Repository navigation
test(agent/crdt): property coverage for the follower read leg - #16237
christian-byrne wants to merge 48 commits into
Conversation
Lands the agent shell dark behind the existing PostHog flag agent-in-app-experience: the gate wiring (extension setup syncing the flag into the panel store), the Ask Comfy Agent topbar entry button, a stub docked panel mounted in graph and linear views, and the flag-off E2E proving no agent surface renders while the flag is off. The stub intentionally omits what later slices own: engagement telemetry and detection, panel resize and maximize, the workspace inset, selection-driven chrome hiding, and the full panel content. File paths match the feature branch so later slices replace stub content in place.
Rapid or overlapping workflow opens, closes, and subgraph hash navigations could interleave: a slower earlier load could overwrite a newer selection, closing a workflow could re-activate it, and stale hash navigations could move the canvas after a newer one. Workflow loads now run through a serialized queue with per-path pending tracking and re-entrant close accounting; hash/graph synchronization carries navigation intent ids so only the newest intent can write the route or move the canvas. Test-led split from the agent feature branch: the added service and navigation-store tests pin the contract - 13 of them fail against the previous implementation - and only non-agent hunks are extracted (the new extension load hooks stay on the feature branch with their consumer).
The full-suite baseline caught main's navigateToHash suite failing against the extracted intent-id signature: the feature branch had extended that suite (+557) plus app.test.ts's intent-forwarding case, and the initial extraction missed both. Both are taken now; the app.test.ts case covering the excluded extension load hooks stays on the feature branch with its subject.
Eight-hat review round on the shell PR; the convergent findings: - Register the gate extension in the cloud block, where the PostHog flag actually exists and the SDK is initialized - the misplaced !isCloud registration made the panel unreachable in every shipped build, failed the cloud E2E job, and leaked posthog-js into the OSS dist scan. - Force-enable dev builds BEFORE the SDK import so a blocked chunk cannot take local development down, and add a bounded retry that re-takes the flag subscription when extension setup wins the race against posthog initialization (pre-init subscriptions are dead). The gate marks a body dataset when settled so the flag-off E2E asserts after gating actually ran, anchored by the tab-bar container being visible. - Restructure LinearView like the canvas overlay (row root + column wrapper) so the docked panel is a right-hand sibling instead of stacking below the workspace; pinned by a placement test. - Replace the entry chip's fixed ink-700 fill with a semantic surface token (the fixed hex under a theme-flipping foreground computed 1.11:1 in light themes) and add aria-pressed. - Rebuild the E2E boot mocks on the shared mockCloudBoot helper instead of a hand-rolled route fork. - Trim agentTheme.css to the four consumed tokens (the trimmed rest returns with its consumers and keyframes in later slices), drop the width ref and dead guards from the store, drop the dead shell testid and alias click handler, and merge the async component into the setup block with a complementary landmark role. - New gate composition tests (flag on/off/flip, init race, blocked SDK), a close-button test for the panel root, a legacy-tab-bar gating case, and a real enable/disable transition in the store persistence test.
The cloud E2E run proved the registration fix end to end (the flag-on test passed against the production build) but the flag-off spec timed out awaiting the settle marker: posthog drops false-valued bootstrap flags, so isFeatureEnabled legitimately stays undefined forever when the flag is off and settle only arrived at retry exhaustion (~30s). The retry budget only needs to cover telemetry's init window, so it shrinks to ~5s worst case, exhaustion being the settled fail-closed state, and the spec awaits settle with an explicit window.
Round-2 panel residuals on the gate: readiness now settles on an observed flags delivery OR a definite flag read, the resolved path no longer stacks a second subscription, and the exhausted path re-takes the subscription against the by-now live instance before settling - so a flag that first appears mid-session still propagates instead of waiting for a reload. The unit mock now models the off state the way the real SDK reports it (undefined, never false) and pins the exhausted-then-arrives path. Also: the close button reuses g.close instead of minting a duplicate key, and the store drops the consumer-less open().
…delivery Round-2 panels caught the rework's own regression with the SDK source as receipt: uninitialized posthog invokes the onFeatureFlags callback synchronously with errorsLoading and registers nothing, so the delivered signal fired against a dead subscription and pinned the panel off for the session - the exact race the retry defends. The flag source now filters error-context callbacks before they reach the listener, the unit double models the real pre-init contract (fires once, registers nothing, undefined flags) and the race test asserts a live subscription exists after recovery. The retry constants are exported so tests pace off the real values, the settle window fits inside the cloud project's test timeout with actual headroom, and the E2E mocks drop a hand-rolled auth-token route that shadowed the shared workspace mint.
Eight-hat review round; the convergent findings: - Pass suppressWorkflowReset=false at the two import call sites (importA1111, loadApiJson) - they reset to root with no later workflow-load hash republish, so the armed suppression left the URL on a destroyed subgraph id; pinned by new app tests covering both branches of the clean flag. - Capture the closing workflow's path once at closeWorkflow entry: a rename mid-close mutates workflow.path in place, and the live re-read in the finally would strand a phantom closingWorkflowCounts entry that silently no-ops every future open of that path. - Restore main's last-workflow fallback for an inactive close: the sole open workflow closing while inactive now still creates the default workflow. - Log queued-load failures inside the serialization chain: attaching the internal handler marks the caller's promise handled, which had silenced fire-and-forget failures. - Test hardening: a module-state reset hook with per-test leak assertions, the active-workflow no-op guard pinned from both sides, a reopen-after-close-settles case, the misleadingly named clean-flag test renamed, and a shadowed parameter renamed.
Tree-shaking: the cloud extension block now gates on the literal __DISTRIBUTION__ comparison - the cross-module isCloud const does not dead-code-eliminate, so the gate's posthog-js import was shipping into OSS and desktop bundles (measured +289KB; the localhost build now scans clean and the vendor chunk returns to baseline). The gate's retry/re-take machinery is deleted: the shipped posthog bundle constructs its featureFlags extension in the constructor, so a pre-init subscription registers and survives init() (probe-verified); what remains is the errorsLoading filter, a settle timeout for the no-token path, and a dev force-enable that now honours the ff: localStorage override in both directions. The gate's settle signal moved off document.body onto a store ref bound as a data attribute on the tab-bar actions container, and its constants moved out of extensions/core so nothing new lands on the window.comfyAPI surface. Layout and a11y: the dock width is a proportional cap (min(420px, 33%)) so narrow viewports and the linear arrange panels can no longer be clipped dead; the panel title is a real heading the landmark references via aria-labelledby; borders unify on interface-stroke and the four-token alias theme file is deleted in favour of design-system tokens. Structure: the extension tree flattens to the manager exemplar's shape (stores/, utils/, components/agent/), and the E2E fixture moves to fixtures/ and drives a raw page like the cloud siblings - with the assets mock typed against ListAssetsResponse and the posthog ingest host routed so CI emits no live third-party traffic. Tests: real-contract gate coverage replaces the mock-manufactured race tests, dev force-enable and blocked-import paths are pinned, aria-pressed and the settle attribute are asserted, the graph-side mount invariants are pinned, and the hollow store case is replaced with a writeDefaults pin.
- key the closing-workflow registry by raw workflow object identity - queue default/blank/duplicate loads through the same load queue - bound the close-drain and skip it when no workflow is open - release a failed workflow-load navigation intent and restore the hash - move the workflow-load hash publish into loadGraphData's finally - pin the close-race orderings and the intent release in unit tests
- new-surface mutation battery left this guard undiscriminated: the existing replacement test's interleaving empties the closing window before the shift loop runs - hold the neighbor's close inside the store call so it is still open AND registered at decision time, then assert the fallback lands on the survivor and never the closing neighbor
- a blocked posthog import only reached console.error; it now also calls reportError with the stable agent_flag_gate_load_failure type, so the failure is queryable in Sentry and RUM - fail-closed settling and the dev force-enable are unchanged and both paths pin the report in the gate tests
- swap the store test onto createTestingPinia({stubActions:false})
per the repo's testing-pinia convention
- pin that disabling the flag mid-session never wipes the stored
open preference
- give the async panel root a Suspense fallback so the dock frame
never flashes empty chrome
…w nits - the nav recovery catch and the queued-load catch now also call reportError with stable types (workflow_navigation_failure, workflow_load_failure), keeping each site's redirect and queue-continue behavior - pin that a stale workflow-load publish reapplies the latest graph intent and clears the pending reset suppression (the finally publish is what prevents the suppression leak on a failed load) - detect the default load by its payload, not a missing argument
The Aug-12 review's one Request-Changes gap, still live: closeWorkflow deleted the draft at the top of the try, so a replacement load that fails (which now rethrows) left the tab open with its draft already gone. The draft is removed only after the store close is real, and the regression pin asserts a failed replacement keeps both the tab and the draft.
… source Two comment anchors naming which coordination primitives the ECS per-document milestone deletes and which survive, so the supersession is greppable from the code as well as the decision log and PR body.
Same treatment the crdt files got: every multi-line justification block drops to the one invariant the code cannot show.
A real configure failure is caught inside loadGraphData (dialog, then resolve), so the close path never saw it and deleted the draft of a tab that stayed open over a never-painted graph. loadGraphData now returns an explicit outcome; openWorkflow and loadDefaultWorkflow thread it; closeWorkflow aborts on an explicit failure report through its existing not-closed contract, releasing the load's navigation intent and skipping warnings for a graph that never painted. The regression tests drive the real failure shape: rootGraph.configure throwing, not a mocked rejection.
The docked panel's complementary landmark lost its accessible name while the async panel body loaded; the Suspense fallback now carries a visually-hidden agent-panel-title heading with the same translated string, covered in loading and resolved states. The dock-siblinghood test now mounts the splitter overlay and asserts the rendered structure instead of pinning source text; the linear-mode guard stays a source pin because GraphCanvas cannot mount in the unit environment. The dock width moves from an arbitrary min() to w-1/3 max-w-[420px].
A rejected AgentPanelRoot load left the Suspense fallback up forever with nothing reported. The dock now acts as an error boundary: the failure reaches reportError under agent_panel_load_failure and the panel renders a localized error state that keeps the landmark named. Also drops a non-null assertion in the mount test for optional chaining.
OSS builds now emit no agent chunk and carry no agent code: the dock mounts and tab-bar button sit behind the distribution literal (new useAgentDockMount seam), and the flag gate follows the initTelemetry idiom - the guard lives inside the retained function with every agent module imported dynamically past it, so the dead remainder and its import edges fold away, taking posthog-js with them. Cloud builds inline the gate into the core graph instead of fetching a separate extension chunk, and the panel body chunks load only once the gate enables and the panel opens. Verified by dist scans of both distributions.
The follower doc, wire schema and frame client, mint ports, litegraph mutator and semantic projector, layout follower bridge, op envelope and sender, schema guard, and the dev panel - all flag-gated and inert with no consumer in this slice (activation arrives with the agent stack). Carries the beforeLoadGraph/afterLoadGraph extension hooks, the layoutStore withActor scope, graph traversal support, the four transport prompt-error registrations with their locale entries, the comfy-multi-player dependency with its license-scan exclusion, ADR-0020 and the follower boundary check. Transitional knip entries mark the tree until slices 07+ import it.
…heet - rename ADR 0020 -> 0024 (main already carries 0020-0023), fix its H1, update all references, add the missing README index row - correct the license-exclusion comment: the pinned package declares GPL-3.0-only and ships a LICENSE file - genericize RUN-CLOUD-E2E.md (no internal hosts or names) and note the follower has no mount site until the agent-panel slice - install __agentCrdtPoc only under import.meta.env.DEV - update stale PoC comments; document both load hooks in docs/extensions/core.md including the configure-failure asymmetry
- port the child's graph-load hook test and pin once-each-in-order - pin afterLoadGraph skipped when graph configure fails - cover the four agent prompt-error types in the resolver suite - document beforeLoadGraph/afterLoadGraph pairing on ComfyExtension - dedupe findSubgraphPathById/findSubgraphNodePathById via a shared walk - reject async callbacks in layoutStore.withActor, with a pinning test - exercise the traversal walker's malformed-subgraph guard
- remove the comfy-multi-player license-scan exclusion: the pinned package declares GPL-3.0-only (allowlisted), and keeping the check live guards future pin bumps of a git-ref dependency - render CrdtDevPanel only in dev builds; knip tracks it via ignore instead of entry so unused-export reporting stays live for the tree - reword the ADR and boundary-check lines that asserted an env gate that does not exist on this branch
- correct the afterLoadGraph JSDoc: neither after-hook fires on a configure failure; brackets reset at the next beforeLoadGraph and close on afterConfigureGraph (superset of this hook's cases) - pin the full four-hook load order in the app suite - widen withActor's guard to thenables, report through the assert seam, neutralize the orphaned rejection, and pin actor restoration
- dev panel: init open state and start the poll interval only in dev builds (the render gate alone left prod lifecycle effects running) - symmetric DEV gate on the console-helper teardown delete - ADR: drop the dev-rendered claim, name the stopgap product-flag-gated, future-tense the mount statement - run sheet: point Verify at the dev panel's real testid
…ooks - the guard asserts, not throws: DEV throws, production reports and returns the thenable - say so on the method - the configure-fail test now pins that afterConfigureGraph is also skipped, completing the documented two-hook promise
…nt load a failed configure had already cleaned or partially replaced the shared root graph while the previous workflow stayed selected with a deactivated change tracker; on an aborted open, reload the retained workflow from its just-saved state through the normal load path so selection, canvas, and change tracking agree - no retry loop
- every gate dependency now loads inside the guarded path and setup() hands the gate promise to the extension service, so a failed chunk settles fail-closed instead of hanging as an unhandled rejection - production waits for the first non-error flag delivery before enabling: posthog's persisted cache could otherwise mount the panel and fetch agent chunks on a stale true; the dev override stays immediate - pinned: stale persisted flag ignored until delivery; dependency failure settles closed, reports, and resolves the setup promise
its only readers (the dev taps and the minting console helper) moved to the write-leg slice; the tracking stays there with them
CI's pnpm refuses git-dep build scripts outside the allowBuilds allowlist, so every install-stage check failed; the entry matches the pinned tarball exactly (carried from the reviewed source, where the not-carried verdict was refuted by the first CI install)
Append-only merge of nathaniel/agent-v1/01-workflow-core-split at 9fd0e2a; clean three-way (workflowService restore-path + test).
- an ordinary seq gap retains the doc and replays via its state vector; a missed doc_reset is caught as a seq regression at doc_subscribed, before any folded catch-up update can apply, and refetched from zero - a workflow change replaces document and projection state together, so the next subscribe starts from the new workflow's own baseline - same-lineage reconnect and doc_reset refetch retain the projector: its snapshot records the materialized canvas, so recovery applies deltas instead of rediffing empty-to-full and duplicating the graph - decode and apply failures enter one typed fail-closed path (frame_error) that requests same-lineage replay without replacing the document - the projector commits its snapshot only after applyBatch finishes, so a partial batch failure re-derives the remaining delta on retry - a retyped node id is replaced with ordered remove + add and its links re-attached; previous-only widget keys emit an explicit clear_widget that restores the node-type default - property generators now exercise retype and widget-key drops
# Conflicts: # docs/adr/README.md
The interface stays (the frame_error union member uses it internally); only the unconsumed export goes, clearing the knip red. Owner-ruled in place of a transitional ignore - no consumer exists anywhere in the chain's canonical tree.
Adds fast-check properties over `readDocSnapshot` (QA-8 / QA-12), the two claims the existing example tests in this directory cannot make: - Arrival-order independence: a corpus of real wire ops is applied host-side through the shared package's own `applyOps`, the incremental Yjs updates are captured, and two followers are fed the same update set in two generated permutations (seed state included, so causally-pending structs are exercised the way a reconnect/replay delivers them). Both followers, and the host, must project identical snapshots. Also checks incremental delivery against a single merged update. - Read purity (KA-6 / invariant #6): reading emits no Yjs update and is idempotent. - Fail-soft ingestion: arbitrary JSON values in the `nodes`/`links` root maps, and a hostile overlay on top of a real minted doc, must never throw, and every surviving entry must satisfy the NodeSpec/LinkSpec contract the canvas layer assumes downstream (string type, finite pos, finite slots). The order property carries a vacuity guard: it asserts the corpus actually populated both root maps and that the two permutations really differed, so an all-rejected corpus cannot make it trivially true. No product code changes.
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
🎭 Playwright: ✅ 1886 passed, 0 failed · 1 flaky📊 Browser Reports
🎨 Storybook: ✅ Built — View Storybook📦 Bundle Size
⚡ Performance Report
Absolute values
Raw data{
"timestamp": "2026-08-29T15:12:43.750Z",
"gitSha": "f0a3ce7a1fa1ac925d0d5b9fef1949d8b2339bfe",
"branch": "christian-byrne/tl2-prop-02a",
"measurements": [
{
"name": "canvas-idle",
"durationMs": 2079.9309999999878,
"styleRecalcs": 9,
"styleRecalcDurationMs": 8.333,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 561.0880000000001,
"heapDeltaBytes": 11243944,
"heapUsedBytes": 79873796,
"domNodes": -281,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 8.093,
"eventListeners": -157,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-idle",
"durationMs": 2053.959000000077,
"styleRecalcs": 7,
"styleRecalcDurationMs": 5.888000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 536.1270000000001,
"heapDeltaBytes": -10670216,
"heapUsedBytes": 58942928,
"domNodes": -284,
"jsHeapTotalBytes": 4710400,
"scriptDurationMs": 9.423999999999998,
"eventListeners": -155,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66999999999998,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1875.1159999999913,
"styleRecalcs": 75,
"styleRecalcDurationMs": 37.873000000000005,
"layouts": 12,
"layoutDurationMs": 3.757,
"taskDurationMs": 893.3270000000001,
"heapDeltaBytes": -9329648,
"heapUsedBytes": 59878536,
"domNodes": -281,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 155.154,
"eventListeners": -155,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1846.7990000000327,
"styleRecalcs": 74,
"styleRecalcDurationMs": 36.476,
"layouts": 12,
"layoutDurationMs": 3.3009999999999997,
"taskDurationMs": 884.8340000000002,
"heapDeltaBytes": -8469784,
"heapUsedBytes": 60572080,
"domNodes": -281,
"jsHeapTotalBytes": 5234688,
"scriptDurationMs": 153.81300000000002,
"eventListeners": -155,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1680.641000000037,
"styleRecalcs": 29,
"styleRecalcDurationMs": 15.412,
"layouts": 6,
"layoutDurationMs": 0.528,
"taskDurationMs": 358.314,
"heapDeltaBytes": 4401000,
"heapUsedBytes": 73612656,
"domNodes": 76,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 10.649,
"eventListeners": 19,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1678.863999999976,
"styleRecalcs": 31,
"styleRecalcDurationMs": 17.177,
"layouts": 6,
"layoutDurationMs": 0.6399999999999999,
"taskDurationMs": 403.52700000000004,
"heapDeltaBytes": 4525564,
"heapUsedBytes": 73974940,
"domNodes": 77,
"jsHeapTotalBytes": 6029312,
"scriptDurationMs": 13.403000000000002,
"eventListeners": 19,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-clipping",
"durationMs": 626.9939999999679,
"styleRecalcs": 11,
"styleRecalcDurationMs": 7.728000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 439.013,
"heapDeltaBytes": 18851184,
"heapUsedBytes": 87952812,
"domNodes": 18,
"jsHeapTotalBytes": 4980736,
"scriptDurationMs": 157.865,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-clipping",
"durationMs": 738.2320000000391,
"styleRecalcs": 11,
"styleRecalcDurationMs": 8.662999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 475.806,
"heapDeltaBytes": 18842424,
"heapUsedBytes": 87398332,
"domNodes": 18,
"jsHeapTotalBytes": 4980736,
"scriptDurationMs": 162.3,
"eventListeners": 2,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66999999999998,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "large-graph-idle",
"durationMs": 2031.9450000000074,
"styleRecalcs": 6,
"styleRecalcDurationMs": 5.412,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 597.2959999999999,
"heapDeltaBytes": 10800764,
"heapUsedBytes": 85825292,
"domNodes": -270,
"jsHeapTotalBytes": -2138112,
"scriptDurationMs": 31.14399999999999,
"eventListeners": -185,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-idle",
"durationMs": 2045.5739999999878,
"styleRecalcs": 5,
"styleRecalcDurationMs": 4.900999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 678.699,
"heapDeltaBytes": -13180796,
"heapUsedBytes": 86674656,
"domNodes": -270,
"jsHeapTotalBytes": -1314816,
"scriptDurationMs": 37.50299999999999,
"eventListeners": -153,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "large-graph-pan",
"durationMs": 2178.369000000032,
"styleRecalcs": 62,
"styleRecalcDurationMs": 7.573999999999997,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1550.4940000000001,
"heapDeltaBytes": 18152620,
"heapUsedBytes": 84226292,
"domNodes": 0,
"jsHeapTotalBytes": 225280,
"scriptDurationMs": 763.6350000000001,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-pan",
"durationMs": 2277.8769999999895,
"styleRecalcs": 64,
"styleRecalcDurationMs": 9.391000000000004,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1633.038,
"heapDeltaBytes": 24893056,
"heapUsedBytes": 91181520,
"domNodes": -275,
"jsHeapTotalBytes": 225280,
"scriptDurationMs": 784.144,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-zoom",
"durationMs": 3604.68800000001,
"styleRecalcs": 63,
"styleRecalcDurationMs": 12.734999999999996,
"layouts": 60,
"layoutDurationMs": 7.027,
"taskDurationMs": 1741.595,
"heapDeltaBytes": -860220,
"heapUsedBytes": 75814104,
"domNodes": -276,
"jsHeapTotalBytes": -1875968,
"scriptDurationMs": 753.045,
"eventListeners": -179,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.670000000000012,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "large-graph-zoom",
"durationMs": 3547.370000000001,
"styleRecalcs": 61,
"styleRecalcDurationMs": 11.447999999999997,
"layouts": 60,
"layoutDurationMs": 7.594,
"taskDurationMs": 1715.8479999999997,
"heapDeltaBytes": 9776972,
"heapUsedBytes": 76134500,
"domNodes": 2,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 763.12,
"eventListeners": 8,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "legacy-node-drag",
"durationMs": 3068.321000000026,
"styleRecalcs": 40,
"styleRecalcDurationMs": 4.550999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1902.9279999999999,
"heapDeltaBytes": 21967564,
"heapUsedBytes": 89096228,
"domNodes": 0,
"jsHeapTotalBytes": 417792,
"scriptDurationMs": 946.2120000000001,
"eventListeners": 188,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "legacy-node-drag",
"durationMs": 3152.774000000022,
"styleRecalcs": 40,
"styleRecalcDurationMs": 5.096,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1960.5410000000002,
"heapDeltaBytes": 24898520,
"heapUsedBytes": 91800248,
"domNodes": -281,
"jsHeapTotalBytes": 679936,
"scriptDurationMs": 958.909,
"eventListeners": 188,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666696,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "minimap-idle",
"durationMs": 2041.5009999999825,
"styleRecalcs": 5,
"styleRecalcDurationMs": 4.328000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 584.096,
"heapDeltaBytes": -14341040,
"heapUsedBytes": 86990276,
"domNodes": -271,
"jsHeapTotalBytes": -1052672,
"scriptDurationMs": 33.947,
"eventListeners": -155,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "minimap-idle",
"durationMs": 2005.4439999998976,
"styleRecalcs": 1,
"styleRecalcDurationMs": 0.8739999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 616.2800000000001,
"heapDeltaBytes": 21804092,
"heapUsedBytes": 87228676,
"domNodes": 2,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 38.290000000000006,
"eventListeners": 4,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66999999999998,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 609.6870000000081,
"styleRecalcs": 47,
"styleRecalcDurationMs": 10.936000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 433.186,
"heapDeltaBytes": 14769136,
"heapUsedBytes": 85609300,
"domNodes": 20,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 150.311,
"eventListeners": 8,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 576.9590000001017,
"styleRecalcs": 47,
"styleRecalcDurationMs": 9.937,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 410.74,
"heapDeltaBytes": 14061084,
"heapUsedBytes": 85716408,
"domNodes": 20,
"jsHeapTotalBytes": 5767168,
"scriptDurationMs": 138.314,
"eventListeners": 8,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-idle",
"durationMs": 2022.566999999981,
"styleRecalcs": 9,
"styleRecalcDurationMs": 7.978000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 487.84400000000005,
"heapDeltaBytes": 235188,
"heapUsedBytes": 71680200,
"domNodes": -297,
"jsHeapTotalBytes": 4710400,
"scriptDurationMs": 8.03,
"eventListeners": -231,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-idle",
"durationMs": 2021.1630000000014,
"styleRecalcs": 10,
"styleRecalcDurationMs": 9.512000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 539.916,
"heapDeltaBytes": -9491900,
"heapUsedBytes": 61682464,
"domNodes": -297,
"jsHeapTotalBytes": 6021120,
"scriptDurationMs": 9.341,
"eventListeners": -201,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1739.313999999979,
"styleRecalcs": 77,
"styleRecalcDurationMs": 36.484,
"layouts": 16,
"layoutDurationMs": 4.539000000000001,
"taskDurationMs": 844.198,
"heapDeltaBytes": 4739112,
"heapUsedBytes": 69567156,
"domNodes": -299,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 114.61200000000001,
"eventListeners": -233,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1736.5800000000036,
"styleRecalcs": 75,
"styleRecalcDurationMs": 37.424,
"layouts": 16,
"layoutDurationMs": 4.405,
"taskDurationMs": 871.329,
"heapDeltaBytes": -12624636,
"heapUsedBytes": 58945528,
"domNodes": -297,
"jsHeapTotalBytes": 6807552,
"scriptDurationMs": 113.721,
"eventListeners": -201,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-transition-enter",
"durationMs": 1418.4629999999743,
"styleRecalcs": 19,
"styleRecalcDurationMs": 33.062,
"layouts": 14,
"layoutDurationMs": 14.696999999999997,
"taskDurationMs": 981.434,
"heapDeltaBytes": 2123932,
"heapUsedBytes": 99218196,
"domNodes": 13671,
"jsHeapTotalBytes": 11010048,
"scriptDurationMs": 34.69300000000001,
"eventListeners": 2383,
"totalBlockingTimeMs": 143,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "viewport-pan-sweep",
"durationMs": 8404.798999999968,
"styleRecalcs": 246,
"styleRecalcDurationMs": 32.952,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5686.277999999999,
"heapDeltaBytes": 12554992,
"heapUsedBytes": 88295708,
"domNodes": -274,
"jsHeapTotalBytes": -1126400,
"scriptDurationMs": 2587.002,
"eventListeners": -169,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.670000000000012,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "viewport-pan-sweep",
"durationMs": 8735.521000000062,
"styleRecalcs": 248,
"styleRecalcDurationMs": 36.898,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5974.08,
"heapDeltaBytes": -18508528,
"heapUsedBytes": 81062464,
"domNodes": -266,
"jsHeapTotalBytes": -1875968,
"scriptDurationMs": 2655.9060000000004,
"eventListeners": -137,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "vue-large-graph-idle",
"durationMs": 16762.654,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 16278.551,
"heapDeltaBytes": -26820348,
"heapUsedBytes": 175837164,
"domNodes": -6548,
"jsHeapTotalBytes": 5013504,
"scriptDurationMs": 254.89000000000001,
"eventListeners": -7473,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.77333333333336,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-idle",
"durationMs": 17888.034999999945,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 17082.540999999997,
"heapDeltaBytes": -32188580,
"heapUsedBytes": 176506020,
"domNodes": -6548,
"jsHeapTotalBytes": -659456,
"scriptDurationMs": 279.13,
"eventListeners": -7477,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.80000000000291
},
{
"name": "vue-large-graph-pan",
"durationMs": 20842.186000000027,
"styleRecalcs": 178,
"styleRecalcDurationMs": 17.337000000000046,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 20420.594,
"heapDeltaBytes": -41996744,
"heapUsedBytes": 160777776,
"domNodes": -6548,
"jsHeapTotalBytes": -1277952,
"scriptDurationMs": 912.8030000000001,
"eventListeners": -7467,
"totalBlockingTimeMs": 129,
"frameDurationMs": 18.33666666666674,
"p95FrameDurationMs": 16.80000000000291
},
{
"name": "vue-large-graph-pan",
"durationMs": 21115.700999999943,
"styleRecalcs": 162,
"styleRecalcDurationMs": 17.798000000000037,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 20781.974000000002,
"heapDeltaBytes": -29362856,
"heapUsedBytes": 180269468,
"domNodes": -6548,
"jsHeapTotalBytes": -7950336,
"scriptDurationMs": 871.612,
"eventListeners": -7467,
"totalBlockingTimeMs": 17,
"frameDurationMs": 17.776666666666763,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-node-resize-workload",
"durationMs": 6118.4099999999835,
"styleRecalcs": 20,
"styleRecalcDurationMs": 79.40599999999998,
"layouts": 20,
"layoutDurationMs": 21.255999999999997,
"taskDurationMs": 5971.421,
"heapDeltaBytes": -26638932,
"heapUsedBytes": 180382396,
"domNodes": -6548,
"jsHeapTotalBytes": 14213120,
"scriptDurationMs": 110.444,
"eventListeners": -7471,
"totalBlockingTimeMs": 382,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-node-resize-workload",
"durationMs": 6314.09199999996,
"styleRecalcs": 20,
"styleRecalcDurationMs": 78.37099999999997,
"layouts": 20,
"layoutDurationMs": 22.506999999999984,
"taskDurationMs": 6253.820999999999,
"heapDeltaBytes": -28707440,
"heapUsedBytes": 181349984,
"domNodes": -6548,
"jsHeapTotalBytes": 13475840,
"scriptDurationMs": 113.92300000000002,
"eventListeners": -7471,
"totalBlockingTimeMs": 413,
"frameDurationMs": 17.776666666666703,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "workflow-execution",
"durationMs": 469.60999999998876,
"styleRecalcs": 19,
"styleRecalcDurationMs": 20.345000000000002,
"layouts": 3,
"layoutDurationMs": 1.2650000000000001,
"taskDurationMs": 117.07399999999998,
"heapDeltaBytes": 5301220,
"heapUsedBytes": 72615344,
"domNodes": 152,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 11.256,
"eventListeners": 99,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "workflow-execution",
"durationMs": 104.66299999995954,
"styleRecalcs": 7,
"styleRecalcDurationMs": 15.449000000000002,
"layouts": 2,
"layoutDurationMs": 0.971,
"taskDurationMs": 79.42600000000002,
"heapDeltaBytes": 3108900,
"heapUsedBytes": 69748652,
"domNodes": 126,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 7.252000000000002,
"eventListeners": 25,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
}
]
} |
Codecov Report✅ All modified and coverable lines are covered by tests. @@ Coverage Diff @@
## nathaniel/agent-v1/02-crdt-foundation #16237 +/- ##
=========================================================================
+ Coverage 80.78% 82.81% +2.02%
=========================================================================
Files 2333 1961 -372
Lines 119358 113055 -6303
Branches 32749 31335 -1414
=========================================================================
- Hits 96425 93627 -2798
+ Misses 22363 19084 -3279
+ Partials 570 344 -226
Flags with carried forward coverage won't be shown. Click here to find out more. 🚀 New features to boost your workflow:
|
There was a problem hiding this comment.
🔍 Cursor Review — Consolidated panel
Triggered by @christian-byrne.
Found 10 finding(s).
| Severity | Count |
|---|---|
| 🟡 Medium | 7 |
| 🟢 Low | 3 |
Panel: 8/8 reviewers contributed findings.
| } | ||
| for (const [id, spec] of snapshot.links) { | ||
| expect(spec.id).toBe(id) | ||
| expect(Number.isFinite(spec.originSlot)).toBe(true) |
There was a problem hiding this comment.
🟡 Medium — These slot assertions are unfalsifiable for this corpus: array elements come only from fc.integer()/fc.string()/null or fc.jsonValue() (finite doubles), and the Infinity/NaN constants appear only as top-level scalars, which readLink drops as non-arrays. That hides the reader's real passthrough — Number(tuple[2]) || 0 leaves ±Infinity unchanged because it is truthy, and Yjs encodes float64, so a host writing [id, 'a', Infinity, 'b', 0] yields originSlot: Infinity. Add non-finite numbers to the array element arbitrary, and tighten the assertion to non-negative safe integers since Number.isFinite alone also admits negative and fractional slots.
Raised by 6 of 8 reviewers (claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case, gpt-5.6-sol-max adversarial, gpt-5.6-sol-max edge-case, kimi-k3-max adversarial, kimi-k3-max edge-case).
| doc.getMap<unknown>(NODES_KEY).set(at(realIds, 0) ?? 'x', hostile) | ||
|
|
||
| const snapshot = readDocSnapshot(doc) | ||
| expect(snapshot.nodes.size).toBeLessThanOrEqual(realIds.length) |
There was a problem hiding this comment.
🟡 Medium — toBeLessThanOrEqual(realIds.length) is satisfied by dropping every node, which is the more dangerous regression than the one it is meant to catch: if a single bad entry made readDocSnapshot return an empty snapshot or abandon the rest of the nodes loop, this passes and the loop below then iterates zero times. The overlay always invalidates exactly one readable entry, so assert the exact surviving set — size realIds.length - 1 (0 when the 'x' fallback fires) and that every other real id is still present.
Raised by 5 of 8 reviewers (gemini-3.1-pro adversarial, claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case, gpt-5.6-sol-max edge-case, kimi-k3-max edge-case).
| expect(spec.id).toBe(id) | ||
| expect(Number.isFinite(spec.originSlot)).toBe(true) | ||
| expect(Number.isFinite(spec.targetSlot)).toBe(true) | ||
| expect(spec.originId).not.toBeNull() |
There was a problem hiding this comment.
🟡 Medium — These two assertions can never fail: toNodeId is String(value), so originId/targetId are always non-null strings for any surviving link. They mask the gap they were written to cover — readLink accepts any endpoint that is not == null, so '', true, {} or [] (all reachable from hostileValueArb) produce NodeIds like '', 'true', '[object Object]', with '' being a value parseNodeId explicitly rejects. Assert the endpoint is a structurally valid node reference (e.g. via parseNodeId) or that the entry was dropped.
Raised by 4 of 8 reviewers (gpt-5.6-sol-max adversarial, claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case, gpt-5.6-sol-max edge-case).
| fc.string({ minLength: 1, maxLength: 6 }), | ||
| fc.array( | ||
| fc.tuple( | ||
| fc.constantFrom('type', 'pos', WIDGETS_KEY, 'flags', 'mode'), |
There was a problem hiding this comment.
🟡 Medium — hostileValueArb only yields plain JS values, never Yjs types, so the reader branches most likely to throw are unreachable from a test whose stated purpose is "never throws": the raw instanceof Y.Array paths in readPos/readLink and the named instanceof Y.Map + toJSON() path in readWidgets. Per this file's own schema note, a real host writes pos, inputs/outputs, flags and widgets as Y types, so those are precisely the reachable shapes. The field-key list also omits OPAQUE_WIDGETS_KEY, leaving the positional-array widget fallback unexercised.
Raised by 4 of 8 reviewers (claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case, gpt-5.6-sol-max adversarial, gpt-5.6-sol-max edge-case).
| } | ||
| } | ||
| const CLASS_TYPES = ['LoadImage', 'KSampler', 'SaveImage'] as const | ||
| const WIDGET_NAMES = ['image', 'seed', 'steps', 'filename_prefix'] as const |
There was a problem hiding this comment.
🟡 Medium — Widget names are drawn only from this fixed safe list, and the hostile corpus never builds a widgets Y.Map, so host-influenced widget names are never fuzzed even though set_widget.widget is wire data. Two concrete defects stay masked: readWidgets accumulates into {}, so a widget named __proto__ reassigns the prototype instead of creating an own key, and diffSnapshots.clearedWidgets tests !(name in next.widgets), which is false for inherited toString/valueOf/constructor, so such a widget is never cleared and keeps a stale canvas value. Building the record with Object.create(null) closes both.
Raised by 3 of 8 reviewers (claude-opus-5-thinking-max adversarial, gpt-5.6-sol-max adversarial, gpt-5.6-sol-max edge-case).
| doc.on('update', () => { | ||
| wrote++ | ||
| }) | ||
| const first = normalize(readDocSnapshot(doc)) |
There was a problem hiding this comment.
🟡 Medium — The purity property performs two reads with no intervening doc mutation and the corpus limits widget values to numbers/strings, so it cannot detect snapshot aliasing. readWidgets returns any non-Y.Map value by reference, including objects/arrays and shared Y.Array/Y.Text, so a later CRDT edit mutates the previously returned snapshot in place and diffSnapshots' Object.is comparison misses the widget change. Add an object-valued widget, then mutate the doc (or the returned snapshot) between reads and assert isolation.
Raised by 2 of 8 reviewers (gpt-5.6-sol-max adversarial, gpt-5.6-sol-max edge-case).
| const host = mint({ nodes: [], links: [] }, CATALOG) | ||
| const updates: Uint8Array[] = [Y.encodeStateAsUpdate(host)] | ||
| host.on('update', (update: Uint8Array) => updates.push(update)) | ||
| for (const op of ops) applyOps(host, [op]) |
There was a problem hiding this comment.
🟡 Medium — applyOps(host, [op]) discards the per-op outcomes and omits the widget catalog that the sibling followerSeam.integration.test.ts passes whenever set_widget is involved. Since roughly a quarter of the minted edits are set_widget, silently rejected ops would leave normalize() comparing always-empty widget records on both sides, so widget projection is never actually exercised — and the vacuity guard tracks only nodes/links/reordering, so it cannot see that collapse. Assert the applier accepted the ops it was given.
Raised by 2 of 8 reviewers (kimi-k3-max adversarial, gpt-5.6-sol-max edge-case).
|
|
||
| const snapshot = readDocSnapshot(doc) | ||
|
|
||
| for (const [id, spec] of snapshot.nodes) { |
There was a problem hiding this comment.
🟢 Low — Unlike the arrival-order property, this one has no vacuity guard despite the same failure mode: a node survives only when the drawn field key happens to be 'type' with a non-empty string value, and a link only when the value is an array of length >= 5 with non-null entries at indices 1 and 3. If a generator or seed change makes those coincidences rare, both contract loops iterate zero times and the property silently degrades to "did not throw". Track surviving node and link counts across the run and assert both are greater than zero.
Raised by 2 of 8 reviewers (claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case).
| import type { DocSnapshot } from './docSchema' | ||
| import { LINKS_KEY, NODES_KEY, WIDGETS_KEY, readDocSnapshot } from './docSchema' | ||
|
|
||
| const FC_OPTIONS = { seed: 0x7b2c0de1, numRuns: 100 } as const |
There was a problem hiding this comment.
🟢 Low — Pinning seed freezes the corpus to 100 fixed cases forever, so the "arbitrary values" ingestion claim is only ever checked against what this one seed produces and never explores new inputs across CI runs; the sibling property tests in this repo (draftCacheV2.property.test.ts, the asset-utils ones) re-randomize per run. It is also fragile with a caret-ranged fast-check: a minor generator change reshuffles the frozen corpus and can shift coverage or flip the vacuity guards with no code change.
Raised by 2 of 8 reviewers (claude-opus-5-thinking-max adversarial, gemini-3.1-pro adversarial).
| }) | ||
|
|
||
| function at<T>(values: readonly T[], index: number): T { | ||
| return values[index % values.length]! |
There was a problem hiding this comment.
🟢 Low — at declares it returns T, but for an empty array 0 % 0 is NaN and values[NaN] is undefined, so the non-null assertion is a lie. This is reachable at line 409: when the corpus deletes every minted node, at(realIds, 0) is undefined and the ?? 'x' fallback writes the hostile value to a fresh key, so the stated "overwrites a live entry" scenario silently never runs. Guard the empty case explicitly and require a live target for that property.
Raised by 3 of 8 reviewers (claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case, kimi-k3-max adversarial).
b047b50 to
107bb96
Compare
|
Closing this dead-base foundation PR as superseded for merge purposes. The property-baseline portion now lives in the main-based port #16353, and the broader agent/follower substrate has been split across active main-based PRs including #16340, #16365, #16364, and #16359. Ledger: christian-byrne/in-app-agent-program reports/review/overlap-dedupe-2026-08-31.md. |
Property-based coverage for the CRDT follower read leg on this slice. Tests only — no product code changes.
docSchema.test.tspins the schema v1 shape with hand-built fixtures. It cannot pin the two things the follower is load-bearing for, so this addsdocSchema.property.test.ts(fast-check, matching the repo's existing*.property.test.tsconvention):add_node/connect/set_widget/delete_node, including resurrect-after-delete and mint-time-recordedremoved_links) is applied host-side through the shared package's ownapplyOps; the incremental Yjs updates are captured and delivered to two followers in two generated permutations, seed state included so causally-pending structs are exercised the way a reconnect/replay actually delivers them. Both followers — and the host itself — must project identical snapshots. A second property checks incremental delivery against one merged update. A reader that folded arrival order into its output (an accumulator, a first-write-wins cache, an iteration-order dependency) passes every example test in this directory today; it would fail this one.logging.js#6). Reading emits no Yjs update and is idempotent across repeated reads.readDocSnapshotis where untrusted host bytes become FE domain objects, and nothing validates entry shapes in front of it (assertReadableSchemagates the schema version, not shapes). Arbitrary JSON values in thenodes/linksroot maps — and a hostile overlay dropped on top of a real minted doc — must never throw, and every surviving entry must satisfy theNodeSpec/LinkSpeccontract the canvas layer assumes (stringtype, finitepos, finite slots). A raw TypeError escaping here kills the follower's observer callback, not just one node.The order property carries a vacuity guard: it asserts the corpus actually populated both root maps and that the two permutations really differed, so an all-rejected corpus cannot make it trivially true.
Verification at this head (
b047b50d1a)vitest run src/workbench/extensions/agent/crdt/vue-tsc --noEmiteslint+oxlint+oxfmt --checkon the new fileknip(pre-push hook)🤖 Generated with Claude Code