> ## Documentation Index > Fetch the complete documentation index at: https://docs.trainy.ai/llms.txt > Use this file to discover all available pages before exploring further. # Changelog > Release history for Pluto SDK and Pluto Server. ## Changelog ### v0.0.41 (Server) — 2026-04-24 Side-by-side comparison performance fixes, runs-table bugfixes, and backend memory hardening. **Changes:** * Fix side-by-side view lag: hidden dashboard tree was still mounted, so chart queries, dynamic-section regex queries, and timers kept running in the background * Fix Config section in side-by-side rendering every row up front (thousands of `JSON.stringify` and inline-diff passes per click) — now lazy * Fix runs table where deselecting a URL-linked run could leave selection in a broken state * Backend: trim `runs.list` responses to only return field values for visible columns (avoids \~17 MB responses on Hydra-style configs); kill duplicate `runs.list` queries from a localStorage/saved-view pageSize mismatch and from `refetchType: "all"` over stale cache entries; pause auto-refresh when the browser tab is hidden * Runs table pageSize is now stored per saved view (DB) instead of in browser localStorage — every saved view (Default and custom presets) carries its own `config.pageSize`, and pageSize changes in the dropdown only persist when you save the active view *** ### v0.0.20 — 2026-04-23 wandb dual-logging, run-lifecycle improvements, and broader display-ID acceptance in the query API. **SDK Changes:** * **wandb dual-logging via a `.pth` import hook** — set `PLUTO_API_KEY` (and either `PLUTO_PROJECT` or your existing `WANDB_PROJECT`) and every `wandb.log()` also logs to Pluto, with no code changes. Set `DISABLE_WANDB_LOGGING=true` for Pluto-only mode. See [Weights & Biases Migration](/pluto/wandb-migration). * **New `run.close()`** — local teardown without marking the run complete on the server. Unregisters the `atexit` hook so a side process that attaches to an active run can exit without ending it. See [Run Lifecycle](/pluto/run-lifecycle#ending-a-run). * **Display IDs accepted in every query helper** — `get_metrics`, `get_statistics`, `get_files`, `download_file`, `get_logs`, and `compare_runs` now take a display ID like `"MMP-42"` in addition to numeric IDs. Previously only `get_run()` did. * **Deep merge for config inheritance** — `pluto.init(fork_run_id=..., inherit_config=True, config=...)` now deep-merges your override into the parent's config (was shallow). `run.update_config()` is also deep-merge now. * Run forking via `pluto.init(fork_run_id=..., fork_step=...)` (preview) — info-log emitted at init time. * Internal: scope Sentry exception reporting to pluto-touching code only. *** ### v0.0.40 (Server) — 2026-04-16 Ops-only release hardening backend memory under load. **Changes:** * Cap L1 cache memory to prevent backend pod OOM under bursty load * Add `runs.list` memory-pressure regression test that fails on dead bursts (not isolated 503s) * Add hybrid-auth test infra for L1 cache regression testing *** ### v0.0.39 (Server) — 2026-04-15 Run status-transition timeline, per-widget best-step pinning, multi-sample media nav, and chart-perf improvements. **Changes:** * Add durable run-status transition timeline on the run summary page; new `GET /api/runs/status/history` endpoint and `runs.statusHistory` tRPC procedure * Apply terminal-status precedence (`FAILED > CANCELLED > TERMINATED > COMPLETED`) for `api`-source writes so DDP fan-out collapses to a single timeline entry and a late `COMPLETED` can't clobber a `FAILED` * Per-widget best-step pinning so each image widget gets its own argmin/argmax step per run (no more "No image at step N" placeholders when one widget happens to lack the metric's argmax step) * Scoped unpin for cross-panel and best-step pins: dropdown with **Unpin this image** vs **Unpin across all panels** * **Multi-sample media nav (`◀ i / N ▶`)** for wandb-style list-of-media logging across images, audio, and video — in both the Charts tab and the Dashboards tab. Previously the multi-group renderer dropped all but one sample per `(run, step)` * Inline ↔ fullscreen sample-index sync; pin remembers the active sample index across Charts ↔ Dashboards tab switches * Image fullscreen: zoom floor lowered to 10% (20% step), `Ctrl + scroll` for fine zoom; centered images instead of stuck at the top of the container; phantom-scrollbar fix * Chart performance: zoom + bounds CTE, metric-discovery cache, TTL memoization * Replace `smartDateFormatter` with `formatAbsoluteTimeTooltip` for tooltips * Patch 23 high-severity dependency security alerts; bump ClickHouse query log line length *** ### v0.0.38 (Server) — 2026-04-10 Per-run image pinning and metric-driven best-step pinning land. **Changes:** * **Per-run image pinning**: pin individual runs at specific steps in multi-run image comparison. Pinned runs stay frozen while the global stepper drives others. Two scopes — pin in a single panel (◇) or across all panels (◈) * **Best-step pinning**: in the runs table, click a metric column header → "Pin steppers at min/max value" → each run gets pinned at its own argmin/argmax step (★). Backed by ClickHouse `argMin` / `argMax` aggregates * **"With image" variants** of the best-step actions, restricting argmin/argmax to steps where an image actually exists for that run * **Find best step** submenu rename with a Target icon and clearer wording * **Clear All Pins** button in the Image Settings popover * UI polish: smaller run-name labels (truncate + title tooltip), max 2 images per row, default stepper to max step, distinct pin-type colors and symbols, fullscreen pinned indicator *** ### v0.0.19 — 2026-04-09 Internal robustness fixes. **SDK Changes:** * Handle headless sessions during run init * Suppress Sentry breadcrumbs from internal Pluto trigger endpoint requests so user telemetry isn't polluted *** ### v0.0.37 (Server) — 2026-04-08 Chart bugfixes, performance improvements, and a new step deduplication feature for line charts. **Changes:** * Add query-time step deduplication toggle for line charts to eliminate duplicate values from distributed training * Add maintenance window banner with localStorage-based dismissal * Fix Y-axis not rescaling when hiding/unhiding runs via the eye icon * Fix X-axis not shrinking when hiding runs with wider data ranges * Fix cursor sync not working across charts until first click * Fix tooltip search breaking the table layout * Fix pinned tooltip not updating highlight when moving mouse vertically or across charts * Fix hidden runs with no data still appearing in the tooltip * Fix tooltip pinning failing in sparse chart regions * Performance: reduce tooltip DOM elements from N (one per chart) to a single shared instance * Performance: fix fullscreen open/close lag on dashboard widgets by replacing React state with DOM class toggle * Fix fullscreen legend sidebar not highlighting the hovered run * Fix dynamic section widget bugs (stale widgets on last-run deselect, widget count mismatch when hiding runs) * Fix dashboard search false positives on collapsed dynamic sections *** ### v0.0.18 — 2026-04-06 Added Sentry telemetry for error tracking and pinned httpx to fix a compatibility issue. **SDK Changes:** * Add Sentry telemetry for error tracking (disabled in CI) * Pin httpx \< 1.0 to fix TypeError on `verify` keyword argument *** ### v0.0.36 (Server) — 2026-04-03 Run forking, saveable table views, MCP OAuth support, and dashboard bugfixes. **Changes:** * Add Neptune-style run forking with inherited metrics * Add saveable default run table views with view presets * Add OAuth 2.1 authorization server for MCP client authentication * Add downsampling algorithm parameter to lineage metrics query * Fix dashboard charts incorrectly showing in dynamic sections * Resolve various search issues and apply UI polish * Fix fullscreen image scrolling and add dynamic tab titles * Fall back to PUBLIC\_URL for Pluto API URL in MCP server *** ### v0.0.15 — 2026-03-27 Removed the SDK signal handler to fix cleanup issues in multi-process training setups. Server v0.0.33 brings major performance improvements across the board. **SDK Changes:** * Remove signal handler to prevent interference with user process signal handling in distributed training **Server Changes (v0.0.33):** * Performance: Add multi-metric batch endpoint for faster chart data loading * Performance: Add columnar JSON serialization, Redis caching, and zoom query optimizations for reduced latency * Add incremental DOM updates for tooltips to eliminate re-render jank * Fix NaN marker colors to match their parent series colors * Fix step sync and fullscreen behavior for non-chart media widgets (images, audio, video) * Fix ClickHouse regex compilation errors from dashboard metric patterns *** ### v0.0.14 — 2026-03-24 Added `pluto sync` CLI command to recover data from crashed runs, and fixed sys metrics stepping independently from training metrics. Server v0.0.31 is a major UI release with Y-axis drag zoom, NaN/Inf markers, chart resolution controls, and significant performance improvements. **SDK Changes:** * Add `pluto sync` CLI command to retry uploading queued logs from crashed or interrupted runs * Fix system metrics to use an independent stepper (no longer tied to training step cadence) * Upgrade git detection log messages from debug to warning level **Server Changes (v0.0.31):** * Add Y-axis drag-to-zoom support for line charts * Add visual markers for non-finite values (NaN, Inf, -Inf) on chart series * Add chart resolution control to reduce implicit smoothing from downsampling * Add hovered series highlighting across tooltip and runs table row * Add step navigation to fullscreen image viewer with sticky steppers * Add page jumping support for runs table pagination * Add Python repr parsing and JSON pretty-print by default in side-by-side comparison * Add pagination support to the `query_logs` MCP tool * Add AbortSignal support for cancelling in-flight queries * Persist legend-toggled series visibility across chart zoom interactions * Make tags cell responsive with dynamic tag visibility based on column width * Fix notes popover vanishing before user can type * Fix zoom reset for hidden runs and `?runs=` deep links for off-page runs * Fix display ID resolution in batch graph procedures during zoom refetch * Fix legend value cell layout to prevent flicker on hover * Fix focus and interaction issues in fullscreen chart dialogs * Performance: group-level chart virtualization and query deduplication * Performance: fast deep pagination via cursor lookup * Performance: prefetch chart data when run IDs are present in URL * Performance: suppress background chart painting during fullscreen mode *** ### v0.0.30 (Server) — 2026-03-19 Run comparison and charting improvements with cross-panel step sync, fullscreen legend sidebar, and inline config diffs. **Changes:** * Add cross-panel step sync with stepper UI for navigating training steps across charts * Add fullscreen legend sidebar for chart comparisons * Add configurable max series limit for line charts * Add inline text diff and JSON pretty-print to side-by-side run comparison view * Make table pagination input editable for direct page navigation * Hide files and media for hidden runs * Make matplotlib an optional dependency in the MCP server * Fix hidden run state not persisting across component remounts * Fix config data missing from table views and side-by-side comparison *** ### v0.0.13 — 2026-03-16 Added Claude AI skills integration for experiment tracking and fixed a bug with config updates. **SDK Changes:** * Add Claude AI skills for experiment tracking * Fix JSON serialization bug in `update_config` that yielded 400 error on ingest **Server Changes (v0.0.28):** * Add chart visualization tools to MCP server for metrics and run comparisons * Add drag-and-drop section reordering in dashboard edit mode * Separate selected vs visible runs for independent chart visibility control * Use distinct colors for metrics in single-run multi-metric charts * Fix phantom runs appearing in pagination after deselecting all runs * Fix fullscreen chart dialog closing when interacting with tooltip * Fix metric column sorting breaking pagination * Fix race condition in run selection between URL params and cache * Fix relative time zoom sync and add step range translation * Fix relative time offset by parsing ClickHouse timestamps as UTC * Fix zoom group isolation and cross-axis sync issues * Add error logging for column key extraction failures *** ### v0.0.12 — 2026-03-12 Fixed a bug where runs that ended with an unhandled exception were incorrectly marked as COMPLETED instead of FAILED. **SDK Changes:** * Fix runs with unhandled exceptions incorrectly marked as COMPLETED *** ### v0.0.11 — 2026-03-10 Fixed a collision bug where user-seeded random state could cause duplicate run external IDs. Server adds server-side downsampling, dashboard copy/paste, and major charting improvements. **SDK Changes:** * Fix run external ID collisions when users seed global random state **Server Changes (v0.0.26):** * Add server-side bucketed downsampling for charts * Add widget copy/paste functionality to dashboard builder * Add log-scale scientific notation formatting for chart axes * Add customizable tooltip columns with search filtering * Add optimistic concurrency control for dashboard edits * Improve chart emphasis visibility and color palette * Improve log scale axis label formatting for wide value ranges * Show tooltip for long tags that get truncated in the runs table * Fix `?runs=` URL deep-link for runs not on the first page * Fix side-by-side config diffs not loading on initial page load * Fix relative time scale consistency across multiple runs * Fix dashboard widget height regression * Fix selected runs pagination, cross-page hover, and search/dashboard persistence * Fix dashboard concurrency warning not triggering * Remove cross-axis zoom translation in run comparison view * Resolve high and critical security alerts *** ### v0.0.10 — 2026-03-05 Fixed a bug where heartbeat retries could cause `finish()` to hang when the server returned errors. Added support for resuming runs by display ID or numeric ID. **SDK Changes:** * Fix heartbeat retries causing `finish()` to hang on server errors * Support resuming runs by display ID or numeric ID **Server Changes (v0.0.25):** * Add resume run endpoint to allow re-opening completed runs * Add write tools (`update_tags`, `update_notes`) to MCP server * Add toggle buttons and keyboard shortcuts to hide/show the run list and graphs panels * Pin selected runs to top of the runs table * Add Run ID column to Linear issue sync table * Add display ID search support to run search queries * Add skip-missing-values option to line charts * Improve fuzzy search matching precision and reduce false positives * Replace print statements with logging and fix session management in Python client *** ### v0.0.24 (Server) — 2026-03-04 Charting improvements including custom x-axis metric selection and persistent zoom across data refreshes. **Changes:** * Auto-hide empty pattern-matched chart widgets * Fix run-scoped metric queries hitting the wrong table * Add tooltips and improve text overflow handling across UI components * Add custom x-axis metrics support for dashboard widgets with searchable combobox * Fix zoom not persisting on virtualized charts after data refresh *** ### v0.0.23 (Server) — 2026-03-02 Added widget search to dashboards and a metrics display in the file tree view, plus several bugfixes. **Changes:** * Add widget search functionality to dashboard builder (with fix for hidden widget deletion) * Fix Y-axis range calculation when chart passes null data bounds * Add metrics display to file tree view * Run page toolbar parity improvements, histogram canvas unification, and bugfixes *** ### v0.0.9 — 2026-02-26 **Breaking change:** The environment variable `PLUTO_API_TOKEN` has been renamed to `PLUTO_API_KEY`. Update your configuration accordingly. **SDK Changes:** * Rename `PLUTO_API_TOKEN` to `PLUTO_API_KEY` for consistency **Server Changes (v0.0.22):** * Overhaul custom dashboards with dynamic sections, widget improvements, and UX polish * Fix duplicate tooltip appearing when another tooltip is pinned * Fix widget config properties being stripped by schema validation * Add per-chart log scale overrides to chart settings popover *** ### v0.0.8 — 2026-02-24 Added Neptune query compatibility and automatic git commit/version logging. Major charting performance improvements on the server side. **SDK Changes:** * Add Neptune query compatibility shim for easier migration * Automatically log git commit hash and SDK version with each run **Server Changes (v0.0.20, v0.0.21):** * Use virtualized charts in custom dashboards to unmount off-screen charts for better performance * Fix tooltip flickering on scroll events * Fix log scale Y-axis with manual bounds and zoom behavior * Fix cross-chart highlight dimming all series in single-metric charts * Fix chart zoom synchronization across X-axes * Add tooltip to display all tags on overflow * Fix CORS headers on error responses * Pin select, status, and name columns in runs table *** ### v0.0.19 (Server) — 2026-02-23 Performance fix for slow tooltip rendering on the dashboard view. **Changes:** * Fix slow tooltips on dashboard view *** ### v0.0.7 — 2026-02-20 SDK now automatically captures system information including OS, CUDA, NCCL, and InfiniBand details. Server adds multi-metric chart support with glob and regex search. **SDK Changes:** * Automatically capture OS, CUDA, NCCL, and InfiniBand system info * Fix `run_id` not being passed in Neptune compatibility wrapper **Server Changes (v0.0.17):** * Add multi-metric charts with glob search, regex search, and dynamic metric selection * Support NaN, Infinity, and -Infinity in metric values *** ### v0.0.16 (Server) — 2026-02-19 Major improvements to search, charting, and editing workflows. **Changes:** * Add two-stage fuzzy search with backend similarity matching and client-side Fuse.js filtering * Add min/max envelope downsampling and zoom-aware chart rendering * Add optimistic updates for tag and notes mutations * Add display ID support for run labels in charts * Add localStorage draft save and navigation guard for dashboard editing *** ### v0.0.15 (Server) — 2026-02-18 Chart export, dashboard view on the run detail page, and a revamped runs table layout. **Changes:** * Add chart export to clipboard and PNG download (with title and legend) * Refactor recent runs display from cards to a table layout * Add diff-only filter and git-style highlighting to side-by-side run comparison * Add dashboard view support to run detail page * Enable automatic data refetching on window focus and reconnect *** ### v0.0.14 (Server) — 2026-02-16 Runs table polish: fixes for trailing zeros, column header UX, and number filter input. **Changes:** * Fix runs table bugs: trailing zeros, column header UX, number filter input *** ### v0.0.13 (Server) — 2026-02-14 Migrated the Linear integration to OAuth and added URL auto-resolution for org-less paths. **Changes:** * Add display ID support for run identification in MCP tools * Migrate Linear integration to OAuth app * Add readiness probes with PostgreSQL and ClickHouse health checks * Add URL auto-resolution to redirect org-less paths to the active org * Fix `?runs=` URL param not selecting runs on page load and being overwritten by stale cache *** ### v0.0.12 (Server) — 2026-02-13 Linear integration reliability improvements and Stripe promo code support. **Changes:** * Fix duplicate Linear comments via idempotent sync with concurrency control * Allow free checkouts with promo codes by making payment optional *** ### v0.0.11 (Server) — 2026-02-13 Major feature release: Linear integration, display IDs, MCP leaderboard, and free PRO for .edu users. **Changes:** * Add Linear integration with write-back and issue picker * Add Neptune-style sequential run numbering with display IDs * Add metric names and leaderboard endpoints for MCP run ranking * Grant free PRO plan to users with `.edu` email addresses * Fix data table layout minimum width *** ### v0.0.10 (Server) — 2026-02-11 Dashboard and runs table overhaul with grid/free mode toggle, Y-axis bounds, and full run table customization. **Changes:** * Add Grid/Free mode toggle for dashboard layout * Add chart Y-axis bounds, fullscreen mode, and clipping indicators * Add run table customization with sorting and filtering * Fix pagination state management in data tables *** ### v0.0.6 — 2026-02-10 Server-side improvements to chart configuration and Konduktor job integration. **Server Changes (v0.0.8, v0.0.9):** * Add configurable line width setting for charts * Fix zoom sync corruption in virtualized line charts * Add Konduktor job info card to run summary page * Add notes column to runs table for experiment annotations * Increase run list limits for table and plots *** ### v0.0.5 — 2026-02-09 Improved sync reliability with batched flushing and backoff to reduce SQLite lock contention. **SDK Changes:** * Add batching and regular flushing of the sync queue with backoffs to reduce WAL/SQLite lock contention * Add `sys/` metrics for Pluto SQLite queue monitoring *** ### v0.0.4 — 2026-01-26 Initial server infrastructure release. *** ### v0.0.3 — 2026-01-14 Fixed a distributed training bug where DDP processes could hang indefinitely. **SDK Changes:** * Fix DDP bug by enforcing thread join timeouts