paqad-ai 1.75.3 → 1.76.0

package/CHANGELOG.md

@@ -1,5 +1,28 @@
 # paqad-ai
 
+## 1.76.0
+
+### Minor Changes
+
+- 2cbd0de: Begin the interactive visual site map (#466). The dashboard "Site map" area becomes an explorable, non-technical diagram of the app rendered statically from a single AI-authored YML at `docs/site-map/`, and creating the map is gated on the documentation family. This first slice lays the data foundation: an honest 5-level trust tier on every map element (surface, transition, guard, journey), a reader for the canonical `docs/site-map/` location, and a write path that stamps each element's earned trust tier into the stored map so the dashboard renders proven tiers statically with no work at view time. The `sitemap run` verb now wires that proof in: each run re-earns the canonical map's trust tiers and its map-vs-code freshness from the same code evidence it already resolves and writes them back when either moved, so the dashboard honesty strip and the freshness gate read earned proof instead of "not yet checked". The run now also reads the map and journeys it verifies from that same canonical `docs/site-map/` location, so it checks, resolves evidence for, and restamps exactly the map the dashboard renders rather than a separate copy. The one-step creation flow can now stamp each recorded answer's provenance back onto the surfaces it settled: a surface a person decided about reads as human-confirmed and a defaulted decision reads as a low-confidence guess, so the map is honest about which choices had a human behind them. That one-step creation flow now has a surface the agent can drive: `sitemap questions` lists only the closed-list questions the map still needs (with each question's plain reason, its `file:line` evidence, and a recommended default), and `sitemap answer` records the person's decisions to `docs/site-map/answers.yaml` and stamps their provenance onto the map in one step, so a settled human answer is never re-asked and a question whose code moved is reopened. The superseded report machinery is retired: a run no longer dumps timestamped reports under `docs/site-map/` or publishes derived index/overview/registry views under `docs/instructions/site-map/`, the `sitemap retest` verb and its dashboard action are gone (a re-run is simply the same run again), and the freshness gate now reads one deterministic signal, the map-vs-code freshness stamped into the stored map. The dashboard's Site map pulse section likewise reads the stored canonical map statically instead of the latest report dump. Everything stays inert while the `site_map` flag is off.
+
+## 1.75.4
+
+### Patch Changes
+
+- 92b4710: fix(#205): never count generated `.paqad/` artifacts as changed code in verification
+
+  When the runtime self-hosts, its generated artifacts land under a nested
+  `.paqad/` home (`runtime/base/.paqad/`) that the root-only managed gitignore does
+  not cover, so they leak into the working tree. The change-evidence classifier
+  treated anything under `runtime/` as code, so the change-completeness gate demanded
+  test evidence and doc updates for generated logs and evidence JSON. `isCodeFile`
+  now excludes every `.paqad/` home (root or nested) via the new
+  `isPaqadArtifactPath` predicate, and the verification changed-file scan strips
+  `.paqad/` artifacts entirely, so they never reach `code_changed`, the test-evidence
+  preview, the quality ratchet, or scope drift. (The invisible Stop-hook loop and the
+  silent exit-2 from the original report were already resolved by #303/#368.)
+
 ## 1.75.3
 
 ### Patch Changes