v2.0.x

​

API & Services

• Files can now be addressed by their fil_ id: New endpoints look up a file by its TypeID instead of its filename: GET /api/files/{fil_id} returns its metadata, GET /api/files/{fil_id}/content downloads it, and DELETE /api/files/{fil_id} deletes it. A new GET /api/files lists your files. The existing endpoints that look up files by filename are still available.
• File downloads support byte ranges: Both download endpoints now honor a single Range: bytes= header (bytes=a-b, bytes=a- or bytes=-n) and answer 206 with Content-Range. A range that selects no bytes returns 416, and multi-range or malformed headers get the full file with 200. Downloads now carry Accept-Ranges: bytes and a strong ETag, and If-Range is honored so a resumed download can’t mix two versions of a file.
• Downloading a file that isn’t available yet now returns 409: A download of a file that is still uploading, corrupt, or has no chunks yet returns 409 instead of 400. The error code stays file_not_available.
• Filenames are limited to 512 characters: Every upload path now rejects a filename longer than 512 characters with 400. Before, the simple upload paths failed with 500, and direct-upload initiate accepted names up to 1024 characters that could never be stored.
• New endpoint lists a fine-tuning job’s input files: GET /api/fine-tuning/jobs/{id}/inputs returns the job’s training and validation files, newest first, with cursor pagination. Filter with role and narrow by file_id with a case-insensitive search.
• Batch registry lists now use the standard pagination envelope: GET /api/batch/registry/pipelines, GET /api/batch/registry/components and their /key/{key}/versions lists now return {data, has_more, next_cursor, prev_cursor}. The default limit is now 100 instead of 50. Breaking change: the pipelines and components arrays are renamed to data, and the limit field is no longer returned in the body.
• The Python client now uses the /api/<domain> routes: The default endpoint is now https://platform.u1.archetypeai.app/api. Breaking change: endpoints in the old format, such as https://api.u1.archetypeai.app/v0.5, no longer work, and the files_endpoint, lens_endpoint and agent_endpoint overrides are removed. The capabilities, sensors, messaging, kafka and files.s3 APIs are removed, and client.lens is deprecated with a FutureWarning.
• The Python client supports packages: client.agents.packages adds create, list, get, delete and wait_until_ready. client.agents.instances.create runs a ready package as a batch or serving agent, wait_until_running waits for a serving agent to start, and query sends it a payload. wait_until_done now raises an error for a serving agent, which never completes.
• New Python client examples for RED and Anomaly Discovery: examples/agents/agent_red_lifecycle.py (pump sensor data) and examples/agents/agent_ad_lifecycle.py (Tennessee Eastman Process data) each optimize the canonical blueprint, promote the best trial, eval it, then bundle and run it.

​

Agents

• Projects group blueprints, evals, optimizations and packages: The new /api/projects endpoints create, list, get, update and delete projects. A project is visible to its creator and to organization admins, and viewers have read-only access. Blueprints can be registered in a project or copied into one with POST /api/agents/blueprints/{reference}/copy. Evals, optimizations and packages made from a project blueprint land in its project, and their list endpoints accept a project_id filter. Agent runs inherit the project of the package they run, report its project_id, and can be filtered by it on GET /api/agents/instances. An eval, optimization or run in a project may only read input files the project’s creator can see, and any other input file returns 400.
• Packages are the new way to deploy a blueprint: A package is an immutable build of a blueprint that resolves every model slot once. Create and manage packages with POST /api/agents/packages, GET /api/agents/packages and GET or DELETE /api/agents/packages/{package_id}. Run a ready package as a batch or serving agent with POST /api/agents/instances, and list its runs with the new package_id filter on GET /api/agents/instances.
• Serving mode is now available to agents: When starting an agent, you can now specify that it should start in batch mode (process the inputs then end) or serving mode (which stays active to answer queries). An agent in serving mode lasts as long as the agent exists.
• Anomaly Discovery can be optimized: The canonical ad blueprint’s detector is now fit by an optimization from normal-only training examples, and threshold, n_neighbors, metric and pca_components are tunable.
• RED is fit by optimization and declares its metrics: The canonical red blueprint now has a metrics section, and its classifier is fit by an optimization instead of shipping pre-fit. Optimizations must set its new events and background values, and its metric is tunable.
• New objectives score detection against a background class: When a category target names a background label, foreground_macro_f1, detection_recall and detection_precision are available as objectives and as the primary metric. Binary targets get them too, with the non-positive class as the background.
• Eval and optimization lists can be sorted: GET /api/agents/evals and GET /api/agents/optimizations take a sort of created_at (the default), name or score, and an order of asc or desc. Rows with no score sort last. The after and before cursors are now opaque and tied to the sort they came from, and replaying one under a different sort returns 400.
• A blueprint’s source says how it was derived: source now has a kind field: promotion for a blueprint promoted from an optimization trial, or copy for one copied into a project. The promotion fields are unchanged.
• Optimization trials run in parallel: Trials that don’t share encoder settings now run across up to eight jobs at once, instead of one after another in a single job.
• Blueprints can pin a model by its mdl_ id: A model slot in a blueprint or bundle can now be a model registry mdl_ id, as well as a registry tag or an aar_ artifact ref. A pinned id keeps resolving after its tags are archived. Canonical blueprints and bundles still require tags.
• Fine-tuned Newton-C LoRA models can be served by id: A blueprint can name a fine-tuned Newton-C model (variants newton-c-2.5-8b-lora and newton-c-2.5-2b-lora) by its mdl_ id or a tag that points at one. The platform resolves the base model and LoRA adapter and merges them at load. Naming a LoRA model in a slot whose node can’t apply an adapter returns 400.
• Optimizations take targets overrides and a seed: targets overrides work the same way they do on eval requests, and the response includes resolved_targets. An optional non-negative seed makes trial sampling reproducible for the same search space.
• Evals and optimizations can score per-class precision, recall and f1: These objectives take a class parameter, for example { name: f1_bearing, f1: { class: bearing } }, and can be used as the primary metric. An eval’s per-target objectives override now replaces the blueprint’s list instead of only narrowing it.
• Examples can hold several input files, and blueprints can have several targets: The new per_input ground-truth form labels each input file of an example. A blueprint is no longer limited to one target.
• Prediction files map each line back to its inputs: Each line of an eval’s predictions file now has an input field naming the input it came from and a keys field with the predicted and expected pairing keys. The abstained field is removed; a prediction with no value now has predicted set to null.
• Eval results appear while the eval runs: Per-example reports are now saved as each example completes. Partial results are visible during a run and are kept if the eval fails late or is cancelled.
• One bad example no longer fails a whole eval: An example whose inputs or ground truth can’t be read now fails on its own, with a failed status and an error, and is left out of the pooled metrics. The eval fails only if every example fails. In an optimization, a failed example still fails its trial.
• New canonical blueprint, omega-embeddings: It produces Omega embeddings from multi-channel numeric time-series data. It runs in batch mode only.

​

Console & UI

• Evals and optimizations have views for detection agents: The eval and optimization pages now show results for Anomaly Discovery and RED agents.
• Optimization trials open in a side panel: Selecting a trial opens its results and config in a side panel, where you can promote it. The new optimization form also has an optional seed field.
• The fine-tuning job’s Training files tab loads from the server: The tab uses the new inputs endpoint for paging, with role and search filters.
• File Manager asks before deleting a file: The delete button now opens a confirmation dialog.
• The login page has a new design.
• API-key logins show N/A as the role: The sidebar profile no longer shows Viewer for every API-key login, because the role isn’t available for those sessions.

​

Bug Fixes

• Cancelling an agent run frees its slot right away: Cancelling a run now also cancels its backing job. Before, an orphaned job could keep holding a slot of the organization’s concurrency limit, so new runs were skipped.
• A batch run with no source file is refused: A batch run whose connectors.source has no file ref now returns 400. Before, it was accepted and stayed running for hours.
• Ground-truth labels for the canonical AD blueprint are no longer refused: Optimize, eval and training jobs refused constant and per_input labels such as normal when a target’s labels held ${values.*} references, as the canonical AD blueprint’s do. These labels are now checked after the references are filled in.