v2.26.10.26.1

What's Changed

• [ENG-2807] Parse the datasource block and provide the typed cloud helper by @tino097 in #487
• ENG-2689: Classify an oversized-image refusal as permanent, not transient by @alecantu7 in #484
• [ENG-2175] Check decks and Word files before they are handed over by @Dav1dF in #492
• [ENG-2121] Tie each completed turn to the account that ran it by @Dav1dF in #493
• ENG-2892: treat a gateway alias echo as "served model not disclosed", not a repoint by @alecantu7 in #490
• ENG-2419: Name the kill in single-line readers of a scratchpad cell error by @alecantu7 in #497
• fix: pin click to >=8.3.3 to clear command-injection CVE by @martyna-mindsdb in #494
• fix: pin anyio to >=4.14.2 to clear IDNA TLS cert-spoofing CVE by @martyna-mindsdb in #495
• fix: floor pydantic-settings to >=2.14.2 to clear secrets_dir symlink traversal CVE by @martyna-mindsdb in #496
• ENG-1827: emit the turn's compaction result from the cloud entrypoint by @ea-rus in #479
• feat: generate_artifact — one-tool web artifact pipeline (ENG-968, ENG-969) by @StpMax in #335
• ENG-1816: MCP client in anton (Stage 1) by @martyna-mindsdb in #478
• [ENG-2807] Reach the datasource gateway through the turn's own inference connection by @tino097 in #501
• ENG-2863: run the verifier live matrix only when the verifier changed by @alecantu7 in #486
• ENG-2690: Stop reporting explainability bookkeeping errors as scratchpad tool failures by @alecantu7 in #499
• Scrub Google AQ. authorization keys at chat input (ENG-2684) by @hamishfagg in #498
• test(skills): drop the duplicated BROWSER STORAGE marker (ENG-2875) by @StpMax in #503
• fix(cloud_turn): never rate-limit tool progress lines (ENG-2981) by @StpMax in #507
• fix(llm): raise one exception type per MindsHub billing stop (ENG-813) by @lucas-koontz in #505
• fix(minds_client): keep only chat models from the catalog (ENG-1958) by @lucas-koontz in #508
• Use Jev for the completion check on MindsHub by @ZoranPandovski in #510
• Skip pip seeding for cloud scratchpad venvs by @ZoranPandovski in #512
• [ENG-2193] Let a host share the completion-verifier latch across sessions by @tino097 in #511
• Match example.com's current body text in the live fetch test by @martyna-mindsdb in #515
• Pass Google Ads' developer token through the cloud data vault by @martyna-mindsdb in #514
• Ship compiled bytecode in the scratchpad image by @ZoranPandovski in #516
• Render connector usage notes for connected engines by @StpMax in #519
• ENG-1985: single source of truth for artifact non-content names by @StpMax in #521
• feat(cloud_turn): interactive ask_user in cloud turns (ENG-3049) by @StpMax in #513
• fix(scratchpad): report venv failure facts instead of a misleading fix hint (ENG-1239) by @StpMax in #524
• Responses API: report truncation and failures, forward tool images, send trace headers by @ianu82 in #525
• fix(scratchpad): build a fresh model client for each helper call (ENG-3191) by @lucas-koontz in #528
• Feature/eng 3093 first token latency by @ZoranPandovski in #518
• Skip verify single tool round by @ZoranPandovski in #535
• Aartifact performance improvments by @ZoranPandovski in #534
• [ENG-2422] Tell a web turn where credentials go instead of asking for them in chat by @tino097 in #466
• Bump docosaurus by @ZoranPandovski in #541
• fix(llm): report silent model calls to hosts and end one after 10 minutes of no output (ENG-3281) by @lucas-koontz in #539
• fix(scratchpad): run a pad's venv work off the event loop, and end the turn when Stop lands in a cell (ENG-3189) by @lucas-koontz in #537
• feat(publish): keep the password hash while the password is unchanged and never copy another owner's artifact (ENG-3225) by @StpMax in #543
• fix(logging): omit provider payloads from warnings and cloud diagnostics (ENG-3304) by @lucas-koontz in #542
• Send the turn's trace on a cell's model call by @ZoranPandovski in #551
• Bugfix/verifier honest data gaps by @ZoranPandovski in #555
• perf(datasources): parse each YAML text once and reuse the package probe (ENG-3362) by @lucas-koontz in #558
• test: name incident fixtures and example hosts by what they show by @lucas-koontz in #559
• ENG-3067 Preload the cloud turn in a resident pod process by @ea-rus in #520
• fix(generate_artifact): keep the brief in the language the user writes in (ENG-3283) by @StpMax in #562
• feat(artifacts): act first, ask later in generate_artifact discovery (ENG-3206) by @StpMax in #533
• Keep owner-side files and files outside the artifact folder out of published bundles (ENG-3346) by @StpMax in #544
• Reject root-relative paths in generated pages (ENG-3226) by @StpMax in #552
• Tell the agent what to pass to publish_or_preview, not where artifacts live (ENG-3294 1/2) by @StpMax in #548
• Put report_tools blocks and table rows on their own lines (ENG-3358) by @StpMax in #550
• fix(product): answer local-model setup questions per surface (ENG-3321) by @pnewsam in #546
• Fail the turn instead of showing raw provider errors as a reply (ENG-3410) by @pnewsam in #569
• fix(session): name the limit that stopped a turn and let the user continue past it (ENG-3343) by @pnewsam in #547
• Replace open_artifact with read_text_file (ENG-3307 1/3) by @StpMax in #545
• Merged Main into Staging by @MinuraPunchihewa in #575
• [ENG-3408] Redact OAuth tokens in scratchpad output before they reach the model or storage by @tino097 in #567
• Don't guess at the work limit by @ZoranPandovski in #566
• [ENG-1706] Show an unnamed connection's engine as its label in the datasource context by @tino097 in #563
• ENG-3349: Fullstack artifact generator: correct STATE usage, no invented page content, stable API step by @ea-rus in #553
• ENG-3452: Trace the Jev completion check like the LLM verifier it replaces by @alecantu7 in #578
• docs: note each piece of text cowork-server's figure sources read (ENG-3433) by @lucas-koontz in #582
• ENG-3456: Attribute end-of-turn memory calls to their turn by @alecantu7 in #583
• Wait out Azure rate limits instead of failing the answer by @ZoranPandovski in #565
• fix(session): keep a turn's tool rows through compaction and hard truncation (ENG-3433) by @lucas-koontz in #585
• Handled Content Filters on Hand-back Call by @MinuraPunchihewa in #584
• fix(session): name a content-filter stop in the turn's reason and retry a hand-back with no answer by @lucas-koontz in #589
• feat(memory): add an optional read-only memory layer to Cortex (ENG-3422) by @lucas-koontz in #572
• fix(session): fire the scratchpad observers on streaming turns, so the cerebellum learns (ENG-3422) by @lucas-koontz in #590
• [ENG-3448] Redact a printed token value that contains a quote or a space by @tino097 in #574
• feat(sandbox): lock a person's agent code to their own folders with Landlock (ENG-3476) by @lucas-koontz in #588
• fix(session): give a progress report when the content filter stops the hand-back twice (ENG-3478) by @lucas-koontz in #593
• fix(identity): honor self-hosted product and memory context (ENG-3501) by @lucas-koontz in #592
• fix(llm): treat a 429 with only a detail body as a rate limit (ENG-2867) by @lucas-koontz in #595
• fix(tools): hold file tools to the asker's sandbox, and stop the key-shape filter reading task- ids as keys (ENG-3428) by @lucas-koontz in #598
• fix(sandbox): keep pip index credentials out of sandboxed agent code (ENG-3305) by @lucas-koontz in #596
• fix(scrub): hide the whole unquoted value after a token key (ENG-3448) by @lucas-koontz in #597
• fix(scrub): hide saved secrets in their encoded forms too (ENG-3428) by @lucas-koontz in #599
• fix(scrub): hide placeholder-led unquoted values, and scrub each line once (ENG-3448) by @lucas-koontz in #600
• fix(scrub): refuse a credential holding a scrub marker, and leave markers alone (ENG-3428) by @lucas-koontz in #601
• fix(scratchpad): drop saved variables of a connection that's gone, and keep a removed value scrubbed for the rest of its turn (ENG-3305) by @lucas-koontz in #602
• feat(scratchpad): hold a cell that changes connected data or existing files until the user approves it (ENG-3515) by @lucas-koontz in #594
• test(identity): assert the whole browser line, not a bare URL (ENG-3501) by @lucas-koontz in #603
• fix(write-check): no false cards on read-only work, and a card that shows exactly what runs (ENG-3515) by @lucas-koontz in #604
• Ignore Python bytecode in artifact folders (ENG-3176 1/2) by @StpMax in #580
• perf(sandbox): prepare cell launches off the event loop (ENG-3362) by @lucas-koontz in #616
• ENG-3351: Log per-call LLM usage for evals when ANTON_USAGE_LOG is on by @ea-rus in #570
• fix(ask_user): end the turn on an unanswered question instead of assuming by @pnewsam in #612
• chore: merge main into staging after the render-check hotfix (ENG-3539) by @lucas-koontz in #618
• chore: record main's commits in staging after the squashed sync (ENG-3539) by @lucas-koontz in #619
• [ENG-3393] Let Collection.put accept an item read back with Collection.get by @ea-rus in #586
• Size output budgets by reasoning effort and cap effort in generate_artifact (ENG-3282) by @StpMax in #554
• Retry a model stream that drops mid-response (ENG-3483) by @StpMax in #587
• fix(logging): keep agent diagnostics quiet and correlated (ENG-3579) by @lucas-koontz in #621
• Write generate_artifact documents in the user's stated language (ENG-3283 2/2) by @StpMax in #624
• chore: record main's commits in staging after the ENG-3283 hotfix (ENG-3283) by @lucas-koontz in #626
• build(release): preserve reviewed dependency selections (ENG-3603) by @lucas-koontz in #629
• feat(artifacts): run live apps in app slots outside the server process (ENG-3566) by @lucas-koontz in #620
• fix(artifacts): log each live-app launch decision with its outcome and app (ENG-3566) by @lucas-koontz in #637
• feat(generate_artifact): follow the project's rules and the host's artifact skills in every step (ENG-3616) by @lucas-koontz in #630
• fix(llm): send a structured call again without the forced tool choice its model refused (ENG-3624) by @lucas-koontz in #633
• fix(generate_artifact): merge staging, keep known_data with its run, and ask the agent for the rows (ENG-3615) by @lucas-koontz in #642
• fix(generate_artifact): let the data checks see the connections and the passed-in data, and try one fetch before "not enough data" (ENG-3615) by @StpMax in #635
• fix(llm): run page builds and the completion check at the configured effort (ENG-3719) by @lucas-koontz in #641
• test(artifacts): check a live app's blocked-resource warning by exact value by @lucas-koontz in #643
• fix(artifacts): launch a live app whose requirements.txt leaves out mangum (ENG-3698) by @lucas-koontz in #638
• fix(session): give the completion check room to answer at max effort (ENG-3719) by @lucas-koontz in #657
• fix(artifacts): offer html-apps only where nothing starts an app backend (ENG-3493) by @lucas-koontz in #645
• feat(artifacts): send the artifact builder's calls to a model and effort of its own (ENG-3623) by @lucas-koontz in #634
• fix(approval): quote each statement on the card for a SQLite or DuckDB file change (ENG-3740) by @lucas-koontz in #659
• feat(generate_artifact): start every view through one helper so a failed view fails alone (ENG-3622) by @lucas-koontz in #631
• fix(approval): refuse a cancelled database file write as a change, not an overwrite (ENG-3740) by @lucas-koontz in #661
• fix(generate_artifact): read public-data-sources only from trusted folders and skip unreadable legacy skills (ENG-3616) by @lucas-koontz in #662
• fix(memory): keep a rule given in a project in that project's memory (ENG-3739) by @lucas-koontz in #660
• feat(ask_user): end the turn when a question goes unanswered (ENG-3751) by @lucas-koontz in #663
• fix(approval): close function, report, and compaction gaps (ENG-3529) by @lucas-koontz in #613
• feat(generate_artifact): hand a page that failed only its browser check back to the agent to repair (ENG-3621) by @lucas-koontz in #632
• Release: staging → main (216 commits) by @mindsdb-release-train[bot] in #610

Full Changelog: v2.26.10.21.3...v2.26.10.26.1