0.4.0-rp

AXP 0.4 now leans heavily into remote runs on the platform: runs can be dispatched, monitored, canceled, and cleaned up; uploads and derivations are scoped per run; large job specs are staged safely; and remote execution is the default path. The release also adds local querying and result access, improves experiment results views with version-aware grouping and filtering, and updates onboarding and docs to better explain the first-run experience.

CLI

• axp local query now queries local derived tables, replacing the old local use of axp query.
• axp query is available for querying platform tables through the proxy.
• axp experiment results get <experiment id> is available for fetching experiment results.
• axp update is available to update AXP.
• axp intro now provides a guided onboarding flow for first-time users.
• axp-build-loop is available as a workflow skill to drive the experiment build loop from a coding-agent chat.

See the AXP CLI reference for the current command surface.

For most remote-run workflows, the loop now looks like this:

axp run experiment.yaml
axp runs status <run-id> --watch
axp query "select * from runs limit 10"

Platform

• Remote runs now use the platform by default, with dispatch, sandbox orchestration, progress/logs, cancel support, fan-out, concurrency caps, and cleanup for orphaned runs. See Runs.
• Per-run uploads and derivations now stay isolated, so each run keeps its own identity from upload through results.
• Local WIP files can now be staged into variant sandboxes before setup runs. See Experiment YAML.
• Large job specs are now staged to a file instead of being passed through the environment, avoiding dispatch failures on bigger experiments.
• Remote sandbox usage is tracked for billing.
• Sandbox timeout behavior was re-evaluated.
• axp run now writes artifacts locally and uploads them automatically.
• Platform SQL proxy access now includes Source, Facts, and Data Products tables.
• get-result support was added on the server, along with the MCP result tool and run result link handling.

Web

• Experiment results now show version-aware grouping, better summary metrics, and richer drill-down into runs and variants. See Results.
• Results filtering and grouping now support variant and run ID controls with table and chart updates.
• Winning-variant selection now follows completion rate, tokens, and time.
• Experiment versioning now segments runs by the exact experiment definition used at execution time. See Experiments.

Documentation

• The getting-started experiment docs now explain what the experiment tests and use a clearer prompt. See Getting Started.
• Docs now support version-gated pages with selective per-page forking.
• User-facing docs were migrated to the new showcase format.