feat: add opt-in tldraw skill and local snapshot store

.agent/skills/_index.md

@@ -56,3 +56,13 @@ Constraints: prefer DESIGN.md tokens over invented values, do not modify
 DESIGN.md unless the user explicitly asks, preserve unknown sections when
 an edit IS authorised, validate with `npx @google/design.md lint DESIGN.md`
 when available.
+
+## tldraw
+Draw, diagram, sketch, or lay out ideas on a live tldraw canvas.
+Worthwhile drawings snapshot into this skill's local store
+(`skills/tldraw/store.py`) for recall across sessions.
+Triggers: "draw", "diagram", "sketch", "wireframe", "flowchart",
+"mind-map", "visualize", "whiteboard"
+Constraints: get_canvas before edits; max 200 shapes per create_shape call.
+Requires: tldraw MCP server wired in the harness's MCP config; user has
+http://localhost:3030 open. Opt-in via `.features.json` (`tldraw: true`).

.agent/skills/_manifest.jsonl

@@ -6,3 +6,4 @@
 {"name":"data-layer","version":"2026-04-26","triggers":["data layer","dashboard","show me the dashboard","what did my agents do","agent analytics","agent status","resource usage","usage report","cron monitoring","daily report","tokens","terminal dashboard","TUI"],"tools":["bash","git"],"preconditions":[".agent exists"],"constraints":["local-only by default","no screenshot delivery without explicit user approval","do not commit private .agent/data-layer exports"],"category":"operations"}
 {"name":"data-flywheel","version":"2026-04-25","triggers":["data flywheel","trace to train","training traces","context cards","eval cases","approved runs","vertical intelligence"],"tools":["bash","git"],"preconditions":[".agent exists"],"constraints":["local-only by default","human-approved runs only","redaction required before trainable","do not train models"],"category":"operations"}
 {"name":"design-md","version":"2026-04-26","triggers":["DESIGN.md","design.md","Google Stitch","Stitch","design tokens","design system","visual design"],"tools":["bash","memory_reflect"],"preconditions":["DESIGN.md exists at project root"],"constraints":["prefer DESIGN.md tokens over invented values","do not modify DESIGN.md unless the user explicitly asks","preserve unknown sections when an edit IS authorised","validate when tooling is available"],"category":"design"}
+{"name":"tldraw","version":"2026-04-21","triggers":["draw","diagram","sketch","wireframe","flowchart","mind-map","mind map","visualize","lay out","architecture diagram","whiteboard"],"tools":["mcp.tldraw.create_shape","mcp.tldraw.update_shape","mcp.tldraw.delete_shape","mcp.tldraw.get_canvas"],"preconditions":["tldraw MCP server reachable","user has http://localhost:3030 open"],"constraints":["call get_canvas before update_shape or delete_shape","at most 200 shapes per create_shape call","coordinates within 0..1600 x 0..900 unless told otherwise"],"category":"visualization","feature_flag":"tldraw"}

.agent/skills/tldraw/SKILL.md

@@ -0,0 +1,117 @@
+---
+name: tldraw
+version: 2026-04-21
+triggers: ["draw", "diagram", "sketch", "wireframe", "flowchart", "mind-map", "mind map", "visualize", "lay out", "architecture diagram", "whiteboard"]
+tools: [mcp.tldraw.create_shape, mcp.tldraw.update_shape, mcp.tldraw.delete_shape, mcp.tldraw.get_canvas]
+preconditions: ["tldraw MCP server reachable via the harness's MCP config", "user has http://localhost:3030 open"]
+constraints: ["call get_canvas before update_shape or delete_shape to discover real ids", "at most 200 shapes per create_shape call", "coordinates within 0..1600 x 0..900 unless the user asks otherwise"]
+category: visualization
+---
+
+# tldraw — draw on a live canvas
+
+The tldraw MCP server exposes a live canvas at `http://localhost:3030`.
+You draw into it; the user watches it fill in. Worthwhile drawings can
+be snapshotted to disk and recalled in future sessions via this skill's
+local store (`store.py`).
+
+## When this skill loads
+
+Any time the user asks to visualize, diagram, sketch, lay out, or map
+something graphically. If they are clearly asking for prose, do not draw.
+
+## Before drawing, once per session
+
+Tell the user:
+
+> Open `http://localhost:3030` to see the canvas.
+
+If any tool returns `No tldraw browser connected`, repeat the hint and
+stop until they confirm.
+
+## Opt-in MCP setup
+
+This beta does not install MCP wiring during default adapter setup. After the
+user enables `tldraw` in `.agent/memory/.features.json`, they must add the
+server to their harness MCP config. Use this local block as the source of truth:
+
+```json
+{
+  "mcpServers": {
+    "tldraw": {
+      "command": "npx",
+      "args": ["-y", "@tldraw-mcp/server"]
+    }
+  }
+}
+```
+
+For Claude Code and Antigravity this usually lives in `.mcp.json`; for Cursor
+it usually lives in `.cursor/mcp.json`. If a config already exists, merge the
+`tldraw` server entry rather than overwriting the file.
+
+## Tools
+
+| tool | purpose |
+|---|---|
+| `create_shape({ shapes: [...] })` | add new shapes |
+| `update_shape({ updates: [{ id, props }] })` | change shapes by id |
+| `delete_shape({ ids: [...] })` | remove shapes |
+| `get_canvas()` | return all current shapes |
+
+Always `get_canvas` first when the user says "add to", "next to",
+"modify", or refers to something already drawn — you need the real ids.
+
+## Coordinate system
+
+- Origin `(0, 0)` top-left. `+x` right, `+y` down.
+- Stay inside `0 <= x <= 1600`, `0 <= y <= 900` unless told otherwise.
+- Default sizes: boxes ~160x80, icons ~60x60.
+
+## Shape vocabulary
+
+| type | required | optional |
+|---|---|---|
+| `geo` | `x, y, w, h` | `geo` (rectangle/ellipse/triangle/diamond/star/...), `color`, `fill`, `text` |
+| `text` | `x, y, text` | `color`, `size` (s/m/l/xl) |
+| `arrow` | `x, y, end:{x,y}` | `color`, `text` (label) |
+| `line` | `x, y, end:{x,y}` | `color` |
+| `draw` | `x, y, points:[{x,y}]` | `color` (freehand, at least 2 points) |
+| `note` | `x, y, text` | `color` (sticky note) |
+
+Colors: `black, grey, light-violet, violet, blue, light-blue, yellow, orange, green, light-green, red`.
+Fills: `none, semi, solid, pattern`.
+
+## Persisting drawings
+
+When a drawing is worth keeping across sessions (architecture decisions,
+recurring diagrams, reference material), snapshot it:
+
+```bash
+# fetch current shapes via MCP get_canvas and pipe the JSON in
+python3 .agent/skills/tldraw/store.py snapshot \
+    --label "auth-flow-v1" --tags architecture,auth \
+    --note "login + refresh token flow agreed 2026-04-21"
+```
+
+The store reads canvas JSON on stdin, writes a snapshot file under
+`snapshots/`, appends metadata to `snapshots.jsonl`, and re-renders
+`INDEX.md` — all within a single file lock. Later sessions recover a
+drawing with `list` / `load`. Use `archive` to retire a snapshot; the
+JSONL is append-only semantic, so archived records move to
+`snapshots/archive/` rather than being deleted.
+
+## Pitfalls
+
+- `text` shapes need non-empty `text`.
+- `arrow.end` is an absolute point, not a delta.
+- Split large scenes into multiple `create_shape` calls (<= 200 each).
+- Always `get_canvas` before an edit; never assume ids.
+
+## Self-rewrite hook
+
+After any failure, or every 5 uses:
+1. Read the last N tldraw-tagged entries from `memory/episodic/AGENT_LEARNINGS.jsonl`.
+2. If a constraint was violated (shape cap, id-before-edit rule), escalate
+   a candidate lesson to `semantic/LESSONS.md` via `tools/learn.py`.
+3. Commit: `skill-update: tldraw, <one-line reason>`.