AI readiness toolkit

/assess: closing the loop on its own feedback

This release acts on #39 — six refinements surfaced by a v1.10.0 run of /assess against this repo — plus a tempered report-prose fix from #37. The headline is the new install-offer flow that lets the Layer 1 dead-code scan resolve its own tool_absent state without the user having to know which tool fits each language.

Per-language dead-code install offer (new Step 2b)

• When the Layer 1 scan reports tool_absent for vulture / ts-prune / staticcheck, /assess now mirrors the Step 2a scc-install pattern: detect the languages present, ask Install / Skip-now / Skip-permanently per absent tool, and write .assess/.no-<tool> markers for permanent declines. Multi-language repos get a single batched AskUserQuestion call rather than serial prompts.

Layer 0 + 1 refinements driven by the v1.10 self-assessment

• stale_hubs.confidence — hubs whose subject_method is repo-baseline (ratio computed against repo-wide churn — a coarse proxy shared across every baseline entry) now carry confidence: "low". The sort de-weights low-confidence priority by 0.5 so a precise-subject hub at half the raw priority of a baseline one still outranks it.
• base_doc_coverage is now size-weighted — counts code files under a maintained base doc, not directories. Nested internal/ / adapters/ dirs no longer drag the headline toward zero on real monorepos. The un-weighted base_doc_dir_ratio is reported alongside so the denominator stays auditable.
• Layer 1 row template caveat — the Evidence cell now requires the agent-environment caveat when scoring Present (rung 3). Reachability from the repo is conditional on the cited CLIs being in the agent's execution context; the row says so by default.
• Score derivation worked examples in Step 4 — four worked cases anchor the math (raw sum capped at 8 to match the 0-2 / 3-4 / 5-6 / 7-8 band table), including the 7 Present + 2 Partial → 7.0 / 8 case the v1.10 run hit.
• finalize-input.json is transient — written to .assess/.cache/finalize-input.json and deleted on success. No more noisy diffs from a one-off LLM input. Legacy in-tree path still read for backwards compat and cleaned up if present.

Also in this release

• Report-prose fixes (#37): navigability now measures curation (wayfinding) rather than access (an agent can always ls docs/); Layer 0 prefers Partial over Missing when a committed passing instruction file sits beside an advertised-but-broken reference; the deterministic, low-token nature of the assessment is now explicit in the report and README.
• Refreshed example SVGs (#40) generated from a v1.10 run, including the new doc-navigability graph.

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 v1.11.0 - 6 refinements from v1.10.0 feedback by @bjcoombs in https://github.com/bjcoombs/ai-native-toolkit/pull/41

Fixes

• fix(assess): temper overstated report claims and surface deterministic approach by @bjcoombs in https://github.com/bjcoombs/ai-native-toolkit/pull/37

Documentation

• docs: refresh example SVGs from v1.10 + add doc-navigability graph by @bjcoombs in https://github.com/bjcoombs/ai-native-toolkit/pull/40

Full Changelog: https://github.com/bjcoombs/ai-native-toolkit/compare/v1.10.0...v1.11.0