learning-context

name	learning-context
description	Code-generation context for metaoptics inverse-design MCE runs.

Skill Overview

Use rollout telemetry to improve inverse-design code-generation context and reduce repeated failures.

Function Contract

The agent writes a Python function:

def solve_inverse_design(*, device: str = "cuda") -> SolverResults:

• Input: keyword-only device parameter (string, defaults to "cuda") specifying compute device placement.
• Output: torchrdit.results.SolverResults object from solver.solve(source).
• The function calls TorchRDIT raw solver/builder APIs directly.
• The query context is provided in the prompt but not passed as a function parameter.

Working Rules

1. The agent produces a complete .py file via the write_file tool.
2. The host runner imports solve_inverse_design from the written file and executes it with an explicit device parameter.
3. When executing Python scripts or modules in this project, always use uv run python ... instead of bare python ... so commands run inside the repo-managed environment.
4. Scoring is deterministic: evaluate_gt_eval() evaluates Python lambda expressions against the returned SolverResults.
5. Treat success_goal as primary metric; use success_exec to isolate runtime failures.
6. Always call .with_device(device) early in the builder chain to ensure all solver tensors are placed on the correct compute device.
7. TorchRDIT provides local gradient-based optimization; for global optima, the agent should write explicit global exploration code (for example multistart or population screening) and then locally refine top candidates.
8. Keep context concise and actionable; remove duplicated or contradictory guidance.

Improvement Loop

1. Group failures by error type (import errors, shape mismatches, solver crashes, goal misses).
2. Add solver stability hints when execution succeeds but goal metrics are not met.
3. Preserve patterns associated with success_goal=true and avoid broad rewrites.
4. Keep guidance minimal and executable; avoid references to nonexistent utilities or helpers.

Working with Learned Context

The context/ folder contains accumulated knowledge from previous training iterations:

• rules.txt - Actionable rules for handling common failure patterns
• analysis.md - Root cause analysis from training data
• examples.json - Examples of successful patterns

Before implementing interfaces:

1. Check what context files exist
2. Read relevant guidance
3. Build upon existing knowledge rather than starting from scratch
4. Update context based on your analysis of train.json