Skip to content

Latest commit

 

History

History
39 lines (29 loc) · 2.17 KB

AGENTS.md

File metadata and controls

39 lines (29 loc) · 2.17 KB

Project agent memory

This file is the project's committed home for project-intrinsic agent knowledge: build, test, release, architecture, and sharp-edge notes that should travel with the code.

Checks

cargo fmt --all, cargo clippy --workspace --all-targets (CI treats warnings as errors) and cargo test --workspace must all be clean. CI additionally runs the test suite with network egress blocked — crates/reify/tests/offline.rs fails the build if a networking crate enters the dependency tree.

The standard this repository is held to

crates/reify-bench is the most load-bearing thing here, and its value is its intellectual honesty rather than its numbers: steel-manned baselines, falsification conditions written down before the run, Wilson intervals reported next to the admission that they overlap, provider failures excluded rather than scored as misses, and one rule from metrics.rs — a metric that cannot be defined precisely does not get reported. Metric definitions live in docs/metrics.md. Reports are generated, never hand-edited; a number that was not re-measured is corrected with a dated note rather than silently regenerated (see the top of benchmarks/REPORT-medusa.md).

When a fitted parameter fails held-out validation, the fit is published and the default reverts — HISTORY_PRIOR_WEIGHT in crates/reify/src/context.rs is the worked example.

Decisions already measured

reify verify — a post-flight check reporting what an agent's patch missed — was measured before being written and failed its pre-registered condition on Rust, Python and Go. benchmarks/REPORT-verify.md has the numbers; reify-bench verify-eval reproduces them in about two minutes with no model. Do not build it on the CALLS graph alone without re-running that benchmark and beating those numbers.

Maintaining this file

Keep this file for knowledge useful to almost every future agent session in this project. Do not repeat what the codebase already shows; point to the authoritative file or command instead. Prefer rewriting or pruning existing entries over appending new ones. When updating this file, preserve this bar for all agents and keep entries concise.