decisions.jsonl

{"id":"DEC-001","title":"Accept explicit equivalent benchmark trees","status":"accepted","ts":"2026-09-02T06:02:34Z","context":"A single exact expected tree produced false negatives when authored HTML visibly justified an extra stat-card or download-card group. Ignoring group wrappers globally would hide genuine missing-card regressions in pricing, feature, and bento layouts.","decision":"A benchmark spec may list a small number of explicit acceptedTrees. Each entry is a complete reviewed tree, is available to every producer and engine, and is scored with the same structure and attribute assertions as the canonical tree. Producer-specific branches and global wrapper forgiveness are prohibited.","consequences":"Source-driven equivalent structures can score correctly without weakening strict container requirements elsewhere. Review pages must display every accepted tree, suite hashes must include them, and fixture authors must justify each alternative as contract-permitted ambiguity rather than use it to rescue one model output.","supersedes":null}
{"id":"DEC-002","title":"Use one canonical cross-harness skill bundle","status":"accepted","ts":"2026-09-02T06:23:28Z","context":"The shipped skill source lived under a parent directory that did not match its frontmatter name, while the installer defaulted only to Claude Code despite the guide being harness-neutral. Current clients converge on .agents/skills, but Claude Code still documents .claude/skills, and unknown harnesses may use arbitrary roots.","decision":"Keep one specification-valid source at skills/block-runner with detailed material under references. The skill installer recursively copies that bundle to project-scoped .agents/skills and .claude/skills by default, supports explicit target, scope, and directory overrides, and never detects models or calls an LLM. Installed instructions pin the invoking package version; managed hashes permit safe updates while local or unmanaged changes require --force.","consequences":"One npm artifact can serve current cross-agent clients, Claude Code, arbitrary skill roots, and stdout-only harnesses without vendor-specific source forks. The two default installed copies may duplicate bytes, but installation is explicit and deterministic; symlink projection, plugins, lockfiles, and marketplace packaging remain deferred until they solve a demonstrated distribution need.","supersedes":null}
{"id":"DEC-002","title":"Use one canonical cross-harness skill bundle","status":"accepted","ts":"2026-09-02T06:40:00Z","context":"The shipped skill source lived under a parent directory that did not match its frontmatter name, while the installer defaulted only to Claude Code despite the guide being harness-neutral. Current clients converge on .agents/skills, but Claude Code still documents .claude/skills, and unknown harnesses may use arbitrary roots.","decision":"Keep one specification-valid source at skills/block-runner with detailed material under references. The skill installer recursively copies that bundle to project-scoped .agents/skills and .claude/skills by default, supports explicit target, scope, and directory overrides, and never detects models or calls an LLM. Installed runtime commands pin the invoking package version while the explicit update command stays on @latest; managed hashes permit safe updates while local or unmanaged changes require --force.","consequences":"One npm artifact can serve current cross-agent clients, Claude Code, arbitrary skill roots, and stdout-only harnesses without vendor-specific source forks. The two default installed copies may duplicate bytes, but installation is explicit and deterministic; symlink projection, plugins, lockfiles, and marketplace packaging remain deferred until they solve a demonstrated distribution need.","supersedes":null}