Skip to main content

Grokky changelog

v.next​

  • AI: Project saves now drive the platform's Save Project dialog through new dialog AI functions (getProjectSaveInfo, setProjectName/Description/SaveMode/PresentationMode, entity-list listEntities/setEntityAction/setDataSync; Share dialog: getShareInfo/addShareGrantee/setShareAccess/setShareMessage) — the datagrok-projects skill was rewritten around this flow, replacing hand-written grok.dapi saves

  • AI: Fixed duplicate action execution (e.g. the same viewers added twice). Root cause: verify assertions written as bare expressions evaluated to undefined inside the async wrapper, reading as a failed verify with no diagnostics, and the "action did NOT take effect" gate feedback then pushed the model to redo the (already applied) action. Single-expression assertions are now auto-wrapped in a return, an assertion returning undefined reports a self-explanatory error, and the verify-gate feedback plus the datagrok-exec skill now forbid redoing state-changing actions before re-reading live state

  • AI: A passing in-round-trip exec self-verify now clears the Stop gate like a standalone datagrok_verify pass — turns with an earlier unverified action no longer pay a pointless hidden revision after a later verify already passed

  • AI: Conversations restored from history now hand their transcript (including the code each datagrok_exec actually ran) to the runtime on the first prompt — "reproduce what we did" works after a page reload instead of "I don't see any prior conversation"; loading a conversation also starts a fresh runtime session so histories never mix

  • AI: Answering an assistant question now gives immediate feedback — the Submit button disables ("Submitted") and the working spinner reappears at once instead of dead air until the next runtime event; the spinner hides while the question awaits the user

  • AI: The turn watchdog no longer kills turns that are legitimately waiting on the browser — a user taking >90 s to answer an AskUserQuestion, or a long-running datagrok_exec, previously hit the 90 s idle kill ("The assistant stopped responding")

  • AI: The system prompt now (softly) prefers a targeted widget's own quick functions for viewer-targeted requests — "zoom in to the scatter plot" resolves via list_view_widgets → the viewer's zoomIn rather than property edits or datagrok_exec code. Backed by new Dart-side quick commands (short names + anchored searchPattern regexes, surfaced through getFunctions()) on ten d4 viewers: scatter plot, line chart, bar chart, histogram, pie chart, box plot, density plot, PC plot, tree map, and 3D scatter plot (zoomIn/zoomOut/resetView where the viewer supports it, plus xBy/yBy/colorBy/sizeBy/splitBy/stackBy/valueBy/categoryBy/aggregationType/chartType/bins per viewer)

  • AI: EVERY d4 viewer (28 types verified in a live browser sweep) now carries an aiDescription briefing, and the remaining interactive ones got commands too — xBy/yBy (matrix/trellis/confusion matrix), innerViewer (trellis), pivotBy/groupBy (pivot), predictionBy (ROC), fromBy/toBy (network diagram), dateBy (calendar), lanesBy (tiles), colorBy (sparklines, shape map), addFilter/resetFilters (filters panel), refresh (correlation)

  • AI: Core widgets are now AI-drivable beyond dialogs/tabs/accordions — column selectors (setColumn), range sliders (getRange/setRange), viewer legends (listCategories/selectCategory), property grids (getProperties/setProperty), per-column filters (resetFilter), membership editors (listMembers/addMember/removeMember); plus xamgle surfaces: Browse panel tree navigation (listNodes/expandNode/selectNode), chemical sketcher (getMolecule/setMolecule/clearMolecule), console, favorites, functions widget (searchFunctions). Conventions documented in d4 CLAUDE "AI Integration" — key structural fact: inputs, tree nodes, tag editors, and popups are not Widgets, so they are projected through their host widget's getFunctions()

  • AI: Widget-level introspection — aiDescription and getFunctions() moved from views down to the platform Widget (Dart d4 Widget, js-api Widget/DartWidget/Viewer; the legacy grok_View_Get/Set_AIDescription interop is removed in favor of the Widget pair), so any widget can brief the assistant; new list_view_widgets meta-tool returns the current view's widget tree (ref/type/briefing/function count), and list_view_functions / get_view_function_result / call_view_function take an optional widget ref to target a sub-widget; buildWorkspaceContext() now includes sub-widget briefings

  • AI: Standard core widgets are AI-drivable — dialogs brief themselves with their title and expose getDialogInfo / setInput / clickButton (so the assistant can fill and confirm dialogs opened by entity commands), tab controls expose listTabs / selectTab, accordions listPanes / expandPane; open dialogs appear in list_view_widgets as dlg<N> roots even though they live outside the view's DOM subtree

  • AI: Entity gallery views (Dart DataSourceCardView subclasses — users, groups, roles, projects, dashboards, connections, queries, scripts, dockers, packages, files, and 20+ more) ship shared AI functions from the base class (listItems, searchItems, selectItem, listItemCommands, runItemCommand, refreshItems) plus per-view briefings and extras (e.g. connection details, role membership ops, container status), all reachable through the view-function meta-tools

  • AI: The grounding gate's post-answer Stop-block is now content-aware, after a 36-turn A/B (4 help + 2 data prompts × 3 reps × gate on/off) showed 0 of 14 blocks ever changed an answer — the system-prompt rule makes the model read the docs in both arms, while the block doubled trivial contextual answers (2 API calls → 1, 7.6 s → 3.6 s measured) and flashed "Revising…". The block now fires only when the visible answer makes platform UI-instruction claims (click/menu/dialog/"go to"…) without a source opened this turn — the one case it was built for. Skill invocations now count as grounding (the prompt names skills a source of truth, yet skill-grounded answers were paying a block), and a per-turn gates: {grounding, verify} override supports future A/Bs without a rebuild

  • AI: Context diet, measured end to end on the 52-prompt suite (benchmark-diet-full vs benchmark-sonnet-clean, 44 shared prompts): median turn 26.6 s → 17.9 s (−33%), TTFT 17.9 s → 14.4 s, median cache-read per turn 92 k → 56 k, prompt prefix 27,977 → 17,637 tokens/call — at unchanged accuracy (39/44 in both arms), plus 8/8 on the new entities category. Three levers, below

  • AI: Pure small talk ("hello", "thanks", "ok") no longer arms the grounding gate. The gate block cost a full hidden revision call just to hear NO_REVISION, and past 1.2 s the panel showed "Revising…" over a greeting — with a dashboard open the panel prepends workspace context, so the check strips that block (anchored on its fixed header, fail-closed for user-typed ---) before classifying. Measured on "hello": 2 API calls → 1, 9.3 s → 6.0 s, no revision phase

  • AI: Restricted the DECLARED built-in tool set (tools, not just allowedTools — which only pre-approves) to the ten the runtime actually uses. Every other built-in schema — Task, TodoWrite, NotebookEdit, plan-mode tools — was shipping in the prompt prefix of every turn unused: 7,572 tokens/call, the single largest diet win (SDK floor 17,493 → 9,921)

  • AI: ClaudeRuntimeClient.query() now times out (default 120 s, above the runtime's 90 s watchdog) and aborts the session instead of waiting forever — when the watchdog kills a wedged turn, no final ever arrives for that sid, and an untimed query() hung its caller permanently (observed hanging a full benchmark arm in its judging phase). The benchmark judge also retries once so a transient judge failure cannot flip a correct answer to failed

  • MCP: Collapsed the 34 datagrok MCP tools into five domain tools — datagrok_functions, datagrok_files, datagrok_projects, datagrok_spaces, datagrok_platform — each dispatching on an op argument, mirroring the meta-tool pattern the view functions already used. Every operation schema previously sat in the model's prompt prefix on every turn; now the description carries only the operation names and a call with no op returns that domain's full parameter schema, so nothing became undiscoverable. Measured with dev/harness/context-probe.mjs: 3,847 → 1,061 tokens per API call, a 72% cut of that block

  • MCP: Tool results are now compact JSON, paged (50 items, offset/limit), and capped. A result is not a one-off cost — it stays in the transcript and is re-read on every later hop of the turn, so one naive unpaged list used to tax the whole turn

  • AI: The verify gate now reads the op to decide whether an MCP call mutated anything. With one tool per domain the tool name no longer says: datagrok_spaces is both list and delete. Classification stays fail-closed — an unrecognised op counts as an action, so the gate over-asks rather than letting an unverified mutation through

  • Benchmark: Added an entities category (8 prompts) covering the server-side MCP surface, which the suite had no coverage of at all — the tool consolidation could have broken every server-side operation and an arm would still have reported 39/44. Their asserts read the server through grok.dapi rather than the view, and delete anything they create

  • Benchmark: Added dev/harness/context-probe.mjs, which attributes the per-turn prompt prefix to its contributors by ablation. Two corrections were needed to make its numbers mean anything: the prefix is input + cacheRead + cacheCreation (reading cacheRead alone measures how the cache broke, not how much context there was), and it must be divided by numTurns because SDK usage is cumulative over every API call in the turn — before normalizing, thinking looked like a 2× context regression when it was really a second API call

  • AI: Fixed the inlined-skill frontmatter stripper, which was written LF-only while the skills are checked in with CRLF — it matched nothing, so every inlined skill's YAML header was shipped in the system prompt on every turn

  • AI: Updated the datagrok-projects and datagrok-chem-toolkit skills, which still instructed the model to call tools removed by the consolidation (create_project, share_project, get_function). Skills are prose and nothing type-checks them, so a test now asserts no skill names a retired tool and that every skill has a routing description

  • Dev: Added dev/mcp.mjs, a local mcp-server container built from the working tree, so the op registry can be iterated on without an image rebuild and redeploy; dev/check.mjs now builds and tests the mcp-server too

  • Dev: dev/runtime.mjs up restores ~/.claude.json from the CLI's own backup when it is missing. That file lives outside the bind mount, so recreating the container lost it, and the resulting failure surfaced as a 90-second watchdog kill rather than as an auth error

  • Dev: Added a headless dev loop for claude-runtime (dev/) — runtime changes no longer need an image rebuild, a browser, a Datagrok instance, or an Anthropic API key. dev/seed-creds.mjs copies only the claudeAiOauth block out of the developer's credentials (leaving unrelated MCP tokens behind) into a gitignored, read-write-mounted copy so the container renews its own token; dev/runtime.mjs runs a local container with dist/ and plugin/ bind-mounted; dev/check.mjs is one command for compile + unit tests + suite lint + a live turn (~8s, or 4s with --fast)

  • Dev: Added dev/harness/drive.mjs, a fake-browser WebSocket driver that exercises the real runtime (system prompt, MCP servers, gates, revision protocol, session resume) while answering browser tool calls from a per-case stub table. Reports which tools a turn called in order, browser round-trips, the code each datagrok_exec ran, TTFT/total, and SDK token/cost

  • Dev: Added the first unit tests for the runtime (49) covering the Verifier and GroundingGate state machines, the fence-aware stream filter, and help-index generation. Extracted src/stream-filter.ts out of session.ts so the filter is testable without importing the ESM-only Agent SDK

  • Benchmark: Fixed scoring defects that made runs non-comparable — a prompt failing every rep scored as unscored rather than failed (dropping it from the denominator); the judge saw only the last rep while latency was measured across all of them; await in an assert was a silent failure and an async IIFE a guaranteed false positive; an assert that threw was indistinguishable from a model failure

  • Benchmark: Asserts now run after a settle delay — a Filters-panel filter applies asynchronously, so every t.filter assertion was previously a race

  • Benchmark: tool_activity now carries the bare tool name, so assertions can check which path a turn took (e.g. that entities were rendered as cards rather than prose-listed)

  • Benchmark: runBenchmark(label, reps, model) can pin every turn to one model, producing control arms for a model-routing comparison without redeploying the runtime; reports gained a per-difficulty rollup, cost-per-passed-prompt, raw assert pass rates, and a broken-assertion section

  • AI: A wedged turn can no longer take a runtime container down with it. The agent CLI can spin at 100% CPU indefinitely (observed on help/grounding turns: its rg --files child fills the pipe buffer and blocks, and the CLI busy-spins instead of draining it); because container.json allocates 1 CPU, that starves the Node server which would otherwise process the browser's abort, so the session — and every other session on that container — hangs with no way to recover. A watchdog now bounds silence between SDK events (not total turn length, so a legitimately long turn is unaffected): 90 s with nothing emitted ends the turn with a clear message to the user, aborts, and SIGKILLs the process the abort could not reach

  • AI: Added a reaper for agent processes orphaned by a finished turn — the watchdog only covers a turn still in flight, but a wedged process can also outlive the query that spawned it, leaving nothing to time it out (observed holding the container's only CPU for eight minutes with no session running). The interlock is a count, not an idle check: the SDK runs one agent process per in-flight query, so any surplus beyond the active-query count is by definition orphaned and is killed oldest-first. An idle-gated sweep would never fire under continuous load, which is exactly when it is needed

  • Dev: Added dev/seed-platform-creds.mjs — the Dockerfile seeds ~/.claude/.credentials.json with {}, and on a subscription real credentials only ever arrive via the in-app claude auth login, which writes into the running container. Anything that recreates the container (image rebuild, docker rm, an on-demand container respawning after an idle-out) silently resets it, after which every full-prompt turn returns in ~1 s with zero tokens and no error — indistinguishable from a model regression. --check reports auth status for every running runtime

  • Benchmark: Added dev/harness/introspect.mjs, which evaluates one expression in a logged-in tab — the way to settle "is the assert wrong or is the model wrong?" against the live API. It immediately found two harness bugs the suite had been attributing to the model: a box-plot assert reading a non-existent categoryColumnName (the real properties are categoryColumnNames / category1ColumnName), and a rubric demanding the answer 1274 where the correct count is 878. lint-suite.mjs now warns on any rubric that hardcodes a number

  • Benchmark: Added dev/harness/run-benchmark.mjs — drives a full arm headlessly through Playwright against a running instance (dev-key login, page-context call, progress heartbeat from the balloon feed) and writes the report into files/benchmarks/, so results land in the repo instead of a Downloads folder. The benchmark can't use the browserless T2 driver: its asserts read live DG objects that only exist in a tab

  • Benchmark: A turn that streamed nothing, called no tool and reported no usage is now an error, not a scored failure — an on-demand container that is still booting answers instantly with nothing, and the old harness recorded that as a model regression. Runs now also block on a readiness probe before the first measured turn

  • Benchmark: runBenchmark takes an only filter (comma-separated categories, difficulties, or prompt substrings) so validating a suite change or probing one tier no longer costs a full 44-prompt arm; partial runs are stamped in the report so they can't be mistaken for full ones

  • Benchmark: compareBenchmarks is N-way (comma-separated labels) rather than pairwise — a model-routing question has at least three arms, and reading it as two diffs hides where routing landed. Adds accuracy per difficulty tier, and shows the pass mark next to each per-prompt time so an arm that got faster by quietly failing is visible

  • Benchmark: Reports gained a per-category rollup and a failure section naming what failed and why (every-rep error / assert pass rate / judge score and reason, plus the tools the turn used) — a red cell said only that something failed

  • Benchmark: Rewrote the suite (19 → 44 prompts, tiered trivial/standard/hard) and strengthened every assertion from an existence check to a correctness check — previously every scored prompt passed at rate 1.0 in every recorded run, so the accuracy axis carried no information. Added dev/harness/lint-suite.mjs to validate the suite before a run spends quota

  • AI: Fixed markdown tables rendering as glued pipe-text in chat — models emit **Section** directly followed by | header | rows, which the platform's spec-strict GFM parser refuses (a table cannot interrupt a paragraph); the renderer now inserts the missing blank line before table blocks (fence-aware), in both streaming and final render. Also added a paragraph break between pre-tool and post-tool text segments ("…check those.No, there…")

  • AI: Gate blocks (grounding / verification) now run a background revision instead of double-answering or flashing drafts — the streamed answer stays visible, the revision streams hidden (runtime suppresses its chunks after the block), a "Revising…" status appears only if it takes a while, and the model decides the outcome: it writes a complete replacement (swapped in atomically on final.revision: 'replaced') or replies NO_REVISION to keep the original untouched ('kept'). Block-reasons are reframed as internal feedback (never mention/quote it — fixes the leaked "the hook feedback doesn't apply…"), backed by an "Internal feedback" system-prompt rule

  • AI: Help grounding is index-first — the runtime generates workspace/help/INDEX (path + title + description + keywords per published page, from the docs frontmatter; regenerated at startup and after each workspace sync), and the system prompt + GroundingGate now prescribe Read-the-index → Read-the-page instead of grepping 500+ files; WebFetch/WebSearch and js-api/ reads now count as grounded, and the repo's CLAUDE/.claude/.kg are excluded from staged user workspaces (a measured help-turn timeout mechanism). Targets the measured 8–33 tool round-trips per help turn (see docs/LATENCY)

  • AI: datagrok_exec self-verifies — a new optional verify: {assertion, description} parameter runs the verification assertion in the browser right after the action code (fresh scope, same round-trip) and returns it as verified; a passing one satisfies the Verifier, dropping the separate datagrok_verify model round-trip that made 2 round-trips the floor for every action turn

  • AI: Enriched all ~440 published help pages with frontmatter keywords (searchable synonyms and task phrases) and one-line descriptions; ~110 legacy pages got proper frontmatter with their body H1 promoted to title

  • AI: Added a latency/accuracy benchmark harness (docs/BENCHMARK) to put numbers on every LATENCY lever — Grokky:runBenchmark(label) drives a golden prompt suite (files/benchmark/suite.yaml, ~20 prompts across help/visualization/analysis/codegen/multitool/query) through the real runtime pipeline, timing TTFT/total/tool-round-trips per turn, reading the SDK token/cost usage now forwarded on the final event, scoring each prompt (deterministic assert + Haiku judge), and downloading a JSON+Markdown report; Grokky:compareBenchmarks(a, b) diffs two runs into a delta table. Latency/context metrics work immediately; token/cost columns need a claude-runtime image rebuild

  • AI: Prompt suggestions (empty-state cards + wand icon) now appear on any view, not only table views — the always-applicable "Anywhere" and "Code generation" blocks show from the home screen; view/column blocks still self-add when a table is open

  • AI: Expanded and diversified the table-view prompt suggestions (empty state + wand menu) — added semantic blocks for sequences, 3D structures, potency (IC50/EC50/concentration), geospatial, money and free text, plus themed whole-table blocks (Visualize, Analyze & model, Time trends, Clean & transform); blocks and individual suggestions self-hide when the data lacks the required column, and a new isDateTime slot filter enables time-series suggestions

  • AI: Ambiguous suggestions no longer make the assistant guess — parameter-dependent ones (substructure/motif search, similarity reference, calculated-column formula, custom viewer, choropleth metric, data search) now post their clarifying question straight into the chat as the assistant's reply via a new immediateResponse suggestion field (no AI round-trip); the user's answer then runs the task with the intent carried forward as context. "Open a demo dataset" uses a new client-side action field to render an inline choice block in the panel (Demographics / SPGI / curves / beer / random walk) that reads like an assistant reply; picking one removes the block, opens the table, and posts an "Opened …" note, and the block is dismissed if the user moves on to a new prompt

  • AI: Views brief the assistant — new View.aiDescription (js-api ViewBase/View, Dart View with JsViewHost forwarding) carries a short what-this-view-is + which-functions-to-call note, prepended to every prompt as "About this view"; query and script editors ship one

  • AI: list_view_functions matching is now OR-ranked (functions matching more query words rank higher) — AND-matching returned nothing for multi-word queries; a zero-match response now says how many functions the view actually has

  • AI: View functions replace view AI tools — the assistant now works with the platform's getFunctions() (registered DG.Funcs applicable to the view: meta.viewType functions, Dart view overrides, JS views via view.jsView) through three static meta-tools (list_view_functions search with ≤10 results, get_view_function_result, call_view_function), so views with hundreds of functions no longer bloat the context; getAITools/AIViewTool/viewAIToolsProvider are removed

  • AI: Restored empty-state prompt suggestions on table views — the singleton panel now resolves suggestion cards against the live current view (was pinned to its creation view) and refreshes them on view switches

  • AI: Query editor — removed the dedicated DB panel and catalog selector; the assistant now works on the query view through tools: getQueryInfo / setQueryAndRun (Dart-native) plus SQL schema exploration and test-execution tools (list_db_*, get_db_table_details, get_sql_test_result)

  • AI: Script editor — removed the dedicated scripting panel and language selector; the assistant reads/writes the open editor via getScriptCode / setScriptCode and infers the language from the script header

  • AI: The AI panel is now a single persistent assistant — switching views no longer swaps panels or resets the conversation; workspace context (current view, all open views, all tables) is rebuilt fresh on every prompt

  • AI: datagrok-exec / datagrok_verify blocks run against the live current view, so code targets views Claude just opened (e.g. a joined table) instead of the view the prompt started from

  • AI: Query-editor AI assistant no longer force-opens the AI panel when a query view opens — it registers and shows only on its toggle icon; it's disposed with its view

  • AI: Script-generation panel became a singleton rebound to the script view whose AI icon invoked it — one scripting conversation across script views

  • GROK-18695: Dependency security updates — refreshed lockfile so dev-only puppeteer no longer pins the runtime ws at a vulnerable 8.17.0 (DoS/memory-disclosure advisories)

  • GROK-18695: Docker: raised security floors (VEX) — claude-runtime: deb12 upgrade, npm@latest refresh, pip/setuptools bump, hono/@hono/node-server/ws floors, dropped unused workspace JDBC jars; mcp-server: express-rate-limit/fast-uri/path-to-regexp/qs/ip-address overrides

  • GROK-20054: Report: Error: Claude runtime container is not running

  • AI: Fixed query/script view going blank (tab unselected) on first click when switching back to it with the AI panel open — defer the panel dock/undock until after the view switch settles

  • AI: Replaced the 15-min skill/agent sync poll with an on-demand, TTL-gated refresh of packages + shared connections — idle sessions no longer sync

  • AI: Fixed shared-connection files that fail to download (e.g. invalid JSON) re-downloading on every sync

  • AI: Voice input — while a prompt is being processed the loader shows a Say "cancel" to stop hint, and saying "stop"/"cancel" aborts the run instead of being sent as a new prompt

  • AI: Fixed Enter key not submitting prompts in Chrome ≤ 50 / Dartium (no KeyboardEvent.key)

  • AI: Prompts claimed by Datagrok's built-in handler now show a green "Handled natively" check instead of a Responses block

  • AI: Ctrl+[ / Ctrl+] cycle through the current session's prompt history

  • AI: Replaced the "Responses" accordion with a hover-only minimize icon that collapses a reply block to one line

  • AI: Chat panel CSS now works in Chrome 50 — replaced flexbox gap with sibling margins, added -webkit- fallbacks for filter / user-select / mask-image

  • AI: Unified the bottom controls row icons — same size, larger spacing, subtle hover background

  • AI: Copy / thumbs-up / thumbs-down icons under a reply now appear only on hover of that message

  • Added src/polyfills.ts (Chrome 50 / Dartium) — crypto.randomUUID, Object.values/entries/fromEntries, String.prototype.trimStart/trimEnd, Array.prototype.flatMap, Element.prototype.append/prepend/replaceWith; routed clipboard writes through a copyToClipboard() helper with an execCommand fallback

  • AI: Added a Run button to the ribbon of file views opened from MyFiles/agents/scripts/

  • AI: Added wand icon to open curated prompt suggestions menu (loaded from suggestions.yaml)

1.0.4 (2026-02-01)​

  • Grokky: Azure compatible client
  • AI: Made a namespace, made a nested config, improved namings
  • Got rid of "latest" versions
  • Grokky: Refactor to Vercel AI

1.0.1 (2025-12-18)​

  • Grokky: Move stuff to proxy

1.0.0 (2025-12-17)​

Features​

  • Grokky: Invoking function and chains of functions
  • Grokky: Fuzzy matching
  • AI: Powersearch Search AI features
  • Grokky: Add structured output and reuse Gemini session to avoid recreating on each call
  • Grokky: AI sql

Bug fixes​

  • GROK-19198: Duplicating AI icons issue