Skip to content

Latest commit

 

History

History
38 lines (24 loc) · 7.22 KB

AGENTS.md

File metadata and controls

38 lines (24 loc) · 7.22 KB

Repository maintenance

Layout

  • skills/numericbridge/ is the entrypoint for the full workflow. It owns stage transitions and handoffs and links to the installable framework-adapt, precision-compare and precision-repair skills for execution methods. Load those methods on demand. Export the suite with python -B -m scripts install --output NEW_DIRECTORY to bundle the shared tools/ once in Compare. Reuse workload execution, verified resources, inputs/state, collectors and results. skills/operator-optimize/ is a reserved placeholder with no implementation.

  • research/ is retained local RSI research, Git-ignored and excluded from both the public repository and normal Skill installation. Preserve its local code, state and recorded experiments. It owns task/iteration orchestration over existing workflow tools and an external harness. Index successful/incomplete task evidence; keep full traces and datasets private. The independent improver builds candidates. Evaluate exact tool/context/procedure bundles on independent task families before activation; label dataset outcomes at task scope. Pin task versions, preserve raw usage and reuse comparable baseline receipts. Native checks remain task-owned. Agent-weight training and operator search are extension points.

  • skills/framework-adapt/ owns native framework deployment and model/backend adaptation. Keep the entrypoint concise; retain environment declarations and framework-specific implementations with their projects and cases. Reuse comparison tools when numerical validation is needed.

  • skills/precision-repair/ is the installable repair skill. Its entrypoint owns correctness evidence, causal diagnosis, patching and original-workload verification. Reuse comparison tools from skills/precision-compare/; keep framework-specific fixes and experiments with their cases.

  • skills/precision-compare/ is the installable comparison skill. Keep routing and common collection steps in SKILL.md, the three task recipes and comparison.md in references/, and artifact interfaces in script docstrings. Optional internal sampling and its msProbe capability summary live in references/layerwise.md. Task recipes own capture requirements, framework integration and metric selection. Define shared training stages in SFT, distribution metrics in inference, and common coordinate/metric/tool rules in comparison.md.

  • Maintain matching English README.md and Chinese README.zh-CN.md. Technical documentation and maintenance instructions use English. Follow the Agent Skills specification.

  • skills/precision-compare/examples/ contains concise runnable artifact examples. Extract task-independent operations into composable tools with explicit interfaces and meaningful regressions; keep task-specific settings in adapters and environments.

  • tests/ contains retained local developer regressions, is Git-ignored, and is not distributed in the public repository or skill export. Preserve these local files. Public CI uses the supported validation and smoke-install commands. Artifact tools require NumPy; the target-score comparator uses the standard library.

  • Private projects/<project>/environments/ owns reusable environment declarations. Cases own configuration, collectors, tests and execution artifacts. Keep case-specific code there; extract shared code when reused.

  • Keep current guides focused on task semantics and tool usage. Work logs and runs/ own dated decisions, one-off execution failures and results; resolve tool defects in code and regressions.

community/records/ owns public record and relation validation independently of RSI. scripts/ owns reusable collection, integration, export, publication and installation tools; .github/ owns GitHub platform configuration only; docs/ owns current guides and reports/ owns final reports and scientific figures. Maintenance tools run as modules; retired .github/*.py commands are documented in scripts/README.md. Follow docs/publication.md: publish final results and minimal reproducibility evidence; keep intermediate execution artifacts and duplicate report exports local. Do not introduce research dependencies into community/shared tools or restore bare script imports via sys.path.

community/site/ hosts the GitHub Pages evaluation and repair histories. Append evaluations to community/site/data/evaluations.json with peer configurations, comparison scope, measured metrics, coverage and immutable evidence. Keep repair experiments in community/site/data/repairs.json with their existing library/fix/verification hierarchy. Preserve recorded IDs, measurements and evidence; keep full execution artifacts in private cases. Use community/site/data/relations.json for explicit, evidence-bound links between comparisons, investigations, issues, candidates and original-workload verifications. Do not infer historical causal links. Run python -B -m community.records validate and python -B -m scripts validate-immutability --base origin/main for history changes; the second rejects edits or removals of recorded IDs, measurements, dates and evidence files.

The RSI dashboard lives in community/site/rsi.html. After hourly maintenance, refresh its allowlisted snapshot with python -B -m scripts rsi-snapshot --monitor projects/rsi-system/monitor.json --output community/site/data/rsi.json --refresh and publish changed public summaries. It reads frozen runs without modifying them. Preserve unknown usage, separate reused baseline costs, show observation timestamps, and keep traces and connection details private.

Changes

Use the configured, authorized human identity for commits. Do not add AI tools or assistants as commit authors or co-authors, or list them as contributors in project materials. Check authorship and co-author trailers before pushing.

Confirm the approach before editing and change at most three files per batch. For a bug fix, reproduce the behavior temporarily, add a meaningful regression and remove the temporary reproducer. Preserve permanent tests.

Create new runs exclusively as <side>-<task>-<UTC YYYYmmddTHHMMSSZ>, using the task's side ID or compare. Keep model/data/environment resources in place and preserve historical run bytes and paths. Use a new run for later validation. For undated migrations, use <side>-<task>-migrated-<UTC migration timestamp> and retain the path mapping.

For skill/no-skill benchmark iterations, reuse a verified no-skill baseline while the task and model/runtime conditions remain comparable; rerun only the skill arm. Keep baseline artifacts outside evaluator inputs, record the reused baseline identity, and retain raw usage for consistent repricing.

Validation

Install maintenance dependencies with python -m pip install -r requirements-dev.txt. Run python -B -m scripts validate and python -B -m community.records validate from the repository root. When the retained local tests/ suite is present, also run python -B -m unittest discover -s tests -v; a public checkout does not contain it. With Torch installed, run python -B -m scripts smoke-install to verify clean skill installations, companion links and the CPU gradient example. Validate skill metadata, local links/anchors and changed script behavior. For documentation edits, check documented fields and commands against their owning scripts.