what shipped, week by week.

A chat you don't own opens read-only, with a button to fork it into your own copy. The Retry button resubmits a failed job with its original inputs.

chat#

• A chat you don't own opens with a read-only composer and a "fork to continue" button. Forking copies the whole conversation into a chat of your own, so you can pick up a colleague's work without changing theirs.
• A failed job's in-progress card turns into an error card in place, showing the error, the tool, the parameters and the time. This applies to guest and signed-in runs.
• The jobs menu and active-job chips keep running jobs in view until they finish, and each finished job announces itself once.

jobs#

• Retry resubmits a failed job with its original inputs, project and chat, from canvas cards, the jobs dashboard and job pages.
• Retry is offered for failures that can change on a second try. A job refused for bad input asks you to fix the input instead.

tools#

• DiffDock docks against the receptor chains you name. For peptide and biologic ligands it points you to Boltz-2 or Chai-1 before any compute runs.
• Boltz-2 returns the structure when a ligand is too large for an affinity estimate, rather than failing the whole run.
• MM/GBSA scores a complex already in your workspace, as well as pasted structure text.
• The tool catalogue lists BoltzGen and Chai-1 by name, with the plan each one needs.

canvas#

• Cards show "not measured" where a tool did not measure a value.

sharing#

• A "specific results" share link shows only the results you picked.

pricing#

• The pricing page compares plans across every workflow domain.

HTTP API runs return what Ivy called, the jobs it started and what it cited. Shared chats show their images and downloads, and R scripts and notebooks run in signed-in chats.

http api#

• A response carries each tool call's id and arguments, the job ids it started, and its citations. You can trace an answer back to the runs behind it.
• A request can name the tools Ivy may use, or the tools to leave out.
• A names-only tool catalogue lists what you can ask for.
• Download links for a finished job's files are one call away, so you don't need the CLI to fetch outputs.
• A turn ends when the caller disconnects.

sharing#

• A shared chat shows its images and offers its files as downloads.
• A share link shows only the cards from the chat you shared.

sandbox#

• R scripts and notebooks run in signed-in chats, and Octave runs in guest chats.
• R and Octave errors come with hints written for those languages.

cli#

• ivybloom can stop a running turn and quote a job's credit cost before you submit it.
• Duplicating a campaign with its children copies its chats.
• The CLI stream's final frame reports the turn's usage, including for turns you stopped.

tools#

• AlphaFold DB ids resolve to structures.
• GWAS Catalog lookups use the current catalogue, and very small p-values stay marked significant.
• A PheWAS source failure reports as an error rather than an empty result.

canvas#

• LigandMPNN designs render on their canvas cards.

Every pose from a DiffDock run opens on the canvas, ESM3 results show as sequence cards, and Studio tool panels have Cancel, Retry and Download.

docking#

• A DiffDock result opens every pose, ranked, with the receptor structure alongside. You can step through the poses instead of seeing only the top one.

canvas#

• ESM3 generate, inverse-fold and embedding results show as sequence cards.
• GenMol results that arrive late as a CSV fill their card instead of leaving it empty.
• P2Rank cards load their predictions table.

studio#

• Tool panels have Cancel while a job runs, Retry or Restart after a failure or cancel, and Download when it finishes.
• Batches submitted from Studio run as a batch, and PDB2PQR takes its inputs from the Studio form.

chat#

• Ivy shows short progress notes between tool steps, collapsed on saved messages, so you can see what it is doing on a long turn.
• The chat page shows one loading screen when it opens.
• When a tool needs a higher plan or you run out of credits, an upgrade card links to plans.
• An answer that finishes in a background tab is saved when you close the tab.

tools#

• BoltzGen runs against a bare PDB id, and PDB2PQR, PROTAC and SolubleMPNN runs receive the inputs each tool reads.
• Workflow help suggests up to three possible next steps rather than one fixed chain.

sandbox#

• Structures you import appear in the sandbox workspace, so code you run there can open them.

cli#

• ivybloom run shows the real error when a submit fails, and a retried submit returns the same job instead of creating a second one.

Failed jobs explain their cause in plain language, and HTTP API callers get one clean answer, streamed or not. Lab-automation tools run in the chat sandbox.

jobs#

• A failed job's card shows the cause in plain language, so you can decide whether to change the input or try again.
• Batch status lists the error for each failed row.

http api#

• Streaming and non-streaming calls return the same answer text, once. A stream carries only answer text and ends on exactly what a non-streaming call returns.

tools#

• Octave, Opentrons protocol generation and PyLabRobot run in the chat sandbox, alongside your other code and files.
• ProteinMPNN keeps the structure you named through to the run.
• ClinPGx filters match on the variant fields in the data, and RxNorm lookups return the identifiers and ingredient concepts.

chat#

• A tool call that has already returned its result shows as finished, not in progress.
• The quick-action chips on the homepage show only actions a guest can run.

exports#

• Markdown chat exports label each source, so merged sources stay distinguishable.

site#

• Public pages have wider text columns, page titles keep their written casing, and prices group thousands.

A canvas card now shows when a job's result came back empty or unusable, and answers that contain SMILES strings or tables arrive whole.

canvas#

• A card for a job that returned an empty, placeholder or implausible result shows the failure message and a failed state, instead of a green status and a metric chip. You can tell at a glance which results to trust.
• The small preview a card shows when it first appears uses the same failed state.

chat#

• Lists of SMILES strings and markdown tables arrive intact. Short answers full of bond and ring notation are no longer replaced by a cut-off notice.
• An answer that is cut short gives the right reason for it.
• Long turns that chain several tools keep running when a background job reports back partway through.
• The end-of-turn summary counts the analysis that ran, and says when a turn was interrupted.

If you name a tool and the result came from a different method, the answer says so. A reply cut short by your next message is marked as partial.

chat#

• When you name a tool, it isn't run, and another method produces the result, the answer adds a note saying so. You know which method the result actually came from.
• A reply that ended early because you sent your next message is marked "reply ended early - partial answer preserved", so it doesn't read as a finished answer.
• Queued prompts are editable before they send.

file explorer#

• Chat cards stay steady while the list updates, so clicks land where you aim them.
• A chat card shows as selected after a single click.

tools#

• PROTAC and degrader requests take the fields the tool expects, so a request phrased in your own terms reaches the tool intact.

exports#

• Exported transcripts no longer repeat tool-call blocks.