bhoot found that download_cdfer 416-loops when a complete cdfer.sqlite3 already
exists in the cache (e.g. a prior run downloaded it but died before
conversion/cleanup): resuming a complete file sends Range: bytes=<size>- which
the server answers with HTTP 416, and raise_for_status treated that as a
retryable error, spinning until max_retries then failing.
Fix: HEAD once for Content-Length up front and short-circuit when the local file
already matches the full size; and in the loop, treat a 416 as "complete" when
the size matches (else drop the stale partial and restart fresh). Remove the now
-unused _head_last_modified helper (HEAD is done inline). Add a no-network test.
Verified live: happy path still downloads+converts 616k parts (~50s); a
pre-existing complete file now short-circuits in ~0.4s instead of looping.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The background warm-cache change made list_symbols() guard cache writes with
self._cache_lock. _manager_for_fixture() builds the manager via __new__ (to skip
disk I/O to system libs), bypassing __init__ where the lock is created, so the 8
SPICE-parsing tests raised AttributeError. Set the lock in the fixture. The
string-aware parser itself is correct — all 8 pass once the lock exists.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
cacb27b omitted the closing parenthesis anchor in _extract_lib_property_positions, leaving lib_props empty and falling back to generic resistor positions.
- New angle parameter for component rotation (degrees, KiCad CCW-in-screen)
- New mirrorY parameter for horizontal symbol flipping (transistor orientation)
- Property positions (Reference, Value, Footprint) automatically rotated
- Property text effects preserved from library definitions
- Updates to add_schematic_component tool with new parameters
- Implements proper S-expression formatting for (mirror y) attribute
- Text angles and offsets derived from library properties for correct placement
People want to be able to pull the whole catalog, and prior downloads stalled or
repeated the same parts. Two fixes:
- Full catalog correctness: yaqwsx's cache.sqlite3 (the full ~10GB set) stores
category/manufacturer as IDs with no v_components view, so the convert left
those fields blank. Build an equivalent v_components join for yaqwsx-style
sources so the full catalog converts with category/subcategory/manufacturer
populated. Clarify in the tool schema that source="yaqwsx" = full catalog
(needs 7z) vs cdfer = in-stock subset.
- Resumable downloads: CDFER stream download now uses a read timeout and resumes
from the partial file via HTTP Range on interruption (up to 5 retries); the
yaqwsx curl calls use -C - / --retry. Addresses the stall/partial-download
complaints.
Adds a yaqwsx-schema conversion test.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Per review feedback: CDFER's upstream scraper pipeline has stalled for weeks
(broken cart API + a bug in their scraper), so a "fresh download" can still be
old data. Compute the catalog age from the source Last-Modified header, expose
catalog_age_days, and emit a stale=True + warning when older than 14 days that
points users to source='yaqwsx' (fresh, needs 7z) or source='official'. Surface
the warning in the MCP tool output.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The download_jlcpcb_database tool paged the community JLCSearch API with an
offset parameter, but that endpoint is a search front-end that ignores offset
and returns the same first 100 parts on every page, so a full catalog download
was impossible.
Add commands/jlcpcb_downloader.py with a layered strategy that reuses prebuilt
catalogs the whole ecosystem already trusts:
- CDFER single-file SQLite (primary; no 7z/zip, reliable on Windows)
- yaqwsx/jlcparts split 7z (fallback; only if a 7z CLI is present)
- official JLCPCB API (optional; cursor pagination, if credentials set)
Conversion reads CDFER's v_components view (or sniffs the largest table for
yaqwsx), C-prefixes integer lcsc, derives library_type from basic/preferred,
maps mfr->mfr_part, and normalizes price JSON to the manager's [{qty,price}]
shape. Rewire _handle_download_jlcpcb_database to use it (closing/reopening the
manager connection so the on-disk db can be rewritten on Windows). Remove the
broken offset loop from jlcsearch.py (client kept for interactive lookups).
Reduce download_jlcpcb.py to a thin CLI wrapper and update the TS tool schema.
Verified end-to-end against live CDFER: 616k parts downloaded + converted in
~40s, FTS search and price-break parsing correct. New unit tests cover the
conversion and source fall-through; no network in tests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three-phase MCP server startup: wait for Python READY handshake, send
_warmup command (pcbnew.BOARD() triggers wxApp init on macOS), connect
to MCP transport only after warm-up completes.
Also pre-populate symbol library cache during SymbolLibraryManager
init so the first search_symbols call doesn't parse 241 .kicad_sym
files from disk (30-120s). Both warm-up and cache population happen
before tools are registered with the MCP client.
Fixes#195
The dehydration recovery test (test_auto_save_recovers_when_save_leaves_
board_dehydrated) builds a KiCADInterface via __new__, skipping __init__.
After #173 was rebased onto the #151/#172 auto-save guard, _auto_save_board
reads self._board_disk_signature (the content-hash conflict guard), which
__init__ normally sets to None. The fixture didn't, so the test tripped an
AttributeError before reaching the recovery logic it exercises.
Initialise _board_disk_signature = None in the _make_iface helper, matching
the pattern already used in test_auto_save_guard.py. Test-only change.
Co-authored-by: mixelpixx <11727006+mixelpixx@users.noreply.github.com>
KiCAD nightly builds occasionally return a SwigPyObject from
pcbnew.LoadBoard, and SaveBoard can leave self.board with no method
dispatch as a side-effect of certain sequences (delete_trace + auto-save
is the one users have hit in the wild). Before this fix, open_project
would keep reporting "Opened project: foo.kicad_pcb" while every
subsequent board operation failed with AttributeError on
GetDesignSettings / GetBoardEdgesBoundingBox / GetCurrentViaSize, with
no path to recovery short of restarting the MCP server.
Add two helpers in KiCADInterface:
* _is_board_healthy(board=None) probes for stable BOARD methods
(GetDesignSettings, GetBoardEdgesBoundingBox, GetFileName) — these
are missing on a dehydrated SwigPyObject, so hasattr() catches the
state without segfaulting.
* _safe_load_board(path) wraps pcbnew.LoadBoard, checks health, and
on dehydration reloads the pcbnew module via importlib.reload and
retries once. Returns None when recovery is impossible so callers
surface real failure rather than fake success.
Wire the helpers in:
* handle_command's open_project / create_project path validates the
loaded board and either recovers (with a warnings[] entry) or
returns success=False with an explicit "restart the MCP server"
errorDetails — never claims success when the board is unusable.
* _auto_save_board now detects dehydration introduced by SaveBoard
itself and reloads from disk so the next command sees a usable
proxy. This is the post-delete_trace failure mode users hit.
* _handle_place_component, _handle_sync_schematic_to_board,
_handle_import_svg_logo and _handle_refill_zones all go through
_safe_load_board instead of bare LoadBoard, surfacing real errors
consistently.
Also fix two adjacent issues observed in the same incident:
* _handle_check_kicad_ui used to call manager.is_running() and
manager.get_process_info() separately, with different detection
methods. They could disagree, producing the confusing
running=True, processes=[] state users hit after manually
quitting KiCAD. processes is now the single source of truth and
running is derived from len().
* run_drc accepts a timeoutSec param (default 600s, clamped to
[10, 1800]) so callers with smaller MCP transport budgets can
bound the kicad-cli subprocess. Same timeout is applied to the
optional report-generation subprocess. Error message names the
actual timeout that fired.
Tests: tests/test_swig_dehydration.py adds 17 unit tests covering
detection, recovery, the open_project surfacing path, the auto-save
post-recovery path, the check_kicad_ui consistency, and the run_drc
timeout clamping. Full suite: same 12 pre-existing failures both
before and after this change, +17 new tests passing.
Note: the SWIG dehydration is fundamentally a pcbnew memory bug
exposed by repeated LoadBoard calls in a single Python process;
this PR is a defensive recovery layer, not a fix to the underlying
binding. Complementary to PR #151 (auto-save-guard), which expands
the LoadBoard call rate by refusing saves on external file change
and forcing re-open_project cycles.
* feat(units): add mil unit support across all position/coordinate commands
KiCad natively supports mils, so the MCP server should too. Added "mil"
as a valid unit option in tool schemas and updated all unit-to-nanometer
scale conversions across component, routing, outline, view, and IPC
handler code paths. 1 mil = 25400 nm (0.0254 mm).
Also fixes a pre-existing mypy overload error in pin_locator.py (str cast
on dict.get key) that was blocking pre-commit on any Python file change.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(units): add mil to TypeScript tool schemas
The Python-side mil support was added but the actual input validation
happens in the TypeScript/Zod schemas. Updated all z.enum(["mm", "inch"])
to include "mil" across board, component, routing, design-rules, and
export tool definitions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(tools): replace CP-1252 mojibake with correct Unicode in board.ts
Replace U+00C3 U+00D7 (×) with U+00D7 (×) in add_logo size output string.
Character was mangled when file was saved as CP-1252 instead of UTF-8.
* fix: restore em-dash and fix pre-commit mypy in component/routing
component.py: replace CP-1252 mojibake (â€") with correct Unicode
em-dash (—) in the 'Add to board first' comment. Addresses
maintainer review on PR #162.
routing.py: annotate ex/ey as float at first assignment site in
_point_to_segment_distance_nm so mypy pre-commit hook passes
cleanly on this branch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Drop GND stitching vias across the board with collision checking
against every non-GND segment, via, and pad on every copper layer.
PTH vias penetrate the full stackup, so an F.Cu-only check (the most
common shortcut) silently creates shorts on inner / B.Cu copper —
this implementation explicitly walks all layers.
grid Regular grid across the board interior. Default
spacing 5mm.
around_refs Densify around specified footprints (e.g. MCUs,
switching regulators, RF parts). Configurable
density via densifyRadius.
in_zones Restrict placements to candidates inside the filled
polygons of GND copper zones, so each new via lands
on copper that's already a GND equipotential.
Recommended on boards where the GND zone is fragmented:
these vias actually stitch real polygons rather than
floating on silkscreen.
All three strategies use the same collision check + intra-call
clump-prevention, so passing `["grid", "around_refs", "in_zones"]`
is a safe kitchen-sink configuration.
- Auto-detect GND net (tries GND / GROUND / VSS / /GND in order)
OR explicit `gndNet` parameter.
- Per-via geometry control: viaSize, viaDrill, clearance.
- edgeMargin: keep-out distance from board edge.
- maxVias: cap on total placements (useful for incremental work).
- dryRun: return placements without modifying the board — for
previewing before committing.
- Validates viaDrill < viaSize, rejects unknown strategy names,
surfaces clear errors when GND net can't be resolved or the
board outline is missing.
Approach ported from morningfire-pcb-automation
(https://github.com/NiNjA-CodE/morningfire-pcb-automation,
scripts/ground/add_gnd_vias.py). The original parses the PCB text
with regex and writes vias by string concatenation; this port reads
obstacles via the pcbnew API (handles rotated footprints, integrates
with the live in-memory board so two sequential calls see each
other's placements, picks up net codes from the loaded board) and
adds the in_zones strategy, the maxVias cap, and dry-run mode.
Credit is in the docstring, the TypeScript wrapper comment, the MCP
tool description (visible to clients), and the CHANGELOG entry.
tests/test_add_gnd_stitching_vias.py — 18 cases, all passing.
Uses mocked pcbnew objects so the suite runs under both the conftest
stub and a real pcbnew install.
- grid strategy fills empty board with correct count
- collision blocks via near a signal track (with extent assertion)
- GND-net obstacles are correctly ignored
- around_refs densifies near footprints with bounded extent
- in_zones rejects candidates outside HitTestFilledArea
- dryRun does NOT call board.Add
- actual run calls board.Add per placement
- maxVias caps total placements
- intra-call clump prevention (asserts pairwise distance)
- viaDrill >= viaSize is rejected
- unknown strategy name is rejected
- missing GND net returns clear error payload
- no board loaded returns clear error
- named GND net (e.g. VSS) is honoured even when GND also exists
- direct unit tests for _point_to_segment_distance_nm helper
Real-board smoke test on TuneForge_TF001 (4-layer, 44 footprints):
- GND net auto-detected
- grid spacing 4mm: 141 placements, 129 blocked by collision
- grid + in_zones: 140 placed, 15 rejected by zone membership,
115 blocked by collision
python/commands/routing.py (+impl, ~370 LOC)
python/kicad_interface.py (+handler registration)
python/schemas/tool_schemas.py (+MCP schema)
src/tools/routing.ts (+TypeScript surface, builds clean)
tests/test_add_gnd_stitching_vias.py (+18 tests)
CHANGELOG.md (+Unreleased -> New MCP Tools)
Components now include boundingBox (min/max X/Y, width, height) in
get_component_list and get_component_properties responses. The SWIG
get_component_properties also includes courtyard dimensions when the
footprint defines a courtyard layer.
SWIG backend uses GetBoundingBox() and GetCourtyard(). IPC backend
tries get_item_bounding_box(), then pad-extent fallback, then falls
back to SWIG backend data when available. This ensures bounding box
data is present regardless of which backend handles the query.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Import-LIB-KiCad-Plugin documentation registers third-party libraries
using `${KICAD_3RD_PARTY}` (without a KiCad-version prefix). KiCad accepts
both unprefixed and version-prefixed forms in lib-tables, and the analogous
`KICAD_SYMBOL_DIR` was already handled in `library_symbol.py`'s env-var
dictionary.
`KICAD_3RD_PARTY` was missing from both `_resolve_uri` env-var dictionaries
(in `python/commands/library.py` and `python/commands/library_symbol.py`),
so lib-table rows authored as `${KICAD_3RD_PARTY}/Foo.kicad_sym` or
`${KICAD_3RD_PARTY}/Foo.pretty` failed to substitute, were treated as
non-existent paths, and disappeared from `list_symbol_libraries` /
`list_libraries` results — even though KiCad's GUI showed them correctly.
Changes:
- Add `KICAD_3RD_PARTY` to both `_resolve_uri` env-var dictionaries.
- Add `KICAD_3RD_PARTY` fallback in
`SymbolLibraryManager._find_3rd_party_dir`.
- Refactor `LibraryManager._find_kicad_3rdparty_dir` to check all four
env-var forms (KICAD10/9/8_3RD_PARTY + KICAD_3RD_PARTY) consistently.
- Add regression tests in `tests/test_kicad_3rd_party_env_resolution.py`.
Reproduces with:
- Set `KICAD_3RD_PARTY` env var in the MCP server's environment.
- Register `(lib (name "Foo") (type "KiCad") (uri "\${KICAD_3RD_PARTY}/Foo.kicad_sym") ...)` in the global sym-lib-table.
- Place a real `Foo.kicad_sym` at the resolved path.
- Before: `list_symbol_libraries` does not return `Foo`.
- After: `Foo` is listed.
The existing single-shot autoroute leaves 1-7 nets unrouted on dense
boards in my testing. Best-of-N drives that to 0 most of the time by
running Freerouting a few times with varied --max-passes and keeping
the SES with the best routing score.
New optional parameters (all backward-compatible):
attempts: int, default 1 (unchanged behaviour). When > 1, run
Freerouting N times and pick the highest-scoring SES.
targetNets: list of critical net names. An attempt that routes all
of them earns a 50,000-point scoring bonus.
passSchedule: list of --max-passes values to cycle through across
attempts. Default: [50, 60, 65, 70, 75, 80, 85, 90, 55,
95] (wraps if attempts > len). Ignored when attempts=1
(legacy maxPasses still used).
Scoring contract (pinned by tests):
score = nets_routed * 1000 + segments
if targetNets and all routed: score += 50_000
- +1 net always beats any segment-count delta (1000 pt step).
- Segments break ties at equal net count.
- Target bonus dominates net-count gains from unrelated nets.
## Implementation notes
- When attempts > 1, each attempt runs with `-mt 1` (single-thread
optimisation). Freerouting 2.x's multi-threaded optimiser is
documented to introduce clearance violations, so forcing
single-thread during scoring keeps the comparison apples-to-apples.
- One failed attempt does not abort the whole best-of-N run. The
failure is recorded in the response under attempts[] with ok=False,
and the remaining attempts compete for best. If every attempt fails
the response surfaces a clear error.
- The winning SES is preserved as <stem>_best.ses next to the
canonical <stem>.ses so the caller can inspect it after the run.
- Response shape:
attempts == 1: unchanged (no attempts/best_attempt fields)
attempts > 1: adds attempts[], best_attempt, best_score,
best_ses_path
## Attribution
Scoring approach and default pass schedule ported from
morningfire-pcb-automation
(https://github.com/NiNjA-CodE/morningfire-pcb-automation,
scripts/routing/freeroute_runner.py). Credited in the function
docstring, the TypeScript wrapper comment, the tool description (visible
to MCP clients), and the CHANGELOG entry.
The MCP version adds: cleaner per-attempt result reporting, automatic
single-thread optimisation, graceful degradation on partial failure,
and explicit validation that surfaces clean error payloads for invalid
attempts values.
## Tests
tests/test_autoroute_score.py 8 cases, scoring contract
tests/test_autoroute_best_of_n.py 6 cases, orchestration logic
All 14 passing. Tests are pure-Python: subprocess is mocked so the
suite runs in any environment (no Java / Freerouting / KiCad required).
- Single-attempt response shape unchanged
- Best-of-three picks the highest-scoring SES
- One nonzero exit attempt doesn't abort the run
- passSchedule wraps when attempts exceeds len
- targetNets bonus wins over higher raw net count
- attempts=0 rejected with clean error before DSN export
- +1 net (1000 pts) dominates any segment delta
- Segments tiebreak at equal net count
- Quoted net names in SES are normalised vs unquoted targets
TypeScript builds clean.
Detects courtyard overlaps between footprints and flags courtyards that
extend past the board outline. Returns overlap pairs with intersection
extents (mm), per-component boundary violations, and a placement summary.
The killer feature for AI-driven workflows is the `positions` parameter,
which accepts hypothetical placements `{ref: [x, y]}` or
`{ref: [x, y, rotation_degrees]}`. The tool evaluates the proposed
placement WITHOUT writing to the board file — so an AI agent can validate
a move_component / place_component before committing it, instead of the
current loop of write -> run DRC -> parse violations -> revert.
## Implementation
- Uses the real courtyard polygons from pcbnew (`fp.GetCourtyard(F_CrtYd)`
or B_CrtYd) for accurate AABBs even on custom and rotated footprints.
- Falls back to `fp.GetBoundingBox()` when no F/B.Courtyard polygon is
present.
- For virtual rotation, rotates the four AABB corners and re-axis-aligns.
Conservative: the rotated-AABB is always >= the rotated-polygon, so
overlap reports are never false-negatives (may be marginally
over-cautious on diagonal rectangles, which is the right error bias
for a placement validator).
- Optional `margin` parameter expands every courtyard by N mm — useful
for enforcing a manufacturing keepout wider than the symbol's
declared courtyard.
## Attribution
The approach is ported from morningfire-pcb-automation
(https://github.com/NiNjA-CodE/morningfire-pcb-automation), specifically
`scripts/placement/check_overlaps.py`. The upstream uses a static
per-footprint-type courtyard lookup table; this implementation reads
the real polygons from pcbnew so it works on any footprint without
maintaining a table. Attribution is in the function docstring, the
TypeScript wrapper, the tool's description (visible to MCP clients),
and the CHANGELOG entry.
## Tests
12 pytest cases in tests/test_check_courtyard_overlaps.py, all passing:
- No overlaps when spaced; overlap detected on intersect
- Margin pushes borderline pairs into overlap
- `refs` filter restricts the check
- Boundary violations are flagged; `include_boundary=false` suppresses
- Virtual position does not mutate the footprint (asserts
`SetPosition` is never called)
- Virtual rotation swaps a tall-narrow courtyard's x/y extents
- No-board-loaded returns clean error payload
- Bad position spec (wrong arity) returns clean error payload
- GetCourtyard() OutlineCount=0 -> fallback to GetBoundingBox()
- `board_outline` override replaces the Edge.Cuts bbox
Tests use mocked pcbnew objects so they run under both the conftest stub
and a real pcbnew install. Real-board smoke test on a 44-footprint
production board succeeds: 1 known overlap detected (SW1<->SW2), 0
boundary violations, virtual placement test reports 6 expected overlaps.
## Files touched
- python/commands/component.py (impl + helpers)
- python/kicad_interface.py (tool registration)
- python/schemas/tool_schemas.py (MCP schema entry)
- src/tools/component.ts (TypeScript surface, builds clean)
- tests/test_check_courtyard_overlaps.py (12 cases)
- CHANGELOG.md (Unreleased -> New MCP Tools)
The existing y_flip tests cover Device:R (symmetric two-pin) and Device:C
rotated 90° (still electrically symmetric). The original bug — a double
Y-flip in get_pin_location — was invisible on symmetric passives because
pin 1 and pin 2 are interchangeable; it only showed up on asymmetric
multi-pin ICs like RF_Module:ESP32-WROOM-32, where labels meant for pin 3
(EN) silently landed on pin 35 (TXD0).
This adds a third test using an inline 6-pin asymmetric symbol with pins
at both positive and negative library Y on both sides. It asserts every
pin lands at the formula-predicted (symbol_x + lib_px, symbol_y - lib_py)
position, with an explicit cross-check that lib +Y pins resolve *above*
the placement centre in schematic Y-down space.
No system-library dependency — the symbol is constructed inline so the
test runs anywhere pytest does.
Fixes#135
* Fix: IPC rotate_component now uses absolute angle as documented
The IPC rotate handler was adding the angle to the current rotation
(relative), but the schema documents it as absolute. This caused
unexpected behavior where setting angle=0 had no effect on a component
already at 180°. Now correctly sets the rotation to the exact angle
specified, matching the SWIG backend behavior.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs(changelog): add unreleased entry for rotate_component absolute-angle fix
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Closes#126.
A placed schematic symbol carries its reference designator in two places:
(symbol
(property "Reference" "R5" …) ← what eeschema renders
(instances
(project "MyProject"
(path "/sheet-uuid/symbol-uuid"
(reference "R5") ← what netlist + PCB sync read
(unit 1) )))
…)
Before this change, `edit_schematic_component` with `newReference` updated
only the (property "Reference" …) field. The (reference "…") leaves inside
(instances) → (project) → (path) kept the old value. eeschema rendered the
new reference correctly and ERC passed, but `kicad-cli sch export netlist`
and "Update PCB from Schematic" both read from the (instances) block and
silently used the OLD reference — producing destructive PCB-sync diffs on
what users thought was a clean rename. Severity was high for anyone running
batch renames because the symptom only surfaces at PCB-sync time, by which
point many renames may be queued.
Walk the (instances) subtree within the matched symbol block after the
property update and replace every `(reference "OLD")` leaf with the new
value. The regex matches `(reference "X")` specifically (not
`(property "Reference" "X"`), and the walk is constrained to the
(instances …) range via the existing _find_matching_paren helper so other
(reference …) tokens elsewhere in the file can't be affected.
Adds tests/test_edit_schematic_component_instances.py covering:
- Single-instance rename updates both property and instances leaf
- Hierarchical case with multiple (path …) entries all updated atomically
- No-instances-block schematics don't crash (older KiCad / partial files)
- The regex doesn't clobber (property "Reference" …) on the instances pass
- Other field values (Value, Footprint) are left intact
- The response payload's updated.reference reflects the new ref
All 6 tests fail on main without the fix (3 fully, 3 on the instances
assertions only) and pass on this branch.
The pre-existing TestAddComponentMirrorParam failures in
test_add_schematic_component.py are unrelated and present on main —
documented in inktomi's PR #169.
Co-authored-by: mixelpixx <11727006+mixelpixx@users.noreply.github.com>
query_traces silently omits PCB_ZONE_T objects, so layer-usage audits
miss power planes and GND pours entirely. query_zones complements it by
iterating board.Zones() and returning each zone's net, layers, priority,
fill state, min thickness, bounding box, and filled area, with the same
net/layer/boundingBox filter surface as query_traces.
JLCPCBPartsManager defaulted db_path to a "data/" directory computed
relative to __file__, which fails with read-only filesystems when the
package is installed to a system-managed prefix (e.g. /nix/store, an
immutable container image, or /usr/lib). The same pattern in
download_jlcpcb.py would silently scatter the ~1.5 GB JLCPCB cache
inside the install tree even when it is writable.
The original integration plan (docs/archive/JLCPCB_INTEGRATION_PLAN.md)
called for a per-user database under ~/.kicad-mcp/. This change moves
the default to the platform-appropriate user data directory by adding
a new PlatformHelper.get_data_dir() helper that mirrors the existing
get_config_dir() / get_cache_dir() conventions:
- Linux: XDG_DATA_HOME/kicad-mcp or ~/.local/share/kicad-mcp
- macOS: ~/Library/Application Support/kicad-mcp
- Windows: %USERPROFILE%\.kicad-mcp\data
Both JLCPCBPartsManager and download_jlcpcb.py now resolve their
database paths through this helper. ensure_directories() and
detect_platform() include the new directory. Unit tests parallel to
the existing config_dir/cache_dir cases cover platform-appropriate
paths and the relative-XDG_DATA_HOME edge case.
`create_netclass` previously called legacy NETCLASSES.Find/.Add APIs that
were removed in KiCad 10. NETCLASS getters like GetMicroViaDiameter that
also no longer exist crashed any subsequent edit. The schema accepted
`traceWidth` but the handler only read `trackWidth`, so requests using
the documented field silently produced no-op netclasses.
- Add a KiCad-version-defensive shim around netclass creation that
prefers the new netclasses_map dict-style API and falls back to legacy.
- Introduce _safe_get/_safe_set helpers so missing getters/setters on
KiCad 10 NETCLASS objects fail gracefully instead of raising.
- Accept both traceWidth and trackWidth in the request payload.
Net classes still need to be written into .kicad_pro directly because
KiCad 10 stores them in net_settings.classes and the MCP only writes
to .kicad_pcb; that's a separate fix.
* Feat: save 2D board view to file instead of returning base64
The 2D view was returning base64-encoded image data in JSON, which
often exceeded token/message size limits. Now saves the rendered
image (PNG/JPG/SVG) next to the PCB file and returns the file path.
This makes the output usable by tools that can read image files
directly.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(board): add opt-in responseMode param to get_board_2d_view
Add a responseMode string parameter (enum: inline | file, default inline)
so callers can choose how the rendered image is delivered.
- inline (default, pre-PR behavior): image bytes are base64-encoded
and returned in the imageData response field -- backward-compatible.
- file: image is written next to the .kicad_pcb as
<board>_2d_view.<ext> and filePath is returned -- resolves the
MCP message-size limit problem on large boards.
Rendering logic is shared between both modes; only response packaging
differs. Updated tool schema (Python + TypeScript) and replaced the
existing test file with 5 focused unit tests covering inline/file modes
for PNG and SVG formats plus the default-is-inline contract.
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: guard SWIG auto-save against external file changes
After every board-mutating SWIG command, kicad_interface._auto_save_board()
unconditionally calls pcbnew.SaveBoard() with the in-memory board. When the
on-disk .kicad_pcb has been modified externally between our LoadBoard and
SaveBoard (KiCad GUI's own save, git checkout, another process), the
in-memory state silently overwrites those external changes - losing data
the user can't see was at risk.
This change records the file's mtime_ns + sha256 at LoadBoard and verifies
the signature matches before each auto-save:
* If the signature has diverged, refuse the save and attach a structured
warning to the command result so callers know their mutation is
in-memory only and they need to reload before retrying.
* If it matches, copy the existing file to .mcp-backups/<name>.<ts>
(rotating, keeps last 20) before overwriting.
* Update the recorded signature after our own writes so subsequent
saves are not falsely flagged.
Backwards compatible:
* No tool schemas changed.
* Successful saves return as before, with an extra `autoSave` field
when the wrapper observed something noteworthy.
* Refused saves return success: true (the in-memory mutation did
succeed) plus warnings: [...] and autoSave.diskChangedExternally,
so callers can detect the situation programmatically.
Adds tests/test_auto_save_guard.py (10 tests, all passing) covering:
signature math, refusal on external change, backup creation + content,
backup rotation, first-save semantics (no recorded signature proceeds
normally), and skip cases (no board / no path).
Motivation: the aircam-pdb fork-user lost ~480 traces and the full
footprint layout to a silent overwrite incident on 2026-05-03; recovery
was only possible because VS Code's local-history extension happened to
have a snapshot from a few minutes earlier. This guard makes that class
of incident loud and locally recoverable.
* fix(auto-save-guard): refuse only on content divergence, not mtime
The guard added in 9ba0010 records `(mtime_ns, sha256)` as the file's
disk signature and refuses auto-save when the recorded tuple no longer
matches the current one. Comparing the full tuple meant any mtime delta
fired the refusal — including a bare `touch` of the file, an atime-style
backup tool, or any MCP read path that opened the .kicad_pcb between
load and save. Users were trapped: every write needed an explicit
save_project call to bypass the false positive (documented as a
workaround in fork users' notes).
Compare on sha256 only; mtime is incidental. The actual data-loss
scenario the guard is meant to catch — an external write that genuinely
changed the file — produces a different content hash, which is what the
guard now keys off. After a touch-only mtime advance with content
unchanged, refresh the recorded signature so we don't re-hash on every
subsequent call.
Drops the mtime-equality fast path on _disk_signature: a filesystem with
coarse mtime resolution (FAT32, some network mounts) could accept two
writes inside one mtime tick; trusting mtime as a hash cache key would
re-introduce a class of silent overwrites the guard exists to prevent.
The hash itself is cheap (sha256 over a typical .kicad_pcb completes in
tens of milliseconds).
Adds 4 regression tests in test_auto_save_guard.py:
- touch-only mtime advance proceeds and refreshes the signature
- content change at the same mtime is still refused (hash divergence
must drive the decision, not tuple equality)
- the user-facing warning calls out "contents", not "mtime"
- _disk_signature returns the same hash when content is unchanged
even after the file's mtime advances
The handler iterated `board.GetFootprints()` and assigned nets to existing
pads, but had no path to *add* footprints for schematic symbols whose
Reference was not yet on the board. New parts placed on the schematic
landed in the net list with no PCB representation — the rats nest had
nowhere to terminate, and place_component / route_pad_to_pad would fail
because the target footprint did not exist.
KiCad's "Update PCB from Schematic" (F8) implicitly adds the missing
footprints; bring the MCP's behaviour in line.
Implementation:
* `_extract_components_from_schematic` runs `kicad-cli sch export netlist
--format kicadxml` (the same path `_handle_generate_netlist` already
uses) and returns a flat `[{reference, value, footprint}]` list. Walks
hierarchical sub-sheets transparently because kicad-cli does.
* `_add_missing_footprints_from_schematic` resolves each missing component
against the project's fp-lib-table via `LibraryManager`, calls
`pcbnew.FootprintLoad`, sets reference / value / FPID, and places the
footprint at the board origin (the user / autoplacer can position it).
Power and flag references (`#PWR…`, `#FLG…`) are excluded — they have
no PCB footprint.
* The pad-net assignment loop now runs *after* the add path, so newly
placed footprints get their nets assigned in the same call.
* Response payload gains `footprints_added` and `footprints_skipped`
diagnostic lists. The textual `message` field reports both the new
footprint count and the existing net / pad counts.
Adds tests/test_sync_schematic_to_board_footprints.py — 9 unit tests
covering the add path (missing ref, already-present ref, power refs,
empty footprint, unknown library) and the kicad-cli helper (XML parse,
missing kicad-cli, non-zero exit).
WireManager.add_wire and add_polyline_wire bailed out with
"No sheet_instances section found in schematic" when called on a
hierarchical sub-sheet — that block only exists in the root .kicad_sch.
The handler in kicad_interface.py converted the False return into a
flat "Failed to add wire" message, leaving callers with no diagnostic.
Apply the same fallback that WireManager.add_label has used since the
hierarchical-design support: when (sheet_instances ...) is absent,
append the new (wire ...) item at the end of the outer (kicad_sch ...)
form.
Adds tests/test_add_wire_sub_sheet.py with 6 regression tests covering
both add_wire and add_polyline_wire on a sub-sheet, including paren
balance, sexpdata round-trip, and segment count for an N-point polyline.
DynamicSymbolLoader.create_component_instance only handled root
schematics: it located its insertion point via `(sheet_instances`,
which exists only in the root .kicad_sch. Adding a component to any
hierarchical sub-sheet raised "Could not find insertion point in
schematic" and aborted the call.
Fall back to inserting just before the closing paren of the outer
(kicad_sch ...) form when the marker is absent. WireManager.add_label
already uses the same fallback for hierarchical sub-sheets.
Adds three regression tests in TestCreateComponentInstanceSubSheet
covering successful insertion, paren balance, and sexpdata round-trip
on a sub-sheet without (sheet_instances).
`BoardOutlineCommands.add_mounting_hole` produced footprints with an empty
library:name id (`(footprint "" ...)` in the .kicad_pcb), which KiCad's GUI
Move tool refuses to select — users couldn't drag the resulting MHs in the
editor. It also emitted the pad with the default `*.Cu` + `*.Mask` LSET on
NPTHs; with `padDiameter > diameter` that creates phantom copper annular
rings on every Cu layer and trips clearance DRC against neighbouring nets.
Repro: call `add_mounting_hole` with `position={x:117,y:84.5,unit:"mm"},
diameter:3.2, padDiameter:3.5`. The resulting MH is unselectable and DRC
reports phantom F.Cu pad shorts to neighbouring component pads.
Changes:
- Set a real FPID via `module.SetFPID(pcbnew.LIB_ID(lib, name))`. Default
is `MountingHole:MountingHole_<diameter:g>mm` (e.g. 3.2 → 3.2mm); a new
optional `footprintLibId` parameter overrides.
- For NPTH (the default `plated:false`), restrict the pad's LSET to
`F.Mask` + `B.Mask` only. PTH path is unchanged — the default Cu+Mask
LSET is correct there.
- Update the schema in `tool_schemas.py`: previously advertised
`x`/`y`/`diameter` at the top level, but the impl reads
`position={x,y,unit}`, `padDiameter`, `plated`. Schema now matches the
implementation and exposes the new `footprintLibId` param.
- New `tests/test_add_mounting_hole.py` regression suite (7 tests) asserting
SetFPID is invoked with non-empty lib:name (default + override forms),
NPTH SetLayerSet excludes any Cu layer, and PTH does not call
SetLayerSet (preserves default Cu+Mask).
