opendfm/xcientist/resource-acquisition

Resource Acquisition

Mission

Prepare real, verifiable resources for the current experiment.

CRITICAL: Code in `project/` must be SELF-CONTAINED

• The code in project/ must be COMPLETE and INDEPENDENT.
• It must NOT depend on repos/ for core functionality.
• Repos are for REFERENCE ONLY - implement actual experiment code in project/.

Protocol

• Read idea.json or agent_reports/prepare/artifacts/idea.md first.
• Discover repositories, datasets, and benchmarks before writing documentation.
• Treat pre-existing directories as unverified until file checks pass.
• Return machine-readable status in the final worker JSON; the runtime persists the stage worker report.
• Record an explicit real experiment target inventory in the managed prepare artifacts and final worker JSON:
• exact models/checkpoints or API-backed models
• exact dataset files/directories
• exact benchmark entrypoints
• exact API env vars and invocation surface
• an explicit description of what later phases must run
• If the real experiment path requires local model checkpoints, acquire them during prepare rather than deferring them to science.
• The dataset acquisition step may also download model checkpoints when those checkpoints are part of the formal benchmark target.
• Create agent_reports/prepare/artifacts/idea.md only after verification work is complete, not as an early summary artifact.
• For HuggingFace downloads, use the official HuggingFace endpoints and inherit the current shell network environment instead of rewriting proxy settings inside the agent.
• agent_reports/prepare/artifacts/idea.md must use these EXACT headings:
• ## Idea Summary
• ## Idea JSON Components
• ## Code Implementation Guidance
• ## Component Correspondence
• ## Dataset Usage Guidance
• ## Environment Variable Usage Guidance
• ## Resource Acquisition Log
• ## Repository-to-Dataset Mapping
• ## Real Experiment Targets
• ## Canonical Idea Components
• Under ## Canonical Idea Components, copy every idea.json.components entry exactly once, in the same order, preserving the component name and explanation.
• After writing agent_reports/prepare/artifacts/idea.md, reopen it and confirm those headings exist before claiming completion.

Report Requirements

The runtime owns reports. Do not write report files with tools. Each stage keeps worker/review/hook attempts under agent_reports/prepare/{worker,review,hook}/<stage>/.

Required Outputs

• Final worker JSON response requested by the planner; the runtime persists the worker report.
• Verified repository list with local paths
• Verified dataset list with expected_paths
• Verified local model/checkpoint list when the formal path depends on local models
• Verified benchmark list with runnable entrypoint or explicit blocker
• Real experiment target declarations with concrete model/API/data choices
• agent_reports/prepare/artifacts/idea.md with the required headings and verified content
• Canonical component handoff copied from idea.json.components

Failure Conditions

• Resource directories exist but no expected files are checked.
• A benchmark is claimed runnable without an entrypoint or smoke command.
• A dataset is claimed downloaded without explicit file evidence.
• Model/API/data targets are described vaguely instead of concretely.
• An API-backed dependency is needed but no usable invocation surface is documented.
• agent_reports/prepare/artifacts/idea.md is only a generic project summary and does not contain the required protocol headings.
• agent_reports/prepare/artifacts/idea.md omits, renames, duplicates, or reorders the canonical idea components.
• The agent prints completion before performing a final read-back validation of agent_reports/prepare/artifacts/idea.md.