Live KiCad saves without refilling.live apply --save writes the
board as it is; a part moved into a pour keeps stale fill over its pads
until live refill. The API gives no access to the interactive router.
Text is placed by hand.existing free finds room for a part's copper
and courtyard, not for its silkscreen; a MOVE can place the reference and
value, but where is your call (11).
Freerouting is the weak link. Not repeatable on dense boards; ignores
polygons; necks tracks; its log names nets as unrouted that are routed; on
an existing dense board it could not route a 3.4 mm net. Worth trying:
its job/API mode and the Topola router.
Schematic editing is narrow.boardwright.sch changes fields, removes
parts and adds a part, copied from one on the sheet or taken from a library
(single-unit only), on the root or a sheet used once, connected by label or
power symbol (09). It does not wire point to point or add sheet pins.
existing update carries a change to the board and swaps a footprint
only for one with the same pads; what it adds lands beside the board for
free and MOVE, and footprints whose symbol is gone wait for REMOVE.
The pad-to-pad router is new and narrow. It routed two nets on the
StickHub demo (one that Freerouting could not), applied live and DRC-clean
against the baseline, and five of six layer-changing nets on a 6-layer demo
by its own clearance measure (10), and the four connections of the 4-layer
rehearsal with vias, judged clean by KiCad (11). Rules in a .kicad_dru
file are not read. No push-aside, no pairs, no length tuning; multi-pad
nets are routed as the pairs you list.
Alternatives to Freerouting were not tried. Its job mode and the
Topola router both need installing or configuring, which is a decision
about the machine, not this tool.
si.pair and the impedance numbers have not been used on a real board.
Stackup figures are from memory of the fabs' tables.
02's bottom pour stays in five pieces (78 %). Tried FR_BOTTOM_COST = 20
on it in a throwaway copy: three attempts left 19, 8 and 1 connections
open; the completion pass finished two of them. A 0.4 mm QFN-56 on two
layers needs the bottom layer for its escape. The honest fix is four layers.
The completion pass is slow. It is capped at 12 open connections and 90
seconds; before the cap it ground for seven minutes on a board with 19 open.
existing verify does not know what the edit touched. A new DRC
finding on a net the edit never named fails verify the same as one on
its own nets; a refill can move findings around a large board. apply,
reroute, update and a staged KiCad edit do not record the nets and
references they changed, so there is nothing to split on (#44).
One flat schematic sheet, label-per-pin, for generated boards.
No current or thermal sizing of copper.
No part sourcing. Konnect's approach: the JLC parts feed in local SQLite
and rule-based footprint-to-package matching.
Boards 01-05 pass with the current tool's checks but were last generated
before the staging refactor. 03 was regenerated in a throwaway copy
afterwards and came out DRC-clean and equivalent.
No way to re-point a placed footprint's library link, or take a footprint off the board -> update swaps footprints with the same pads, REMOVE (12, #37, #38).
Live apply, selection, refill, move, rip-up and save not run against KiCad -> all run on KiCad 10.0.8 (#4).
Parts moved onto other copper without a word -> apply's clash check, existing free (11, #24).
Stacked vias from routes sharing a pad -> routes reuse their net's through copper (11, #23).
No way to move a footprint's text -> MOVE text placement (11, #25).
Generated boards unlinked from their symbols -> gen and update link them (11, #26).
MOVE rewrote the board file -> moved blocks edited in place (11, #32).
Schematic adds only by copying, root sheet only, no way to the board -> library symbols, global tables, sub-sheets, existing update (#3).
ESP32 footprint warning: the module's keep-out zone was left beside the part -> board.place_footprint_zones (05, #6).
Pour necks not measured by the current check -> review.neck (#5).
Pours invisible in renders -> copper plots (03).
Legal-but-split power polygons -> review (03).
Hand routing, vias, polygons, keep-outs, silkscreen text -> copper.py (03, 05).
Length figures doubled by duplicate segments -> de-duplication (04).
Crossed hand tracks -> check_tracks (05).
Autorouter stubs and dangling vias -> removed after routing (04).
Starved thermal reliefs that moved every run -> one spoke suffices (04).
Bottom pours shredded by the autorouter -> coverage review, FR_BOTTOM_COST (04, 05).
Skew as total copper -> match on routed pad-to-pad paths (06).
Stock libraries only -> LIBS (08).
Two layers only -> LAYERS, PLANES (06).
No return-path check -> ref, via_gnd (06).
Orientation found after the fact -> pad_at (06).
Minutes per look -> place + preview, 5 s (06).
Generated boards only -> boardwright.existing (07).
Hand routes as typed coordinates -> ROUTES, pad to pad (07).
Autorouter leaving random nets open -> completion pass and retry (06).
Decouplers with no via of their own -> via_near, hand ground vias (06).
No current sizing -> current, per path, IPC-2221 (03).
Values from memory -> every board and the stackups read from the documents (references.md).
No check of an edited live board -> stage and verify --staged (07).