paqad-ai 1.83.0 → 1.85.0

package/CHANGELOG.md

@@ -1,5 +1,86 @@
 # paqad-ai
 
+## 1.85.0
+
+### Minor Changes
+
+- e5e56fc: feat(#551): PR visual evidence — screenshots of documented flows as feature-bundle evidence
+
+  When a feature-development change touches frontend files, paqad now captures ordered
+  screenshots of the documented user flows the change affects, pairs each with a
+  business-language caption, stores them in the feature's git-ignored evidence bundle
+  (latest-run-wins), assembles an overview GIF, and verifies the evidence at end-of-change
+  like any other bundle artifact. Everything at runtime is deterministic scripts — zero LLM
+  calls in the execution path.
+
+  Off by default. Turn it on with the new `visual_evidence` flag (env `PAQAD_VISUAL_EVIDENCE`);
+  `visual_evidence_mode` (`warn`|`strict`, floored) governs how firmly the gate enforces. With
+  the flag off, feature-development is byte-identical to before.
+
+  What is new when `visual_evidence` is on and a change is frontend-triggering:
+
+  - a stack-pack `visual_evidence.frontend_globs` trigger decides when a change is frontend;
+  - compiled capture scripts (`docs/site-map/journeys/<id>.capture.yaml`) derived from
+    **confirmed** site-map journeys drive the capture;
+  - a paqad-provisioned Playwright + Chromium runtime under `~/.paqad-ai/ve-runtime/` (never a
+    hard paqad dependency, never imported from the target project);
+  - per-step `image.png` + `caption.txt` and an `overview.gif` under the bundle's
+    `screenshots/` subtree, with a bundle-integrity carve-out;
+  - a `visual-evidence` verification gate, receipt line, and `report.html` section;
+  - a `paqad-ai visual-evidence` CLI group (`run`, `plan`, `setup`).
+
+## 1.84.0
+
+### Minor Changes
+
+- c6f611c: feat(#547): wire the spec pipeline and ship the domain experts
+
+  The grounded prompt-to-spec pipeline and the expert plumbing were built but idle. This change
+  starts the pipeline from the specification stage, gives eleven domain experts and a chief
+  architect real behaviour, and records what every run costs and changes. It is off for onboarded
+  projects by default; both flags stay off unless a project opts in.
+
+  What is new when `spec_pipeline_enabled` is on:
+
+  - The specification stage runs the pipeline. A shared instruction in the policy and the router
+    says so, a new `spec pipeline start` verb grounds and labels the request in one go, and a new
+    `spec_pipeline_adoption` knob (warn or strict) controls how firmly a hand-written spec is
+    refused at freeze.
+  - Eleven pickable experts and a chief architect become real through one `expert-notes` runner
+    skill with a lens per expert and one `expert-synthesis` chief skill. The model decides which
+    experts a request needs; the script only validates against the roster.
+  - Each expert writes short request-time notes; the chief accepts or declines them, recommends a
+    resolution for each disagreement (which becomes a decision you make), and lists what nobody
+    covered. The spec is crafted with every line traced to its source, and the craft gate refuses
+    an untraced requirement or an accepted finding that never reached the spec.
+  - Freeze copies the run provenance into the record; the readable spec ends with a Provenance
+    section; and every run records full metrics, with a `spec pipeline metrics` verb to read them
+    back. A human's later edit to a frozen spec is captured by section.
+
+  With the flags off, feature-development is byte-identical to before.
+
+### Patch Changes
+
+- 912d9e3: fix(#540): stop a background notification opening a change and orphaning a bundle
+
+  A `<task-notification>` from a background monitor reaches a session as an ordinary user
+  turn, and the extra `Stop` it produces made the marker parser re-read the whole transcript.
+  With the previous change already closed, nothing was active, so every stale `paqad:stage`
+  line read as unrecorded and a phantom untitled `change-<ULID>` bundle was opened for it —
+  stealing the session pointer and failing its own completion gate, which turned a finished,
+  green change into a red one.
+
+  Three seams are fixed:
+
+  - The prompt-routing seam recognises a host-injected system notification and treats it as
+    no request at all, so a CI ping cannot re-route or clobber an in-flight change.
+  - The retrospective marker seam refuses to open a bundle once the session has already
+    closed a change. A session's first change is unaffected, including on the Codex and
+    Gemini completion hooks that share the seam.
+  - `paqad-ai resume --feature <ref>` now resolves a ref against the recorded bundles on
+    disk as well as the session's paused stack, so a displaced change is reachable without
+    hand-editing anything under `.paqad/ledger/`.
+
 ## 1.83.0
 
 ### Minor Changes

package/README.md

@@ -134,6 +134,8 @@ Three steps turn AI coding from a prompt you hope works into a pipeline you can
 | **Follow** | Agents run your workflows and rules the same way every time, instead of guessing.         | Represent a workflow once, and every agent follows it.      |
 | **Prove**  | Automatic checks confirm tests, spec coverage, docs, security, and quality before "done." | "Done" means proven, not claimed.                           |
 
+When you turn the flags on, the spec itself can be produced by a pipeline: it grounds the request in your own docs, brings in the domain experts the request needs, asks one small batch of plain-language questions, and writes the spec with every line traced to where it came from.
+
 ---
 
 ## Set up every agent once
@@ -191,7 +193,7 @@ A built-in router reads each request, judges its complexity and risk, and sends
 - **graduated** lane for everyday features
 - **full** lane for risky or wide-reaching changes, with a frozen spec and story breakdown first
 
-Along the way it coordinates **20 built-in specialist roles**, each with one job and no conflicting incentives:
+Along the way it coordinates **23 built-in specialist roles**, each with one job and no conflicting incentives:
 
 | Group             | Roles                                                                                                                                                                      |
 | ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |