Copilot discovery now drives `copilot --stdio` over JSON-RPC
(connect → models.list) — the same wire protocol (version 3) the
official github-copilot-sdk uses. Speaking the protocol directly
rather than linking the SDK crate keeps the workspace on Rust 1.85
(the SDK crate requires 1.94). Documented model names stay as the
fallback when the CLI is installed but the query doesn't answer.
Codex discovery now drives `codex app-server` over JSON-RPC stdio
(initialize → initialized → model/list), the official App Server
API — more accurate and stable than reading models_cache.json,
which may be stale or absent. The cache stays as a fallback, then
a minimal placeholder when codex is installed but neither source
answers. A reader thread + deadline keeps a hung server from
blocking discovery.
Replace the chat panel's agent-cycling chip with a real model
picker: the chip shows the selected model's provider brand logo +
display name, and clicking it opens an upward dropdown listing
chat.available_models grouped by provider with a selected check
(mirrors the TS ai-chat-model-selector). Selecting a model also
re-syncs chat_selected_agent to its provider so the chat transport
targets the matching CLI. Outside clicks dismiss the picker.
The chat panel never had a model list — shell-core is transport-free
and cannot query a CLI. Add desktop-side cross-platform discovery:
Codex's models_cache.json (real, file-based), `opencode models`
(real, subprocess), and the documented CLI model names for Claude /
Gemini / Copilot which expose no listing interface — each gated on
the CLI actually being installed (PATH walk with platform exe
extensions). Discovery runs on a worker thread and drains into
chat.available_models. shell-core keeps only the ModelEntry type.
Move the per-frame paint pass to frame.rs to keep main.rs under cap.
Node.effects already round-trips through persistence and canonical
import, but the canvas never rendered it. Add a fill_drop_shadow
RenderBackend primitive (default impl approximates with a translucent
fill; native + web will override with skia MaskFilter blur) and paint
each DropShadow behind Frame / Rect / Ellipse fills, offset + blur
scaled by viewport zoom.
Canonical .op shadow colours use functional rgba()/rgb() notation, which
the hex-only parser rejected — valid shadows were imported as opaque
black, then dropped entirely after the prior fix. Add parse_css_color
covering both hex and functional forms. Move the effect-conversion
helpers into persistence_effects.rs so pen_doc_adapter.rs stays under
the 800-line cap.
`shadows_from_canonical` fell back to `[0,0,0,1.0]` (opaque black)
when `parse_hex` rejected a shadow colour — so an `rgba()`,
named, or short-hex colour imported as a wrong solid-black
shadow (codex stop-gate). Switch the fallback to `filter_map`
`None`: an unparseable colour drops that shadow entirely, which
is honest (no shadow) rather than misleading (wrong shadow).
Matches `first_solid_color`'s existing `parse_hex(...)?` discipline.
`pen_doc_adapter::node_to_payload` dropped every `PenEffect` —
a `.op` authored with shadows lost them on import (codex stop-gate;
the inverse of the save-path effects round-trip just added).
`shadows_from_canonical` reads each variant's effects (Frame /
Group / Rectangle carry them on `container`, leaf shapes directly;
IconFont / Ref have none) and maps `PenEffect::Shadow` →
`ShadowPayload`. Blur / background-blur are skipped — the shell's
`Effect` model is drop-shadow-only today.
First slice of the Effects gap — the property panel's 效果 section
was a header-only stub with no data behind it.
- `Effect` / `DropShadow` types in `document.rs`; `Node.effects:
Vec<Effect>` (offset_x/y + blur + color, doc-px). Both `Node`
builders + `deep_clone_with_new_ids` carry it.
- `persistence_effects.rs` — `ShadowPayload` + `effects_to_payload`
/ `effects_from_payload`; `NodePayload.effects` is `#[serde(default)]`
so legacy `.op` files still load. Carved into its own module so
`persistence.rs` stays under the 800-line cap.
- Round-trip tested through the real serde JSON path (2 tests).
858 workspace tests pass. Still pending (own slices): canvas
drop-shadow paint (needs a blur primitive on `RenderBackend`),
the property-panel 效果 editing rows, and a `set_node_shadow`
MCP tool. Shadows currently round-trip through save/load but do
not yet render.
- crates/CLAUDE.md MCP catalog was stale at 21 tools; replace the
outdated table with the current 77-tool categorized summary and
point at mcp_serve.rs TOOL_SCHEMAS as the source of truth.
- Refresh the mcp/ file-layout list (node_attr_tools,
selected_ops_tools, extra_read_tools, json_serializer, …).
- New "AI chat (real provider integration)" section documenting
begin_send / ChatProvider / chat_session / the model chip and
the Codex/OpenCode honest-error behavior.
- chat_session: gate the test-only `finished()` accessor behind
`#[cfg(test)]` to drop a dead-code warning.
When a send for an unwired agent (Codex / OpenCode) landed while a
previous turn was still streaming, `launch_if_pending` wrote the
error bubble but left `current` pointing at the old session — so
the next `pump` streamed the prior agent's deltas straight into
the fresh error message. Clear `current` before writing the error.
`provider_for_agent` silently substituted Claude Code when the
user picked Codex or OpenCode — so the transcript looked like the
chosen CLI answered when it hadn't. Codex stop-gate.
It now returns `Option`: `None` for Codex / OpenCode (HTTP-server
transport not yet bridged). `launch_if_pending` writes an explicit
`error: … not wired yet` into the assistant bubble and starts no
session, so the user knows their pick didn't run.
The chat panel's bottom-toolbar chip was a hardcoded "Default"
label. It now shows the selected CLI agent and cycles through the
connected ones on click.
- `UiState.chat_selected_agent` — index into `AgentProvider::ALL`
(Claude Code / Codex CLI / OpenCode / GitHub Copilot / Gemini).
- `Document::cycle_chat_agent` — advances to the next *connected*
agent (`agent_settings.connected`); walks all 5 when none are
connected so the user can pre-pick before connecting.
- `AIChatHit::CycleModel` — the chip's left 150 px of the input
toolbar; `apply_click` routes it to `cycle_chat_agent`.
- The chip label now renders `AgentProvider::name()` instead of
the static "Default".
- `chat_session::provider_for_agent` routes the selected agent to
its `ChatProvider`: Claude Code → SDK adapter, Copilot / Gemini
→ subprocess transport. Codex / OpenCode (HTTP-server CLIs whose
bridge isn't wired yet) fall back to Claude Code.
856 workspace tests pass (+2: cycle walks connected / wraps,
cycle with nothing connected walks all five).
The chat send path was hardwired to a stub (`ChatState::send` →
"(stub) Got it"). The real providers existed but main.rs never
called them.
- `ChatState::begin_send()` — native-host send: pushes the user
message + an empty assistant bubble, raises `pending_send`. The
web shell keeps the offline `send()` stub.
- `chat_session.rs` — `ChatSession` runs the turn on a background
thread (`ChatProvider::send` is a blocking iterator; draining it
on the UI thread would freeze the window). `poll()` is
non-blocking; `launch_if_pending` / `pump` are the desktop glue.
- main.rs — Enter key + Send-button click drain `pending_send`
into a `ChatSession` against `ClaudeCodeProvider`; the winit
loop pumps deltas into the transcript, waking ~30 fps mid-turn.
A missing / unauthenticated `claude` CLI now surfaces as an
`error: claude query: …` message instead of a fake stub reply.
854 workspace tests pass (+4: chat_session stream/error,
begin_send push/empty).
Append a fresh empty page + switch the active page to it. No
args. Mirrors TS `addPage()`. Applier returns false on
id-space exhaustion at `max_node_id() + 1`.
Wire shape:
args: {}
result: { "wrote": "true" }
command: `McpCommand::AddPage`
Companion to set_active_page (commit 1af880c2): LLMs can now
create + navigate pages without touching the UI. Future patches
can add rename_page / delete_page / duplicate_page to complete
the page CRUD (Document already has the mutators).
Desktop --mcp registry + tools/list schema + exact-count test
updated for 29 tools.
28th tool in the catalog. set_active_page lets LLMs switch which
page is the active target for subsequent inserts / batch_design /
design_* commands. Required arg: `index` (0-based u32). The
applier rejects out-of-range indices.
New `Document::set_active_page(idx) -> bool` mutator in
page_mutators.rs (peer of add_page / duplicate_page / remove_page).
Clears selection on switch so stale-selection ops don't carry
across pages.
Companion refactor: write_tools.rs was at 900 lines (over the
800-line cap) after the components CRUD shipped. Split out
components/page-related tools to `mcp/component_tools.rs`
(InstantiateComponent / CreateComponent / DeleteComponent /
RenameComponent / SetActivePage). write_tools.rs back to 686.
Desktop --mcp registry + tools/list schema + exact-count test
updated for 28 tools.
27th tool in the catalog. Completes the components MCP read
surface: list_components returns the catalog (name + id pairs);
get_component drills into one entry and returns:
- name (display name)
- kind ("frame" / "group" / etc — the component's root NodeKind)
- leaf_count (sum of leaves in the prototype subtree)
LLM workflow: list_components to discover the catalog →
get_component to size each candidate → instantiate_component to
drop a clone.
Wire shape:
args: { "component_id": "<positive u64>" }
result: { name, kind, leaf_count } or ToolFailed when the id
doesn't resolve.
The snapshot factory computes leaf_count at registration time
so the tool stays `&self`. NodeKind enum mapping shared with
the rest of the MCP surface ("frame" / "group" / "rect" / etc).
Desktop --mcp registry + tools/list schema + exact-count test
updated for 27 tools.
Adds the missing update verb to the components registry CRUD.
26 tools total now. The full lifecycle on the MCP side:
- list_components (read)
- create_component (create)
- rename_component (update — name only today; the prototype
subtree stays whatever node was promoted at create time)
- instantiate_component (clone onto page)
- delete_component (delete)
Wire shape:
args: { "component_id": "<positive u64>", "name": "<string>" }
result: { "wrote": "true" }
command: `McpCommand::RenameComponent { component_id, name }`
New `ComponentLibrary::rename(id, new_name) -> bool` mutator
rejects empty / whitespace-only names so list_components never
returns a stub a user can't recognize. Apply layer also
defensively rejects unknown ids.
Desktop --mcp registry + tools/list schema + exact-count test
updated for 26 tools.
Completes the components registry MCP lifecycle:
- list_components (read)
- create_component (create from existing Frame/Group node)
- instantiate_component (drop a clone onto the active page)
- delete_component (remove from registry)
25 tools total. Live instances already on the page are NOT
affected by delete — they're independent clones; the registry
only holds the prototype that future instantiate_component
calls would have used.
Wire shape:
args: { "component_id": "<positive u64>" }
result: { "wrote": "true" }
command: `McpCommand::DeleteComponent { component_id }`
New `ComponentLibrary::remove(id) -> bool` mutator returns
whether a component was removed; apply layer propagates that
so the LLM can tell whether the id resolved.
Desktop --mcp registry + tools/list schema + exact-count test
all updated for 25 tools.
Closes the components MCP lifecycle: list_components (read),
create_component (promote a node to a component),
instantiate_component (drop a clone onto the active page). 24th
tool in the catalog.
Wire shape:
args: { "node_id": "<positive u64>", "name": "<string>" }
result: { "wrote": "true" }
command: `McpCommand::CreateComponent { node_id, name }`
New `Document::create_component_from_node` mutator (peer of
the existing `create_component_from_selected`) takes the target
id explicitly instead of reading `self.selected`, because MCP
sessions have no UI selection state. Both share the same
container-kind guard (Frame / Group only) and replace-on-
duplicate-id semantics via `ComponentLibrary::insert`.
Apply path rejects unknown ids + non-container kinds.
Variables table rejects InstantiateComponent + CreateComponent
in its Pages-level guard (defense in depth).
Desktop --mcp registry + tools/list schema + exact-count test
all updated for 24 tools.
LLM workflow:
1. list_components to see what's registered
2. create_component on a Frame / Group to register it
3. instantiate_component to drop a clone of any registered component
Closes the matching write surface to list_components. 23rd tool
in the catalog. LLM clients can now (a) list registered
components via list_components and (b) drop a clone of any of
them onto the active page via instantiate_component.
Wire shape:
args: { "component_id": "<positive u64 id from list_components>" }
result: { "wrote": "true" }
command: `McpCommand::InstantiateComponent { component_id }`
The apply path routes through the existing
`Document::instantiate_component` mutator, which deep-clones
the component's root subtree with fresh ids past
`max_node_id()`, appends it to the active page, pushes a
history snapshot, and selects the new instance root. Returns
false when the component id is unknown.
Desktop --mcp registry + tools/list schema + exact-count test
all updated; handshake test now expects 23 tools.
Components panel UI is still pending (a future patch adds a
right-rail Components section); the MCP surface is the
LLM-driven workflow alternative until then.
Surfaces the existing Document.components ComponentLibrary via
MCP so LLM clients can discover what reusable Frame / Group
subtrees the user has saved. 22nd tool in the catalog (was 21).
Wire shape:
count — total component count
components — `;`-records of `name|id`, with the standard
2-level escape on each side (matches list_variables /
list_pages wire convention).
Closes the read-only MCP surface for components. Instance
insertion via MCP (the matching write tool) requires a JSON
Node descriptor and is the natural follow-up; the data model +
`Document::instantiate_component` already exist.
Desktop --mcp registry + tools/list schema updated; test
renamed to all_twenty_two_tools with exact-count guard intact.
Codex re-review on 3891ea07 BLOCKed 8 more bypass paths.
Fix them in three groups:
1. input_active() (widget_host/input.rs:13) now includes
variable_row_focus, so apply_copy / apply_cut / apply_paste /
apply_select_all / apply_reorder (which already gate on
input_active) automatically block while a row is focused —
no need for per-method commits.
2. State-mutating shortcuts that DON'T gate on input_active() get
explicit `commit_variable_row_focus_if_any()` at the top:
- apply_undo / apply_redo / apply_group / apply_ungroup
(shortcuts.rs:55–104)
- apply_toggle_code_panel (shortcuts.rs:122)
- apply_toggle_agent_settings (shortcuts.rs:132)
- apply_set_tool (shortcuts.rs:144)
- apply_boolean_op (widget_host.rs:338)
- apply_delete (keyboard.rs:149) — special: pops a char from
the variable draft instead of falling through to selection
delete, matching the apply_backspace behavior
3. Desktop entry points in main.rs that bypass apply_press AND
the shortcut methods (Cmd+S / Cmd+O / Cmd+Shift+S / Cmd+Shift+P):
call new `commit_variable_row_focus_if_any_pub()` proxy on
the host before persistence::handle_*.
The proxy is exposed because the internal helper is
`pub(in crate::widget_host)`.
377 shell-core tests still pass.
Codex re-review on 86167224 BLOCKed two more bypass paths:
- Right-click → `apply_right_press` (press.rs:71) opens layer
context menus and returns without going through apply_press
or apply_text. Add `commit_variable_row_focus_if_any()` at
the top.
- Cmd+J → `apply_toggle_chat` (shortcuts.rs:106) flips chat.focused
directly. Add the same commit before the toggle so the chat
takeover doesn't leave keystrokes routed into the variable
draft.
The same row re-click + escape cascade paths stay clean.
Codex re-review on 939c9742 BLOCKed: piggy-backing the
variable-row commit onto commit_property_focus_if_any only
covered call sites that ran when property_focus was already
set. A standalone variable_row_focus (no property focus) +
click on canvas / toolbar / topbar / chat / layer rail
bypassed the commit entirely, keeping keystrokes routed into
the variable draft.
Move the commit to the top of apply_press, BEFORE every
hit-test branch. Property-focus commit-on-blur still gated
on the property panel's x boundary (different layout — the
property panel sits below the variables panel in the same
right rail; clicks on the property panel area shouldn't
commit a variable edit and vice-versa, but the variables
editor lives within the right rail so any non-right-rail
click is "outside" for it, and any property-panel click is
"outside" for it too via the down-stream hit-test cascade).
Chat clicks enter through apply_press first (they call
apply_click as a sub-step, lines 438 + 527), so the top-of-
apply_press commit also closes the chat-tab focus leak that
codex flagged as B2.
Codex stop-time review on 0ccf10b9 flagged that the new
variable_row_focus state stayed set when the user clicked
outside the variables panel (canvas, toolbar, etc.), causing
subsequent keystrokes to leak into the variable draft instead
of routing to the canvas / chat / other surfaces.
Root cause: the existing 9+ outside-click handlers called
`commit_property_focus_if_any` but not the new
`commit_variable_row_focus_if_any`. Rather than patching all 9
call sites, piggyback on `commit_property_focus_if_any` — it
now commits the variable-row focus first, then proceeds to its
own logic. Every existing outside-click commit path now covers
both editor surfaces in one call.
The explicit double-call at the Number/String row click site
(property_dispatch.rs:308–309) is now redundant but not wrong,
so leave as documentation that the order matters when seeding
a fresh focus.
Closes MINOR #8 non-color variable editors UI gap. Clicking a
Number or String variable row in the VariablesPanel now opens
an inline edit field on that row.
Surface:
- `Document.ui.variable_row_focus: Option<VariableRowFocus>` with
Number(idx) / String(idx) variants — new enum in document.rs
- Reuses `property_input_draft` as the buffer (single caret-blink
anchor + input handler tier serves both PropertyFocus and
VariableRowFocus)
- `VariablesPanel` carries `editing_row` + `editing_draft`
snapshots from doc.ui in `for_document`
- Paint: when the row idx matches `editing_row`, replace the
value-preview label with the draft text + a thin foreground
underline as the active-focus affordance (no caret blink yet —
the property panel's caret machinery isn't shared; the
underline is sufficient discoverability)
- Input dispatch (`widget_host/keyboard.rs`):
* apply_text: new branch BEFORE property_focus. Number accepts
digit / leading `-` / one `.`; String accepts any non-control
char. Both push to property_input_draft.
* apply_backspace: new branch BEFORE property_focus, pops draft
* apply_send: routes to new `commit_variable_row_focus_if_any`
* apply_escape: clears focus + draft (before property_focus
tier so the escape cascade still feels right)
- Commit helper `commit_variable_row_focus_if_any` in
`property_dispatch.rs`: parses the draft per kind (f64 parse
for Number — rejects NaN/Inf; raw string for String), calls
`VariableTable::set_scalar`, pushes a history snapshot on success
so undo restores the prior value.
Color rows still open the color picker (unchanged); Boolean rows
still toggle on click (unchanged); Number / String now have an
inline editor instead of a no-op.
377 shell-core tests still pass.
Coarse perf regression test for the MINOR "1000+ node doc render
pan/zoom not benchmarked" gap. Three integration tests in
`crates/openpencil-shell-core/tests/large_document_perf.rs`:
- max_node_id × 100 on 1000-leaf doc < 50 ms (debug build)
- node_at_doc_point × 100 across the grid < 25 ms (debug build)
- apply_mcp_command(BatchInsert) of 1000 descriptors < 100 ms
Bounds are deliberately generous on debug builds — the job is to
catch a future 10× regression, not measure wall-clock precision.
Real benchmarks via criterion still pending (criterion is not a
workspace dep yet); these guards are the bounded alternative.
All three pass on the current build.
Bounded slice of the non-color variable editor UI gap. Clicking
a Boolean-kind variable row in the VariablesPanel now toggles
its value through `VariableTable::set_scalar`, honoring the
active-theme routing (subset match / no default clobber / no
other-axis shadow) so a flip under theme=dark writes only the
dark entry.
Color rows still open the color picker; Number / String rows
are explicit no-ops (the MCP path covers writes today via
set_variable_number / _string — an inline text input on the
row needs a draft buffer + caret + input dispatch that's not
worth the scope tonight).
Each toggle pushes a history snapshot so undo restores the prior
value. 377 shell-core tests still pass.
Closes MINOR #9 theme axis dropdown gap. Chip-click semantics
change from "cycle to next value" to "toggle dropdown for direct
pick", matching TS's behavior where users see all axis values in
a menu and can pick any one.
Surface change:
- `Document.ui.axis_dropdown_open: Option<String>` (axis name)
- `VariablesPanelHit::AxisDropdownItem { axis, value }` new hit kind
- `VariablesPanel` carries `dropdown_open` snapshot from doc.ui
- Paint: dropdown overlay anchored under the open chip; popover
card with rounded border, per-value rows, active-value
highlight via theme.muted
- Hit-test: dropdown overlay takes top-most priority so a click
on a value row wins over the chip / row beneath
- Host wiring (`property_dispatch.rs`):
* AxisChip click toggles axis_dropdown_open (same chip closes,
different chip switches)
* AxisDropdownItem click calls VariableTable::set_active_theme +
closes the dropdown + pushes a history snapshot
Test: `axis_dropdown_hit_routes_to_named_value` asserts that with
themes=[mode: light/dark/system] and the dropdown open, hits on
rows 0 and 2 return the correct (axis, value) pair.
377 shell-core tests pass (was 376).
Closes the auto-update banner MINOR gap. TS's electron-updater
shows a status banner with a dot + label; the Rust shell now
matches the visual surface while staying honest about the
backend:
- Green status dot + "Up to date" label paint side-by-side with
the "Auto-update" header in the System tab card.
- Description line: "Running the latest installed build. No
update channel is wired in this build — the next release
ships through your package manager / .dmg / .exe."
- Existing "not yet wired" sub-line kept as the smaller
follow-up so a user wondering "why no Check button?" sees
the answer in-place.
- Three new i18n keys (settings.system.upToDate / .upToDateDescription
/ .checkForUpdates) with EN + ZH translations.
- Card height grew 64 → 88 px to fit the new lines.
No togglable switch — the previous module comment was explicit
that flipping a boolean would lie to the user. No real
network check happens. When the updater backend lands, this
can grow back into a check-now button + real status field on
Document.ui.
Codex stop-gate review on 7cee053d flagged two concerns:
1. design_skeleton/content/refine tools/list schema descriptions
could mislead LLM clients into inferring differentiated apply
semantics from the phase label. Per the internal comment in
batch_design.rs, all three currently dispatch to BatchInsert
with phase metadata only — but the schema description is what
clients see, not the comment. Add explicit "Apply behavior is
identical to batch_design today (phase is metadata only)" to
each of the three schema descriptions.
2. The new tools_list test only verified each tool name via
`contains`. A future tool addition could silently pass this
test without updating the expected count. Add an
`assert_eq!(TOOL_SCHEMAS.len(), 21)` exact-count guard with
a clear error message ("add the new tool to this test").
Closes the layered design workflow MAJOR gap line. Three new
write tools (13th / 14th / 15th in the write catalog, 21 tools
total) mirror TS's pen-mcp phased generation workflow:
- design_skeleton — phase 1, structural scaffolding
- design_content — phase 2, fill content
- design_refine — phase 3, polish details
Each tool shares batch_design's wire shape (nodes_json carrying a
JSON array of leaf descriptors). Today every phase emits the same
McpCommand::BatchInsert — the phasing is metadata only, stamped
into the response payload as `phase=skeleton|content|refine` so
LLM clients can correlate the call back to its workflow phase.
A future patch may grow per-phase apply semantics (design_refine
patching existing nodes via UpdateNode batches, etc.) once the
richer commands exist.
mcp_serve.rs wires the three new snapshots into rebuild_registry
+ adds their tools/list schemas. The handshake test is renamed
`tools_list_response_includes_all_twenty_one_tools`.
Smoke test: three back-to-back tools/call invocations (skeleton
→ content → refine) return 3 valid responses each carrying the
correct phase label, and all 3 nodes persist to disk.
Codex stop-time review on d422f5ef flagged that the Color+Str
arm of VariableTable::set_scalar wrote any string into a Color
variable without hex validation. The tool snapshot for
set_variable_string filters Color names out, but a misrouted
command (or a forged McpCommand) would still mutate the Color
variable to garbage at apply time. Defense in depth required.
Fix: forbid Color from set_scalar entirely. Color variables MUST
go through set_color_hex which validates the hex up front.
McpCommand::SetVariableScalar against a Color variable now
returns false instead of corrupting the stored value.
Regression test `apply_mcp_command_rejects_color_through_scalar_path`
stamps a Color variable with `#abcdef`, fires a forged
SetVariableScalar with `String("garbage")`, and asserts the
Color value is untouched.
376 shell-core tests pass (was 375).
Closes the MCP side of the "non-color variable editors" MINOR
gap. Three new write tools (10th / 11th / 12th in the write
catalog) mirror set_variable_color for the Number / String /
Boolean variable kinds.
All three route through one shared apply-time mutator:
\`VariableTable::set_scalar(name, VariableScalar)\` (new in
\`document/variables.rs\`). This generalizes set_color_hex's
theme-routing logic (subset match / no default clobber / no
other-axis shadow) over any VariableScalar variant. Kind/scalar
mismatch is rejected at apply time as defense in depth even when
the tool's snapshot would have admitted it.
One unified McpCommand variant \`SetVariableScalar { name, scalar:
VariableScalarPayload }\` plus a wire-friendly payload enum
(Number(f64) / String(String) / Boolean(bool)) so the wire layer
doesn't need to reach into shell-core's document API.
The three tools share an arg shape (\`name\` + \`value\`):
- number: value parsed as finite f64, rejects NaN / Inf / non-numeric
- string: value taken verbatim
- boolean: value must be "true" or "false" (case-sensitive)
Code lives in a new \`mcp/scalar_vars.rs\` sibling so
\`write_tools.rs\` stays under the 800-line cap; tests in
\`mcp/scalar_vars_tests.rs\` (5 tests).
\`openpencil-desktop --mcp\` registry + tools/list schema updated
so external CLIs see all 18 tools (was 15). Handshake test
renamed to match.
375 shell-core tests pass (was 370 — 5 new scalar_vars tests).
Closes MAJOR item from the 2026-05-14 gap list. The ninth MCP
write tool. Mirrors TS \`batch_design\` for the leaf subset of
NodeKind (frame / group / rect / ellipse / polygon / line /
text / path).
Wire shape: scalar arg \`nodes_json\` carrying a JSON array of
\`{kind, name, x, y, width, height, fill_hex?}\` descriptors. The
shell-core parser rejects structured args at the top level (so
an LLM can't sneak nested objects past scalar contracts), but a
JSON array wrapped in a quoted string round-trips cleanly — the
tool parses the inner JSON itself.
Implementation lives in \`crates/openpencil-shell-core/src/mcp/
batch_design.rs\` (sibling carve-out keeps write_tools.rs under
the 800-line cap). New \`McpCommand::BatchInsert { items:
Vec<BatchInsertItem> }\` variant + apply branch in mcp_apply.rs
that:
- validates EVERY descriptor before any mutation (kind /
geometry / fill_hex)
- allocates fresh ids past max_node_id() up front for the whole
batch (checked_add per id; bail on overflow)
- only after all validation + allocation passes, attaches each
new node to the active page
A single bad entry rejects the whole batch — the LLM never sees
a partial design tree (defense in depth: same check at tool layer
+ applier).
Tool also handles wire-level JSON escaping: the shell-core
parser stores string values with escape sequences intact (\\\"
arrives as backslash-quote), so the tool runs an unescape pass
before its hand-rolled JSON-array parser. Handles \\\" / \\\\ /
\\n / \\t / \\r / \\/.
Tests in \`mcp/batch_design_tests.rs\` (9 tests):
- requires nodes_json arg
- rejects empty array (tool) + empty items (applier)
- rejects unknown kind, negative geometry, missing required keys,
malformed JSON (6 separate cases)
- parses minimal 2-node array + applies it with fresh ids
- atomicity on bad descriptor at the apply layer
\`openpencil-desktop --mcp\` registry + tools/list schema updated
so external CLIs see batch_design as the 15th tool.
Smoke test: \`tools/call batch_design\` with 3 leaf descriptors
(2 rect + 1 text) → \`{count:"3",wrote:"true"}\` response + all
3 nodes persisted to disk.
370 shell-core tests pass (was 361 — 9 new batch_design tests).
Codex stop-time review on 23b0bfba flagged that real MCP clients
(Claude Code, Codex, etc.) won't dispatch tools/call cold — they
open with \`initialize\` and \`tools/list\` to discover the
server. The previous --mcp implementation responded only to
tools/call, so real clients hung at handshake.
Add in-binary handlers for the MCP control-plane methods:
- \`initialize\` → protocolVersion "2024-11-05" + capabilities
(tools.listChanged=false) + serverInfo
- \`notifications/initialized\` (and bare \`initialized\`) →
absorbed silently (spec: notifications have no response)
- \`tools/list\` → response with all 14 tools + their JSON
inputSchemas (name / description / required args / enums for
kind + drop_children)
- \`ping\` → empty result
- everything else falls through to shell-core's
run_stdio_with_applier so the parser hardening + apply
discipline still apply
Top-level method/id sniffing uses the same key-walker pattern as
shell-core's arguments_field, so a nested key called "method" or
a string value of "initialize" can't shadow the real top-level
method. Four sniff/response unit tests cover the discipline.
End-to-end roundtrip verified with real stdin/stdout:
initialize → notifications/initialized → tools/list → tools/call
insert_node → tools/call list_pages → all five frames return
correct JSON-RPC + the inserted rect persists to disk.
Closes BLOCKER #1 from the 2026-05-14 gap report: the Rust
shell's MCP ToolRegistry now has a host. \`openpencil-desktop
--mcp <path>\` skips the winit event loop and runs a JSON-RPC
stdio server backed by that .op file. External CLIs (Claude
Code / Codex / Gemini / Copilot) can spawn the binary in this
mode the same way they spawn TS pen-mcp today.
Implementation (\`crates/openpencil-desktop/src/mcp_serve.rs\`,
~95 lines):
- Loads the .op file via the existing canonical persistence path
(made \`load_from_path\` pub).
- Re-builds the ToolRegistry against the live document between
every dispatched call so read-tool snapshots reflect the
latest state (write commands mutate the doc; reused registry
would have stale snapshots).
- Dispatches one JSON-RPC line at a time through
shell-core's run_stdio_with_applier so the parser-level
hardening (structured-arg rejection / top-level walker / no-
hang error path) protects this binary too.
- Apply closure mutates the doc + saves to the same path on
every successful write command. Save failures surface to
stderr but don't crash the loop.
Smoke test passed: list_pages → insert_node → list_pages
roundtrip through real stdin/stdout returns three JSON-RPC
responses, and the inserted rect persists to disk (verified
by grepping FromMCP out of the saved file).
361 shell-core tests still pass; workspace build clean.
Adds a new "MCP server (shell-core)" section to
crates/CLAUDE.md covering:
- The 14-tool catalog (6 read + 8 write) with arg shapes and
emitted McpCommand variants per tool.
- Pre-validate-then-mutate discipline at the apply layer.
- `replace_node`'s `drop_children` destructive-swap guard.
- Wire-format stop-gates (structured-arg rejection, top-level
`arguments` walker, no-hang on parse failure, read-only path
refusing write tools).
- mcp/ file layout including the sibling test files for
copy_node and replace_node.
- Pending: host wiring, batch_design / design_skeleton tools,
HttpServer transport.
Codex stop-time review on 4ff2b87f flagged that the stricter
parser created a new failure mode: when parse_tool_call returns
None (structured args, unparseable JSON, etc.), run_stdio just
\`continue\`s — JSON-RPC clients waiting on a response correlated
by id never get one and hang.
Fix:
- Expose \`parser::extract_request_id\` so the dispatcher can
recover the id when the full parse fails. Helper does only
what parse_tool_call's first three lines used to do; parse_tool_call
now calls it internally so there's a single source of truth.
- In \`run_stdio_with_applier\`, on \`parse_tool_call → None\`,
attempt to extract the id; if found, write a typed error
response (\`InvalidArgument\` + "malformed tool call:
unparseable or structured arguments"). If even the id is
missing, drop the line silently — there's nothing to
correlate against.
- Drop the now-dead \`extract_object_body\` helper (\`arguments_field\`
replaced it in the previous commit).
Tests:
- \`run_stdio_emits_error_when_parse_fails_so_clients_dont_hang\`
— structured \`drop_children\` arrives with valid id 42;
response echoes id 42 + carries an \`error\` field naming
"malformed tool call".
- \`run_stdio_skips_lines_without_an_id\` — id-less lines stay
silent (existing behavior, explicitly covered now).
361 shell-core tests pass (was 359).
Codex BLOCKed ee0671a7 on the substring-find inside
arguments_field: `body.find("\"arguments\"")` matched any
nested key or string value containing the literal `"arguments"`,
not just the top-level params.arguments field. Two valid-but-
unusual JSON shapes bypassed the rejection contract:
1. `meta:{"arguments":{}}` alongside `"arguments":"oops"` — the
substring scan hit meta.arguments as a Body and parsed it,
skipping the real top-level string-typed arguments which
should have rejected.
2. `name:"arguments"` with no real arguments field — false
positive Malformed instead of Missing.
Replace with a proper top-level key walker that:
- skips whitespace + commas between top-level pairs;
- reads each quoted key + its value, classifying the value
(string / object / array / number-bool-null);
- only treats `arguments` as a real hit when seen at depth 0;
- categorizes the matched value: object → Body, anything else
→ Malformed; missing entirely → Missing.
Three new regression tests in `mcp_tests.rs::
parse_tool_call_arguments_lookup_is_top_level_only` cover the
nested-shadow, string-value-collision, and deep-nested-key
cases.
359 shell-core tests pass.
Codex BLOCKed 0a434e6b on three items:
1. MCP arguments non-object downgrade — \`"arguments":"oops"\`
used to flow through extract_object_body returning None and
parse_tool_call downgrading to empty args. Replace the call
site with a tri-state \`arguments_field\` helper:
- Missing → empty args (legit MCP can omit \`arguments\`).
- Body(s) → parse the object body.
- Malformed → reject the parse.
2. Test coverage gap — the structured-rejection tests only
exercised the legacy/direct line-protocol shape. Add two new
tests under mcp_tests:
- parse_tool_call_rejects_structured_values_in_mcp_tools_call_shape:
nested object + array inside MCP \`arguments\` both reject.
- parse_tool_call_rejects_non_object_arguments_field: string /
number / array \`arguments\` reject; missing \`arguments\`
still parses as empty args.
3. Stale doc comment on parse_flat_object_body — said nested
values are "skipped", now correctly describes the wire-layer
rejection contract.
358 shell-core tests pass (was 356 — two new MCP tools/call
shape tests).
Codex re-review on c6e6a308 flagged the sentinel approach's
collision risk: a variable literally named "{...}" — or any
string-accepting arg with that value — was indistinguishable
from wire-malformed input. Move the rejection up to the wire
layer so no tool sees a structured value at all.
- `parse_flat_object_body` now returns `None` the moment it
sees `{` or `[` for any value (was: insert a sentinel).
- `parse_tool_call` previously masked None with
`.unwrap_or_default()`, which would have silently swallowed
the rejection back into empty args. Replace
`extract_params_object` with a tri-state `ParamsResult`
(Missing / Body / Malformed) so we can distinguish "no params
key" (legit empty args) from "params present but malformed"
(full call rejection). MCP `tools/call` path treats a missing
`arguments` key the same way (empty args), and propagates a
`None` from the inner parse as a full call rejection.
Tests:
- `parse_tool_call_rejects_structured_arg_values` (in
`mcp_tests.rs`) asserts both object and array values cause
`parse_tool_call` to return None, while scalar-only still
parses.
- `parser_refuses_structured_drop_children_at_wire_layer` (in
`mcp/replace_node_tests.rs`) end-to-end: real MCP JSON with
structured `drop_children` and structured `name` never reach
the tool layer.
- `replace_node_rejects_malformed_drop_children` drops the
obsolete sentinel cases — they can no longer reach the tool.
356 shell-core tests pass.
