamps data-node-id on every element (renderer.ts:233); detail page iframes are same-origin and Phase 19's highlightNode already reaches into them
  • repo write-back path exists (archiveRepoCanvas pattern in aggregate.ts) — spec extracts a generic updateRepoCanvas
  • ensureFresh mtime reload delivers viewer-written comments to the running MCP server with zero new plumbing

Review focus — two ★ forks

  • C1 storage: on-canvas metadata.feedback[] (recommended) vs separate feedback.json
  • C2 tool surface: two tools get_feedback/resolve_feedback (recommended) vs one action-enum tool

Everything else has a recommended default a review comment can flip. Slices: A (model/core/tools) → B (viewer UI) → C (loop integration).

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