valent-pipeline 0.19.61 → 0.19.63

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "valent-pipeline",
-  "version": "0.19.61",
+  "version": "0.19.63",
   "description": "v3 multi-agent AI pipeline for software development lifecycle",
   "type": "module",
   "bin": {

package/pipeline/orchestrators/claude-code/sprint.workflow.js

@@ -1951,7 +1951,7 @@ async function runStory(story) {
     // reworks via the QA-B re-execution alone, then falls through to JUDGE at the cap as usual.
     const evGate = await runEvidenceGate(storyId, [])
     recordStage('evidence', evGate)
-    const visual = await runVisualGate(storyId, profiles)
+    const visual = await runVisualGate(storyId, profiles, itemType)
     recordStage('visual', visual)
     phase('Judge')
     const decision = await runJudgeGate(storyId, visual, evGate)
@@ -2234,8 +2234,9 @@ async function runStory(story) {
   recordStage('evidence', evGate)
 
   // Visual validation (revived PMCP): for ui-profile stories, produce pmcp-evidence.md so JUDGE's
-  // evidence pass has the visual artifact it requires. No-op (returns null) for non-ui stories.
-  const visual = await runVisualGate(storyId, profiles)
+  // evidence pass has the visual artifact it requires. No-op (returns null) for non-ui stories, and a
+  // non-blocking N/A for itemType:bug (the QA-A-less bug-fix path mints no checklist — M-192).
+  const visual = await runVisualGate(storyId, profiles, itemType)
   recordStage('visual', visual)
 
   phase('Judge')
@@ -2826,11 +2827,23 @@ async function runStory(story) {
   // exact trigger QA-A uses to produce the checklist (qa-a/write-spec.md Step 9: `ui` in profiles).
   // The agent starts its own dev server: unlike the old always-on topology where QA-B triggered an
   // idle PMCP while QA-B's server was up, here the stages are sequential, so PMCP owns its server.
-  async function runVisualGate(sid, storyProfiles) {
+  async function runVisualGate(sid, storyProfiles, sidItemType = 'story') {
     if (!storyProfiles.includes('ui')) {
       log(`${sid}: no ui profile — visual validation N/A (QA-A wrote no checklist; nothing for PMCP to run)`)
       return null
     }
+    // QA-A-less bug-fix path (owner-ruled M-192): a bug-fix item runs bug-intake -> Build -> CRITIC ->
+    // QA-B -> Judge and STRUCTURALLY SKIPS the Spec phase, so QA-A never mints a
+    // visual-validation-checklist.md. Running PMCP against a checklist that can NEVER exist yields a
+    // 0/0 block-fail (onMissingMcp inherited from the parent story) on a path with no checklist to
+    // drive — a config contradiction, not a real visual defect. Treat it as a non-blocking N/A: no
+    // visual coverage is lost because the parent story's `revalidate` re-drives the FULL visual gate
+    // on the merged fix (its checklist exists). A FULL story always runs QA-A, so a missing checklist
+    // THERE stays a real, blocking fault — this scopes the gate down ONLY for itemType:bug.
+    if (sidItemType === 'bug') {
+      log(`${sid}: ui profile but itemType:bug — the QA-A-less bug-fix path mints no visual-validation-checklist.md; visual validation N/A (owner-ruled M-192). The parent story's revalidate re-drives the full visual gate on the merged fix.`)
+      return { schema: 1, agent: 'pmcp', story: sid, verdict: 'na', skipped: true, mcpAvailable: null, totalCheckpoints: 0, passed: 0, failed: 0, naCheckpoints: 0, files: [], notes: 'QA-A-less bug-fix path: no visual-validation-checklist.md minted (Spec phase skipped for itemType:bug); visual gate N/A (owner-ruled M-192). The parent story revalidate re-drives the full visual gate on the merged fix.' }
+    }
     phase('Visual')
     const naMode = visualOnMissingMcp === 'na'
     const result = await spawn('PMCP', 'pmcp.md',
@@ -2920,7 +2933,13 @@ async function runStory(story) {
       : ''
     // A ui story has a PMCP visual pass; tell the evidence agent the result so Step 9 (Verify PMCP
     // Evidence) is grounded rather than guessing whether visual evidence should exist.
-    const visualNote = visual
+    const visualNote = visual && visual.skipped
+      ? `\n\nThis is a UI-profile BUG-FIX item: the QA-A-less bug-fix path mints no ` +
+        `visual-validation-checklist.md, so PMCP did NOT run and there is NO pmcp-evidence.md. Treat ` +
+        `PMCP evidence as N/A in Step 9 (no checklist ⟹ visual N/A) — do NOT flag its absence as a ` +
+        `missing required file (owner-ruled M-192). The parent story's revalidate re-drives the full ` +
+        `visual gate on the merged fix.`
+      : visual
       ? `\n\nThis is a UI story: a visual-validation (PMCP) pass already ran and wrote ` +
         `pmcp-evidence.md (verdict: ${visual.verdict}, ${visual.passed || 0}/${visual.totalCheckpoints || 0} ` +
         `checkpoint(s) passed${visual.mcpAvailable === false ? `, browser MCP unavailable` : ''}). ` +

package/pipeline/prompts/pmcp.md

@@ -63,7 +63,25 @@ Follow the on-missing-MCP behavior the orchestrator passed in your trigger:
   `mcpAvailable: false`, and return — the story proceeds without visual evidence.
 
 ### Step 3: Execute checkpoints (MCP available)
-Start the dev server. For each checkpoint: navigate to its route, establish the required state (seed
+Start the dev server.
+
+**Pre-flight — assert an INTEGER `devicePixelRatio` before driving any interaction.** Once the server
+is up and you have navigated to the app, evaluate `window.devicePixelRatio`. If it is NOT an integer
+(commonly `1.5` when the host display runs Windows/macOS scaling at 150%), the browser's
+coordinate-based `click()`/`fill()` land OFF-TARGET: Playwright reports the click succeeded but the
+app's `onClick`/`onInput` handlers never fire, so you can screenshot STATIC renders yet can NEVER
+drive an interaction or a multi-step / WS flow. **Do NOT misread this as "the browser MCP cannot drive
+WS events," and do NOT FAIL or `na` the interaction checkpoints** — it is an infrastructure config
+bug, not an app or reachability defect (the e2e test runner pins the scale factor, which is why an
+acceptance e2e can drive the same flow the MCP cannot). STOP and record an **infrastructure blocker**
+in `pmcp-evidence.md` (`verdict: fail`, `mcpAvailable: true`) naming the cause and the fix: the
+browser MCP must launch with **`deviceScaleFactor: 1`** (passing a fixed `--viewport-size` defaults
+the scale factor to 1), a project `.mcp.json` / browser-MCP-config change that needs an MCP/session
+**RESTART** to take effect. Static render-only checkpoints may still be captured; mark every
+interaction-dependent checkpoint blocked-by-DPR, not FAIL. A non-integer DPR is the single most common
+cause of "PMCP can render but can't drive."
+
+For each checkpoint: navigate to its route, establish the required state (seed
 data, auth, form values), wait for the expected state using element/network conditions — **never hard
 waits** — capture a screenshot into `{story_output_dir}` using the checklist's filename convention,
 and evaluate it against the pass criteria. Record PASS or FAIL per checkpoint. Stop the server when done.

package/pipeline/steps/judge/evidence-review.md

@@ -6,7 +6,7 @@ Runs after bug review (Steps 1-4) approves with no P4->P1-P3 escalations.
 
 Read all inputs, record in `inputsRead`. If any REQUIRED file missing, set blocker and STOP.
 
-Required: `execution-report.md`, `traceability-matrix.md`, `bugs.md`, `qa-test-spec.md` — all in `{story_output_dir}/`. Also `pmcp-evidence.md` **if this is a UI story** — i.e. `ui` is in `{testing_profiles}` (equivalently, `visual-validation-checklist.md` exists). If there is no `ui` profile, PMCP evidence is N/A — do NOT treat its absence as a missing required file.
+Required: `execution-report.md`, `traceability-matrix.md`, `bugs.md`, `qa-test-spec.md` — all in `{story_output_dir}/`. Also `pmcp-evidence.md` **if this is a UI story** — i.e. `ui` is in `{testing_profiles}` **AND** `visual-validation-checklist.md` exists. The checklist's existence is the operative trigger: a `ui`-profile item with **no** `visual-validation-checklist.md` has PMCP **N/A** — do NOT treat the absence of `pmcp-evidence.md` as a missing required file. This is exactly the **QA-A-less bug-fix path** (an `itemType:bug` fix runs bug-intake → Build → CRITIC → QA-B → Judge and structurally skips the Spec phase, so QA-A mints no checklist); its visual coverage is re-driven in full by the parent story's `revalidate` on the merged fix (owner-ruled M-192). If there is no `ui` profile at all, PMCP evidence is likewise N/A.
 
 **Machine evidence (when `stories/{story_id}/evidence/` exists — required in that case):** the newest JSON under `evidence/asserts/` (the deterministic green-gate's verdict), `evidence/pins.jsonl` (which commit each gate reviewed), and the run records under `evidence/runs/` (frozen junit reports + hashes, written only by the CLI). **ATDD stories add** `evidence/atdd-red.json` (the pre-implementation failing run: per-case statuses, the pre-dev SHA, the acceptance-source snapshot hashes) and — for a proven story — `evidence/proof.json` (red run + green run + the CLI-hashed diff between: the story's headline evidence). These are the authoritative numbers — the markdown artifacts are the human-readable companion. If the evidence directory does not exist, this is a LEGACY story: apply the legacy fallbacks below and flag every count as self-reported.