Schema already in 001_init.sql; this is just the Go store layer.
IO markers are project-scoped wall-outlet terminators (a cable's
"this end plugs into a wall socket outside the diagram" endpoint).
Power-by-convention; no schema-level type enforcement.
- CreateIOMarker validates frame_id is in the same project (cross-project
ref → ErrInvalidInput), defaults label to "IO" when blank.
- GetIOMarker is project-scoped — wrong-project read returns ErrNotFound.
- UpdateIOMarker uses the FrameRef tri-state for frame_id (same as
DeviceUpdate) so callers can clear it explicitly.
- DeleteIOMarker is direct delete — ON DELETE SET NULL from the schema
drops the io_markers.frame_id ref cleanly when the frame is deleted
(verified by TestDeleteFrame_SetsIOMarkerFrameIDToNull).
Snapshot now populates IOMarkers from the store; field type tightened
from []any to []IOMarker.
7 new table-driven tests, all green with -race.
m's review of v4 locked 6 answers. Tight doc pass:
- Schematic-only bundling: dropped trunk-segment/frame-edge/cable-tray
language. v3 endpoint-pair rule is the only bundle rule.
- Setup templates folded in (not post-MVP): migration 004 with 3
built-ins (Living Room, Home Office, Server Rack) + 3 new device
types (Screen, Keyboard, Mouse) + apply-template API in slice 6.
- Unmet-requirement quick-fix: combo endpoint adds a missing port and
re-solves in one server roundtrip.
- Solver still button-only, catalog still SQL-seeded, promote still
explicit on cable inspector.
All 9 §9 questions resolved.
Tight pass on m's review of v4 (single commit per head's instruction).
Six locked answers integrated:
1. mCables is a schematic, not a physical-routing tool. Stripped
'trunk', 'frame-edge corridor', 'cable tray', 'path optimisation'
from §5b.1, §5b.2, §7, §8, §9. Bundling reduces to the v3 endpoint-
pair rule: ≥2 cables between the same A↔B endpoint pair → group as
one bundle. Anything path-shaped is "out of scope, period" (§8).
2. Solver button-only for v0 (no change). Live-solve parked at 9+.
3. Unmet-requirement quick-fix: red badge on the affected device in the
inspector with a single "+ Add <type> port to <device> and re-solve"
button per §5b.4. New endpoint
POST /api/projects/:pid/devices/:id/ports-and-resolve chains the
port insert + the solve re-run in one transaction.
4. Setup templates fold INTO v4.1. New §2.4 with the schema for
setup_templates + setup_template_devices + setup_template_requirements
(migration 004), 3 built-in templates seeded (Living Room, Home
Office, Server Rack). New API: GET /api/setup-templates,
POST /api/projects/:pid/apply-template. New UI flow: "or start from
a template" section in the New Project modal + an "Apply template"
action on empty projects. Built-in catalog grows to 14 types
(adds Screen, Keyboard, Mouse).
5. Catalog SQL seed in migration 002 (no change).
6. Promote-to-manual: explicit button on cable inspector (no change).
§8 slice 6 absorbs the templates work alongside the solver MVP.
§9 closes all six v4 questions; no open design questions remain.
Trailer changes to "DESIGN v4.1 READY FOR REVIEW".
CLAUDE.md mirrors: schematic-only framing, 14-type catalog, setup
templates as a first-class feature, quick-fix UX note.
Big rescope driven by m's product-vision clarification: mCables is a
cable-management framework with a solver as its core value prop, not a
manual draw-and-click editor. m declares devices + required connections
between them; the solver emits the cable plan + bundle recommendations,
optimising for maximum bundling.
Schema additions (migrations 002 + 003):
- device_types (catalog) — built-ins (project_id NULL) + project-custom
(project_id non-null). 11 built-in types seeded with default port
profiles (NAS, PC, Mac, TV, Soundbar, Switch, fritz, ChromeCast,
SteamLink, IOx-3/6/8, Notebook).
- device_type_ports (profile rows: cable_type × count × edge).
- devices.type_id (nullable). Picking a type seeds ports once;
instance-owned thereafter (no retroactive re-seed).
- connection_requirements (per-project, from/to device + preferred type
+ must_connect flag, with order-normalised pair_lo/pair_hi for
duplicate prevention).
- cables.auto (slice 5.5 migration) — distinguishes solver-owned cables
from user-drawn ones.
API additions:
- GET /api/device-types (built-ins only, read-only) and
GET /api/projects/:pid/device-types (built-ins + project-custom merged)
- POST/PATCH/DELETE under /api/projects/:pid/device-types (project-custom
only; built-ins are 403)
- /api/projects/:pid/connection-requirements full CRUD
- POST /api/projects/:pid/solve with ?preview=1 — pure-function solver
(greedy port allocation, endpoint-pair bundling for v0); returns
add[], remove[], bundles_added[], unsatisfied[], warnings[]
Solver algorithm (§5b):
- Read project devices + ports + connection_requirements + manual cables
- Assign each requirement a (port_a, port_b) using the preferred cable
type (or auto-pick if exactly one type matches both ends)
- Bundle by endpoint-pair (v3 rule, applied to auto cables only)
- Surface unsatisfied requirements per class (no compat type / ambiguous
type / no free port) — does NOT auto-add ports; UI quick-fix instead
- ?preview=1 returns the diff without writing; default applies in a tx
UI additions:
- Device-create modal: type dropdown (built-ins grouped by kind, then
project-custom, then "Custom (no type)" for the v3 freeform fallback)
- Left-sidebar Requirements section with + Requirement button
- Header Solve button (S keybinding) → preview modal → Apply
- Inspector for selected device: type, ports grid, unmet requirements
with red badges + quick-fix actions
- Inspector for selected auto cable: driving requirement, parent bundle,
Promote-to-manual button
Slice reshape (§8):
- Slices 1, 2 shipped. v4 inserts: 4 = catalog + type-aware device create,
4.5 = catalog management, 5 = requirements CRUD + UI, 6 = solver MVP +
Solve button. Old "manual port + manual cable draw" slides to slice 7
as a tweak path on solver output. Export becomes slice 8.
Six new open questions (§9) for m to gate before slice 4:
1. Path source (auto-route through frame edges / user cable-trays /
Steiner-tree)?
2. Live-solve vs. button-only?
3. UX when solver has no compatible port pair?
4. Setup templates in v4 or post-MVP?
5. Catalog as code seed or JSON file?
6. Auto-promote vs. explicit Promote-to-manual on solver cable edits?
CLAUDE.md updated to reflect the solver-core framing, hybrid catalog,
connection-requirements model, and auto/manual cable distinction.
Trailer changes to "DESIGN v4 READY FOR REVIEW".
startDrag set state.selection but never re-rendered. One render() call
after the assignment fixes it. Now selecting a device or frame
populates the inspector with name/dims/delete-button as designed.
startDrag set state.selection but didn't render until pointerup's onUp
ran — and onUp can throw on `e.currentTarget.classList.remove` if the
event reference is stale after pointer capture release, which leaves
the inspector stuck on 'Nothing selected.'
One-line fix: call render() right after state.selection assignment so
the inspector + halo update from pointerdown, independent of whether
onUp completes cleanly. The drag-completion render at the end of onUp
stays — when both fire it's idempotent (renders are pure functions of
state).
Primary fix: e.preventDefault() on the pointerdown for both armed-tool
branches in onCanvasPointerDown. Without it, the browser's default
mousedown action blurs the freshly-focused input in promptInline
(the SVG click target isn't focusable), and the blur handler calls
done(null) before m can type.
Secondary fix: clear activeNamer before fo.remove() in done(), to
prevent a re-entrant pageerror when Enter triggers blur synchronously.
Root cause traced by sherlock with Playwright (docs/sherlock-+dev-bug.md
on the sherlock branch). The previous routing fix at 94869f3 was
necessary but not sufficient: placeDeviceAt() now reaches promptInline()
correctly, but the synchronous input.focus() is undone ~6ms later by
the browser's compatibility-mousedown default — which blurs the active
element when the mousedown landed on a non-focusable target (SVG rect /
SVG root). The blur listener then ran done(null) and ripped the
<foreignObject> out before m could type a name.
Primary fix: e.preventDefault() at the top of both armed-tool branches
in onCanvasPointerDown. Suppresses the focus-shifting default so the
input keeps focus.
+Frame is also wrapped for symmetry. It wasn't strictly affected (its
namer runs from pointerup, not pointerdown) but preventDefault avoids
a subtle text-selection side effect during rubber-band drag.
Secondary fix in promptInline.done(): clear activeNamer *before*
fo.remove(). Enter-key triggers a synchronous blur listener which
re-enters done() — if remove() ran first, the re-entry hit a
"node no longer a child" pageerror. Reordering makes the re-entry a
no-op (activeNamer is already null).
Verified locally: served /main.js shows e.preventDefault() inside both
tool branches and the reordered done() body. go test -race ./... still
green.
- CSS: .canvas-wrap.tool-{frame,device} #canvas, #canvas * { cursor:
crosshair !important } so frame/device rects don't display grab while
a tool is armed
- Server: Cache-Control: no-cache on embedded static handler so browsers
revalidate via ETag instead of serving stale main.js after redeploy
Issue 1 — cursor lies about armed tool. .svg-draggable { cursor: grab }
on frame/device rects beat the .canvas-wrap.tool-device #canvas {
cursor: crosshair } rule because element-level wins over descendant.
m saw "grab" hovering a frame with +Dev armed and thought the tool was
broken even though clicks routed correctly after the previous fix. Add
a descendant rule with !important so tool-armed wraps any child cursor:
.canvas-wrap.tool-frame #canvas *,
.canvas-wrap.tool-device #canvas * { cursor: crosshair !important; }
Issue 2 — stale browser cache after each redeploy. http.FileServerFS
served embedded assets with no Cache-Control header, so browsers held
on to the previous main.js/style.css until hard-reload. New noCache
middleware on the static handler emits Cache-Control: no-cache. Note:
embedded FS files have zero ModTime, so http.FileServer suppresses
Last-Modified — every fetch is a fresh 200 rather than a 304. Fine at
~30KB of JS+CSS, and fixes the staleness problem completely.
Middleware is wrapped only around the static handler. /api/* responses
write their own headers and aren't touched.
Verified locally:
curl -I /main.js → Cache-Control: no-cache
curl -I /style.css → Cache-Control: no-cache + contains the new rule
curl -I /api/healthz → unaffected (no Cache-Control from us)
go test -race ./... still green.
Move the [data-frame-id]/[data-device-id] early-return below the
tool-armed branches in onCanvasPointerDown. With a tool armed,
the canvas-level handler always wins; without a tool, the original
behaviour (frame/device pointerdown handlers capture for drag/select)
is restored.
onCanvasPointerDown returned early whenever the click landed on a
[data-device-id] or [data-frame-id] element so the per-element drag
handlers wouldn't get hijacked. Problem: this early-return fired BEFORE
the tool check, so clicking +Dev inside an existing frame never reached
placeDeviceAt().
Reordered: tool-armed branches run first and short-circuit. Only when
no tool is armed does the "click started on a child element — leave it
alone" guard kick in. End behaviour:
- +Dev anywhere (incl. inside a frame) drops a device. frame_id
auto-resolves via the existing frameAt() point-in-rect.
- +Frm anywhere (incl. inside an existing frame) starts a rubber-band;
rare but not harmful.
- No tool armed: clicking a device/frame still goes to its own handler
(drag / select). Clicking empty canvas still clears selection.
Hand-tested via the served /main.js + the equivalent backend POST/PATCH
sequence: device-in-frame, device-outside, device-drag, frame-drag with
cascaded device patches — all work.
Renders the slice-2 backend on the empty canvas from slice 1.
Canvas:
- Frames render as dashed-stroke rects with top-left label, slightly
tinted fill. Devices render as solid-stroke rects with centred label
in device.color.
- Selection halo via .selected class (stroke-width bump).
- Empty-state hint disappears once any geometry exists.
Tools (left sidebar + keyboard):
- F / + Frame — rubber-band rect on the canvas. <80×60 cancels. On
release, inline foreignObject namer → POST /api/projects/:pid/frames.
- D / + Device — single click places a 100×35 device centred at the
click. Inline namer → POST devices. Drop-point determines initial
frame_id via point-in-rect against all frames (smallest bbox wins).
- Esc cancels active tool / inline namer / clears selection.
Drag (pointer events + svg getScreenCTM):
- Devices: drag updates x/y live via transform, persists via
PATCH .../devices/:id on pointerup. Also recomputes frame_id from
drop point and includes "frame_id": null|<id> if it changed.
- Frames: dragging a frame moves its contained devices visually too;
on pointerup, single PATCH for the frame + one PATCH per moved device.
Children-batch is computed at pointerdown and only sent on release —
no per-pointermove network traffic.
Inspector:
- Frame selection: name (debounced rename), x/y/w/h, device count,
Delete button (confirm prompt — devices keep existing, frame_id → NULL
via the schema's ON DELETE SET NULL).
- Device selection: name (debounced rename), colour picker
(change-event PATCH, no debounce), x/y/w/h, current frame, Delete.
- Background click clears selection.
devicePatch wire format uses tri-state frame_id: key absent = leave,
key:null = clear, key:<int> = move. Frontend uses `null` explicitly
when a device drops outside all frames.
All 8 endpoints (list, create, patch, delete) for both resources. Path
params parsed via Go 1.22 ServeMux PathValue.
devicePatch uses json.RawMessage for frame_id so the wire format
distinguishes:
- key absent → leave as-is
- "frame_id": null → clear (device leaves all frames)
- "frame_id": 42 → move to that frame
parseFrameRef translates that into the store's db.FrameRef tri-state.
Sentinel-error mapping unchanged (writeError covers ErrInvalidInput,
ErrConflict, ErrNotFound, etc.). Cross-project frame_id refs surface as
400.
Snapshot now populates frames + devices from the DB (slice 1 left them as
empty arrays).
Frame store:
- CreateFrame requires positive width/height; rejects empty name; UNIQUE
(project_id, name) collisions surface as ErrConflict via mapWriteErr.
- GetFrame is project-scoped — wrong-project read returns ErrNotFound.
- UpdateFrame applies a partial; project_id is not exposed (moving a
frame across projects would orphan its devices).
- DeleteFrame relies on the schema's ON DELETE SET NULL to drop
devices' frame_id refs cleanly; verified by test.
Device store:
- CreateDevice defaults color to #1e1e1e if blank; rejects empty name,
non-positive size; validates frame_id is in the same project (returns
ErrInvalidInput on cross-project ref).
- UpdateDevice uses a FrameRef tri-state for frame_id so callers can
distinguish "leave alone" from "clear to NULL" from "move to frame X".
- Cross-project frame_id on PATCH is rejected with ErrInvalidInput.
- ListDevices supports an optional frame_id filter.
13 new table-driven tests, all green with -race.
mAi got admin on m/mCables but Gitea container packages are
user-namespace-scoped — repo-collab perm is insufficient. Pushed
once using m's ~/.netrc token, deleted the mAi/mcables stub.
Compose now references mgit.msbls.de/m/mcables:latest.
m granted mAi admin on m/mCables, but Gitea's container registry is
user-namespace-scoped (not repo-collab-scoped) so the push had to go
through m's own credentials for this one administrative move:
docker login mgit.msbls.de -u m -p <m's token>
docker push mgit.msbls.de/m/mcables:latest
Image digest sha256:76624f17… is identical to the one previously living
at mgit.msbls.de/mai/mcables:latest — same build, just retagged.
Drops the workaround comment from the compose file. The mai/mcables
package will be deleted via API after the deploy verifies.
picasso shipped (commit 8a31f0a on mai/picasso/deploy-mdock):
- Dockerfile: multi-stage golang:1.23-alpine -> distroless/static
- docker-compose.yml at repo root (raw-docker pattern, not Dokploy)
- .dockerignore
- README deploy section
Live: http://mdock:7777 (image sha256:76624f17, 12.2MB).
Persistence verified across compose restart.
Note: mAi lacks write on m/ in Gitea, so image lives at
mgit.msbls.de/mAi/mcables:latest. m can retag once mAi gets write
on m/mCables (see docker-compose.yml comment).
Pulls the deploy infra forward from §10 so m can see slice 1 on his LAN.
- Dockerfile: multi-stage golang:1.25-alpine → distroless/static-debian12.
CGO_ENABLED=0 (modernc.org/sqlite is pure Go). USER 1000:1000 so the
bind-mount on mDock (owned by m:m) is writable without chowning the
host dir. -trimpath + -s -w; 12.2MB final image.
- docker-compose.yml: matches the mDock convention surveyed earlier
(container_name explicit, restart: unless-stopped, env_file in
/home/m/secrets/mcables/.env, bind-mount /home/m/stacks/mcables/data,
port 7777 exposed on LAN). Image temporarily under the mai/ namespace
on mgit.msbls.de because mAi doesn't have write access to m/* today —
documented in a comment so retagging is one line when permissions land.
- .dockerignore: keeps .git, .worktrees, .m, data/, docs/, *.md,
editor cruft out of the build context.
Manual deploy verified end-to-end:
- docker build → image sha256:76624f17 (12.2MB)
- mAi-authenticated push to mgit.msbls.de/mai/mcables:latest
- ssh mdock anonymous pull works (registry allows public reads on this
namespace)
- POST /api/projects {"name":"LOFT"} returns the row, GET /api/projects
shows it; docker compose restart preserves it on disk; second GET
still shows LOFT.
Gitea Actions auto-deploy left for a follow-up task per the head's
instruction — gets us the moving parts right first.
Adds table-driven store tests:
- projects: drawing_name auto-default, explicit-name accept, empty-name
reject, duplicate-name conflict, ordered list, GetProject not-found,
partial PATCH semantics, blank-drawing-name re-default on PATCH,
?confirm=<name> guardrail (wrong / empty / correct), snapshot returns
the 5 globally-seeded cable_types
- cable_types: 5 seeded with the legend colours, global UNIQUE(name),
rename + recolour, RESTRICT-blocked delete when a cable references the
type (with count surfaced via CountCablesUsingType), unused delete
succeeds, project deletion does NOT cascade into cable_types
go test -race ./... passes. Updates README.md with run instructions,
env vars, the slice-1 API surface, and the slice roadmap.
Tight pass on round-4 answers (single commit per head's request):
- cable_types is GLOBAL — drop project_id, UNIQUE(name). Migration 001
seeds the 5 defaults once; POST /api/projects no longer seeds them.
API moves to top-level /api/cable-types. Renaming/recolouring affects
every project. CASCADE from projects does not touch cable_types.
- devices: UNIQUE (project_id, name) added.
- projects: drawing_name defaults to "<name>.excalidraw" server-side
on POST when omitted; editable via PATCH.
- DELETE /api/projects/:pid requires ?confirm=<name>; server checks
name match, returns 400 if missing or mismatched.
- io_markers: no type_id (Power-by-convention, UI soft-warn). Confirmed
v0 stance.
- Bundles ignored on export — carries over from v2.
- §0 changelog rewritten as "what changed in v3 / what carried over".
- §2 schema rewritten; FK-shape paragraph updated to call out the one
global table.
- §3 endpoints: cable-types moved to top level; POST/DELETE projects
show new defaults + guardrail semantics.
- §4 export table notes cable_types pulled from global.
- §7 "edit cable type" flow gains the cross-project-effect banner +
ON DELETE RESTRICT inline-error UX.
- §8 slice 1 rewritten: no per-project seeding; legend reads global.
- §9 all six v2 questions marked resolved with the v3 answer per item.
- Trailer changes to "DESIGN v3 READY — coder shift gated".
- CLAUDE.md mirrors: global cable_types, device UNIQUE per project,
drawing_name default, delete guardrail.
Sync project instructions with design v2:
- Framework framing: top-level `projects` table, LOFT/OFFICE/… as
separate projects, frames as sub-zones inside a project.
- DB path moved from ~/.m/mcables.db to ./data/mcables.db (gitignored).
- Frontend stack locked: vanilla JS modules + SVG, no build step,
TypeScript types via JSDoc, Preact-via-CDN-ESM as fallback.
- Deploy: raw docker on mDock under /home/m/stacks/mcables/ — explicitly
NOT Dokploy. Port 7777, no reverse proxy, no auth (LAN-trusted).
- mExDraw access: raw HTTP API (mcp__mexdraw__* not configured for this
project), one-way export only.
- Seed drawing reframed as visual-grammar reference, NOT a runtime
import target. IO markers documented as wall-outlet terminators
(type=Power), not inter-frame bridges.
- Out-of-scope list updated: no auth, no inventory fields, no runtime
import. Worker-preference slices re-aligned with the new design.
Revision after m's answers (2026-05-15):
- mCables is a framework. Top-level `projects` table; LOFT and OFFICE
are separate projects, each backed by one drawing. project_id is
denormalised onto every row for cheap project-scoped queries; CASCADE
from projects wipes a project's whole subgraph.
- IO diamonds are wall-outlet terminators (type=Power), not inter-frame
bridges. paired_with_id removed.
- No runtime importer. The seed Cable-Management.excalidraw is the
visual-grammar reference for the exporter only. /api/sync/import is
dropped from MVP; only /api/sync/export remains (one-way, manual).
- No cable inventory fields. Strictly visual structure for v0.
- DB at ./data/mcables.db (project-local, gitignored).
- Deploy: raw docker on mDock under /home/m/stacks/mcables/ (NOT Dokploy).
Conventions verified live (mgreen, mgeo, msports-garmin patterns).
Port 7777, container_name mcables, image from Gitea registry, Gitea
Actions self-hosted runner builds + deploys on push to main.
- Bind 0.0.0.0:7777 on the LAN. No auth.
- UI gains a projects picker; all CRUD endpoints scoped under
/api/projects/:pid/.
- Slices re-planned: empty bootstrap → frame+device → port+cable →
IO+cable-type editing → export.
- Open questions trimmed; six new ones (drawing-name policy, device
uniqueness, non-Power IO, bundle export, cross-project cable types,
delete guardrail).
Ends with DESIGN v2 READY FOR REVIEW.
Inventor shift 1. Reads the live Cable-Management.excalidraw on
mxdrw.msbls.de, lands on:
- vanilla JS modules + SVG diagram (no build step) served by a single
Go binary with embed.FS
- SQLite schema with frames, devices, ports, cable_types, cables,
io_markers, bundles (cgo-free driver)
- HTTP API under /api/, /api/state as the editor's one-shot loader
- importer that respects port-on-edge geometry (ports are positional,
not containerId-bound) and resolves arrow endpoints to port/device/IO
- bundle detection MVP: same-endpoints-pair → suggestion
- sync model: DB authoritative, Excalidraw is a projection, manual
import/export buttons with element-ID stability and conflict flagging
- five vertical slices for the coder shift, smallest end-to-end first
- nine open questions for m before code starts
Ends with DESIGN READY FOR REVIEW.
