paqad-ai 1.85.0 → 1.87.0

package/CHANGELOG.md

@@ -1,5 +1,38 @@
 # paqad-ai
 
+## 1.87.0
+
+### Minor Changes
+
+- 533cf1b: Require and evidence rule-loading for feature-development changes (#557).
+
+  paqad enforced that the process ceremony ran (plan, spec, review, checks) but never that the
+  project's rules were actually loaded — so an agent could complete a fully green change with
+  every non-scripted rule unread, and code written on a non-feature route loaded no rules at
+  all. Rule-loading is now a required, deterministic (no-LLM) part of any feature-development
+  code change:
+
+  - **`paqad-ai rules load`** prints the full text of the rules that apply to your changed
+    files and writes `rules-loaded.json` into the feature bundle (which rules apply to which
+    files, plus a content hash of the loaded text). It works on any route.
+  - **Edit-time:** a new `rules-loaded` kernel capability blocks the first feature-dev source
+    edit until that record exists — catching code written on a non-feature route at edit time.
+  - **Completion:** a new gate fails a feature-dev change whose applicable rules were never
+    loaded (the same way a missing plan or spec fails) and reads Inconclusive when the load is
+    stale.
+
+  No new config knob — rule-loading is required, gated only by paqad being enabled and the
+  change being feature-development. The record attests that rules were loaded and acknowledged,
+  not that they were comprehended.
+
+## 1.86.0
+
+### Minor Changes
+
+- 13efb31: Make the feature-development `checks` stage fast without lowering the bar (#554, closes #505).
+
+  The check runner now executes each mapped command as a quote-aware `&&` argv chain (in-process `mkdir -p`, no shell interpretation), captures stdout, and parses the test command into per-test results — so a profile command like `mkdir -p .paqad/test-results && ./vendor/bin/pest …` can no longer report green without running a test. The test suite runs with the runner's own parallel mode when the project has it (detected per runner from the stack pack and the ecosystem lockfile, recorded in the project profile and the stack doc, with a sequential fallback for small machines and harness failures — paqad installs nothing). Check commands are overlapped by a scheduler (formatters first, build before test, everything else concurrent), and every failure under parallel is confirmed alone before it blocks: fails-alone blocks with test name, file and line; passes-alone is quarantined in the flaky registry and never blocks. New `checks_parallel`, `checks_max_processes` and `checks_flaky_under_parallel` knobs, a doctor check, a `checks plan` / `checks record-runner` verb pair, and a Checks section on the feature report page. Everything at runtime is deterministic; no LLM call from Node.
+
 ## 1.85.0
 
 ### Minor Changes