Cloud catalog records never carry [tool.orcaslicer.plugin.settings], so the
metadata merge wiped the locally-parsed settings and plugins silently ran on
their built-in defaults (ctx.params arrived empty).
The M620→M621 firmware toolchange block weight (500x) was only applied
to sparse track sample lines. Hundreds of G1 moves between samples
inside the M620 block got weight=1, causing firmware toolchange to
appear compressed on the timeline plot.
Build continuous M620→M621 line ranges from track samples and apply
weight=500 to ALL lines within those ranges. This makes toolchange
and wipe tower zones proportionally accurate on the timeline.
Add two standalone Python tools for deep comparison and analysis of .3mf
slicing project files:
- compare_slices.py: comprehensive slice comparison with filament usage,
nozzle mapping, tool change sequences, prime tower analysis, temperature
timeline, and automatic critical discrepancy detection
- show_temp_plot.py: interactive HTML temperature timeline plotter for
visualizing heater profiles during multi-nozzle prints
Both tools use only Python stdlib (no external dependencies).
Primary use case: regression testing H2C carousel purge volumes and
BBS compatibility verification.
The upstream _make_wipe_tower() tracked purge volumes per-extruder (2 slots),
which collapsed all H2C carousel filaments into one slot and caused massive
redundant AMS flushing (~40g instead of ~0.4g).
Changes:
- Print.cpp: Replace per-extruder nozzle_cur_filament_ids with BBS
NozzleStatusRecorder that tracks per group_id (carousel slot 0..6).
Use get_nozzle_for_filament() to resolve physical slot per layer.
Select filament_prime_volume_nc for nozzle changes, filament_prime_volume
for filament changes (BBS pattern).
- PrintConfig.hpp/cpp: Register filament_prime_volume (per-filament EC prime
volume, default 45mm³) matching BBS PrintConfig.
- Preset.cpp: Add filament_prime_volume to preset keys list.
Safe for non-carousel printers: group_id == extruder_id when each extruder
has one nozzle, so NozzleStatusRecorder behaves identically to the original
per-extruder tracking.
Reference to BBS: BambuStudio/src/libslic3r/Print.cpp _make_wipe_tower() L3341-3392
Drop the orphaned PluginCallbackList (dead
after the PluginManager migration) and hop
run_on_*_callbacks onto the UI thread via
CallAfter, snapshotting under the mutex on the
worker first. Keeps wx subscribers off the
detached load/unload workers.
Only one action source ever existed, so the
IActionSource interface and ScriptActionSource
are gone. ActionRegistry now subscribes to the
plugin loader and enumerates actions directly
in init() - no polymorphism for one impl.
Replace the opaque FNV-hash SpeedDialActionId
with a readable composed id of the form
prefix:title:source_key. Split AppAction's
single source field into source_key (stable
identity, e.g. plugin_key) and source_name
(display), so identity and display no longer
share one field.
* fix: impl refactor
* fix: unload/load python module, race conditions, freezes
* remove dead code
* remove extra hook
* remove more dead code
* fix gil run script
Resolve five conflicts, all of which needed both sides rather than a pick:
- BackgroundSlicingProcess: ours was a pure tabs->spaces reformat of base, so
keep main's per-filament volume/nozzle map read-back (its only change here).
- GUI_App: main's #12506 else-if attached to an `if` this branch deleted;
re-expressed onto the same-agent early-return path (the agent factory caches
per id, so pointer equality is the same predicate).
- MainFrame: both sides relocated Sync Presets independently; keep main's
push_notification plus the branch's Plugins menu items.
- Tab: the "TODO: Orca: Support hybrid" blocks were unchanged base, not a branch
decision; take main's enabled Hybrid to match the already auto-merged siblings.
- test_config: union of both sides' cases (6 plugin + 9 multi-nozzle).
The filament-group golden harness landed with H2C/A2L support (#14685). Its
"FilamentGroup golden regression" / stress_66 case fails intermittently on
Windows x64, on main and on unrelated PRs alike. The test depends on how fast the
runner is.
The k-medoids clustering these goldens exercise is an anytime search bounded by a
3 second wall clock. Every restart is seeded from its own index, so nothing about
it is random. What varies is how many restarts fit in the budget, and the best
cost is a minimum over completed restarts, so a slower runner is never better.
Grading a score produced that way measures the machine as much as the code.
Add a ClusteringBudget struct and let the tests set it. The defaults are the
current 3 seconds and 30 restarts, so slicing behavior is unchanged. A
non-positive timeout removes the wall clock and bounds the search by restart
count alone.
The goldens are then graded under a fixed budget of four restarts, where every
one of them reaches the BambuStudio reference within 3%, so the score becomes a
property of the code. This retires the machine-specific 125103 lock on stress_66.
The default wall-clock path keeps its own test, asserting the grouping is valid
and the search does not run away. It makes no score assertion, because under a
wall clock that number is not a property of the code.
The golden test also checks the run fits in ten times the default wall clock.
Slicing quality depends on how many restarts fit in the budget, so a search an
order of magnitude slower would degrade real groupings while a fixed-budget score
gate stayed green.
The 3% tolerance stays as the parity allowance against the goldens. It also
covers a small spread across standard libraries: the k-medoids search seeds each
restart with std::shuffle, whose algorithm the C++ standard leaves unspecified,
so libstdc++, libc++ and the MSVC STL permute the same seed differently, start
from different medoids, and settle on slightly different groupings, about 3e-4
apart and only on the goldens heavy enough to reach the k-medoids search.
The Unit Tests job sparse-checks-out only .github/scripts/tests, so the
baked-in absolute PROFILES_DIR was missing at runtime; the shipped-profile
test then read a non-existent JSON and null-dereferenced in opt_string.
Check out resources/ in the unit-test job, and guard the test helper to
skip when the profile is absent and require the key before dereferencing.
Follow-up to the cross-gantry mirror axis change, from an adversarial review of it.
GCodeViewer builds its own copy of the zone grid to place the sequential-preview
carriage markers, and it must reproduce PartPlate::calc_imex_zones exactly or the
markers drift away from the ghosts they are meant to track. It did not, in two ways,
because sizing the grid and placing a cell answer different questions:
- Sizing: calc_imex_zones counts every Copy/Mirror tool's OWN column, including the
non-representatives of an aggregated (Span) gantry -- they still donate a column.
GCodeViewer only ever saw the representative, so on an aggregated gantry it could
count fewer columns than the plate and lay its markers out against wider strips.
- Placement: calc_imex_zones PINS an aggregated cell to the primary's column, because
that row-strip spans the full bed and has no column of its own. GCodeViewer used the
representative's own column, which put the marker a strip away from the ghost
whenever the representative was not column-paired with the primary.
Track the two sets separately: grid_tool_ids sizes the grid from own columns, eff_col_of
pins only aggregated tools when placing. Out-of-grid tool indices are deliberately left
unfiltered -- calc_imex_zones drops them while calc_imex_ghosts keeps them, so no policy
here can agree with both, and a comment says so rather than pretending otherwise.
Also:
- The mirror-axis rule lived in three copies (two PartPlate lambdas plus an inline
re-derivation here). Hoist it to imex_mirror_axis_for() so the ghosts and the markers
cannot drift apart, and unit-test it, including degenerate tools_per_gantry.
- Drop imex_head_transform's mirror_axis default. A defaulted axis silently hands a
forgetful caller the X reflection, which is wrong for every cross-gantry tool and
fails silently -- exactly how a stale test kept certifying the old rule.
- Replace that stale test: it asserted a diagonal mirror "flips X only", the rule this
work overturned, and stayed green only because of the default.
- A secondary sharing the primary's gantry now takes the primary's Y box facing. The
box shows the side a tool could be hit from, and two tools on one beam can only be
hit by the same other gantry. No-op on the rear-* layouts, where the hardcoded value
already matched; on front-* layouts it pointed the box away from the only tools that
could reach it.
- Correct two comments that described the aggregated X-frame substitution as a
reflection plane. It is not one: it exists to zero gantry_offset.x, and removing it
would push aggregated ghosts a column off their strip.
Verified: 385/385 tests; CLI slice of the IMEX regression project is byte-identical to
the previous commit's G-code apart from the timestamp, confirming this is
visualization-only and cannot affect sliced output.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
catch_discover_tests registers each Catch2 test with ctest by name, and ctest then
re-invokes the binary passing that name back as a -# filter. The IMEX test names
contain em dashes (U+2014). They survive discovery, but on Windows the round trip
through the console codepage mangles them, so the filter matches nothing:
Filters: "imex_head_transform G-- copy mode is pure translation"
No test cases matched -> No tests ran -> exit != 0 -> ctest: Failed
All 85 IMEX tests were reported as failures on both Windows x64 and arm64 without a
single one of them ever executing. Linux and macOS are UTF-8 end to end and were
unaffected, which is why this went unnoticed since the names were introduced in
77c32a2e15.
These were the only non-ASCII test names in the whole tests/ tree. Renaming them to
plain hyphens costs nothing and keeps the suite portable.
Test names only -- no assertion, no logic, no comment is touched.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A Mirror tool reflects across the boundary it shares with the primary's zone, and
which boundary that is depends on where the tool sits:
- Same gantry: the tools are side by side along X, so the shared boundary is
vertical and the reflection negates X. This is what single-gantry IDEX does,
and it was the only case the code modelled.
- Different gantry: the zones are stacked along Y (front strip vs rear strip),
so the shared boundary is horizontal and the reflection negates Y. The part
that comes off gantry 1 is a Y-reflection of the tool directly behind it.
imex_head_transform() hardcoded diag(-1, 1, 1) for every mirror, as its own TODO
acknowledged. Lift the axis to a caller-supplied ImexMirrorAxis; PartPlate picks it
from the tool's gantry row. Both reflections keep det = -1, so a mirrored part stays
a true mirror image rather than a 180-degree rotation, which would print the
primary's part merely turned around.
The correct axis removes two workarounds. Both ghost paths special-cased aggregated
mirrors to "drop the X reflection, translate 1:1 and bake the flip into the mesh"
because reflecting X pushed the ghost off-bed as the primary was dragged. With a Y
reflection the X translation is already zero for aggregated tools, so that falls out
for free and the special cases are deleted.
Preview markers follow the same rule, which also fixes two placement bugs:
- Mirrors reflected across a Copy tool's zone edge, falling back to the primary's
column when a row had no Copy. In iq-mirror (0:P,1:C,2:M,3:M) the front row has
no Copy, so t2 and t3 both fell back and computed the identical X — both drawn
on top of each other in t3's zone. A mirror now reflects within its own zone.
- The toolhead footprint box flipped to the far side of the nozzle for any mirror
right of the primary. That only holds for an X-axis mirror, which reverses the
carriage's orientation; a cross-gantry mirror keeps the X orientation of the
tool behind it, so its box stays on the same side.
Tests cover the cross-gantry and diagonal cases, that the axis is caller-supplied
rather than inferred from the offset vector, and that both axes are reflections
(det = -1) rather than rotations.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Conflicts were all co-located additions rather than design collisions:
- GCode.cpp: adopt upstream's toolchange(filament_id, nozzle_id) signature and
per-variant set_config_index() while keeping the IMEX bare-T<n> suppression;
rebase the second-layer temperature loop's non-IMEX branch onto upstream's
get_filament_config_index() resolution.
- Preset.cpp / PresetBundle.cpp / PrintConfig.cpp: keep both sides' option-list
and enum-map entries.
- GLCanvas3D.cpp: upstream's printable_heights argument plus the IMEX ghost pass.
- PartPlate.cpp: keep <set> (still used).
- test_gcodewriter.cpp / test_3mf.cpp: keep both sides' test cases.
Resolves two conflicts:
- GLCanvas3D.hpp: keep both the IMEX ghost render declarations and
upstream's _render_wireframe_overlay().
- test_gcodewriter.cpp: both sides appended test cases to the same
region; keep upstream's origin/machine-limit tests alongside the
pressure-advance and temperature scenarios.
Keep action identity and display metadata
constructor-set so registry keys cannot drift
from the objects they index.
Require sources to publish unique ownership while
the registry retains shared keepalive for runs.
Preserve opaque IDs and persisted state keys.
Lifecycle callbacks can be registered while
another thread dispatches them, risking races
and iterator invalidation.
Snapshot shared callback objects under a mutex,
then invoke them unlocked to preserve reentrancy
and mutable callback state.
Move plugin capability enumeration, loader
subscriptions, and action construction into
ScriptActionSource. Keep ActionRegistry focused
on generic action state, dispatch, and snapshots.
Use source rather than package for generic origin
metadata so future non-plugin providers share the
same interface. Action IDs and persisted
configuration remain unchanged.
Subscribe before initial enumeration so plugin
events cannot be missed between the snapshot and
callback registration. Add a focused test for
source startup.
A dual-nozzle H2C print with support filament hangs at its first nozzle
switch. The emitted file shows the change-filament block's M620 O ordinal
jumping from O1 straight to O230, plus a duplicate "M1020 S<n>" toolchange
command right after every change block. Two causes, fixed together because
they interlock (the ordinal check keys off the same toolchange detection
that suppresses the duplicate):
- append_tcr incremented m_toolchange_count once per prime-tower visit
(roughly once per layer), while the change-filament template only emits
its M620 O{toolchange_count + 1} line on real filament changes. With 229
change-less sparse tower layers below the first support layer, the first
real change reported ordinal 230. The counter now advances only when the
expanded change block really contains a toolchange command, and the
placeholder exposes the upcoming change's ordinal (count + 1). The
set_extruder path already counted per real change and is unchanged.
- toolchange_prefix() returned "M1020 S" for BBL printers, so the
custom_gcode_changes_tool() dedup could never match the stock profiles'
line-leading "T[next_filament_id] ..." commands and the writer's own
toolchange was appended after every change block on dual-extruder
machines. The prefix is now the plain "T" (the manual-filament-change tag
branch stays first), and the M1020 form moved into GCodeWriter::toolchange()
as an explicit branch that also carries the nozzle:
"M1020 S<filament> H<nozzle>". The nozzle parameter is signed on purpose:
the null-safe nozzle lookup legitimately yields -1, matching the stock
templates' own H-1 convention.
The prefix change also lets the CoolingBuffer recognize the change blocks'
T commands as tool boundaries on BBL printers (its per-filament attribution
previously keyed off the duplicate M1020, or nothing at all on
single-extruder models); its existing out-of-range guard ignores
T1000-class machine commands.
Verification: full suites green (libslic3r 48998 assertions / 169 cases;
fff_print 692 / 65 including three new scenarios - writer emission per
printer kind, dedup + ordinal progression on sequential prints, and a
prime-tower regression scenario verified to fail against the old per-visit
counting). Byte gate: 18 of 20 fixtures bit-identical; the sequential repro
differs by exactly its 3 removed duplicate M1020 lines, deterministic
across two runs. Reslicing the field project that exposed the hang yields
M620 O1 followed by a gapless O2..O59 and zero duplicate M1020 lines.
Co-authored-by: songwei.li <songwei.li@bambulab.com>
The time estimator's speed/acceleration limits were indexed by time
mode only, reading slot 0 of the per-(extruder x volume-type) arrays
the multi-extruder profiles already carry (H2C 0.4: 8 entries, H2D
0.4: 10). Every move was therefore modelled with the first machine
slot's limits regardless of which nozzle variant was printing -
estimation fidelity only, since emitted feedrates/accelerations are
decided on the slicing side.
Now the estimator resolves the machine slot of the nozzle currently
mounted in the active extruder: the nozzle grouping context is handed
to the processor BEFORE the streaming replay (new member + setter -
deliberately separate from the post-stream result-field handover that
gates the richer change-time model, whose timing is unchanged), the
occupancy recorder is populated on every filament change (bookkeeping
decoupled from the gated time model; recorder writes have no time
effect), and get_machine_config_idx maps (volume type x extruder type
x extruder) to the slot via the printer's variant layout, newly
carried on the processor result. The feedrate/acceleration getters
gain a slot parameter indexing [slot*2 + mode]; jerk and the
print/travel/retract accelerations stay mode-only. Reloaded sliced
projects re-estimate with the result's saved grouping context;
imported bare g-code degrades to slot 0 - the historical read.
M201/M203 write the parsed value into EVERY slot's mode entry (a
firmware envelope change is global), which keeps per-slot reads in
lockstep with the mode-only reads they replace: the fleet emits
envelope lines before any motion, so estimates - hence the estimated
time header, M73 lines, and every other byte - are unchanged (20/20
pinned-slice byte gate bit-identical, incl. the sequential repro
sliced twice). Fidelity improves where envelope emission is off or a
migrating per-layer plan moves filaments across variants.
Tests: a stub-driven processor case proving the slot follows the
active nozzle through the exact production path (T..H.. commands,
fallback recorder bookkeeping, 4x time ratio on the slow variant),
that emitted M201/M203 reach every slot, and that a missing context
degrades to slot 0. Suites green (libslic3r 48998/169, fff_print
667/62).
When a per-layer nozzle grouping migrates a filament across nozzle
variants, the write-back turns two groups of config arrays from
filament-indexed into column-indexed: the per-variant filament options
(one column per variant a filament uses) and the merged extruder
retract overrides (resized to the column count by apply_override).
Export-path readers that still indexed them with the raw filament id
read a neighbor's column for every filament ordered after a migrating
one: toolchange/standby temperatures (M104/M109), retraction lengths
and feedrates, wipe distance, z-hop types, air-filtration keys, and -
through the Extruder's cached flow term - the extrusion E of every
move.
Now every such read resolves its column through the existing
layer-aware resolver (get_filament_config_index ->
Print::get_filament_config_indx), which returns the raw filament id
whenever no per-layer grouping result is published, so static prints
are byte-inert by construction. The Extruder itself has no layer
knowledge, so it gains an injected config column (set_config_index,
default = filament id) that the generator refreshes at the only two
resolution-changing events - layer change and writer toolchange - and
that re-syncs the cached e_per_mm3 flow term. Old-filament reads
resolve at the current layer, which is safe because the per-layer maps
are gap-filled carry-forward. Whole-array placeholder copies
(toolchange temperature overrides) are rebuilt in filament order,
mirroring the existing per-variant placeholder remap. The resolvers
move to the public section so non-friend helpers (ooze prevention) can
resolve too.
Documented, deliberately unchanged: the wipe tower's per-filament
parameter rows (no layer dimension; tower x per-layer grouping is a
follow-up), travel_slope's physical-extruder read, estimator pre-heat
bookkeeping temps, and index-0 header diagnostics.
Verification: new Extruder column-injection scenario (defaults, column
follow + flow-cache rescale, filament-indexed reads unaffected, reset
semantics) and a migrating write-back case proving the column shift for
filaments ordered after a migrator and the resolver tracking it (11 +
14 assertions); suites green (libslic3r 48998/169, fff_print 655/61);
20/20 pinned-slice byte gate bit-identical (incl. sequential repro x2
deterministic).
When the per-layer filament selector (enable_filament_dynamic_map)
migrates a filament across nozzle variants (e.g. Standard -> High Flow),
the config write-back only stored the derived extruder map; every
per-variant filament value (retraction, nozzle temperature, flow,
flush...) kept the numbers resolved from the pre-slice static mapping.
Now both dynamic write-back sites (the by-layer branch and the
sequential stitch) branch on the result's dynamic support. Migrating
results run a mixed-filament expansion that regathers every
filament_options_with_variant key from the pristine per-variant
superset, giving a migrating filament one config slot per (extruder
type x nozzle volume type) it lands on - filament_self_index,
filament_extruder_variant, and all value arrays grow in lockstep - and
recompute the retract overrides with per-slot machine indices so a nil
slot falls back to its own variant's machine value. Non-migrating
dynamic results take the merged three-map write-back so re-applies
reproduce from the written maps. Unrouted filaments resolve from the
result's own default map, so slot resolution never depends on
filament_map round-tripping through the plate config.
Print::apply reproduces the identical expansion from the persisted
group result (shared dedupe helper, expansion function, and slot
indices on both sides): the expanded keys sit in the psWipeTower /
psGCodeExport invalidate lists, so without the reproduction every
re-apply after a selector slice would diff non-empty and permanently
invalidate. cal_non_support_filaments now resolves the extruder per
layer from the published result for dynamic groupings.
filament_map_2 keeps its apply-time static derivation; nothing on the
dynamic path reads it (the per-slot machine indices key the override
merge), and per-(extruder x volume-type) machine limits in the g-code
processor remain a documented follow-up.
Every change is gated behind is_dynamic_group_reorder() or a persisted
result with dynamic support; no profile sets the flag, so the static
fleet's instruction stream is unchanged (20/20 pinned-slice byte gate
identical, incl. the sequential repro sliced twice, deterministic).
Tests: expansion unit coverage (migrating slots, unrouted fallback via
the default map, mis-sized volume map ignored, nullable retract keys in
lockstep, slot machine index layout), an end-to-end stub-driven
write-back asserting expanded slots, per-layer config-index resolution,
the override merge incl. the nil-slot variant fallback, and re-apply
stability, plus a real selector slice staying valid across re-apply.
Suites green (libslic3r 48987/168, fff_print 633/60).
Sequential (by-object) prints were incoherent with the per-layer filament
selector (enable_filament_dynamic_map): the by-object branch published a
static grouping while each per-object ToolOrdering independently ran the
dynamic planner from an empty nozzle status and wrote its own map to the
config (one write per object, last object wins). The exported toolchange
sequences then disagreed with the published result that drives the
per-layer maps, placeholders, and selector emission.
Now the by-object branch, when the selector is enabled, plans each unique
object once — threading the physical nozzle occupancy and the previous
object's last filament into the next plan — stitches the per-object
per-layer nozzle maps into one print-wide result (gap-filled by the new
normalize_nozzle_map_per_layer so any layer index resolves a filament's
nozzle consistently), publishes it, and writes the derived extruder map
back once. The plans are cached on the Print and g-code export consumes
the cache: the ToolOrdering seed changes the plan input (dontcare
assignment, first-layer reorder), so a fresh export-time construction
could re-plan differently from the published stitch. The per-object
dynamic write-back is gated off for sequential prints.
Every change is gated behind is_dynamic_group_reorder(); no profile sets
the flag, so the static fleet's instruction stream is unchanged (20/20
pinned-slice byte gate identical, incl. the by-object repro sliced twice).
Tests: normalize unit coverage (carry-forward, back-fill, ragged input),
stitched-blocks selector detection, and an end-to-end by-object selector
slice (apply -> process -> export) asserting the published stitched
result, one cached plan per object, the config write-back, and a clean
export. Suites green (libslic3r 48958/165, fff_print 633/60).
Experimental fuzzy on geometry
Mirrors libslic3r's fuzzy_polyline on the slice contours at Step.posSlice,
demonstrating the count-changing mutation idiom (rebuild ring via
Polygon.append, write back via ex.contour / ex.set_holes). C++ analogue
test proves area preservation, cascade, and bounded displacement.
The Print-level LayeredNozzleGroupResult had a single producer, the
by-layer branch of ToolOrdering, which is gated to non-sequential prints.
The by-object branch in Print::process computed a grouping only in auto
map modes and never stored it, so a sequential slice exported with a null
group result: the per-nozzle placeholder tables came up empty and any
start g-code indexing nozzle_diameter_at_nozzle_id[] aborted with
"Indexing an empty vector variable". A prior by-layer slice masked the
bug by leaving its (never cleared) result on the Print.
Now the by-object branch runs get_recommended_filament_maps in every
static map mode (in manual modes the result mirrors the user's
assignment, deviations throw as in by-layer) and publishes it
print-wide. The config write-back stays gated to auto modes: in manual
modes it would only re-store the pre-slice values.
Regression test: a two-object by-object print must publish a non-null
group result and resolve nozzle_diameter_at_nozzle_id[] in start g-code
(both fail without the fix). Suites green (libslic3r 48929/162,
fff_print 633/60); 18-fixture byte gate identical; the by-object repro
project goes from the export error to valid g-code, determinism x2.
SoftFever