v1.96.0

📦 Standalone skill ZIPs (claude.ai, Claude Desktop, Cowork, any Agent-Skills assistant): https://github.com/bjcoombs/ai-native-toolkit/releases/tag/standalone-skills-v1.96.0

Highlights

/assess now checks the things recent research and agent-vendor guidance say matter most for coding agents. All new checks are deterministic and add no model tokens to a run.

• Nested and path-scoped instruction files (#518): subdirectory CLAUDE.md / AGENTS.md / GEMINI.md, .claude/rules, Copilot .instructions.md and Cursor .mdc files are found and graded, with the instruction budget each tool loads at start-up. Dead globs, ignored Cursor .md files and a CLAUDE.md shadowing a different AGENTS.md are reported.
• Instruction files graded on quality, not volume (#517): credit for commands that resolve to real scripts, targets or standard CLIs; a size curve from 200 lines; penalties for overviews, directory trees, README copies and stale references.
• Committed agent config that runs code (#516): an inventory of hooks, MCP servers, env and apiKeyHelper, plus findings for tracked secrets, a committed settings.local.json, broad Bash allows, devcontainer credential mounts, hidden Unicode and comment-triggered deploys. Secret values are never printed.
• Agent working environment in Layer 5 (#520): committed setup, a named check command, lockfiles per manifest and CI duration.
• Hotspot formula unchanged (#519): a co-change breadth term was tested on three repos and did not hold across churn windows; see skills/assess/references/hotspot-cochange-experiment.md.

Upgrade notes

Scores can move on an unchanged repo. These are method changes, not regressions:

• Instruction-file grades can drop. Delegating to skills no longer waives the size penalty (it halves it), and a file over about 400 lines or 4,800 words is capped at B (#517).
• Layer 5 can drop to Partial for a repo with CI but no instruction-file-named check command, or with neither committed setup nor a lockfile for every manifest (#520).
• run-context.json gains nested_instructions, agent_harness and agent_environment; artifact_schema_version is 1.4.0.

This repo applied the new rules to itself: uv.lock files and a Makefile with make setup / make check (#521), and CLAUDE.md cut from 5,202 to 1,760 words (#522). Its score holds at 7.5 / 8 (#523).

What's Changed

Features

• feat(assess): Discover nested and path-scoped agent instruction files by @bjcoombs in #518
• feat(assess): Report committed agent config that executes code or exposes secrets by @bjcoombs in #516
• feat(assess): Grade instruction files on verified commands and size, not directive counts by @bjcoombs in #517
• feat(assess): Score the agent working environment within Layer 5 by @bjcoombs in #520

Documentation

• docs(assess): Record the co-change breadth hotspot experiment and keep the formula by @bjcoombs in #519
• docs: Slim CLAUDE.md to a map of commands and invariants by @bjcoombs in #522
• docs: Codebase assessment after the instruction-file and agent-environment checks - 2026-10-11 by @bjcoombs in #523

Other Changes

• chore: Add uv lockfiles and a Makefile with setup and check targets by @bjcoombs in #521

Full Changelog: v1.95.0...v1.96.0