CHANGELOG.md

Changelog

All notable changes to dikw-web are documented in this file.

The format is based on Keep a Changelog. Version numbers are standard three-digit SemVer (MAJOR.MINOR.PATCH); package.json.version is the single source of truth. See [0.0.2] below for why the four-digit VERSION file format introduced in [0.0.1.0] was dropped.

[Unreleased]

[0.8.10] - 2026-07-05

Changed

• dikw-core verification pin single-sourced in scripts/live-core/harness.mjs. live-integration.yml no longer sets a DIKW_CORE_VERSION env — CI/nightly falls back to the harness's DEFAULT_CORE_VERSION, the same value local runs use — and bump-dikw-core.yml now edits (and reads the current pin from) only the harness file. This unjams the auto-bump loop discovered while bumping to core 0.6.5 (#150): a bump commit that edited a workflow file was doubly blocked — GitHub rejects such a push from a PAT without the Workflows permission, and the gate-integrity required check flags any .github/workflows/** edit without a maintainer's gate-change label. With the pin out of the workflow file, future auto-bump PRs need neither. The DIKW_CORE_VERSION env override still works everywhere; docs/integration-verification.md no longer hardcodes the pin value (one less thing to drift per bump).

[0.8.9] - 2026-06-29

Added

• Automated watch-CI + fix (dikw-web-watch-ci skill) + loop observability (scripts/loop-log.mjs). Turns delivery-loop step 8 from prose into a bounded, self-logging loop: watch gh pr checks --watch, route a real failure to the fresh-context fixer, rerun the one known flake (graph.spec.ts Pixi) at most once, always pull the review prose before merge, and merge explicitly (never --auto, which outraces the review) — gated by MAX_ROUNDS = 3 and a same-failure circuit breaker. Each transition is appended as a structured JSON line to a gitignored .loop-log.jsonl (node scripts/loop-log.mjs <event> [detail]) so an autonomous/background run is diagnosable after the fact (runaway / stuck / flake / silent death). Wires into dikw-web-delivery-workflow step 9 and CLAUDE.md. See docs/adr/0005-delivery-loop-hardening.md.

[0.8.8] - 2026-06-29

Changed

• dikw-web-verify-frontend gains a measured perf + a11y pass (Step 2.5). The frontend-verify skill previously eyeballed the docs/ui-checklist.md a11y / contrast / perf items. It now uses the already-installed chrome-devtools-mcp: lighthouse_audit for accessibility + best-practices (the tool excludes performance), plus a performance_start_trace/stop_trace for Web Vitals, scored to a rubric — a11y ≥ 0.9 with no new violation, CLS ≤ 0.1 (cross-checking the perf.spec.ts gate), LCP recorded as a soft budget — turning the qualitative items into numbers. Locally-measured, not a new CI gate (Lighthouse + trace timing is runner-dependent). The #graph Pixi route audits a11y normally but skips the background-tab perf trace. See docs/adr/0005-delivery-loop-hardening.md.

[0.8.7] - 2026-06-29

Added

• Reward-hacking gate (gate-integrity CI job / npm run check:gate). A deterministic check the in-loop agent cannot fool, hardening the delivery loop against an agent weakening the verification to make a check go green. scripts/check-gate-integrity.mjs diffs the branch against its merge base and fails the PR if a coverage threshold was lowered, the coverage exclude list grew, a bundle budget was raised, e2e retries were raised, a test was deleted/disabled or stripped of assertions, or the gate/CI machinery itself (the script, .github/workflows/**, fixer.md's forbidden list) was edited. The good direction (raising a threshold, adding tests) is always allowed; a deliberate weakening is allowed only when a maintainer adds the visible gate-change PR label. Until now "don't weaken the tests" was prose only in CLAUDE.md / docs/review-rubric.md / fixer.md — the weakest defense. The job is PR-scoped and a required status check. See docs/adr/0005-delivery-loop-hardening.md.

[0.8.6] - 2026-06-29

Fixed

• Flaky graph e2e (graph.spec.ts > renders a nonblank Pixi graph canvas) — root-caused, not retried away. In React StrictMode's dev double-invoke, the graph mounts two createPixiGraphEngine() inits concurrently. The canvas was attached via mount.replaceChildren(app.canvas) inside the async factory, outside the effect's active guard, so a slower-resolving stale init could clobber the live engine's canvas and then empty the mount when its destroy() fired — leaving data-ready="true" over a canvas-less stage. GraphCanvas now attaches the canvas only inside the active guard (the engine exposes a canvas getter), so the discarded init never touches the DOM. Production rendering is unchanged (no StrictMode double-invoke there → single attach). The spec also gates on data-render-count >= 1 before reading the canvas contract.
• Flaky Mermaid disclosure clicks (wiki.spec.ts, theme.spec.ts). A bare getByText("flowchart") also matched the <style> Mermaid injects (its CSS mentions flowchart), so once any diagram had rendered the locator resolved to two nodes and .click() threw a Playwright strict-mode violation. Both specs now scope to the <summary> element (locator("summary", { hasText: … })).

[0.8.5] - 2026-06-28

Fixed

• Chat agent tool calls fail (fetch failed) in a proxied dev/live setup. dikw-core has no CORS, so the browser keeps serverUrl at the default and routes its /v1 reads through the same-origin Vite proxy (VITE_DIKW_PROXY_TARGET → the real, dynamically-ported core). But the sidecar's /agent core calls run server-side and bypassed that proxy — they dialed the default port the browser sent (http://127.0.0.1:8765), where nothing listens, so every agent tool (retrieve_knowledge, dikw_health, …) returned fetch failed and the chat could not read the knowledge base. The sidecar now mirrors the Vite /v1 proxy: agentSidecarPlugin injects the Vite-resolved VITE_DIKW_PROXY_TARGET (honoring .env.local, not just a shell export, and warning on a malformed value) and applyDevProxyTarget (server/agent/http.ts) routes the default core URL to it. Dev-only by construction — the standalone production sidecar injects no target, so it always dials the configured Server URL verbatim; a custom, directly-reachable serverUrl is also left untouched. Unit guards: server/agent/http.test.ts "applyDevProxyTarget…" + the injected-target handler test; server/agent/vitePlugin.test.ts "resolveDevProxyTarget…".

Changed

• live:verify agent↔core check now reproduces the real browser path. scripts/live-core/verify-agent.mjs previously POSTed /agent with the live core's real dynamic URL, bypassing the proxy the browser depends on — so it passed even while the actual chat was broken. It now sends the default core URL (like the browser) and asserts a core tool reached status:"succeeded" (not merely that one was invoked — a fetch failed call still emits a tool_event). This catches the bug above: it fails against a sidecar that can't reach core and passes once the proxy is mirrored.

[0.8.4] - 2026-06-27

Fixed

• Hero metric value is the documented 27px, not a drifted 30px. A token-migration override silently pinned .metric-card__value to 30px over the 27px base rule, contradicting DESIGN.md §3 (which documents the hero-number as 27px twice). Dropped the override so the documented size ships.

• Three sans-uppercase "eyebrows" now use the mono label voice (Mono-Only-Uppercase). .wiki-info-grid dt (every sibling field label was already mono), .wiki-preview__header, and MB-Web's .mb-lab bilingual column header (the twin of the workbench's mono .bi-colhead span) were sans uppercase — the exact AI "eyebrow" tell the type system forbids. Routed onto the mono label role at the shared 0.04em tracking.

• Three mono labels declared a font-weight: 600 no-op. IBM Plex Mono ships only 400/500, so 600 silently rendered as 500 and diverged from the canonical label weight: .bi-colhead span, .import-step__marker, .trace-invocation__dur now declare 500 (no visual change — they already rendered at 500).

• User Markdown table headers render verbatim. .markdown-table-wrap th no longer force-uppercases the author's header row (a content-fidelity breach — a header like pH/mRNA was mangled — and a sans-eyebrow tell). The row stays distinct by weight + surface, not case. The label role's "table heads" remain the workbench's own data-table column labels (mono uppercase); DESIGN.md §3 now spells out the distinction.

• .wisdom-dialog__header h2 rides the title role. It was an untokenized font-size: 17px that inherited the UA bold 700; now sans 600 with the title tracking, byte-for-byte parity with its sibling .wisdom-card h2.

• .markdown-chart__caption is on the role ladder (body-sm 13px), not an off-scale 0.85em.

  e2e guards: typography.spec.ts "type-system rules hold on sweep-unreachable surfaces (workbench)" + "MB-Web column labels use the mono uppercase voice" — these inject the auth/hover/MB/state-gated surfaces the passive invariant sweeps can't render, so the rules are now guarded there too.

[0.8.3] - 2026-06-27

Fixed

• AI-translate dual columns fill the reader pane. The bilingual view's root inherits .markdown-body's single-column max-width: 72ch, so the source / translation columns crushed into one measure on the left and left the pane's right half empty. .bilingual-cols now overrides that cap — it fills the reader width (each column gets its own measure) and, past two measures on an ultra-wide pane, caps and centers. Below 1100px, where the pairs collapse to one stacked column, the cap reverts to the single-column 72ch so a stacked paragraph never runs past a comfortable line length. e2e guards: bilingual.spec.ts "dual-column view fills the reader pane…" + "…restores the single-column measure when columns stack…".
• DESIGN.md typography frontmatter matched to the implementation. The machine-readable typography: block still named the pre-v0.8.0 families (Source Serif 4 / Inter Tight / JetBrains Mono) and carried swapped weights (display 500, label 600) — contradicting the shipped IBM Plex superfamily and the §3 prose table (display 600, label 500). Frontmatter families, weights, and the stray §3 "JetBrains Mono" mention now read IBM Plex.

Changed

• MB-Web radii and colors consolidated onto the shared tokens. src/mb/mb.css kept a parallel light + dark palette whose values already duplicated the workbench tokens. Differently-named locals (--ac / --acb / --border / …) now alias their styles.css equivalent and same-named ones (--bg / --surface / --text / …) inherit the global values, so the parallel dark block is gone; the off-scale radii snapped to the 4 / 6 / 7 / 8 / 999 scale (cards 8px; the badge / chip / filter / toast / sync pills fully round); the note highlighter yellows became named --mb-mark / --mb-mark-staged tokens. Only the few genuinely MB-specific literals with no shared equivalent (answer-blue, on-accent foreground, AA-darkened faint, hover wash) stay local. Zero-visual-change for colors; verified in the browser across library, reader, and notes in light + dark.

[0.8.2] - 2026-06-27

Fixed

• Settings appearance copy no longer mixes languages. The zh-CN 外观 subtitle read "System 会跟随操作系统偏好" — the English enum label System inside Chinese chrome, breaking the single-language-chrome rule. Now "「跟随系统」会随操作系统的明暗 偏好自动切换。".
• Editorial / responsive heading sizes snapped onto the role ladder. The tail the 0.8.1 small-text pass left — .wiki-preview-card h2 (19 → title 17), .import-done-banner__headline (20 → title 17), .wisdom-popover__title h2 (15.5 → body 15), and the responsive .metric-card__value downscales (24 / 26 → 22, the reader-h2 editorial step) — is now on the ladder.

Changed

• MB-Web type migrated onto the role tokens. src/mb/mb.css dropped its self-contained half-step scale (11.5 / 12.5 / 13.5 / 14.5 …); 59 of its 60 font-size declarations now resolve to --type-* (label / body-sm / body / title) by nearest role, made available by the app-wide styles.css import. The one kept literal is .mb-r-title (22px), the paper-reader title. MB-Web radii/colors remain a tracked follow-up (DESIGN.md §3). Verified in the browser across the library, reader, and notes views.

[0.8.1] - 2026-06-27

Fixed

• Small-text ladder closed (post-v0.8.0 scale consolidation). A residual 12px tier crammed between label (11) and body-sm (13), plus stray 14 / 16 sizes beside body (15) / title (17), survived the v0.8.0 pass. All 46 12px sites now snap to body-sm (dense metadata — IDs, paths, durations, pills, table cells, hints, segmented controls) or label (the uppercase label voice). The off-scale outliers resolve too: .page-header__description (14 → body-sm), .settings-panel__header h2 (16 → title, matching every other panel title), the import section-head titles + .section-title (14 → body), .answer-text (16 → body), and the reader markdown table header/cell inversion (th 12 / td 14 → both body-sm). The small-text band (11–17) on every workbench route now lands on the {11, 13, 15, 17} run — no 12px tier, no stray 14 / 16. A few editorial / responsive heading sizes outside that band (preview-card 19, done-banner 20, wisdom-popover 15.5, responsive metric 24 / 26) are a pre-existing tail and a tracked follow-up (DESIGN.md §3).
• Inline / block code (0.92em) and KaTeX math are documented as deliberate em-relative off-ladder exemptions; a new no UI text renders off the role scale e2e invariant guards the ladder across overview / base / import / chat / settings / tasks / wisdom.

Notes

• MB-Web (src/mb/mb.css) keeps its own half-step subsystem scale for now; migrating it onto the --type-* role tokens is a tracked follow-up (DESIGN.md §3).

[0.8.0] - 2026-06-27

Changed

• Typeface system → IBM Plex superfamily. The three voices are now IBM Plex Sans (UI), IBM Plex Serif (reading), and IBM Plex Mono (machine), replacing Inter Tight / Source Serif 4 / JetBrains Mono. One superfamily means the voices share a single design DNA — coordination by role, not by foundry — escaping the ubiquitous Inter default. CJK still falls through to the system fonts; fonts load from Google Fonts (external, outside the bundle budget). mb.css (MB-Web) tracks the same swap.

Fixed

• Serif sprawl reined in (DESIGN.md §3 made honest). The serif had crept onto metric values, every detail H2, and explorer headers across ~8 ad-hoc sizes while §3 still claimed it was "reserved for the reader." Serif is now used in exactly three places — the document reader prose, the assistant reply prose, and the two stat-number tiers (27px hero, 18px stat). Every panel / card / detail heading reverts to the sans title role (17px). §3 documents the real policy + the editorial serif sub-scale (32 / 22 / 17 / 27 / 18).
• Off-scale chrome + the chat sub-scale snapped onto the six roles. The topbar breadcrumb (12.5px) and connection chip (11.5px) → body-sm; panel titles (14px) → title (17px); the assistant/agent surface (body 14, headings 19/16.5/14.5, code 12.5, blockquote 13.5, role 10.5) folds onto body / serif title-body-bodysm / body-sm mono / 11px label — no private sub-scale. The ~15 distinct font sizes collapse toward the role set.
• Named-rule violations cleared. trace-pane__title, wiki-backlinks headers, import-eyebrow, table captions, and the agent role chip were sans tracked-uppercase (the AI "eyebrow" tell) — all moved to the IBM Plex Mono label voice. The reader header path now renders mono (Mono-for-paths). Every mono label shares one tracking (0.04em), retiring the 0.02 / 0.03 / 0.08 / 0.1 spread.
• Two-Weight Rule restored on #trace. Markdown strong/b is pinned to weight 600; the UA default font-weight: bolder is relative and was compounding to 900 (a fifth weight, past IBM Plex's Bold ceiling) inside already-bold content.
• 11px low-vision floor enforced. The 10.5px mono labels (agent role chip, graph legend, lint fix label, …) move up to 11px; nothing renders smaller. New tests/e2e/typography.spec.ts invariants guard the floor, the Mono-Only-Uppercase rule, the three Plex voices, and the panel-title token. Surfaced by the /impeccable critique typography pass.

[0.7.13] - 2026-06-27

Fixed

• Typography: the cascade base is now the body type role (15px), not the 16px UA default. body had no font-size, so every unsized element — and every font: inherit control (buttons, inputs, selects), the Overview/Settings panel <dd> values, and empty-state titles — silently rendered at the browser's 16px default, bypassing the six-role scale (which puts body at 15px). Setting body { font-size: var(--type-body-size) } pulls all of it onto the scale. UA-default headings sized by 1.5em shift proportionally (~6%; e.g. the task-detail H2 24→22.5px, toward the 22px reader-H2 role) — a deliberate, benign consequence. A new tests/e2e/typography.spec.ts guards the base against regression. Surfaced by the /impeccable critique typography pass.

[0.7.12] - 2026-06-26

Added

• Overview: loading skeletons, a freshness stamp, and clickable metrics. While the first /v1/health|status|info load is in flight the metric strip and the Runtime/Providers panels now render skeleton placeholders (locking the layout height so data swaps in without a pop-in) instead of misleading zeros; the header carries an Updated HH:MM / 更新于 HH:MM freshness stamp; and the Data, Knowledge, and Wisdom metrics link to #base, #graph, and #wisdom respectively (a quiet stretched link — no button chrome). The Refresh button spins and marks aria-busy while loading but stays clickable, so a stalled request can still be retried.

Changed

• Overview metric strip is now a semantic <dl>. Each metric renders as a <dt> label + <dd> value, so a screen reader announces term/value pairs instead of a flat text stream (matching the Runtime/Providers panels). A failed initial load now shows only the error notice rather than a half-rendered strip.

[0.7.11] - 2026-06-26

Changed

• Theming: migrated the genuine remaining chip/graph literals to tokens. The audit's "manual dark-patch block" finding is mostly a false alarm — most of that block is legitimate dark theming (light --surface → dark --surface-2), not redundant literal re-patching. The actual literal re-patches — .frontmatter-chip (#dfe8e5/#f8fbfa/#34423e), .wisdom-chip (#eef3f1/#34423e), and .graph-detail h2/__section strong (#111817/#2d3b37) — now use --surface-2/--line/--text, so they resolve to exactly their old dark-override values (dark mode unchanged, browser- verified) and those four selectors were dropped from the override block. Light chips render as clean recessed warm-stone pills instead of a near-invisible cool near-white.

Fixed

• a11y: MB-Web view switcher. The 研究/笔记 switcher declared role="tablist"
  • role="tab"/aria-selected but had no tabpanels or roving-tabindex keyboard model. Downgraded to role="group" + aria-pressed (matching SegmentedControl and the mobile panel tabs) so the announced semantics match the actual behavior.

[0.7.10] - 2026-06-25

Fixed

• Accessibility round 2 (from a hardened, anti-hallucination /impeccable audit). With the earlier rounds' fixes confirmed in place, the audit surfaced genuinely new items:
  • Label-in-Name (WCAG 2.5.3). Removed hardcoded English aria-labels that overrode the localized visible label on inputs already inside a <label> / Field (Graph "Hide orphans", Settings Server URL / Token), and localized the two search inputs (Base directory, Graph) via a new searchLabel copy key — so the accessible name now matches the visible/zh-CN label instead of announcing English.
  • Touch targets. Extended the @media (pointer: coarse) 44px block to the chat composer send/stop buttons, Base directory-tree rows, and Wisdom editor controls (were 26–28px on touch).
  • MB-Web --faint. Darkened the light-mode token from #95968f (~2.7:1) to #767771 (~4.5:1) so the metadata text it colors clears AA.

[0.7.9] - 2026-06-25

Fixed

• Accessibility & contrast pass (from an /impeccable audit). Resolved the genuine findings of a 5-dimension UI audit:
  • ARIA state on selection controls. SegmentedControl is now a named role="group" with aria-pressed per option (was className-only, so screen readers couldn't tell which theme/locale was active); the active option is a distinct raised chip vs. a faint hover tint. aria-current marks the active sidebar route and the selected master-detail row in Tasks / Trace / Chat. Notice exposes role="alert" (errors) / role="status" (info) so post-action banners are announced. The chat composer label is sr-only (was display:none, leaving the textarea unnamed). MB-Web note-editor and popover textareas regained a real focus ring (WCAG 2.4.7).
  • Dark-mode contrast leaks → tokens. Hard-coded graphite/amber literals with no dark override (.wiki-preview-card/__header, .graph-force-control, .citation-item/.event-tape__item/.result-table body text, .graph-toggle, .nav-item svg, the #b9d9d3 selected-row border, and the .summary-metric--warn/.file-error-card warn cluster) now use theme-tracking tokens (--text/--muted/--accent-border-hover/ color-mix(--amber …)), so dark mode no longer renders dark-on-dark text or glaring near-white cards. Status-pill ok/info text mixes the signal color toward --text to clear 4.5:1 on its tint; Trace timestamps moved off the sub-AA --subtle.
  • Touch targets. @media (pointer: coarse) raises segmented / reader-tab / graph-zoom controls to a 44px tap target (WCAG 2.5.5); desktop is unchanged.
  • MB-Web mobile reflow. Below 900px the library and Q&A panels were display:none'd with no fallback, so a phone-opened #MB-Web link could neither pick a paper nor ask a question. A tab bar (论文库 / 阅读 / 知识问答) now switches which single panel fills the width (WCAG 1.4.10).

[0.7.8] - 2026-06-25

Changed

• Badge + segmented-control components (Phase 2 of docs/ui-refactor-plan.md). Added SoftLabel, FrontmatterChip, and a generic SegmentedControl (<T extends string>) under src/components/ and migrated their call-sites: the .soft-label count/layer/anchor badges across Base, Graph, Wisdom, and Tasks; the .frontmatter-chip tag/source chips in the Markdown reader and the Base info panel; and the Settings theme/locale toggles. The wrappers are DOM-identical (same element, classes, and passthrough attributes), so the e2e console gate and layout are unchanged. StatusPill stays separate and the three CSS classes are not merged; three block uses of .soft-label on a <p> (hints / empty state) stay raw markup by design.
• Scattered color literals → design tokens. Migrated the six hand-picked hex literals to the DESIGN.md Intent-Mapping ramp: the nav / segmented / secondary+icon hover borders (#cbd8d4 / #c5dad5 / #9fb0aa) → --accent-border-hover; the nav resting text (#3c4a46) → --muted; the nav strong label (#1d2926) and the panel + section titles (#25322f) → --text. Collapsed the dead duplicate .page-header h1 rule (the survivor now consumes --type-title-page-*) and dropped three now-redundant dark-mode color overrides whose light base already tracks the theme. The only intended visual change is the nav resting label rendering one step lighter (#3c4a46 → --muted); active markers, fills, and hover/active states are unchanged. Verified route-by-route in light + dark against a live dikw-core.

[0.7.7] - 2026-06-25

Changed

• Shared control components (Phase 1 of docs/ui-refactor-plan.md). Added Button, IconButton, and Field (plus a cx class-joiner) under src/components/ and migrated every hand-rolled .primary-button / .secondary-button / .icon-button / .field call-site across 14 page files to them. The wrappers are DOM-identical (same element, classes, and attributes), so behavior, the e2e console gate, and layout are unchanged — the change is −104 net lines of duplicated markup. Button defaults type="button" (no accidental form submits) and exposes variant="primary|secondary|danger"; IconButton requires a label (→ aria-label) so every icon button keeps an accessible name; Field maps grow/small/inline/token to the existing field--* modifiers. Segmented controls are deferred to a later phase.

Added

• Design-system token foundation. src/styles.css gains additive token groups, all documented in DESIGN.md: a 4px spacing grid (--space-*, with 6/10/14px half-steps), a four-step motion scale (--dur-fast|state|popover| overlay + --ease-standard|emphasis), six typography roles (--type-<role>-{size,lh,ls} across the serif/sans/mono voices), and color-mix alpha layers (--line-alpha, --overlay-hover|active, --accent-border-hover) for borders/hover fills that track the theme over any surface. They are defined now and consumed incrementally — the additions alone change no rendering.
• docs/ui-refactor-plan.md — the phased UI/UX refactor roadmap (token foundation → shared controls → badges/segmented → panels/MB-Web), learned from Vercel's Geist system while keeping the warm-stone + petrol identity.

Fixed

• Destructive control text was invisible. --danger was referenced by .secondary-button--danger and the session ... menu's delete item but never defined in src/styles.css, so the text silently fell back to the inherited body color. It is now var(--red) (defined once, tracks the theme through --red), so destructive affordances actually read red in both light and dark.

Changed

• DESIGN.md restructured to a Geist-informed section order — Intent Mapping table, six Type Roles, a Spacing & Rhythm section, a Motion section, a formal radius scale, per-component spec tables, and a new bilingual Voice & Content section — reconciled to the real shipped values (7px control radius, sans page titles, borderless pills) and kept on the "Quiet Instrument" identity. docs/ui-system.md and docs/ui-checklist.md updated to match.

[0.7.5] - 2026-06-25

Changed

• Chat: Stop is shown only while a reply streams. The idle composer no longer renders a disabled, glyph-less Stop control next to Send (it read as an empty placeholder square). Stop now appears only while a reply is streaming — an honest surface with no dead control.
• Tasks: removed card-in-card nesting in the detail panel. The summary tiles (created / duration / digest) and the Result metric tiles were bordered boxes nested inside the bordered detail panel. They are now borderless recessed tonal wells, and the result-summary grouping is seamless on the panel surface; the semantic --warn (amber) and --error (red) borders are kept explicit. Verified in light and dark against a live core.

[0.7.4] - 2026-06-24

Added

• Chat: send on Enter. The composer now sends on Enter and inserts a newline on Shift+Enter, matching the dominant chat convention (previously only the Send button submitted). The handler skips while an IME is composing (nativeEvent.isComposing) so confirming a Chinese candidate with Enter never sends mid-composition.

[0.7.3] - 2026-06-24

Fixed

• Retrieve layout collisions. Page Refs rendered the page title and layer label with no separator (Knowledgeknowledge); the layer is now a muted mono tag in a flex row. In the Chunks table a long unbreakable path overflowed its grid track and collided with the Seq column; the path cell now shrinks and wraps (min-width:0; overflow-wrap:anywhere).
• Reader line length. Prose ran edge-to-edge on wide panes; .markdown-body is now capped to 72ch (DESIGN.md body 65–75ch). Narrower surfaces (chat replies, bilingual columns) already sit under this width, so it's a no-op there.
• Wisdom empty state. An empty wisdom base showed "No wisdom pages match" (implying a filter miss); it now distinguishes the true zero-state ("No wisdom pages yet" + a hint to use New or run synth) from a filtered no-match.
• MB-Web identity drift. MB-Web shipped its own green palette (forest-green accent, green-tinted neutrals); its mb.css neutrals, accent, and danger are realigned to the workbench warm-stone + petrol tokens (One Indicator Rule), keeping the answer-blue semantic.
• MB annotation side-stripe. .mb-anno used the banned 3px accent border-left callout pattern; reworked to a full accent-tinted border.
• Copy polish. Graph stats read "{n} links" (was singular); the Retrieve "No page refs" empty state gained a teaching detail line.

[0.7.2] - 2026-06-24

Changed

• Single-source keyboard focus from the accent token (src/styles.css). Two new tokens — --focus-ring (a soft 3px box-shadow ring for text fields) and --focus-outline (a crisp 2px solid outline for other controls) — both derived from --accent via color-mix, so the focus indicator now tracks light/dark automatically. A zero-specificity :where(button, a[href], [role="button"], summary, [tabindex]):focus-visible rule gives buttons, links, and other interactive controls a consistent branded keyboard-focus ring (WCAG 2.4.7), while existing per-element focus styles (graph nodes, reader controls, resizer) still win. Fixes a dark-mode defect where text inputs rendered a hardcoded light-petrol focus ring (rgba(11,124,112,…)) that was nearly invisible on the dark surface. See docs/ui-system.md.

Added

• Impeccable design baseline + UI audit backlog. PRODUCT.md (strategic: register, users, brand personality, anti-references, principles) and DESIGN.md (visual: tokens, typography, components, do's/don'ts, in Stitch format) capture the design system, with .impeccable/design.json as the machine-readable sidecar. docs/ui-audit-backlog.md records findings from a full live-core interaction walkthrough (light + dark) for follow-up.

[0.7.1] - 2026-06-24

Added

• Auto-track new dikw-core releases (.github/workflows/bump-dikw-core.yml). Weekly (and on demand) the workflow resolves dikw-core's latest GitHub release and, if it's newer than the pinned DIKW_CORE_VERSION, opens a chore/bump-dikw-core-<version> PR editing both pin sites (live-integration.yml + scripts/live-core/harness.mjs) and labels it live-integration so the full real-core verification runs on the PR. The pin stays explicit + reviewed (reproducible nightly); upgrades are surfaced automatically instead of silently followed. Needs a DIKW_BUMP_TOKEN fine-grained PAT (contents + pull-requests write) — the default GITHUB_TOKEN can't open a CI-triggering PR; without the secret the job no-ops with a warning. See docs/integration-verification.md.

[0.7.0] - 2026-06-23

Added

• Live integration verification (npm run live:verify). End-to-end verification of the working tree against a real dikw-core (the published GHCR image, e.g. ghcr.io/opendikw/dikw-core:0.6.1) backed by Postgres/pgvector — the gap the rest of the stack can't cover (Playwright mocks /v1; smoke:core needs a core you seeded yourself). One command boots core + Postgres on dynamic ports under a unique compose project (so multiple core versions coexist), seeds it through dikw-web's own write pipeline (import → ingest → synth → lint, reusing buildImportBundle + DikwClient via tsx so the wire shape can't drift), then verifies the read contract (smoke-core), the browser read routes (a new opt-in Playwright live project on real data with the console gate), and the agent↔core path (a real chat turn must invoke a core-backed tool, read off the agent's observable event stream). New: docker-compose.live-core.yml, scripts/live-core/*, scripts/seed-core.mts, tests/fixtures/live-base/, tests/e2e/live/, .env.core.example, and a non-gating .github/workflows/live-integration.yml (dispatch / nightly / live-integration label). Sub-commands: live:up / live:seed / live:smoke / live:down. The default verify gate is untouched. See docs/integration-verification.md.

Changed

• README polish. Added a centered header (logo, tagline, status + tech badges) and a compact section nav; pointed the dikw-core link at the public GitHub repo instead of a sibling-checkout relative path.
• Corrected "read-only" doc drift against the code. dikw-web is no longer read-only over dikw-core: ImportPage and MB-Web upload papers (/v1/import → ingest → synth), the Wisdom editor saves/favorites (POST /v1/base/wisdom), the Tasks page fires maintenance ops, and the agent proposes maintenance. Dropped the inaccurate whole-app "read-only" framing from the GitHub About, README.md, CLAUDE.md, CONTEXT.md, docs/core-contract.md, and docs/ui-system.md. Genuinely read-only contexts (the Graph view, the reader's Read/Frontmatter tabs, the top-bar connection strip) keep the wording.
• Genericized the white-label branding example. public/config.example.json, the README branding snippet, and the branding tests now use neutral placeholders (Acme-DIKW / 示例知识库) instead of a real downstream brand name. No runtime behavior change — the default brand stays OpenDIKW and config.json remains gitignored per-deployment.

[0.6.1] - 2026-06-21

Security

• Clear the OpenTelemetry moderate advisories (12 of them). Bumped the experimental OTel deps we import directly (@opentelemetry/api-logs, exporter-trace-otlp-http, sdk-logs) to ^0.219.0, and added overrides forcing the stable OTel packages (@opentelemetry/core, resources, sdk-metrics, sdk-trace-base, sdk-trace-node) to ^2.8.0 tree-wide so the copies pulled transitively by @google/adk are also patched — clearing the @opentelemetry/core <2.8.0 W3C Baggage unbounded-memory advisory and every advisory chained to it. Also overrode dompurify to ^3.4.11 (the IN_PLACE/SAFE_FOR_TEMPLATES XSS fixes) for the copy pulled by mermaid. The remaining moderate advisories (uuid / googleapis / protobufjs / @google-cloud/*) are all transitive via @google/adk's unused GCP exporter path and have no clean fix (npm's only "fix" is a @google/adk major downgrade to 0.1.3) — left as accepted risk. None are reachable in dikw-web's runtime and none gate the prod high+ audit.
• Drop npm from the production image. The runtime stage only ever runs node dist-server/standalone.mjs, so npm is removed (rm -rf of /usr/local/lib/node_modules/npm + the npm/npx bins). This clears the Trivy HIGH for the npm-bundled undici < 6.27.0 (CVE-2026-12151, WebSocket fragment-count DoS) — which npm i -g npm@latest would not fix, since the latest npm still bundles undici 6.26.0 — and keeps future npm-bundled-dep CVEs out of the image entirely.

[0.6.0] - 2026-06-17

Added

• Retry a failed PDF/Office conversion in place. A failed conversion row in the import converting stage now offers a Retry button beside Skip that re-runs convertSource for just that file — no re-selecting or re-uploading the whole batch. A single transient failure (e.g. a mineru quota/5xx hiccup that fails every file at once) can now be recovered per row.

Changed

• Conversion no longer auto-advances on partial success. Previously, once any file converted the batch jumped straight to the bundle preview and silently dropped the still-failed files. Now a still-failed row keeps the converting stage open (with Retry/Skip) until it is resolved; the batch advances to bundling only once every row is converted or skipped with at least one success. Skip now drops only that one row.

[0.5.1] - 2026-06-16

Changed

• MB-Web appearance unified with the workbench. MB-Web no longer keeps a standalone theme preference (dikw-mb.theme); it now reads the same dikw-web.theme key the workbench Settings appearance panel writes, resolving a stored system preference to light/dark via the shared resolveTheme (lifted into src/i18n.ts) and following live OS prefers-color-scheme changes while it stays system (parity with the workbench). MB-Web keeps its one-tap header light/dark toggle, but the toggle now commits an explicit light/dark to that shared key — so a flip in MB-Web is reflected in the workbench and vice versa, and switching between #MB-Web and the workbench no longer flips the theme between two independently stored values. Opening MB-Web (or the OS flipping while it's open) never overwrites a stored system preference (it applies the resolved value to the DOM but only persists on an explicit toggle).

Removed

• MB-Web's standalone dikw-mb.theme localStorage key and its private theme initializer — superseded by the shared dikw-web.theme read above.

[0.5.0] - 2026-06-15

Changed

• Connection settings unified into the workbench Settings page. The Server URL
  • Token now commit on an explicit Save (replacing the previous keystroke-by- keystroke write), with an inline "unsaved changes" hint and a "Saved" confirmation beside the button; the existing clear-connection button is renamed Clear and now resets to the default URL + empty token immediately. The connection persists to localStorage instead of per-tab sessionStorage, so a saved connection is shared across tabs and survives a browser restart — the bearer token is therefore at rest in localStorage by default. The shared connection keys + default URL are single-sourced in src/config/connection.ts.
• MB-Web reads the connection instead of owning a settings panel. Its header gear and the actionable library-load error CTA both navigate to the workbench #settings page; a cold-opened #MB-Web link recovers the connection from the shared localStorage without a per-tab handoff (the #97 cold-open fix, now via shared storage rather than an opt-in mirror).
• Page titles use the UI sans family. .page-header h1 no longer renders in the editorial serif (which fell back to a serif CJK face, clashing with the sans body); page chrome now reads as one cohesive type family. Reader / card headings and big metric numbers keep the editorial serif.

Removed

• The MB-Web connection panel (MbConnectionPanel) and its opt-in dikw-mb.rememberConn localStorage mirror — superseded by the unified Settings page + always-on localStorage persistence above.

[0.4.1] - 2026-06-14

Fixed

• MB-Web could not reach dikw-core on a cold open (#97). The #MB-Web variant read the core Server URL + Token only from per-tab sessionStorage and shipped no settings UI of its own, so a shareable #MB-Web link opened in a fresh tab (or after a browser restart) silently fell back to http://127.0.0.1:8765 + an empty token and showed a generic "论文库加载失败" with no way to recover. MB-Web now:
  • has its own connection settings panel (Server URL + Token), reachable from a header gear and from the library error CTA — so a cold-opened link can be configured/recovered without bouncing through the workbench Settings page;
  • can opt in to "记住连接", mirroring the connection to localStorage (dikw-mb.rememberConn) so it survives a tab close / browser restart; off by default, so the bearer token only lands at rest when the user explicitly enables it;
  • replaces the generic library error with an actionable one that distinguishes a 401/403 auth failure from an unreachable core, each with a button into the panel; and
  • reports an honest 未连接 · dikw-core header state instead of a stuck "连接中…".

[0.4.0] - 2026-06-14

Added

• 论文知识库 (MB-Web) reading variant — a focused paper-reading UI shipped in the same bundle. src/main.tsx now renders a top-level Root that mounts the new MbApp (src/mb/) for the #MB-Web hash and the original multi-page workbench (App) for every other hash (#chat / #base / …), so both ship together (shareable link http://127.0.0.1:4321/#MB-Web); App only rewrites its own legacy query→chat hash, so the two never fight over the URL. The variant is read-only over dikw-core and reuses the existing /v1, /agent, and /web/translate + /web/mineru infra; its UI lives under src/mb/ with .mb--scoped styles (no new framework), and App + pages/ are untouched. Features: a three-column resizable + collapsible workspace; a Source-layer paper library with browser-side rename aliases; a reader with 原文 / 中英对照 (paragraph-aligned bilingual via /web/translate, persisted per paper); per-paper agent Q&A (chat history kept per paper); 我的笔记 (localStorage card list + detail page, with a delete-confirmation dialog) mirrored to the Wisdom layer (POST /v1/base/wisdom); and upload (PDF/Office → MinerU → import → ingest → synth) with optional pre-translation at import.
• Observability demo stack + docs — Phase 7 (final) of OTel observability. A new docker-compose.observability.yml (separate from the production docker-compose.yml, which it does not touch) brings up an OTel Collector that fans OTLP into Jaeger (traces) + Prometheus (metrics) + Loki (logs), all visualized in Grafana with datasources pre-provisioned. otel-collector-config.yaml wires the three pipelines (OTLP in → Jaeger / Prometheus-scrape / Loki, CORS-enabled for the browser RUM exporter); observability/prometheus.yml + observability/grafana-datasources.yml make the stack work on first up. Default up is backends-only (no LLM creds needed); --profile app also builds + runs the sidecar pointed at the in-network collector (creds from the operator's .env, never inlined — placeholder secrets only, gitleaks-clean). New docs/observability.md documents the full OTEL_* env reference (incl. OTEL_EXPORTER_OTLP_HEADERS for Grafana Cloud / Honeycomb / Datadog), the env-only sampling limitation, the dikw.* metric catalog, the browser-RUM config.json block + its credential-exposure security model, and the privacy posture; .env.example, README.md, and CLAUDE.md cross-reference it. No app behavior change.
• Browser RUM (opt-in frontend traces) — Phase 6 of OTel observability. A new src/telemetry/initBrowserOtel.ts boots the OpenTelemetry web SDK (WebTracerProvider + OTLPTraceExporter, with document-load + fetch instrumentations) so the React app emits RUM spans under service.name=dikw-web-browser. The fetch instrumentation propagates a W3C traceparent to same-origin requests — always the sidecar's /agent + /web, stitching a browser span to the sidecar's Phase 2 SERVER span; a cross-origin core (/v1 in the standalone deployment) gets no header and no CORS preflight, and the exporter's own collector POSTs are excluded via ignoreUrls. Fetch-span URLs are redacted before export — the query string is dropped (it can carry user-derived values such as originalFilename and inputSha) and high-cardinality id path segments are templated (/web/mineru/jobs/:id, /v1/base/pages/:id), mirroring the sidecar's privacy-minimizing SERVER-span posture (serverRoute); this runs in an onEnding span processor so every fetch span (success, error, abort) is covered, not just the default url.full/http.url raw export. The user-interaction instrumentation is deliberately not registered: it only wraps listeners added after it loads, but React attaches its delegated listeners at render — before the async, lazy SDK init resolves — so it would miss them, and the fix (loading zone.js, which globally patches Promise/setTimeout/requestAnimationFrame, before React + Pixi mount) is an unjustified risk for click-span coverage; the default StackContextManager is enough for fetch propagation. It is opt-in and default-off: a new src/config/telemetry.ts reads a telemetry.endpoint from the runtime public/config.json (mirroring loadBranding()); with no endpoint configured the init is a no-op and — because the SDK is loaded via dynamic import() — none of it is downloaded. The OTel web SDK therefore stays out of the entry bundle (it lands in a lazy chunk); scripts/check-bundle.mjs's total JS ceiling was raised 1900 → 1950 KB to count that opt-in chunk while restoring the pre-RUM headroom (the entry budget is unchanged). RUM init is best-effort — any failure is swallowed so it can never break the app or add console noise (the e2e console gate is unaffected). public/config.example.json ships the telemetry block with an empty endpoint (disabled) so copying it for branding never silently enables telemetry. New browser runtime deps: @opentelemetry/sdk-trace-web, @opentelemetry/exporter-trace-otlp-http, and @opentelemetry/instrumentation-{fetch,document-load} (@opentelemetry/api stays single-deduped at 1.9.x). RUM does not touch dikw-core; browser logs/metrics and user-interaction spans are noted follow-ups (RUM convention is traces first).
• OpenTelemetry outbound CLIENT spans (opt-in) — Phase 5 of OTel observability. A new server/agent/instrumentation.ts registers @opentelemetry/instrumentation-undici so the standalone sidecar's outbound fetch/undici calls (dikw-core /v1, MinerU, Tavily/Jina, the translator's Anthropic SDK) emit OTel CLIENT spans and inject a W3C traceparent, continuing the inbound SERVER-span trace through to the upstream. Registration is gated on an OTLP traces endpoint (OTEL_EXPORTER_OTLP_ENDPOINT / *_TRACES_ENDPOINT): with no endpoint env the HTTP layer is not patched at all, so outbound behavior is byte-identical to before. The patch is via undici's diagnostics_channel (dispatcher-agnostic). The query string is redacted from every CLIENT span (startSpanHook keeps scheme://host/path, sets url.query to [REDACTED]) so MinerU's presigned upload/result URLs — whose query carries a bearer credential — never reach the OTLP backend. CLIENT spans (like SERVER spans) are filtered out of the in-memory #trace store (DikwSpanProcessor now keeps only SpanKind.INTERNAL, ADK's own kind), so the #trace waterfall stays agent-only; the spans still export over OTLP. @opentelemetry/instrumentation@^0.219.0
  • @opentelemetry/instrumentation-undici@^0.29.0 are added as runtime deps (@opentelemetry/api stays single-deduped). Note: a fetch using a per-call dispatcher from the npm undici package (the tools.ts HTTPS_PROXY ProxyAgent path) is incompatible with Node's built-in fetch independently of this change (a pre-existing undici-version mismatch); the instrumentation is orthogonal to it.
• Structured sidecar logging + OTel logs bridge — Phase 4 of OTel observability. A new server/shared/logger.ts (createLogger(scope)) replaces the sidecar's ~21 ad-hoc console.* calls with one-line structured records: JSON to stdout (a human text line when stdout is a TTY or DIKW_LOG_FORMAT=text), the active span's trace_id / span_id injected for trace↔log correlation, and a mirrored OTel LogRecord. When OTEL_EXPORTER_OTLP_ENDPOINT (or OTEL_EXPORTER_OTLP_LOGS_ENDPOINT) is set, ADK's maybeSetOtelProviders builds the LoggerProvider (its own OTLP BatchLogRecordProcessor), so logs ship to Loki/etc. alongside traces and metrics under the same service.name=dikw-web resource; with no endpoint env there is no provider and the bridge is a no-op (stdout output only). Field names matching key|token|auth|secret|password|credential are redacted to [redacted] as defense in depth, and there is no arbitrary-object dump path, so a secret can't be logged by accident; Error field values are reduced to name: message (no stack dump). This is a visible stdout format change for npm start — set DIKW_LOG_FORMAT=text for the prior human-readable style. @opentelemetry/api-logs is a new runtime dep (the emit API); @opentelemetry/sdk-logs is added for the test harness (both pinned to ADK's resolved 0.205.0). Logging never touches dikw-core.
• OpenTelemetry metrics — Phase 3 of OTel observability. A new server/shared/metrics.ts records six instruments on the global meter: http.server.request.duration (histogram, seconds; http.request.method / http.route / http.response.status_code, recorded as each SERVER span ends), dikw.job.duration + dikw.job.count + dikw.job.inflight ({dikw.job.family = mineru|translate, dikw.job.outcome = succeeded|failed}, recorded at the detached conversion/translation runner boundaries — inflight is an UpDownCounter toggled +1/-1 across the run), dikw.llm.tokens ({gen_ai.token.type = input|output}, from the token counts DikwSpanProcessor already parses), and dikw.agent.turn.duration ({dikw.agent.turn.outcome = ok|aborted|error}, around one runMessage). Instruments bind lazily on first record, so they pick up the MeterProvider ADK's maybeSetOtelProviders installs when OTEL_EXPORTER_OTLP_ENDPOINT (or OTEL_EXPORTER_OTLP_METRICS_ENDPOINT) is set — metrics then export over OTLP via ADK's PeriodicExportingMetricReader, alongside traces, sharing the service.name=dikw-web resource. With no endpoint env there is no MeterProvider and every record is a no-op (no behavior change). @opentelemetry/sdk-metrics is added for the test harness only (pinned to ADK's resolved 2.7.1); the sidecar never imports an exporter. Metrics never touch dikw-core.
• OpenTelemetry inbound SERVER spans — Phase 2 of OTel observability. Every /agent/* and /web/* request is wrapped in a route-templated SERVER span (server/shared/withServerSpan.ts): it extracts inbound W3C trace context (so a browser → sidecar trace stitches through), runs the handler inside that span's context (the ADK invocation / future outbound spans nest under it), and ends on response finish/close with the http.request.method / http.route / http.response.status_code semantic-convention attributes. Session and job ids collapse to :id (and proposal ids to :proposalId) so span names / http.route stay low-cardinality; the raw request path is deliberately not exported as url.path (it would re-introduce those ids into span data — http.route is the aggregation key). SERVER spans export via OTLP (when an endpoint is configured) but are filtered out of the in-memory #trace store (DikwSpanProcessor skips SpanKind.SERVER; ADK agent spans are all INTERNAL), so the #trace waterfall is unchanged. Because the SERVER span now parents ADK's root invocation span, SpanStore.getSessionTraces normalizes any parent absent from the served view (the filtered SERVER span, or a FIFO-evicted parent) back to null, so the invocation stays a true root and the TraceSpanView contract holds. No new dependencies; no change to the #trace DTO or any /agent/* wire format.
• OpenTelemetry trace export (opt-in) — the first phase of OTel observability. Agent spans now carry a service.name=dikw-web resource (server/agent/telemetryResource.ts; service.version from package.json, a per-process service.instance.id, overridable via the standard OTEL_SERVICE_NAME). Setting OTEL_EXPORTER_OTLP_ENDPOINT (or OTEL_EXPORTER_OTLP_TRACES_ENDPOINT) makes ADK additionally export these spans over OTLP to any collector/backend (Jaeger, Grafana Tempo, Honeycomb, …), with sampling honoring OTEL_TRACES_SAMPLER/_ARG. With no endpoint env the spans stay in the in-memory #trace store only — behavior-identical to before — and OTLP-exported spans inherit the same content-free redaction (ADK_CAPTURE_MESSAGE_CONTENT_IN_SPANS=false). @opentelemetry/resources + @opentelemetry/semantic-conventions are promoted from already-hoisted transitive deps to direct deps (pinned to ADK's resolved versions; no new physical install). No change to the in-memory SpanStore, the #trace DTO, or any /agent/* wire format.
• Automated release tags — a release job in CI (.github/workflows/ci.yml) cuts a GitHub Release tagged dikw-web-v<package.json version> after the full CI gate is green on a push to main. It is idempotent (skips when the tag already exists), so only an actual package.json.version bump produces a new tag — which makes "every version update gets a tag" fall out for free; the current 0.3.0 is tagged on the first run. Release notes are pulled from the matching CHANGELOG section by scripts/changelog-notes.mjs. Job-scoped contents: write keeps the workflow-level read-only default everywhere else. No new dependencies.
• Sidecar translation endpoint (/web/translate, server/web/) — the backend half of Base-reader bilingual reading. A job + poll API (mirroring mineru, issue #60): POST /web/translate/submit ({ blocks, targetLang }, the document's markdown split into text blocks) → 202 { jobId }, then poll GET /web/translate/jobs/{id} and fetch block-aligned { blocks: [{ i, tr }] } from …/result (…/cancel aborts). The text blocks are translated in ordered batches (splitIntoBatches), one streaming @anthropic-ai/sdk call per batch over MiniMax (messages.stream(...).finalMessage(); deliberately not the ADK MiniMaxLlm adapter, which is bound to ADK's BaseLlm interface). After each batch the runner publishes progress: { done, total, blocks } to the poll status, so the browser reveals paragraphs progressively instead of waiting for the whole document; per-batch timing is logged. The model wire protocol is a delimiter — blocks joined by a distinctive sentinel line, reply split back on it — not a JSON array: arbitrary scientific Markdown (LaTeX backslash commands like \circ, unescaped quotes around code identifiers like "scikit-learn", brackets) round-trips verbatim instead of corrupting a JSON array (both modes were live-observed failing the JSON protocol on a real paper). A wrong block count from the model is reconciled by splitting the batch and re-translating the halves down to singletons (not a re-ask — the miscount is often deterministic); two model failure modes a matching count hides are self-healed by repairBlocks (Chinese target, one re-ask per block): a block echoed back untranslated (≥ 6 English words, no CJK), or one whose translation contains its source verbatim (bilingual echo) or is > 2× the source length above a 60-char floor (EN→中 compresses, so this flags hallucinated/continued content — live-observed on test2.md). The flagged block is re-asked alone and the re-ask re-validated: a result still oversized/echoed is rejected for the source text rather than injecting bloat. The system prompt also forbids adding/continuing/summarizing/inventing content. The SDK's own maxRetries is 0 so retry policy is single-sourced in TranslatorClient.translate, which retries retryable transport faults (429 / timeout / 5xx / connection) and an empty reply with exponential backoff + jitter; a truncated (max_tokens) reply and aborts are not retried. Wikilink targets are server-side re-pinned by order (repinWikilinks) so a translated [[target|label]] keeps its destination even if the model rewrites it. Submit enforces a 4 MB / 2000-block cap. Reuses the chat agent's MiniMax credentials (DIKW_AGENT_API_KEY / _BASE_URL / _MODEL, no dedicated key); only the per-call output cap DIKW_WEB_TRANSLATOR_MAX_TOKENS stays translator-specific; missing key → 503 translate_disabled and GET /web/translate/health → { enabled: false } (the reader hides its AI 翻译 entry). Browser client src/utils/translate.ts adds an IndexedDB cache (dikw-translate-cache, 7-day TTL sweep) keyed by a TRANSLATE_VERSION that is bumped when translation logic changes so stale pre-fix translations are dropped. The reader UI lands in a follow-up. No new dependencies.
• Base reader bilingual reading (#base) — the reader half of the feature above. English pages (detected after fetch via CJK ratio) gain an AI translate toggle fused onto the Read tab (aria-pressed, hairline seam) when the sidecar translator is configured; Chinese pages and a disabled translator show no toggle. Toggling on splits the body into blocks, sends the text blocks to /web/translate, and renders a paragraph-aligned dual-column view (BilingualView): source markdown left, Chinese right, with special blocks (code / tables / charts / $$ math / rules / standalone image lines) rendered once and centered, never translated — a figure is no longer duplicated and alt-translated across both columns, even when a hard line break joins it to its caption (the MinerU figure shape). Translated paragraphs stream in progressively as the sidecar's batches land (each pair shows a skeleton only until its own translation arrives), rather than appearing all at once. A status bar exposes translating / cached / re-translate / cancel; narrow screens stack the columns. The shared markdown renderer was extracted from MarkdownView into markdown-runtime (mono and dual-column render through the exact same markdown-it instance, sanitizer, and chart / mermaid / image hydration); the translated column renders through a prefixed heading-slug env so its ids can't collide with the source column's. Clicking a wikilink in the translated column previews the target page with its title + summary translated to Chinese (usePreviewTranslation — two tiny blocks through the same /web/translate + IndexedDB cache — with an AI badge on the card and a silent fallback to the original on failure); a source-column or mono-view click previews the original, and an already-Chinese target is never re-translated. No new dependencies; see docs/adr/0003-bilingual-reading.md.
• ESLint gate (npm run lint → eslint.config.js, flat config, --max-warnings 0, in verify + CI). Covers the lint layer tsc --strict misses: React hook dependency/order rules, unused symbols (tsconfig has no noUnusedLocals), and no raw console in the shipped browser bundle. react-hooks is pinned to its two classic rules (rules-of-hooks + exhaustive-deps) rather than the v7 recommended preset, whose opinionated rules (set-state-in-effect, immutability, …) would force a non-surgical refactor of working code; type-checked rules are omitted to keep it fast. Pre-existing violations were resolved — dead types/imports removed, two useless assignments and an over-escaped regex fixed, and genuinely intentional hook-dep omissions annotated with a reason.
• Prettier (npm run format / format:check, in verify + CI) across code (.ts/.tsx/.js/.mjs/.css/.json); a one-time repo-wide reformat landed in its own commit (printWidth 100 to match the codebase). Markdown is excluded (.prettierignore) — prettier's prose pass is cosmetic churn (e.g. *italic* → _italic_) with no correctness value on the hand-maintained docs.
• Layout-shift (CLS) perf budget (tests/e2e/perf.spec.ts, rides test:e2e). Asserts Cumulative Layout Shift ≤ 0.1 on the primary routes (overview / base / graph / tasks / chat / wisdom) — the one Core Web Vital stable under headless Chromium. LCP and long-task totals are measured and surfaced as Playwright annotations but never gate (runner-dependent timing), the same reasoning that keeps smoke:core out of CI. 2026-06 baselines: CLS 0.00–0.05 across all six routes.
• Bundle-size budget gate (npm run check:bundle → scripts/check-bundle.mjs, wired into CI after the verify gate). Asserts gzipped ceilings for the entry JS (index-*.js), total JS, and CSS against dist/, catching the "a heavy lib crept into the bundle" regression class that typecheck/tests can't see. Budgets are inline with deliberate headroom over the 2026-06 baseline (entry 243 / total 1695 / CSS 25 KB gzip) and raised consciously like the coverage thresholds. The CSS ceiling doubles as a backstop for the "no UI framework" rule. No new deps.
• Review rubric (docs/review-rubric.md) — the project-specific lens for the pre-merge independent review (Delivery Loop step 6): simplicity-first / surgical / TDD, the repo traps generic reviewers miss (server/** .js imports, no UI framework, token never exposed, single-language chrome), and the don't-touch list. Wired into dikw-web-delivery-workflow and CLAUDE.md.
• Live-core contract smoke (npm run smoke:core → scripts/smoke-core.mjs, driven by the new dikw-web-smoke-core skill). The e2e suite mocks /v1 entirely, so it can't see real contract drift; this script asserts the consumed subset of docs/core-contract.md (the wiki → knowledge layer value, the /v1/tasks envelope, PageReadResult.frontmatter, and the health/status/graph shapes) against a reachable core and exits non-zero on drift. Intentionally outside CI (needs a live core); Node fetch is proxy-immune for localhost. Validated green against a live dikw-core 0.5.1. No new dependencies.
• E2E console gate. Every Playwright spec now imports test/expect from a new tests/e2e/harness.ts instead of @playwright/test; the harness fails any test that emits a console.error or an uncaught pageerror, turning "the console stayed clean" from a manual review step into a deterministic gate. Resource-load 404s (the suite intentionally 404s config.json and a missing asset) and AbortError are allowlisted; a test that deliberately drives an error path opts out with test.use({ consoleGuard: false }).
• Project verification skills (.claude/skills/, now tracked in git via a .gitignore exception while local settings stay ignored): dikw-web-verify-frontend encodes the per-route browser-verification pass (Delivery Loop step 5) with the project's hard-won gotchas (graph canvas needs --headed, not Chrome MCP; requestIdleCallback paths only hit the timeout in an MCP tab), and dikw-web-delivery-workflow makes the Delivery Loop an executable orchestration of the existing skills.
• docs/ui-checklist.md — a pass/fail rubric for the qualitative UI rules (single-language chrome, dark reader contrast, small radii, no UI framework, graph filters/legend/no-bloom, the markdown HTML allow-list), run by the dikw-web-verify-frontend skill. Test/tooling/docs only — no runtime, wire, or contract change, so no version bump.
• App sidebar collapses to an icon rail (src/App.tsx). A footer toggle (PanelLeftClose / PanelLeftOpen) shrinks the left sidebar to an icon-only rail; nav labels are visually hidden but kept for assistive tech (native tooltips restore the name on hover), and the choice persists in localStorage (dikw-web.sidebarCollapsed). The rail is desktop-only (gated above the 900px breakpoint); on narrow screens the sidebar keeps its existing behavior.
• Base directory panel is resizable (#base). The directory-tree column can be dragged wider/narrower via a splitter in the layout gap (pointer-capture, so the drag tracks past the handle) or stepped with the arrow keys on a focusable role="separator" (Home/End jump to the bounds), clamped 240–640px and persisted (dikw-web.wikiSidebarWidth), so long filenames fit. Hidden on narrow screens.

Changed

• Bilingual reader keeps an in-flight translation on toggle-off — shared useBilingualReader.toggle() no longer aborts the running translation when switching back to the source view; it lets the job finish and write the IndexedDB cache, so toggling 原文 ↔ 中英对照 (in both the Base reader and the new MB reader) no longer restarts the translation from scratch every click. BilingualPaper (MB reader) is now React.memo'd so a parent re-render doesn't rebuild the dual-column DOM and drop an active text selection / staged highlight.
• Extracted the dotenv helpers shared by both sidecar config loaders (parseEnv / readEnvFile / readOptional) into a new server/shared/env.ts; server/agent/config.ts and server/web/config.ts now import them instead of each carrying a byte-equivalent private copy. Behavior-neutral internal refactor — no env var, wire format, or loadAgentConfig / loadWebConfig contract changed, so no version bump.
• Base reader "AI translate" control is now an on/off switch (role="switch"
  • aria-checked with a sliding knob) rather than a plain aria-pressed button — a clearer affordance for the EN→中 dual-column toggle. Behavior unchanged.
• Base reader tab rename: "Info" → "Frontmatter" (zh-CN 信息 → 元信息), matching what the tab shows (PageReadResult.frontmatter). The en source-tab label stays "Source"; only the zh-CN label changed (原 "源码" → "原文").
• Base reader raw-source panel restyled. The source tab dropped the harsh black-on-white block for the reader token surface (--reader-*, dark-mode safe) and now word-wraps (white-space: pre-wrap; overflow-wrap: break-word), so long lines no longer force a horizontal scrollbar.
• Source import now normalizes filenames and frontmatter (docs/adr/0004). At import, every source page filename is rewritten to a Unicode-kebab name (Han preserved, whole name incl .md < 32 code points; src/utils/kebab-source-name.ts) and a flat provenance frontmatter is injected/merged — original_filename (the true name, always) plus converter: mineru for converted files. title is not injected (a body H1 stays the document title via core's resolution); no original_sha256 or timestamps reach the frontmatter, so the bytes stay deterministic for core's package_sha256 dedup (normalizeForImport + src/utils/frontmatter-merge.ts, which never clobbers author keys). The MinerU sidecar frontmatter changed from the nested source: { converter, original_filename, original_sha256 } block (which the reader rendered as raw JSON) to flat keys, and the <stem>-<sha12>/ converted-page directory is now <kebab-stem>/<kebab-stem>.md. MinerU uploads under the kebab stem + the original extension (kept for format detection); two inputs whose stems kebab to the same root surface a visible duplicate_path skip, and distinct plain names that collapse to the same kebab get a -2/-3 suffix. Removed the now-unused src/utils/shorten-filename.ts.

Removed

• Tasks toolbar "Task running" status pill. The green-dot live indicator next to the maintenance buttons (Ingest / Synth / Lint Propose / Lint Apply) was visually abrupt and redundant — the buttons already disable while any task is running/pending, which is the real safety gate and is unchanged. Only the pill (and its now-orphaned actions.running copy + .task-actions__live style) was removed; the shared .live-dot class stays (RetrievePage still uses it).

[0.3.0] - 2026-06-03

Changed: rename the sidecar dotenv file and the MinerU key (BREAKING)

• .env.agent.local → .env.local. The file feeds both the chat agent (server/agent/config.ts) and the /web/* browser helpers (server/web/config.ts), so the agent-specific name no longer fit. Both loaders now read .env.local; .env.agent.example is renamed to .env.example. Rename your local file: mv .env.agent.local .env.local.
• MinerUAPIKey (and its DIKW_AGENT_MINERU_API_KEY alias) → DIKW_WEB_MINERU_API_KEY. The mineru conversion key is a /web/* concern, so it now follows the standard DIKW_WEB_* prefix, parallel to the DIKW_AGENT_* keys. This is a hard rename with no backward-compatible fallback — update the variable name in your .env.local. A missing key still degrades ImportPage to .md/.pdf only; the 503 mineru_disabled message now names the new variable.
• .env.example documents DIKW_WEB_MINERU_API_KEY (previously the template never listed the mineru key at all), and docker-compose.yml now forwards it into the container alongside the optional Tavily / Jina keys, so the documented deployment env actually reaches /web/mineru/*.

[0.2.0] - 2026-06-03

Added: chat context-window compaction at >50% of the model window

• Long chats now compact their history instead of growing the prompt unbounded. AdkAgentRunner attaches ADK's built-in TokenBasedContextCompactor + LlmSummarizer to the LlmAgent (factory in server/agent/contextCompactor.ts, reusing the agent's own MiniMaxLlm for summarization). When the threshold is crossed, ADK summarizes the oldest events into a persisted CompactedEvent and ContentRequestProcessor rebuilds the prompt as [summary, ...recent raw events], so the prompt actually shrinks; the summary persists to sqlite and carries across turns.
• On by default, env-tunable (.env.agent.local): DIKW_AGENT_COMPACTION_ENABLED (default true), DIKW_AGENT_CONTEXT_WINDOW (default 1048576, the MiniMax-M3 window), DIKW_AGENT_COMPACTION_RATIO (default 0.5), and DIKW_AGENT_COMPACTION_RETENTION (default 8). The trigger is round(contextWindow * ratio) = 524,288 tokens at the defaults. Because ADK's shouldCompact sums per-event prompt-token counts, effective compaction fires somewhat before the live prompt literally reaches the ratio.
• The summary never leaks into the chat UI. AdkSessionStore.projectMessages filters isCompactedEvent so the [Previous Context Summary] event is not rendered as an assistant bubble, and mapAdkEvent emits no live wire event for it. A summarization failure is swallowed (logged) and the turn proceeds with the un-compacted history. See docs/agent.md#context-compaction.

[0.1.1] - 2026-06-03

Added: 7-day TTL cleanup for the browser MinerU convert cache

• The IndexedDB convert cache (dikw-mineru-cache) now prunes entries older than 7 days. Each cached conversion already recorded a cachedAt; on every ImportPage mount, IDBConvertCache.sweepExpired (src/utils/mineru-convert.ts) walks the store with a readwrite cursor and deletes any entry whose cachedAt is more than CACHE_TTL_MS (7 days) old, plus legacy/corrupt rows with a missing or non-finite cachedAt. The TTL is absolute — a cache hit does not refresh cachedAt. The sweep is deferred to requestIdleCallback (with a setTimeout fallback) so its readwrite transaction never queues ahead of the import flow's foreground cache reads, and runs only at the single point the cache is opened, so no ImportPage / wire-format changes were needed.

[0.1.0] - 2026-06-03

Changed: chat agent runtime migrated from Pi Agent to Google ADK

• The chat sidecar's agent runtime moved from @earendil-works/pi-agent-core / @earendil-works/pi-ai to Google ADK (@google/adk). The same-origin /agent/* HTTP API and the AgentStreamEvent NDJSON wire format are unchanged, so the browser / chat UI is unaffected. AdkAgentRunner (server/agent/adkRunner.ts) drives one turn on ADK's Runner and maps each ADK Event → AgentStreamEvent (mapAdkEvent); the DIKW tool set is now ADK FunctionTools (server/agent/adkTools.ts) reusing the existing core / web tool clients.
• The LLM is MiniMax via its Anthropic-compatible endpoint through a custom MiniMaxLlm extends BaseLlm adapter (server/agent/minimaxLlm.ts, using @anthropic-ai/sdk as transport), and the model is upgraded MiniMax-M2.7 → MiniMax-M3 (.env.agent.example). MiniMax thinking content blocks are dropped — only text / tool_use cross the boundary.
• Sessions now persist to local SQLite (.agent-sessions/agent.sqlite, appName dikw-web, userId demo) via ADK's DatabaseSessionService instead of one JSON file per session. AdkSessionStore (server/agent/adkSessionStore.ts) projects ADK events into the existing AgentSession DTO at read time, mirroring title / createdAt / message-count / proposal-status into session.state. Old .agent-sessions/*.json files are not migrated (local demo data; expect a one-time reset).

Added: hidden #trace page (OpenTelemetry span waterfall)

• New hidden route #trace (src/pages/TracePage.tsx) — reachable by URL only, intentionally absent from the sidebar nav (hiddenViewIds in src/App.tsx). It shows a per-session conversation alongside an OpenTelemetry span waterfall served at GET /agent/sessions/{id}/traces.
• Spans are captured by a custom DikwSpanProcessor (registered once per process via ADK's maybeSetOtelProviders) into an in-memory, bounded SpanStore (server/agent/spanStore.ts), which re-assembles flat span rows into the per-session SessionTraceView. Spans are ephemeral — lost on a sidecar restart by design; only the conversation content is persisted (sqlite).

Changed: Docker runtime ships a production node_modules

• build:server now bundles with esbuild --packages=external — ADK + MikroORM + native sqlite3 can't be bundled (dynamic driver import() + native addons), so dist-server/standalone.mjs imports its dependencies from node_modules at runtime.
• The Docker image is rebuilt around node:24-slim (Debian glibc, for reliable sqlite3 N-API prebuilts) with a dedicated prod-deps stage (npm ci --omit=dev) whose pruned node_modules ships into the runtime image alongside package.json. The HEALTHCHECK probes via node -e fetch(...) since node:24-slim ships neither wget nor curl.
• package.json gained overrides (node-gyp, tar) to clear HIGH npm-audit CVEs and an explicit undici dependency.

[0.0.21] - 2026-06-01

Fixed: MinerU conversion survives a request-timeout proxy (job + poll)

• A slow MinerU conversion no longer dies behind a reverse-proxy / tunnel request timeout (#60). The converting pre-stage used to run the whole MinerU pipeline (submit → upload → poll-until-done → download → tar+gzip) inside a single POST /web/mineru/convert request and wrote no response bytes until it finished, so its time-to-first-byte equalled the full conversion time. Behind a proxy that caps request duration (Cloudflare free ~100s, nginx proxy_read_timeout 60s) any conversion slower than that limit had its connection cut mid-flight — the browser saw a transport-level Failed to fetch and the job was aborted, so larger / scanned PDFs (the slow vlm path) failed while small files in the same batch succeeded.
• POST /web/mineru/convert now returns 202 { jobId } immediately and runs the conversion detached in an in-memory JobStore (server/web/jobStore.ts). The browser polls the short GET /web/mineru/jobs/<id> for status and fetches the tar.gz from GET /web/mineru/jobs/<id>/result on completion (served idempotently within the job's TTL, so a transfer cut mid-flight by a flaky proxy is retry-safe rather than a lost conversion), with POST /web/mineru/jobs/<id>/cancel to abort — mirroring the task model already used for ingest / synth / lint. Every request is now seconds-short, so none approaches a proxy timeout, and a failure surfaces as a structured mineru_* code (e.g. mineru_timeout) instead of an opaque Failed to fetch.
• The change is encapsulated in convertSource (src/utils/mineru-convert.ts); the per-file Import UI gains a polling substage while the detached job runs. converting stays non-resumable across a page reload (the in-memory job store is also dropped on a sidecar restart) — unchanged from before.

[0.0.20] - 2026-05-31

Added: configurable logo text + brand-tracking tab title

• The sidebar logo text and browser tab title are now configurable at runtime. They default to OpenDIKW but can be overridden — without rebuilding — by a config.json served from the static root ({ "brand": { "name": { "en": …, "zh-CN": … } } }), fetched once during app bootstrap. A missing or malformed file falls back to the defaults, so existing deployments are unaffected. name is per-locale and the tab title follows the resolved brand name; the logo image and favicon stay fixed. See public/config.example.json and the README "Branding" section.
• The top-bar breadcrumb root is now a fixed Workbench / 工作台 label instead of the hard-coded brand name, decoupling the "workbench home" crumb from branding.
• The agent system prompt is brand-neutral ("a helpful knowledge base agent") so white-labeled deployments don't surface the OpenDIKW name in chat.

[0.0.19] - 2026-05-30

Fixed: import survives transient gateway errors; Wisdom metric aligns

• Import no longer aborts on a transient gateway/network blip during task follow (#56). When dikw-core sits behind a reverse proxy / tunnel (Cloudflare, nginx, Caddy) that occasionally drops or 5xx's a long-lived connection, a single failed poll during a multi-minute stage (typically synth with a slow LLM) used to mark the whole import "failed" even though the task succeeded server-side. DikwClient.streamTaskEvents now silently reconnects on an upstream 5xx (502/503/504) or a network-level fetch error, resuming from the unchanged from_seq cursor with capped exponential backoff (1s→15s, up to 8 retries); cancellation and non-transient 4xx errors still propagate at once. If retries are exhausted, ImportPage's consumeTask extends the existing getTaskFinalEvent reconciliation to a thrown poll, so a stage that finished during the outage lands done instead of a spurious failure. The same resilience benefits the Tasks and Wisdom follow paths.
• Wisdom Overview metric is vertically aligned. The Wisdom card had no caption row, so its number sat ~23px below the others in the metric strip. It now carries a "wisdom items" caption like every other card, restoring a shared baseline.

[0.0.18] - 2026-05-30

Changed: Import upload is file-only, with auto-filtering and shorter MinerU filenames

• Directory upload removed. The Import picker no longer offers a "Choose folder" button, and dropped folders are ignored with a hint — only individual files are accepted (the file input still takes multiple files at once).
• Unsupported formats are filtered at selection. Files whose extension can't be imported are dropped as soon as they're picked or dropped and reported in a short notice, instead of cluttering the bundle preview's skipped column (which now surfaces only content-level issues like empty bodies).
• Long MinerU filenames are shortened before conversion. MinerU errors on very long names, so mineru-bound files are uploaded under a ≤25-char stem (Unicode preserved, extension kept, bytes unchanged → dedup unaffected). The true original filename is forwarded via a new originalFilename query on /web/mineru/convert so the converted page's frontmatter original_filename stays complete.

[0.0.17] - 2026-05-30

Fixed: three UI / path-consistency issues

• Sidebar Settings stays reachable. The left sidebar is now pinned to the viewport (position: sticky; height: 100vh; overflow-y: auto) so its Settings footer no longer scrolls out of view when a page's main content is taller than the window. The mobile (≤900px) horizontal sidebar is unchanged.
• Chat normalizes legacy wiki/ sources. Sessions persisted before dikw-core's 0.4.0 wiki/ → knowledge/ rename still carry wiki/ source paths and a wiki layer; the right-rail Sources list now displays both as knowledge/ / knowledge (paths via a new normalizeKnowledgePath helper) and dedups a page recorded under both prefixes into a single row. Live core already returns knowledge/, so new sessions are unaffected, and a frozen assistant message body is left untouched.
• Tasks Op filter drops the dead distill hint. The Op input placeholder no longer suggests distill (removed from core) and gains a <datalist> of the current ops — ingest, synth, lint.propose, lint.apply — so they can be searched/selected; free-text entry (e.g. wisdom.write) still works.

Removed

• Dropped the dead distill maintenance action from the agent: the proposal tool no longer offers it and /v1/distill (now 404 on core, superseded by /v1/lint/propose) is no longer routed.

[0.0.16] - 2026-05-29

Feature: fire maintenance ops from the Tasks page toolbar

• The #tasks filter-bar gains Ingest / Synth / Lint Propose / Lint Apply buttons to the right of the Status / Op controls. Ingest, Synth, and Lint Propose start their respective core tasks with default params; Lint Apply runs against the currently-selected succeeded lint.propose task and applies all proposals (pick:null, no review gate) — it is disabled unless such a task is selected.
• After firing, the page refreshes the list, selects the new task, and follows its live event stream (reusing the existing follow() machinery).
• A short background poll of /v1/tasks (running, then pending) detects when core is busy — authoritatively, regardless of the active Status/Op filter — and disables the four fire buttons while any such task exists. The gate starts closed on mount and opens only once the first probe confirms core is idle, so it can't be bypassed during the initial network window. A "Task running" indicator marks a real reason (a detected task or an in-flight submit). The detail-panel Stop button changes from a client-only event-stream detach to a real POST /v1/tasks/{id}/cancel, then re-probes the gate authoritatively — the fire buttons re-enable only when no running/pending task remains (a queued task keeps the gate closed). Follow / Load events still only stream events. A running task hidden by an active Status/Op filter can't be selected to Stop until the filter is cleared — it otherwise releases the gate when it finishes on its own.
• TasksPage becomes a second write surface alongside #import (docs updated in CLAUDE.md and docs/core-contract.md).

[0.0.15] - 2026-05-29

Fix: live progress during mineru PDF / Office conversion

• The converting stage no longer looks frozen. Three issues fixed: (1) the <Loader2 className="spin"> icon never animated because .spin had no rule in styles.css (the spin keyframe was only wired to chat-page classes); (2) the single fetch covering upload + the multi-minute mineru server conversion + download was labelled "uploading to mineru" the whole time, while the intended polling / "waiting on mineru" substage was never emitted; (3) no elapsed timer or progress bar, so a long wait showed no motion.
• ConversionProgress now fixes .spin (reuses pr4-spin), shows a per-row live elapsed timer (reuses formatElapsed + a 1s tick like PipelineSteps), shows an indeterminate progress bar on active rows (reuses @keyframes import-indet), relabels the in-flight wait "Converting on mineru…" with a "can take a minute or two" hint, respects prefers-reduced-motion, and styles the rows into a proper panel. ConversionFileState gains an in-memory startedAt. The mineru conversion is a black box (no progress stream), so progress is honestly indeterminate rather than a fake percentage.

[0.0.14] - 2026-05-29

Fix: import pipeline misreporting a succeeded task as failed

• consumeTask now reconciles against the authoritative task row when the event stream drains without a type:"final" event. streamTaskEvents stops polling when task_status goes terminal + has_more:false, but the import pipeline read success from a captured final event — two signals that can disagree within a single /events response (live task_status vs. a lagging events[] tail). The race surfaced as "Import failed / ingest failed" while the Tasks page (which reads GET /v1/tasks/{id}) showed the very same task succeeded. New DikwClient.getTaskFinalEvent reads that authoritative row and synthesizes the terminal verdict, so a succeeded ingest/synth/lint task is no longer misreported as a failure. Mirrors WisdomPage's existing drain-then-read-result pattern.

[0.0.13] - 2026-05-29

Adapt to dikw-core 0.4.0: K-layer wiki → knowledge

• Contract rename across src/types.ts and every consumer to match dikw-core 0.4.0 (refactor!: rename K-layer "wiki" to "knowledge"): Layer is now "source" | "knowledge" | "wisdom"; InfoResponse.wiki_root → base_root; LayerCounts.wiki_pages → knowledge_pages; StorageCounts.last_wiki_log_ts → last_knowledge_log_ts; ApplyReport.wiki_paths_changed → knowledge_paths_changed. No dual-name compatibility — core rebuilds on incompatibility and keeps no shims.
• Dead fields dropped that 0.3.0/0.4.0 stopped returning: StorageCounts.wisdom_by_status (wisdom is a first-class document layer now, with no candidate status) and LlmInfo.max_tokens_query / max_tokens_distill. Overview's stale "N candidates" detail on the Wisdom card is removed with it.
• PageReadResult.frontmatter (new in 0.4.0, server-parsed YAML) is surfaced read-only in the Base reader's Info tab, replacing the prior client-side body parse as the metadata source.
• Base page now shows only source + knowledge; wisdom is reachable exclusively via the dedicated #wisdom page.
• The #wiki route is gone — the Base page lives at #base (matching the "Base" / "知识库" label and the /v1/base/* endpoint family). Legacy #wiki no longer resolves and falls back to #overview (no redirect). The in-page [[wikilink]] syntax is unchanged.
• The sidecar agent list_pages tool's layer enum is knowledge (not wiki).

[0.0.12] - 2026-05-28

Wisdom: real dikw-core wiring (lists, reads, async writes, resume)

• #wisdom is now backed by the live core HTTP API. List comes from GET /v1/base/pages?layer=wisdom&active=true (via useAsyncResource), body + frontmatter from GET /v1/base/pages/{path}, backlinks from GET /v1/base/pages/{path}/links?direction=in. Saves go through the async POST /v1/base/wisdom task: the page POSTs the submit, then drains client.streamTaskEvents(taskId) until terminal and unwraps the WisdomWriteReport via client.getTaskResult. The mock data module src/pages/__mock__/wisdom-data.ts is deleted along with all of its imports.
• Pending draft create flow: the New dialog now constructs a client-only __pending__/{slug} draft (no path in core yet). It appears in the tree marked (unsaved), lives in Edit mode, and the first Save is what POSTs to core. Empty bodies are rejected client- side before the request goes out (min_length=1 is enforced by core).
• Favorite optimistic + rollback: ☆ flips the chip immediately and POSTs the full page with no_embed=true (core lacks a dedicated PATCH/status endpoint). The optimistic update rolls back if the task fails. preStarStatus is kept on a frontend useRef map — not persisted, refresh resets it (acceptable for now).
• Resume on mount: the in-flight write task_id is stored in sessionStorage["dikw-web.wisdomWrite"] (per-core scoped). On mount, if a non-terminal task is found, the page re-enters the saving state and continues polling. Lives in src/state/wisdom-write.ts next to the existing import-pipeline.ts.
• K / D candidates for the Add wikilink / Add source picker are lazy-loaded from GET /v1/base/pages?layer=knowledge and ?layer=source when the popover opens. Client-side filter keeps the search-as-you-type UX; core has no search endpoint yet — that's the next iteration's lever if D-layer page counts get large.
• Tests: wisdom.test.tsx rewritten to mock client.get / post / streamTaskEvents / getTaskResult, covering 14 scenarios (list, detail fetch, pending draft, save round-trip, picker dedup, sources attach, favorite optimistic + rollback, three dirty-edit confirm paths, resume from sessionStorage, empty-body rejection). src/test/mockClient.ts gains a getTaskResult vi.fn stub. tests/e2e/mockApi.ts returns [] for layer=wisdom/knowledge/source so chrome e2e specs still pass.
• Out of scope (next iteration): real status dropdown (draft → published / archived), delete / rename / move, K/D remote search, Playwright spec specifically for the Wisdom write round-trip.

[0.0.11] - 2026-05-28

Wisdom: mock-driven page with three-pane layout, wikilink picker, inline backlinks

• #wisdom is now a fully interactive mock page driven by hardcoded fixtures in src/pages/__mock__/wisdom-data.ts — no /v1/wisdom* requests at all (the endpoint was retired). Three-pane layout: left directory tree (folder-first, alphabetical), middle read/edit tabs, right rail with Linked references + Sources (sources only shown in Edit; Read merges backlinks inline). Reuses the WikiPage buildWikiTree pattern and MarkdownView for the Read tab.
• Read tab runs injectInlineRefs from src/utils/source-inline-refs.ts to splice [[title|literal]] wikilinks into the body at first occurrence of any backlink title; unmatched backlinks fall into a bottom "Linked references" panel — same behavior as source-layer pages. Wikilink clicks resolve against the wisdom multimap; W→W resolves silently, K→K shows a 2.4s amber toast ("[[title]] isn't a wisdom page in this mock.") via a new .wisdom-toast block in styles.css.
• Edit tab toolbar: "Add wikilink" popover merges K + W candidates (deduped by path, per-item layer chips for k/w/d) into one picker matching the "Add source" chip column; cursor-position-aware [[title]] insertion. New dialog enforces kebab-case slug and author with newError.author copy; lowercased path dedup against existing pages.
• Lifecycle: status is draft | published | favorite | archived with a side preStarStatus field so toggling the star ☆ round-trips through favorite without losing the prior state (favorite → ☆ back to published/draft/archived, not always published).
• Header layout: timestamp → ☆ favorite → status pill stacked below the path on the same .reader-header--stacked + .reader-header__meta--inline modifiers WikiPage now uses, so Base reader and Wisdom share chrome.
• Saving is a 800ms setTimeout that mutates the in-memory mock map (no API). The timer lives in a ref so navigation/create cancels the in-flight save instead of racing; popover closes on saving=true; Esc handler is one effect with priority unsaved > newDialog > popover.
• Dead-code cleanup: removed WisdomItem/WisdomKind/WisdomStatus types, the wisdomItemsFixture from src/test/fixtures.ts and tests/e2e/fixtures.ts, and the /v1/wisdom mock route from src/App.test.tsx and tests/e2e/mockApi.ts.
• Tests: 12 component tests in src/pages/wisdom.test.tsx covering tree → read → edit → save, "Add wikilink" K/W picker, "Add source" D picker, dirty-edit confirm + form preservation, and the draft → favorite → draft lifecycle round-trip. e2e: i18n.spec.ts + navigation.spec.ts switched to the new chrome (Filter wisdom pages label, Starred only chip, exact: true heading match).
• Out of scope: real GET /v1/base/pages?layer=WISDOM / POST /v1/base/wisdom wiring, async task polling, unresolved_wikilinks round-trip, asset upload, status mutations beyond ☆. Next iteration plugs the mock data source into real core endpoints once visual/interaction sign-off lands.

[0.0.10] - 2026-05-28

Chat: dedup right-rail Source list across streaming/session boundary

• ChatPage was concatenating streamingSources onto activeSession.sources without crossing the dedup boundary, so when a turn-N stream emitted a source already committed by an earlier turn, React rendered the same composite key (${kind}-${path}-${title}) twice and logged "Encountered two children with the same key". Surfaced under the auto-scroll stress fixture (24 sources × 2 turns → 24 duplicate-key warnings).
• Fix folds streamingSources into the session-base via mergeSources (the same path/title/kind dedup function used inside the streaming buffer), so identical sources appear once and order is preserved (session first, then streaming-only newcomers).
• Tool-call dedup was already safe via mergeTools' id-based merge — only sources had the cross-boundary gap.

[0.0.9] - 2026-05-27

Markdown reader: standard ![alt](path) images resolve against PageAsset.assets

• Standard CommonMark ![alt](path) syntax now routes through the same asset resolver as Obsidian ![[path]] embeds. Previously only the Obsidian variant matched PageAsset.original_paths; standard references like ![Kitchen](./images/scenes/kitchen.png) fell back to markdown-it's default renderer, which emitted a literal <img src="./images/scenes/kitchen.png"> that the browser then tried to load from the SPA host (404). The 10 images in sources/scenes.md and similar locales-tagged corpus pages now load from /v1/assets/<sha256> 200, with token-hydration when a session token is configured.
• Three correctness refinements while we were in the file:
  • alt text passes through self.renderInlineAsText(token.children) so ![**bold** plain](x) renders alt="bold plain" per CommonMark, not the raw source string.
  • title attribute is preserved across remote, plain-src, and token-hydrated branches (previously dropped on the two local paths).
  • Non-ASCII paths resolve correctly: markdown-it normalizeLink percent-encodes Unicode in standard syntax but original_paths stores raw, so the resolver tries the literal src first and falls back to decodeURIComponent(src) at the standard-image call site (scoped to that path only; Obsidian resolution stays verbatim).
• Behavior changes worth calling out:
  • Empty ![]() short-circuits to nothing instead of rendering the default <img src=""> — keeps drafts quiet.
  • Local refs in surfaces that don't pass assets (e.g. ChatPage) now render an explicit .md-broken-image placeholder where they previously emitted an <img> the browser silently 404'd; the explicit failure is intentional.
  • Remote URLs (http(s)://, data:) keep the markdown-image class for consistent CSS; authenticated-hydration stays inert on them because the selector requires data-asset-src.

[0.0.8] - 2026-05-27

Import: PDF / Office support via mineru

• ImportPage now accepts .pdf / .doc / .docx / .ppt / .pptx / .xls / .xlsx alongside the existing .md + assets surface. New files run through a converting pre-stage that POSTs to a new sidecar route, which calls mineru.net's v4 batch API and streams the converted markdown + assets back as a USTAR tar.gz. Once converted, files join the existing bundle → /v1/import → ingest → synth → lint pipeline unchanged. The /v1/import wire shape is unmodified, so core needs no changes.
• New sidecar namespace /web/* for dikw-web's own browser helpers, parallel to /agent/* (Pi Agent chat) and /v1/* (dikw-core). First occupants: POST /web/mineru/convert?inputSha=<hex> (single-file multipart in, tar.gz out) and GET /web/mineru/health. Same Node process as the agent sidecar — mounted in dev via webApiPlugin() in vite.config.ts and in prod via the same /web branch in dist-server/standalone.mjs. The browser only talks same-origin.
• New optional env var MinerUAPIKey (alias DIKW_AGENT_MINERU_API_KEY) in .env.agent.local. The variable name matches the dikw-plugins/.env convention so the same key file can be reused. Missing key → /web/mineru/* returns 503 mineru_disabled, ImportPage shows a Mineru not configured notice, and the file picker accept collapses back to .md/.pdf (PDF still works as a passive asset when referenced from a sibling .md).
• Idempotency is the headline contract: same input bytes → identical package_sha256, every run. Three layers enforce this: (a) mineru cache_tolerance=31536000 + data_id=<sha256[:32]> so the server-side cache returns the same conversion for one year; (b) browser IndexedDB cache keyed by SHA-256 of the input file bytes (LRU, 500 MB ceiling, mineruVersion: 1 cache-bust knob); (c) byte-stable tar packaging — entries sorted, mtime=0, uid=gid=0, mode 0644, frontmatter contains only converter / original_filename / original_sha256 (no timestamps, no batch_id).
• New converting pipeline stage with a ConversionProgress UI surface (src/pages/import/ConversionProgress.tsx). Per-file rows surface substages queued / hashing / uploading / polling / downloading / done / failed with a per-file Skip on failure. Two-concurrent worker pool — conservative on mineru quota while still pipelining well. Refresh during converting is non-resumable in v1 — the pipeline returns to idle, but the mineru server cache + browser IDB cache make re-conversion typically millisecond-fast for the same input bytes. Resume during the existing core-side stages (ingest / synth / lint-*) is unchanged.
• Image references in the converted markdown are rewritten from the mineru ![alt](path) form to project-conventional ![[assets/<rel>|alt]] wikilinks, matching what other dikw-web sources do. Resolution uses a 4-tier match (exact / case-insensitive / basename-unique / basename-folded-unique), mirroring the dikw-plugins/dikw-converter-mineru Python implementation.
• Token redaction: the mineru bearer token never appears in error messages or logs in full; only …<last 4 chars> is ever surfaced.
• New files of interest: server/web/{config,http,mineruClient, mineruConvert,vitePlugin}.ts, src/utils/{tar,tar-reader, mineru-convert}.ts, src/pages/import/ConversionProgress.tsx, tests/e2e/import-mineru.spec.ts. src/utils/tar.ts is a new isomorphic extraction of the USTAR writer that was previously inlined in import-bundle.ts; the writer's behavior is unchanged, it is now just importable from both the browser bundler and the sidecar's tar.gz response builder.

Known follow-ups

• The IndexedDB conversion cache stores cachedAt but does not yet enforce the planned 500 MB LRU ceiling — it relies on the browser's own quota eviction. Tracking as a follow-up so this PR stays scoped.

Manual verification (post-merge)

End-to-end mineru verification is not part of CI because it requires a real MinerUAPIKey and burns mineru quota. The e2e suite mocks the /web/mineru/* wire. Before relying on the feature in a workspace:

1. Copy MinerUAPIKey=… from dikw-plugins/.env (or wherever you keep it) into dikw-web/.env.agent.local.
2. npm.cmd run dev and open http://127.0.0.1:4321/#import.
3. Drop a small PDF or .docx. Watch ConversionProgress walk through hashing → uploading → polling → downloading → done, then watch the regular Bundle preview render with the synthesized markdown.
4. Drop the same file again. The IDB cache should make it instant — no network call to mineru.net (verify in DevTools Network).
5. Compare package_sha256 across two end-to-end runs of the same input — they must be identical for core's dedup to work.

[0.0.7] - 2026-05-26

Sidebar regroup + Base rename

• Split the sidebar's single flat list into three semantic clusters separated by hairline border-top dividers: Overview / Import / Base / Graph / Wisdom, then Retrieve / Chat, then Tasks. Settings stays in the footer group as before. The visible KNOWLEDGE and SYSTEM uppercase group labels are gone; each <nav> keeps aria-label for screen readers.
• Rename the en wiki-route concept from Knowledge to Base to match the /v1/base/* core endpoint family. Touches sidebar (nav.wiki), breadcrumb, wiki page heading (pages.wiki.title), the refresh button (pages.wiki.refresh), the Graph detail panel button (pages.graph.openInWiki → Open in Base), and the Graph canvas aria-label (Base graph). zh-CN keeps 知识库 / 在知识库打开 unchanged.
• zh-CN: rename nav.wisdom 智慧 → 认知 (English Wisdom unchanged) and cascade the wisdom page strings (认知沉淀, 刷新认知条目, etc.). The Work nav group's zh-CN aria-label is 工作 to avoid colliding with the contained 任务 button.
• CSS: drop .nav-group-label rules; extend .nav-main + .nav-main to render the hairline divider. Mobile breakpoint (≤720px) flips the divider to border-left so the new vertical hairline doesn't appear as a stray top border in horizontal-scroll mode.
• Docs: CLAUDE.md, README.md, and docs/graph-view.md updated to the new Base terminology.

Fixes

• OverviewPage: defensive optional chain on data?.status.wisdom_by_status?.candidate. Older / 0-wisdom core payloads omit the wisdom_by_status subkey, which previously threw TypeError and blanked the entire app via React 19's tree unmount.

[0.0.6] - 2026-05-26

Import page redesign

• Slim the 1018-line ImportPage.tsx into a 538-line orchestrator plus seven src/pages/import/ subviews (IdlePicker, BundlePreview, PipelineSteps, LintReview, DoneSummary, plus format.ts + readDroppedItems.ts). All styling reuses the existing src/styles.css token system — no UI framework introduced.
• IdlePicker: dropzone now accepts both files and folders. webkitGetAsEntry walks the directory tree and injects webkitRelativePath so computeProjectRelPath produces the same archive paths as the folder picker.
• BundlePreview: two-column Included / Skipped layout with per-row type icons, byte sizes, and ref counts; skipped rows carry a reason tag.
• PipelineSteps: resumed pipelines show a blue "Resumed your import" banner above the stepper; a 1-second interval drives the elapsed-time text so it no longer freezes on silent stages. 5-step stepper renders progress bars and an active-stage description card.
• DoneSummary: large success banner with Open in Wiki / View graph CTAs (hash navigation), two stat cards (what was added / lint outcome), and a restart tail.
• Tests: ImportPage.test.tsx grows from 4 to 17 cases across five describe blocks (idle picker, pipeline resume, lint review, done summary, failure & cancel).

Correctness fixes (landed alongside the redesign)

• handlePipelineError(err, owner) now compares the throwing controller against controllerRef.current so a late AbortError from a cancelled pipeline cannot clobber the state of a freshly-started one after the user clicks "Start a new import".
• applyLint resets wasResumed, breaking the "refresh into lint-review → click Apply → lint-apply incorrectly renders the resume banner" chain.
• Two hardcoded English strings now route through i18n: stepMeta returned "5 committed" and BundlePreview rendered "refs 3" even in the Chinese locale. Both have proper en and zh-CN translations now.
• Resume path seeds pipelineStartedAt = Date.now() so the stepper's elapsed segment is populated on the most important code path (mid-stage refresh). Combined with the 1s ticker, the clock no longer freezes when a task runs silent for 90+ seconds.
• IdlePicker.onDrop is no longer async. readDroppedItems rejections (permissions errors, browser quirks) now route through onDropError into the existing bundleError Notice instead of vanishing into the React synthetic-event system as unhandled promise rejections.

[0.0.5] - 2026-05-24

Import 页

• 新增 Import 路由(#import),sidebar 加同名入口。这是 web 第一个写入 surface:用户在浏览器内选本地文件或文件夹,自动跑「打包 → ingest → synth → lint(propose + 审阅 + apply)」全管线。任意阶段可取消,任务阶段刷新可 续轮询。core 端零改动,沿用 /v1/import 现有 multipart 协议。
• 浏览器侧打包(src/utils/import-bundle.ts):手写 USTAR tar + 原生 CompressionStream('gzip') + crypto.subtle.digest 算 SHA-256。无新增 npm 依赖。package_sha256(md_sha, asset_shas) = sha256(sorted([md, ...assets]).join("\n").encode("ascii")) 严格复现 dikw-core/src/dikw_core/md_inspect.py:60-66。
• Markdown 引用解析(src/utils/md-asset-refs.ts):正则与 core 的 _IMG_MD / _IMG_WIKILINK 一致,sibling-of-md → project-root 两段式 解析。远程 URL 不上传,缺失 asset 在 pre-flight 阻止导入。
• 管线状态(src/state/import-pipeline.ts):PipelineStage 联合 + sessionStorage 持久化(键 dikw-web.importPipeline,scope 与 serverUrl/token 一致; state 携带 coreUrl 防止跨核重放)。任务阶段(ingest/synth/lint-*)落地 对应 task_id,刷新后从 streamTaskEvents 续跟。上传阶段单次 POST, 刷新即丢,picker 重置。
• Lint 走法:propose 后弹审阅面板让用户勾选要 apply 的 proposal;部分 修复也算完成 —— 只要 task 终态 SUCCEEDED,即便 ApplyReport.skipped 非空,管线进 done,跳过项以 reason 形式展示,不判 failure。
• DikwClient 扩展:postMultipart + importBundle + startIngest / startSynth / startLintPropose / startLintApply + getTaskResult + cancelTask。multipart 上传不注入 Content-Type,让浏览器自填 boundary。
• i18n:nav.import + pages.import.* 全量中英文案。
• 测试覆盖:import-bundle (Python-golden hashes、USTAR 头结构、gzip magic、 端到端管线包)、md-asset-refs(19 个 case)、import-pipeline(状态机 迁移 + 持久化 + cancel)、DikwClient(4 个新方法 wire)、ImportPage (各 stage 渲染 + cancel)、e2e(sidebar 入口 + picker 预览)。

[0.0.4] - 2026-05-24

Source reader

• Source 层 read tab 渲染时,把已有反向边的 K 页 title(backlinks ∪ derived) 在 source body 中首次字面出现位置自动合成 [[title|原文本]] wikilink, 阅读体验向 wiki 页对齐。未匹配上的 K 页留在底部 Linked references panel。 Source tab(raw view)始终用原始 body,不做替换。
• 匹配规则:大小写不敏感、英文要求 \b 边界、CJK 无边界、最小长度英文 ≥3 CJK ≥2、longest-match-first、保留 source 原文字面写法。
• 受保护区段不替换:YAML frontmatter / fenced & indented code(含 mermaid)/ inline code / inline & display math / raw HTML 块(details/table/...)/ existing wikilink(含 image embed)/ markdown link 整体。
• 实现:新增 src/utils/source-inline-refs.ts 纯函数(30+ 单元测试), WikiPage 加 enhancedSourceBody useMemo 串起来,MarkdownView 和 wikilink rule 不动 — click 走现有 previewDoc 通道直达右侧 preview。 零新增 CSS / 零新增 i18n key。
• findPageForTarget 加 K 优先(wiki/wisdom > source)— 同名 K 与 source 共存时,合成的 inline wikilink 始终命中 K 页(否则可能 self-route 回 当前 source)。手写 [[xxx]] 行为保持兼容。
• Cache-lag 防御:injectInlineRefs 调用前先按 pages.data 路径表过滤 refs。resolveDerivedPages 用 wire title 占位的 K 页(还没进 pages.data)不内联,留在底部 panel,由 openBacklink 的 path-based fallback 兜底 —— 否则 inline 按钮的 openWikiLink 会查不到 path → dead link。
• 受保护区段补全(round-2 review):
  • CRLF normalize:injectInlineRefs 起手把 \r\n 归一为 \n, frontmatter / fenced / indented / display math 这几个行向 recognizer 才能正确识别 Windows-edited source 文件。
  • Fenced code 行扫描:替换原 regex 实现,识别 0-3 空格缩进、backtick/tilde 长度 ≥3、closing fence 长度 ≥ opener、EOF unclosed fence 整段保护。
  • Markdown link 兼容:URL 允许一层 balanced parens ([docs](https://example.com/path(v2)));新增 reference-style ([text][label] / [text][])和 link reference definition ([label]: url)整段保护。
• Heading slug 对齐(PR #37 /code-review finding):slugifyHeading 在 lowercase 前先剥掉 [[label|literal]] / [[label]] 语法, 这样 MarkdownView 的 heading_open(看 enhanced body)和 extractHeadingsWithSlugs(看 original body)即使输入不同也能 产出同一 slug。修复 source 页 outline 在"heading 含被内联的 K 页 title"场景下 getElementById 命中 null → 跳转静默失败的回归。

[0.0.3] - 2026-05-24

Added

• Source reader merges body [[wikilink]] backlinks and frontmatter sources: provenance into a single Linked references panel with linked / sourced labels. Consumes GET /v1/base/pages/{path}/provenance?direction=in from dikw-core 0.2.6+; pre-0.2.6 cores 404 (or 405) and the panel degrades silently to /links-only behavior. Other /provenance failures (5xx, network, parse) also clear the sourced channel but log a console.warn so the disappearance is debuggable. A dev-mode console.warn also fires when the core contract's layer-safety invariant is violated (non-empty derived_from on a source page).
• Source labels carry aria-label (Linked via body wikilink / Linked via frontmatter source) so screen readers announce the evidence channel, not just the chip text.

Fixed

• MarkdownView no longer wipes mermaid SVG (or chart / hydrated image output) on unrelated parent re-renders. React diffs dangerouslySetInnerHTML by the wrapping object's identity, not by __html string equality, so a fresh { __html: html } literal on every render was forcing innerHTML to be re-set even when the body HTML had not changed — discarding any post-render DOM mutations. The wrapper is now memoized on html. This was latent before but surfaced when the new source-reader effect started firing a second setState after page load.
• resolveDerivedPages no longer silently drops a /provenance entry when the cached pages list lags behind core — it falls back to the wire title so a just-synthesized K-page is visible in the source reader without a manual refresh. The click handler closes the loop: openBacklink now previews-by-path when the entry is not yet in pages.data, so the cache-lag fallback never ships a dead button. previewDoc and the new previewByPath share the same request lifecycle. The preview's Open as main document button is also hidden for cache-lag stub previews — the selection effect would otherwise round-trip the unknown path back to the default page, silently dropping the click.
• mergeSourceReferences is now pure-functional (no in-place mutation of map entries) and sorts sourced-only above linked- only inside the single-evidence tier so the two evidence channels read as contiguous visual blocks instead of interleaving by title.
• Switching between source pages clears the prior page's Linked references content synchronously, so the panel never shows stale chips during the body-fetch window. Existing race guards in the merge memo remain as defense in depth.

[0.0.2] - 2026-05-23

Removed

• VERSION file (4-digit MAJOR.MINOR.PATCH.MICRO from [0.0.1.0]). The format was gstack /ship-specific and forced package.json.version into non-SemVer territory (0.0.1.0), which kneecaps any tool that validates SemVer (npm version <bump>, semver.coerce(), future Renovate / semantic-release).

Changed

• package.json.version reverted to standard 3-digit SemVer (0.0.2); now the single source of truth for the project's version. package-lock.json re-synced.
• CLAUDE.md delivery workflow step 7 reverted to gh pr create as the primary PR-creation path. /ship is not used in this repo — its 4-digit VERSION requirement is incompatible with valid SemVer in package.json, and the individual sub-skills it orchestrates (/code-review, /review, code-simplifier subagent) can be invoked directly when useful. Bumping package.json.version and writing the matching CHANGELOG.md entry are now manual steps in the workflow.
• CLAUDE.md delivery workflow step 8 rewritten to emphasize active monitoring of CI status AND PR review comments after pushing, with resolve-as-found discipline rather than batching at merge time. Lists the three gh api endpoints (reviews / inline comments / top-level issue comments) plus the gh run view --log-failed pattern for failing CI logs.

[0.0.1.0] - 2026-05-23

Added

• CLAUDE.md "Working principles" section (Think before coding / Simplicity first / Surgical changes / Goal-driven execution) anchored against the project's /v1/base/graph, src/styles.css token system, and #chat canonical route.
• CLAUDE.md "Delivery workflow" — 8-step end-to-end loop from request clarification through TDD, codex-review rounds (max 3), /simplify / /code-review xhigh, Chrome MCP verification, doc walk (CLAUDE.md + README.md + docs/*.md), PR creation, and squash-merge with explicit gh api .../reviews + /comments pull.
• CLAUDE.md "Chat / agent rules" subsection covering right-rail session-scoped context, core-first tool preference, and the maintenance-action confirmation gate.
• VERSION + CHANGELOG.md scaffolding so the /ship delivery workflow can run end-to-end on this repo (bump version, write changelog entry, commit, push, open PR) instead of falling back to gh pr create.

Changed

• package.json version synced to the four-digit VERSION (0.0.1.0); the package is "private": true and not published, so the extra digit has no registry impact.
• CLAUDE.md architecture notes corrected against the actual code: Graph page removed wiki / source / all scope toggle (only search + hide-orphans remain), npm.cmd run build line updated to mention dist-server/standalone.mjs, dist-server/ added to the local/generated list.
• CLAUDE.md delivery workflow step 7 dropped the "this repo doesn't carry /ship scaffolding" caveat now that VERSION + CHANGELOG.md are in place; prefers /ship over plain gh pr create.
• CLAUDE.md delivery workflow step 4 replaced the non-existent /simplify and unsupported /code-review xhigh argument with the actually-installed /code-review plugin (five parallel Sonnet agents with confidence scoring) and the code-simplifier subagent for cleanup passes.
• README.md "Where canonical docs live" entry rewritten to point at CLAUDE.md as the operational guide for Claude Code sessions, and the npm.cmd run build row in the command table updated to match CLAUDE.md by mentioning build:server
  • dist-server/standalone.mjs (addresses CodeRabbit feedback on PR #33).
• README.md "Where canonical docs live" list extended with a docs/adr/ entry so future Architecture Decision Records show up in the doc index.

Removed

• AGENTS.md — superseded by CLAUDE.md now that Claude Code is the sole development surface; unique guidance (sidecar coreUrl rejection, token-never-displayed rule, .env.agent.local containment, generated-artifact ignore list) was integrated into CLAUDE.md rather than left in a parallel file.