ccqa 1.35.0 → 1.36.0

package/README.md

@@ -206,12 +206,15 @@ end-to-end on its own:
   spec, choose deterministic or live mode, record, run to green.
 - **ccqa-rerecord** — bring a flagged or failing test case back to green:
   read the hub's finding, decide what went stale, repair that, re-record.
+- **ccqa-resolve** — clear everything the hub is holding for a person: read
+  the verdict, order the rows, route each to a product fix, a re-recording or
+  an environment repair.
 
 Install them with the [skills CLI](https://github.com/vercel-labs/skills)
 into a consuming project (or `-g` for all projects):
 
 ```bash
-npx skills add <this-repo> --skill ccqa-record --skill ccqa-rerecord
+npx skills add <this-repo> --skill ccqa-record --skill ccqa-rerecord --skill ccqa-resolve
 ```
 
 ## Documentation

package/dist/bin/ccqa.mjs

@@ -16142,7 +16142,9 @@ Ask whether the **intent** the step describes still exists in the product:
 - The intent exists, but the string or selector the spec names is gone or renamed → **TEST_DRIFT**. Cite where the replacement lives.
 - The intent itself is gone, or deliberately different → **SPEC_CHANGE**. Cite the source that shows the new shape.
 
-A renamed button is TEST_DRIFT. A button that no longer exists because the flow was replaced is SPEC_CHANGE. If the source shows a rename you can point at, prefer TEST_DRIFT; if you cannot find where the intent went, SPEC_CHANGE is the most you may claim, and UNKNOWN when even that is a guess.
+A renamed button is TEST_DRIFT. A button that no longer exists because the flow was replaced is SPEC_CHANGE. If the source shows a rename you can point at, prefer TEST_DRIFT.
+
+SPEC_CHANGE is the more expensive answer — it sends a human to rewrite or retire the spec — so it takes the *stronger* evidence, not the weaker. Failing to find where the intent went is not a finding; that is UNKNOWN. Claim SPEC_CHANGE only when you can point at the source that shows the new shape, or at where the implementation would sit if it still existed.
 
 ## Which surface drifted
 
@@ -16175,6 +16177,8 @@ sit nearby.
 
 - **No drift is a claim, not a default.** Make it after picking the concrete strings from *every* surface you were given — the spec's \`expected\` and the generated code's selectors alike — and finding each of them in the source. Clearing the test case because one surface checked out is the most common way to miss a real finding. If you never looked, the honest answer is UNKNOWN.
 - **A finding needs a citation.** Every TEST_DRIFT and SPEC_CHANGE must carry at least one \`evidence\` entry with a real \`file\`, and a line where you can give one. A label with no citation is a guess wearing a verdict's clothes — answer UNKNOWN instead.
+- **A citation must apply to the case at hand.** Finding the string is not the end of it — read what encloses the line before you cite it. A line inside a guard, behind an early \`return\`, or in a branch this spec's steps never enter says nothing about this spec. Name the conditions that must hold for that line to run, and check the spec puts the product in them. A citation that only proves the line exists is not evidence.
+- **A comment is not the code.** Comments in the product's source say what someone intended, and they rarely restate the conditions they sit under. A line reading "this is not supported", sitting inside a guarded branch, is true only inside that branch. Cite the control flow you traced, not the sentence you found.
 - **Do not report style.** Wording you would have phrased differently is not drift. Report only what would make a replay fail, or what asks about something the product no longer does.
 - \`confidence\` is about the label: how sure you are it is the right one, not how bad the finding is.
 
@@ -16184,6 +16188,7 @@ sit nearby.
 2. \`Grep\` the source for them, at the page, component or handler the step is about.
 3. For \`include\` steps, confirm the block exists under \`.ccqa/blocks/<name>/spec.yaml\` and that every \`params\` key is declared on it.
 4. When a string is missing, look for what replaced it before concluding. Where it went is what decides the label.
+5. Before citing any line, read the block that encloses it. Which conditions must hold for it to run, and does the spec put the product in those conditions? A line that only runs in a case the spec never enters proves nothing about the spec.
 
 ${guidance.userPromptBlock ?? ""}${guidance.customPromptBlock ?? ""}## Output (STRICT)
 
@@ -16541,7 +16546,7 @@ function determineExitCode(results, threshold) {
 //#endregion
 //#region src/drift/to-report.ts
 /** Tracks the drift prompt's own version — the two must never drift apart. */
-const DRIFT_REPORT_PROMPT_VERSION = "6";
+const DRIFT_REPORT_PROMPT_VERSION = "7";
 /**
 * Spec-level status under the given threshold, mirroring determineExitCode's
 * per-spec logic (exit-code.ts) but scoped to a single SpecResult.
@@ -19414,9 +19419,11 @@ function computeRerun(input) {
 		const runMoved = specMovedSince(spec.changedAt, coords.lastRun?.deployedSha ?? null, coords.lastRun?.at ?? "", deployTimes);
 		if (auditMoved && audit.audit !== "due") audit = { audit: "due" };
 		if (runMoved && execution.execution === "passed") execution = { execution: "stale" };
+		const manualState = readAttestation(attestations.specs[spec.key], spec, coords.lastRed, range, log, deployTimes);
+		const lapsedOverRed = manualState?.kind === "lapsed" && manualState.because !== "newerRed";
+		if (execution.execution === "failed" && lapsedOverRed) execution = { execution: "stale" };
 		const held = heldBy(locks, spec.key, now);
 		let verdict = decide(audit.audit, execution.execution, held);
-		const manualState = readAttestation(attestations.specs[spec.key], spec, coords.lastRed, range, log, deployTimes);
 		if (manualState?.kind === "covers" && !held && verdict !== "verified") verdict = "manuallyVerified";
 		out[spec.key] = {
 			verdict,
@@ -19525,8 +19532,10 @@ function auditState(drift, key, range) {
 * buckets advance from the same terminal-run trigger, so the run that wrote
 * `run` wrote exactly one of `green` or `red`.
 *
-* `failed` outranks `stale`: a red result is current information whatever has
-* deployed since, and re-running it teaches nothing until someone repairs it.
+* `failed` outranks `stale` here: a red result is current information whatever
+* has deployed since, and re-running it teaches nothing until someone repairs
+* it. Its caller retires a red a lapsed attestation already answered
+* (ADR-0020) — this function does not know about attestations.
 */
 function executionState(coords, since) {
 	const { lastRun, lastRed } = coords;
@@ -24010,6 +24019,13 @@ const CLIENT_JS = `
     return null;
   }
 
+  // Did the last run end red? Read off the ledger the same way the execution
+  // axis reads it, because a red a lapsed attestation retired ships as stale
+  // and the sites that need the failure itself must still find it (ADR-0020).
+  function lastFailed(rr) {
+    return !!(rr && rr.lastRun && rr.lastRed && rr.lastRed.runId === rr.lastRun.runId);
+  }
+
   // The execution axis as one state. "What did it say last time" and "is that
   // still true" are two faces of one question, so the row answers it once and
   // the detail panel keeps the coordinates.
@@ -24540,7 +24556,9 @@ const CLIENT_JS = `
     }
     // Only once the audit has actually cleared the spec: while it is still
     // due, nobody knows yet whether the environment was the only thing wrong.
-    if (rr.audit === "clean" && rr.execution === "failed" && rr.lastRed && rr.lastRed.label === "ENVIRONMENT") {
+    // Read off the ledger, not the axis: an environment red is demoted to
+    // stale so it re-runs (ADR-0020), and the offer has to survive that.
+    if (rr.audit === "clean" && lastFailed(rr) && rr.lastRed.label === "ENVIRONMENT") {
       return buildOverrideOffer("environment", feature, spec);
     }
     return null;
@@ -24580,7 +24598,7 @@ const CLIENT_JS = `
   // report loads — and the whole content if it never arrives.
   function reasonFindingSource(rr, driftEntry) {
     if (auditOpen(rr) && driftEntry && driftEntry.runId && driftEntry.label) return driftEntry;
-    if (rr.execution === "failed" && rr.lastRed && rr.lastRed.runId && rr.lastRed.label) return rr.lastRed;
+    if (lastFailed(rr) && rr.lastRed.runId && rr.lastRed.label) return rr.lastRed;
     return null;
   }
 

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccqa",
-  "version": "1.35.0",
+  "version": "1.36.0",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccqa",
-  "version": "1.35.0",
+  "version": "1.36.0",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {