AI readiness toolkit

A grab-bag release for once: one new skill, twelve fixes, and a unified excludes system

The unreleased v1.12.0 window covered /deslop plus three /assess bug bundles; v1.13.0 lands on top with the custom-exclude feature. Everything from v1.11.1 onward ships in this one tag.

New skill: /deslop (#45)

• Detects and removes the telltale signs of AI writing from prose - the em dashes, the "comprehensive analyses", the "tapestry of interconnected" filler. Ships a reference checklist and a SKILL.md that the agent invokes when asked to clean up AI-flavoured copy.

/assess custom excludes across every scan (#51, closes #50)

• Repos that intentionally track large non-source reference data (regulatory raw exports, vetted-context corpora, seed datasets) now have an escape hatch from the dominance warning that doesn't require forking the script or removing the data from git.
• Two mechanisms, both additive to the built-in defaults:
  • .assess/config.toml with top-level exclude_dirs and exclude_patterns lists. Recommended for durable per-repo excludes.
  • --exclude PATTERN on the treemap CLI (repeatable, auto-classifies glob vs dir name).
• The same excludes apply across every scan - the heatmap, the doc-navigability graph, the doc-staleness pass, and the dead-code/observability scan. A regulatory-raw/ exclude is one statement meaning "this is reference data, not source," not a per-layer toggle.
• .assess/ is self-excluded - prior runs' run-context.json (often 2,000+ LOC) used to surface as a top-large file on every re-run. Circular pollution gone.
• Missing or malformed config files degrade silently with type-guarded loaders; the assessment never blocks on a broken config.

/assess feedback bundles (#48, #49)

• Step 5 now gates the "open a PR?" offer on gh repo view --json viewerPermission,viewerCanPush so read-only targets get a fork-based PR offer or "leave local" with the reason named.
• Step 6a starts with the deterministic git remote -v + gh repo view --json hasIssuesEnabled,nameWithOwner signal before any soft-signal judgment - "no GitHub tracker detected" on a repo with 270 open issues was a real run.
• fd ... "$REPO_ROOT" in Steps 2a and 2b was passing $REPO_ROOT as the regex pattern, not the search path. Both Step 2a (scc-offer) and Step 2b (dead-code offer, added in v1.11.0) silently returned 0 file counts. Real v1.11.0 regression. Fixed with an inline comment so it doesn't return.
• doc_graph.py no longer counts link targets inside fenced code blocks or inline-code spans as dangling. FORMAT specs teaching link syntax and Obsidian-style skills demoing [[wikilinks]] used to inflate dangling_links. New _strip_code_spans() removes both before extraction. Plus seven more deterministic-core fixes documented in #49.

Both the Claude Code plugin and the standalone Desktop/Cowork ZIPs are updated (ZIPs on the standalone-skills-latest rolling release).

What's Changed

Features

• feat(assess): add --exclude flag and .assess/config.toml for custom excludes by @bjcoombs in https://github.com/bjcoombs/ai-native-toolkit/pull/51

Full Changelog: https://github.com/bjcoombs/ai-native-toolkit/compare/v1.12.0...v1.13.0