trajectory-sdk and import Client from trajectory.
Configuration
tj login to save a credential for Client(). Saved credentials are bound to
their API origin; changing a copied client to a different origin drops inherited
credentials. See Authentication for profiles and unattended jobs.
Resources
trajectory.lib provides two forms of the same spec upload workflow:
push() calls start_push() and waits with operation.result() internally. Both accept
the same arguments. start_push() is a regular Python function; background processing
happens on the server. See the spec guide.
To edit tasks within an existing spec version, use client.benchmarks.tasks.append(...)
followed by client.benchmarks.tasks.commit(...) with that version’s spec ID.
Training and evaluation
Both resources acceptcreate(agent_name=..., benchmark_name=..., base_model_slug=..., options=...)
to select the latest spec version under that agent. Pass bench_id instead of the name
pair to select a specific version. Finish a baseline/checkpoint comparison before publishing
another version under the same name.
Use parent_checkpoint_id to resume training or evaluate a checkpoint. Request
options dictionaries follow the shared RunOptionsParam type; run responses expose
typed RunOptions values. Omitted, null, and empty options preserve defaults.
Discover models and options
Before uploading a spec, useclient.training.list_models() or
client.evals.list_models() to choose a supported model and estimate token costs.
Each response’s items contains base_model_slug, model_name, and
pricing.input / pricing.output in USD per million tokens. Training and evaluation
can have different models and prices; use the catalog for the mode you plan to run.
See pricing for training, grading, sandbox, and other charges.
client.training.list_options(...) or client.evals.list_options(...) to
discover supported settings for that mode, optionally filtering by base_model_slug.
The response’s models contains base_model_slug and an options map with types,
bounds, descriptions, and fixed or computed defaults, without pricing. Training and
evaluation may have different bounds for the same setting. list_options() also
works before a spec exists: bench_id is deprecated and ignored.
Training also accepts options={"load_weights_only": True} with parent_checkpoint_id,
even though this option isn’t listed. This loads checkpoint weights with fresh training
state; see resume from a checkpoint.
Run costs
Useclient.billing.list_run_costs() to read run costs:
offset and a
limit of 1–50 (default 20).
Costs may take several minutes to appear after a run finishes and can change as
usage is processed. Charges for deleted runs remain listed under the run ID until
they fall outside the 30-day window.
Pagination
.items is the current page. Iterating the page fetches the rest.
.has_next_page() and .get_next_page() request the next page yourself.