Repository navigation
1.12.0: release - #27
Merged
Merged
Conversation
…(G33) + crossing-hard-cap cross-band relief (G34) FEAT-002 (SEP 09-crosscloud, owner wave-6). Two class-scoped fixes so a dense as-built that reaches several external sibling bands publishes without weakening the collapse guard, and builds under the 300s guard. G33: build_diagram skips the greedy Variant A finishing (caption-safe + crossing-repair + collinear) at or above _FINISH_NODE_GATE (=20, mirroring ordering.FINISH_NODE_BUDGET), keeping the crossing-minimised layout() geometry. On 09 the unbounded pass ran 260s (over guard) and net-worsened the geometry (9 -> 19-22 crossings, account grown into the external bands). 09 now builds ~8s; every flow golden < 20 nodes -> byte-identical. G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G3s G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G3s G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G3s G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G34G3erG34G34G34G34G34G3ef for a diagram with no external sibling bands, so the collapsed quick-1.10.0 landscape (one container, 28/28) stays an ERROR; 09 (22/25, 19 cross-band pairs) drops to the soft WARNING. Corpus byte-identical (make examples-check); CHANGELOG + docs/REVIEW G33/G34/G35. Tests: tests/test_crossband_relief_1_11_0.py.
…x (G35) FEAT-001 (SEP 01-sep-context, owner wave-6). The exporter models a wrapped Flow/Legend box's render with a fixed model-space allowance (_WRAPPED_TEXT_RENDER_OVERSHOOT=15), but the draw.io CLI's real wrap render runs a few px further; the mandatory upscale pushed 01 to 1608px, over the 1600 flow budget. Inflating the global allowance re-scales every within-budget golden (forbidden), so export_one now measures the ACTUAL exported PNG width (_png_width, an IHDR read) and re-exports ONCE at a corrected scale (scale * (budget-1)/actual_px, floored at MIN_EXPORT_SCALE) only when the raster exceeds its class budget. Fires only for an over-budget raster -> every within-budget golden exports once at the unchanged scale and stays byte-identical (make examples-check clean). 01 now 1599px, layout + the e8 lm->sep flip intact. --max-width opts out. Docs: CHANGELOG + docs/REVIEW G35 (committed with G33/G34 in the prior commit). Tests: tests/test_crossband_relief_1_11_0.py (_png_width).
Order the same-side exit SPREAD by the target's direction, not the marker
number, generalising the proven bottom-face ordering (pipeline step
2a-bottom) and right-face route-lane rank (_exit_band_rank, step 2b4) to
every same-side group. New shared key routers.exit_direction_key orders a
group by the target's along-face centre (centre-y on a right/left face,
centre-x on a top/bottom face); _global_contacts step 2 sorts its generic
spread by it and step 2b4 uses it as the same-rank tiebreak. Rewrites only
the along-face BAND; the exit FACE is never moved, so edge-direction and
exit-thirds stay clean by construction.
Resolves owner wave-7 comments 01 (9/10), 02-swap (8/9), 07 (3/4),
09 (2/9, 17/16, 11/18).
Tests/gates run (all pass, from /Users/Anton_Isaiev/Documents/Cloud_Architecture):
.venv/bin/python -m pytest -q \
tests/test_spread_by_direction_sys_b.py (4 passed) \
tests/test_layout_engine.py (129 passed) \
tests/test_layout_package.py (8 passed) \
tests/test_layout_regressions.py (11 passed, 1 xfailed) \
tests/test_solver.py tests/test_genai_hub_routing.py (7 passed) \
tests/test_golden_examples.py tests/test_diagram_layout.py (119 passed) \
tests/test_solver_properties.py tests/test_placement_properties.py \
tests/test_placement_variants_properties.py \
tests/test_crossband_relief_1_11_0.py tests/test_g23_peering_node_1_11_0.py \
tests/test_snapshot_gate.py (115 passed)
make examples-check -> clean (zero golden re-baseline; byte-identical)
Files: src/rule_engine/layout/routers.py (exit_direction_key),
src/rule_engine/layout/pipeline.py (_global_contacts steps 2, 2b4),
tests/test_spread_by_direction_sys_b.py, docs/REVIEW.md (G36),
CHANGELOG.md ([1.11.0] - 2026-10-23).
…R5-masked (G37) SYS-A inverts the old Rule A (always exit RIGHT) so a MOSTLY-below target is left on the BOTTOM face (owner rule R3 "ціль знизу ⇒ вихід вниз"), backed by the already-live ENG-R2 opaque-caption mask. Resolves owner wave-7 comments 03 (13 down), 08a (5 down / drop stair), 10, 09-12/16 (down-left), and the lower half of 02. Changes (contacts/pipeline/variants/routers): - contacts.select_contacts: return bottom-centre exit (0.5,1.0) + top-centre entry (0.5,0.0) when _box_mostly_below (below, |dx|<=COL_STEP/2, clear column = GUARD 1) AND allow_downward clears it (GUARD 2); else keep right-centre. - contacts._box_mostly_below: new helper (GUARD 1 obstacle test over the union x-span of the two glyphs). - pipeline._global_contacts step 1: GUARD-2 pre-pass groups candidate downward edges by source and clears at most ONE per source caption (allow_downward); the R5 mask covers ONE line, not two. - variants._spine_variants: relax the straight-drop gate from strict _box_directly_below to _box_mostly_below so the scored solver weighs the drop against the stair. - routers.route_spine: draw a mostly-below offset drop as a clean drop-then-step-then-drop (no stair past the glyph). - GUARD 3 (over-connected bottom) left to the existing spill/valve; GUARD 4 (target own caption) inherent to a top entry. - geometry.rule_violations UNCHANGED: edge-crosses-own-label stays lint/ finishing-pass-only, so a masked own-caption drop scores no penalty. AUTHORIZED golden re-baseline: ZERO goldens changed. The goldens' below targets are strict same-column stacks already drawn straight by pipeline step 1b/1c; SYS-A generalises the rule-based seed + variant gate, which the solver argmin- selects to the identical route. GUARD 2 produced 0 conscious keeps across the golden corpus (confirmed by the byte-identical re-baseline). Tests/gates (engine .venv/bin/python, per-file batches): - tests/test_downward_exit_sys_a.py (NEW, 8 tests) PASS. - tests/test_layout_engine.py 129 PASS (updated the Rule-A directly-below/obstructed contact tests to the SYS-A behavior). - tests/test_golden_examples.py + test_snapshot_gate.py 137 PASS. - tests/test_layout_regressions.py 11 passed, 1 xfailed. - tests/test_solver.py + test_genai_hub_routing.py + test_crossband_relief_1_11_0.py PASS. - tests/test_solver_properties.py + test_placement_properties.py + test_placement_variants_properties.py + test_g23_peering_node_1_11_0.py 45 PASS. - tests/test_layout_package.py + test_diagram_layout.py 36 PASS. - make examples-check: OK (every generated example and raster is fresh — zero golden re-baseline). No ERROR-level checks flip: edge-direction / edge-float / corridor-sharing / icon-crossing edge-routing all satisfied by a bottom->top vertical. Two goldens visually reviewed (aws/02 HA summary, oci/01 genai stack) — downward drops read correctly. REVIEW.md G37 + CHANGELOG [1.11.0] 2026-10-23 wave-7 entry added.
…l (G38)
Add an optional per-peering caption so a materialized vpc_peering resource
node carries the SHORT peer-resource name (e.g. "Atlas peering"), not the raw
pcx-id (which would trip node-label-length; pcx-id + CIDR stay in the Flow
legend). FEAT-003 of the seventh edge-routing wave.
- model.py EdgeSpec.peering_label: Optional[str] = None (additive, defaulted;
every existing spec byte-identical). Names a caption, never a coordinate;
ignored when peering_as_node is off.
- model.py DiagramSpec.peering_labels: Tuple[(pcx_id, label), ...] = () —
engine-derived mapping the renderer discovers the pcx caption from.
- peering.py materialize_peerings stamps peering_labels (one entry per
materialized edge that carried a peering_label; unlabelled edges contribute
none, so the no-label case stays backward-compatible). Legs stay non-peering
with empty markers; the "peering" word stays suppressed.
- peering.py peering_label_for(spec, pcx_id) -> str | None: the stable accessor
the SEP build scripts (08/10, FEAT-004) use instead of hard-coding
"VPC Peering". Exported from rule_engine.layout.
- tests/test_g23_peering_node_1_11_0.py: label discoverable on the pcx node,
short label passes node-label-length, no-label backward-compatible default,
word-suppression + empty-marker legs still hold, EdgeSpec default identity.
Tests run (engine .venv):
pytest -q tests/test_g23_peering_node_1_11_0.py tests/test_layout_package.py
tests/test_layout_engine.py -> 154 passed
pytest -q tests/test_layout_regressions.py tests/test_solver.py
tests/test_solver_properties.py tests/test_placement_properties.py
tests/test_golden_examples.py tests/test_snapshot_gate.py
tests/test_diagram_layout.py tests/test_crossband_relief_1_11_0.py
-> 234 passed, 1 xfailed
Golden examples unchanged (no re-baseline; additive defaulted fields).
…ANGELOG 2026-10-23 - REVIEW.md: add G38 (ENG-1 VPC-peering node naming via EdgeSpec.peering_label) and G39 (FEAT-004 workspace sweep: SPEC-13 apigw-into-region, ENG-1 wiring, ENG-2 verify-after resolved with no engine change). G36/G37 already present. - CHANGELOG.md [1.11.0] - 2026-10-23: add the ENG-1/G38 Added bullet, widen the wave-7 theme line, and record the FEAT-004/G39 workspace-sweep note incl. the 'no ERROR flips' confirmation and the zero-golden-rebaseline statement. No engine code change in this commit (docs only). Prior wave-7 engine commits 05c0b8e (SYS-B/G36), 6d39778 (SYS-A/G37), fa75a36 (ENG-1/G38) are untouched; no amend, no push. uv.lock left untracked.
Judge a node's geometric containment against its effective role scope, independent of the declared container - the geometry companion to the layout() SpecError. Four reasons: global-in-region / regional-in-vpc / vpc-scoped-outside-vpc / external-in-account. Container kind via the style-signature classifier (no container_kinds map); effective scope from a scope= token else role->ROLE_SCOPE (icon-index reverse map). ERROR on a network diagram, WARNING otherwise; lint-only (not in the solver oracle). Precision-over-noise: provider-independent cases (global-in-region, external-in-account) fire from the role fallback; provider-dependent cases (regional-in-vpc, vpc-scoped-outside-vpc) fire only from an explicit scope= token, so a provider-agnostic ROLE_SCOPE never flags correct goldens. One authorized re-baseline: cross-cloud/01 CDN moved out of the AWS VPC to the edge (a real global-in-region defect); every other golden byte-identical.
…8 ITEM B, G42) Add five neutral EXTERNAL node roles to the bundled mapping tree and an explicit NodeSpec.external placement flag (SEP audit E-02/E-03/F7, corrections R2-2/R2-3): - actor -> AWS General User glyph, users -> General Users glyph (never the IAM Role shield, E-02) - saas -> General Client glyph, on_prem_app -> General Server glyph: the NEUTRAL fallback every un-curated SaaS/on-prem system uses (GitLab/Jenkins/Jira/ Confluence/Slack and any other -> THESE, never a claimed brand mark, never a random AWS service icon, R2-3). The four curated brand roles (mongodb_atlas/redis/postgresql/keycloak) stay the ONLY SaaS brand marks. - egress_only_igw -> IGW-family glyph + egress-only caption (F7, closest-official; ROLE_SCOPE key is new+unused, the AZ/subnet CORE being deferred under G40) NodeSpec.external=True forces a node OUTSIDE every cloud boundary regardless of a declared container (base._assign_nodes_to_leaves carve-outs) and emits an external=1 style token that FEAT-001's scope-placement lint reads for its external-in-account case. E-01 (bedrock_guardrails) confirmed present (bundled), skipped. Additive + default-inert: icon-index gains rows (existing byte-identical), external defaults False and emits nothing, every shipped golden byte-identical (make examples-check clean).
… marker (ITEM C, G43) Wave-8 ITEM C (audit E-08): a long edge's numbered flow-marker defaulted to the geometric midpoint, which can land far from both endpoints (dead space, a container border, an unrelated caption). - geometry.marker_anchor: an unpinned marker on a route > LONG_EDGE_MARKER_LEN (2.5*COL_STEP) anchors at min(0.5*len, 1.2*COL_STEP) from the source; a short or label_pos-pinned edge is unchanged. - resolve_marker_collisions: seed from the same near-source default (_default_label_pos) and EXTEND phases 2-3 to slide an overlapping badge clear (no new phase 4, per R2-med-3). - linter._is_network_component_or_flow: raise marker-on-container-border + marker-label-collision to ERROR on network/component/flow, WARNING on landscape/other (component recognised by diagram_type, R2 Finding 7). Gate (R2 Finding 8): engine goldens 0 marker-on-border/label findings before and after. One authorized golden re-baseline (examples/gcp/01-gcp-vertex-pipeline: four long markers move near-source, both versions lint clean); other 13 goldens byte-identical. docs/REVIEW.md G43 + CHANGELOG + both diagram-lint.md copies + tests/test_markers_1_11_0.py.
…mplates, rule-engine-relayout E-04: export_raster FLOW_SMALL_WIDTH=1000 / FLOW_SMALL_MAX_SCALE=2.5 — a very small flow canvas upscales up to 2.5 instead of 1.5 (scale path only; _canvas_width and the G35 width fix untouched; budget fit term still binds, so no budget bust). Every shipped golden flow is >1000px, so the corpus is byte-identical. E-06: examples/spec_templates/ — one runnable, commented DiagramSpec template per diagram_type (context/c4_container/network/component) + README, each driving the public layout()+build_diagram() pipeline into a publication-eligible triple. E-07: rule-engine-relayout <file>.drawio — best-effort re-flow of an engine-emitted .drawio that is idempotent (byte-identical re-emit) and fails honestly on a hand-authored/hand-pinned source it cannot reconstruct. Hand-authored .drawio is documented as not first-class. Additive + default-inert; make examples-check byte-identical (no golden re-baseline).
…klog (Wave-8 FEAT-005)
Introduce BANDED_KINDS = frozenset({'az','vpc','subnet','public_subnet'}) as a
shared base.py constant and repoint the three literal banded_kinds sites at it.
Add the pure _subnet_ordinals(spec) (mirror of _az_ordinals, keyed on
subnet-within-az parent; declaration order -> public 0, app 1, batch 2, db 3).
Additive + default-inert: no shipped spec declares a subnet container.
… bands
Add the kind in ('subnet','public_subnet') branch to _node_az_band returning
the subnet tier ordinal; keep the existing az service-row offset and the else=0
branch. Widen _band_packing band_has_box so a subnet / public_subnet band is
boxed (four stacked subnet bands sit 3*CONTAINER_PAD apart, matching the two
stacked az bands). Gated on kinds no shipped spec declares -> goldens unchanged.
…king _leaf_region_containers prefers a region's subnet/public_subnet containers as its leaves (else az, else vpc), so _assign_nodes_to_leaves populates leaf_nodes[subnet] and size_containers wraps each subnet around its members. _stack_peer_azs_in_vpc partitions a VPC's az children BY REGION and shares a secondary range only within a same-region group of >= 2: SEP one-az-per-region -> two groups of 1 -> skipped -> East-West preserved; HA two-az-per-region -> each group of 2 still stacks -> byte-identical.
_validate_spec now raises SpecError when a regional/global node declares a vpc/az/subnet/public_subnet container, and its precondition matches any region/vpc/az/subnet/public_subnet container of the node's region (so the guard fires in the locked G40 model, which has no region-kind container). A scope='network' node in a subnet is still accepted. place.py:size_containers is UNCHANGED per design [r3-med1] (_size already recurses depth-first + memoized, raising SpecError for a childless container).
Three tests still carried pre-change expectations and failed on the Wave-8 commits (they passed at v1.11.0 but the commits shipped behavior changes without updating them): - test_summary_marker_on_caption / test_lint_reports_marker_on_icon: assert marker-label-collision severity ERROR (not WARNING) on a flow diagram, per the G43 (ITEM C) escalation of marker overprints to ERROR on network/component/flow diagrams. - test_widening_the_grid_reduces_crossings_on_a_tight_fan -> renamed to test_col_step_lever_reaches_the_router_on_a_tight_fan: the SYS-A/SYS-B exit-ordering changes (G36/G37) altered fan routing so a wider grid no longer yields a strict crossing reduction on this synthetic fixture. The test now proves the lever reaches the router (crossing count responds to col_step), its documented purpose; the golden corpus remains the real validation surface.
The Wave-8 diagram-quality and authoring work was developed on the 1.11.0 line after v1.11.0 was cut (2026-10-22) and tagged, but never published. Ship it as 1.12.0. - Bump every version pin 1.11.0 -> 1.12.0 (pyproject, powers plugin.json, bootstrap.sh, check-workspace-init hook). - Rewrite the 1.12.0 CHANGELOG section as an aggregated, user-facing summary of the new functionality (edge routing, external-system icons, subnet tiers, new authoring templates + relayout command, sharper placement/marker checks), replacing the mislabeled per-day '## [1.11.0]' dev sections. The released '## [1.11.0] - 2026-10-22' section is left intact beneath it. - Complete the G40 subnet-tier feature: the draw.io container style for the subnet / public_subnet leaf kinds, with its test. - Update the SECURITY.md supported-version matrix to the 1.12 minor. version_guard --triple (pyproject / CHANGELOG / VERSION-at-tag) agrees on 1.12.0; release notes extract cleanly for 1.12.0.
Flip docs/REVIEW.md G40 from Deferred to Resolved (1.12.0, FEAT-001) with the edit sites and the four HA landscapes as named golden-safety coverage; add the dated G40 CHANGELOG entry under 1.12.0.
…-9 seam) G40 step 4 broadened the _validate_spec scope guard precondition to match any region/vpc/az/subnet/public_subnet container of the node's region. That also matched a pre-G40 synthetic spec's OWN vpc/az container, wrongly raising SpecError for a regional/global node placed in a vpc/az with no region-kind container and no subnet — breaking 10 pre-existing tests (test_hybrid_1_11_0, test_layout_engine, test_placement_variants_properties) and contradicting the clause's documented 'no region container -> pre-1.10.3 behaviour'. Narrow the precondition to fire only for (a) a node declared directly in a subnet/public_subnet tier (the G40 locked-model violation that test_g40_azsubnet exercises) or (b) a present region-kind container (the classic region tier). G40 subnet enforcement and the region-container case are preserved; the geometry scope-placement lint rule (G41) still judges the vpc/az-only synthetic specs post-layout. Corpus byte-identical.
…before regional-band work)
Add NodeSpec.band / band_row: a region-direct node can opt out of the vertical _REGIONAL_COLS column logic and be laid out as a horizontal row below the VPC inside the Region (band='south'), below the Region inside the Account (band='account-south'), or above the VPC inside the Region (band='north'). Members wrap at the Region width and a declared band group stays contiguous. _assign_nodes_to_leaves wraps a south band in its Region and an account-south band in the Account so the containers size around it. Default-inert: a spec that sets no band lays out exactly as in 1.12.0.
Add DiagramSpec.band_groups (+ band_group_captions / band_group_styles) and diagram_layout.band_group_cells: one boundary box per band group, sized to its members PLUS full-caption width (never truncated), emitted at the back of the document by build_diagram / _assemble_document. The bandgroup- boxes are decorative like flowgroup- boxes: geometry.build_geometry and cli._parse_drawio_page skip them, so they are not counted by node-count, read as edge-pierced, or promoted to boundaries. Default-inert.
compute_scale now fits the canvas to min(width_budget/W, height_budget/H) via a new _canvas_height helper, so a tall-but-narrow landscape auto-fits the class height budget (1600 flow / 3600 landscape) WITHOUT a manual --max-width. The height term only ever reduces the scale and only when the canvas would otherwise overshoot its height budget, so a raster already within budget is byte-identical; --max-width stays a width-only override.
Bump version to 1.13.0, add the CHANGELOG 1.13.0 section, a short ARCHITECTURE.md note on regional-band placement and the height-aware export, and focused unit tests (tests/test_regional_band_1_13_0.py) covering the south/north/account-south bands, wrapping, group contiguity, the captioned band-group box, and the height-aware export scale.
…plugin, bootstrap, workspace hook)
…-in-subnet) Add VPC_EDGE_ROLES + is_vpc_edge(role) in layout/base.py and extend scope-placement with a new edge-service-in-subnet reason: a VPC-edge device (IGW/NAT/VPN/Transit gateway, egress-only IGW) drawn inside a subnet/public_subnet container is flagged (ERROR on a network diagram, WARNING otherwise). The roles stay network-scoped in ROLE_SCOPE; this adds a placement constraint, not a new scope tier. Additive + default-inert: the new is_vpc_edge parameter defaults off, so the four existing reasons and every shipped golden are unaffected.
Add EdgeSpec.peering_id, the engine-derived DiagramSpec.peering_ids (stamped by materialize_peerings parallel to peering_labels), and the peering_id_for accessor (exported from rule_engine.layout). A materialized vpc_peering node can now carry the raw pcx-id on a second caption line below the short peer name, keeping it under node-label-length. Additive + default-inert: an edge with no peering_id yields no peering_ids entry (single-line fallback), so every shipped spec is byte-identical.
…ualize _equalize_peer_widths widens the narrower peer AZ outward AFTER size_containers sized the VPC band around the pre-equalization (narrower) AZ, so an asymmetric two-region VPC (a multi-node service row widening one region) ends with the peer AZ a few px past the VPC (container-overlap, ERROR on a landscape). size_containers now re-wraps every non-account, non-az band around its equalized children before the account is sized. Golden-safe: where the peer AZs already match (every shipped HA golden) equalize is a no-op and the re-wrap reproduces the same box, so all 13 example generators stay byte-identical.
The E-EDGE edge-service-in-subnet rule forbade any VPC-edge device (IGW / NAT / VPN / transit gw / egress-only IGW) inside a subnet OR public_subnet box. But a NAT gateway inside a PUBLIC subnet is the canonical AWS pattern (a NAT needs a public-subnet route to the IGW), so geometry.check_scope_placement now fires edge-service-in-subnet only for a PRIVATE subnet and is suppressed inside a public_subnet. A private/database subnet is still a placement defect. The four existing scope-placement reasons, the is_vpc_edge default, and every shipped golden/raster stay byte-identical (regen_examples.py --check fresh). tests/test_e_edge_scope.py updated: NAT-in-public-subnet clean, NAT-in-private-subnet still ERROR on a network diagram / WARNING otherwise.
…ow backlog G47 refined to note a public_subnet is an allowed home for a VPC-edge device (canonical AWS NAT placement; only a private subnet flags edge-service-in-subnet, per commit 5a2e814). New G50 records the depth-5 limitation that a 3+ device VPC edge row is not cleanly renderable (service-row band 0 collides with the public-subnet tier), with the Wave-11 workaround (egress-only IGW as a Legend note) and the scheduled fix.
Collapse the formerly-separate 1.12.0 + 1.13.0 + Wave-11 work into a single 1.12.0 release. Pin every version reference (pyproject, power plugin, bootstrap, session-start hook) to 1.12.0, align the SECURITY.md supported-version matrix to 1.12.x, and finalize the consolidated CHANGELOG entry dated 2026-10-28.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This release gathers all the diagram-quality, authoring, and network-topology work that continued after 1.11.0 shipped — the formerly-separate 1.12.0 (edge routing, external systems, subnet tiers, sharper checks), the 1.13.0 regional-band / band-group / height-aware-export work, and the Wave-11 SEP network-topology capabilities (E-VPCFIT, E-PCX, E-EDGE) — collapsed into one release. Every change is additive and default-inert: no shipped spec sets any of the new options, so every shipped example renders byte-identically.
Added
New ways to author a diagram. A set of ready-to-run, commented templates now covers the common diagram types — Context, Container, Component, and cloud network topologies — so you can start from a working example instead of a blank page. A new
rule-engine-relayoutcommand re-flows an engine-generated diagram, cleaning up its routing and spacing without changing the content; running it on an already-tidy diagram leaves it untouched.External systems look right. Users, actors, and third-party or on-premises systems that have no official cloud icon now get a clear, neutral glyph (a person, a generic client, or a server) instead of being drawn with an unrelated cloud-service icon. You can also mark a node as external so it is always placed outside every cloud boundary, matching how it actually relates to your account.
Subnet tiers in network diagrams. Diagrams can now express public and private subnet tiers nested inside availability zones, so a network layout can show the real subnet structure rather than stopping at the VPC. Availability-zone stacking is grouped per region, so single-zone-per-region layouts keep their side-by-side arrangement while multi-zone regions stack as before.
Horizontal "regional band" placement (opt-in). A region-direct node can now carry
band="south"to be laid out in one or more horizontal rows directly below the VPC and inside the Region, ordered left→right byband_rowthenslotand wrapping when the band is wider than the Region, instead of stacking in the vertical right-hand regional column.band="account-south"places the band below the Region, inside the Account (for a CI/CD delivery strip). Members of a declared band group stay contiguous. Implemented insrc/rule_engine/layout/base.py(_place_base/_place_regional: band nodes are excluded from the_REGIONAL_COLScolumnar logic and laid out by a new_lay_bandhorizontal packer;_assign_nodes_to_leaveswraps a south band in its Region and an account-south band in the Account so the containers size around it) and the newNodeSpec.band/NodeSpec.band_rowfields insrc/rule_engine/layout/model.py. Default-inert: a spec that sets nobandlays out exactly as before.Captioned band-group container. A new
DiagramSpec.band_groups(withband_group_captions/band_group_styles) draws one boundary box per band group, sized to its members plus enough width to show the FULL caption (never truncated) using the caller's style — unlike the decorativeflow_groupsbox, whose multi-word caption truncates on a narrow run. Implemented asdiagram_layout.band_group_cells, emitted at the back of the document bybuild_diagram/_assemble_document. Default-inert.Height-aware raster export.
rule-engine-export-rasternow fits the canvas to the smaller of the width- and height-budget scales (scale = min(width_budget/W, height_budget/H)), so a tall-but-narrow landscape auto-fits the class height budget (1600px flow / 3600px landscape) without a manual--max-width. The height term only ever REDUCES the scale and only when the canvas would otherwise overshoot its height budget, so every raster already within its height budget is byte-identical.--max-widthstays an explicit width-only override. Implemented insrc/rule_engine/export_raster.py(_canvas_height,compute_scale).VPC band re-wrapped around its AZ children after width-equalization (E-VPCFIT).
size_containerssizes each VPC band around its AZ children, then_equalize_peer_widthswidens the narrower peer AZ outward to match its sibling. When one region is wider than the other (a multi-node VPC service row widening region a), equalize grew the region-b AZ a few pixels past the VPC that was already sized to the narrower AZ, so the VPC no longer wrapped it (container-overlap, ERROR on a landscape).size_containersnow re-wraps every non-account, non-az band around its (now equalized) children before the account envelope is sized (src/rule_engine/layout/place.py). Golden-safe: where the peer AZs already match (every shipped HA golden) equalize is a no-op and the re-wrap reproduces the same box, so every shipped example is byte-identical. Test:tests/test_e_vpcfit_rewrap.py.pcx-id on a two-line peering node caption (E-PCX). A materialized
vpc_peeringresource node can now carry the rawpcx-…connection id on a SECOND caption line below the short peer name (line 1Atlas peering, line 2pcx-0e257b23d337fda0b) — the two short lines keep the node undernode-label-length, where a single line carrying the ~21-char pcx-id would trip it. The additive tripletEdgeSpec.peering_id(src/rule_engine/layout/model.py),DiagramSpec.peering_ids(engine-derived, stamped bymaterialize_peerings), and the accessorpeering_id_for(src/rule_engine/layout/peering.py, exported fromrule_engine.layout) mirrors the existingpeering_labeltriplet exactly. Default-inert: an edge with nopeering_idcontributes nopeering_idsentry, so a peering stays single-line and byte-identical. Test:tests/test_e_pcx_label.py.VPC-edge service placement check (
edge-service-in-subnet, E-EDGE). A VPC-edge device (internet / NAT / VPN / transit gateway, egress-only IGW) attaches to the VPC and routes across its boundary, so it belongs at the VPC edge, OUTSIDE every private subnet. These roles staynetwork-scoped inROLE_SCOPE(so they are not wrongly flagged for sitting outside a subnet); the newVPC_EDGE_ROLESfrozenset +is_vpc_edge(role)predicate insrc/rule_engine/layout/base.pyidentify the subset, andgeometry.check_scope_placementgains ONE new reasonedge-service-in-subnet:<node>fired when anis_vpc_edgenode is drawn geometrically inside a PRIVATEsubnetcontainer (a device inside apublic_subnetis the canonical AWS pattern — a NAT needs a public-subnet route to the internet gateway — and is ALLOWED, per the owner decision; being inside the VPC but not a subnet is clean).linter._check_scope_placementresolves the node's role from the committed icon reverse-map and passes the predicate; the reason rides the same severity join — ERROR on a network diagram, WARNING otherwise. The newis_vpc_edgeparameter defaults to "never an edge device", so the four existing scope-placement reasons and every shipped golden/raster stay byte-identical (scripts/regen_examples.py --checkfresh). Tests:tests/test_e_edge_scope.py(NAT in a public subnet clean end-to-end; NAT in a private subnet still ERROR on a network diagram / WARNING otherwise).AWS AZ + subnet-tier nested container CORE (G40, 2026-10-27). The Wave-8 CORE the earlier releases deferred (REVIEW.md G40): the locked depth-5 model
account ⊃ vpc ⊃ {az-a region "a", az-b region "b"} ⊃ {public, private-app, private-batch, isolated-db subnets}. Each AWS availability zone is an engine region (side-by-side East-West via_region_secondary_offset), a realaz-kind box is drawn per zone, and the four subnet tiers are the new banded leaf kind stacked North-South. Edit sites (allsrc/rule_engine/layout/base.pyunless noted): a sharedBANDED_KINDS = frozenset({"az","vpc","subnet","public_subnet"})constant repointing the threebanded_kindsliteral sites; the pure_subnet_ordinals(declaration order → public 0, app 1, batch 2, db 3); the_node_az_bandsubnet branch (keeping the HIGH R2-1 az service-row offset);_band_packingband_has_boxwidened so a subnet band is boxed;_leaf_region_containerspreferring subnet leaves;_stack_peer_azs_in_vpcpartitioning az children BY REGION (SEP one-az-per-region → side-by-side preserved; HA two-az-per-region → stack preserved byte-identical); the_validate_specscope guard broadened to raise on a regional/global node declaring a vpc/az/subnet/public_subnet (narrowed in the same release so it fires only for a node declared directly in asubnet/public_subnettier — the locked-model violation — or when aregion-kind container is present; the earlier broadening also matched a pre-G40 synthetic spec's own vpc/az container and wrongly flagged 10 pre-existing layout tests, so the fix restores the documented "no region container → pre-1.10.3 behaviour" while keeping the G40 subnet enforcement — the geometryscope-placementlint rule still judges those vpc/az-only specs post-layout); anddraw_cli._container_styleresolving thesubnet/public_subnetkinds (additive fallback tonetwork_boundary).place.py:size_containersis unchanged —_sizealready recurses depth-first and raisesSpecError("… has no children to size around …")for a childless (empty) subnet. Additive + default-inert: no shipped spec declares a subnet container, so every golden stays byte-identical, with the four HA landscapes (each nestingazboxes at depth 5) as the named golden-safety anchor (make examples-checkclean). Covered bytests/test_g40_azsubnet.py.Sharper diagram checks. A new placement check flags a service drawn in the wrong scope — for example a global or edge service (CDN, WAF) sitting inside a VPC, or an external actor drawn inside the account — and blocks it on network diagrams. Numbered step markers that land on top of a service name, an icon, or a boundary line are now blocked on network, component, and flow diagrams, where a misplaced number is genuinely misleading.
AWS network container colours match the official 07/2026 icon pack (G46). The Availability Zone box is now the official teal (
#00A4A6) dashed box instead of the earlier blue#147EBA, and the Region box border is now dotted instead of dashed, matching the AWS Icon-package_07312026 and the awslabsGroups/all.pumldefinitions. VPC (purple), public subnet (green) and private subnet (teal) are already correct and unchanged; the darkened, higher-contrast caption colours are kept by design. This re-baselines three AWS example diagrams (01-agent-platform, 02-ha-multiregion summary+landscape, 03-hybrid-infrastructure) with a stroke/line-pattern change only — no layout changes — and every other shipped example renders identically.Improved
Edges read more naturally. When a service connects to another one directly below it, the connection now drops straight down instead of detouring out to the side. When several connections leave the same side of a node, they fan out in the direction of their targets rather than in an arbitrary order, so the lines cross less and the flow is easier to follow. Long step numbers on long connections now sit near where the flow starts, instead of drifting into empty space or onto a border.
Clearer labels for network peering. A VPC-peering connection can now be labelled with the name of the thing it peers with (for example the peered database or network), instead of a generic placeholder, so the relationship is obvious at a glance.
Small diagrams export larger. A very small diagram (a two- or three-node context view) is no longer exported as a tiny postage-stamp image; it now scales up to a comfortable, readable size while still respecting the file-size budget.