feat(wave3): inlineParallelCapable detection threading + capability badge (#127, #128)

dashboard/src/lib/types.ts

@@ -80,6 +80,7 @@ export interface EndpointRecord {
   costModel?: "per_token" | "per_transaction" | "per_request" | "free";
   frequencyClass?: string;
   batchCapable?: boolean;
+  inlineParallelCapable?: boolean;
   cacheCapable?: boolean;
   streaming?: boolean;
   isMiddleware?: boolean;

dashboard/src/pages/Endpoints.tsx

@@ -76,6 +76,7 @@ function EndpointCard({ ep }: { ep: EndpointRecord }) {
   const caps: string[] = [];
   if (ep.streaming) caps.push('stream');
   if (ep.batchCapable) caps.push('batch');
+  if (ep.inlineParallelCapable) caps.push('inline-parallel');
   if (ep.cacheCapable) caps.push('cache');
   if (ep.isMiddleware) caps.push('middleware');
 

docs/superpowers/plans/PROGRESS.md

@@ -130,7 +130,7 @@ Two trackers in one file:
 
 Post-foundation work organized into 10 waves + 2 standalones, tracked via `wave/*` + `area/*` GitHub labels. **Execution order is severity-weighted** (platform issues outrank accuracy): Wave 9 → 6 → 7 → 8 → 10 → standalones → accuracy waves (5 → 3 → 1 → 2 → 4).
 
-**As of 2026-05-27:** all platform waves (6–10) are SHIPPED, plus accuracy Wave 3. Remaining open work: accuracy waves **5 → 1 → 2 → 4** + standalones #45/#52.
+**As of 2026-05-29:** all platform waves (6–10) are SHIPPED, plus accuracy Waves 3, 5, and 1. Remaining open work: accuracy waves **2 → 4** + Wave 3 follow-ups #127/#128 + standalones #45/#52.
 
 ## Overall Wave Status
 
@@ -143,10 +143,10 @@ Post-foundation work organized into 10 waves + 2 standalones, tracked via `wave/
 | **10** | Config hygiene | #97, #98 | 🟢 | Shipped via PR [#124](https://github.com/recost-dev/extension/pull/124). |
 | — | Standalone #45 | #45 | ⬜ | Opt-in Project ID persistence — independent feature. |
 | — | Standalone #52 | #52 | ⬜ | Dashboard theming — can defer indefinitely. |
-| **5** | Housekeeping (accuracy) | #118, #119 | ⬜ | Cheapest accuracy work. Both in-repo (benchmark runner + `benchmark.yml`). |
-| **3** | Resolver follow-ups (accuracy) | #114, #115, #116 | 🟢 | Shipped via PR [#126](https://github.com/recost-dev/extension/pull/126). Follow-ups #127 (detection threading) + #128 (dashboard badge). |
-| **1** | Findings quality (accuracy) | #84, #85, #112 | 🟡 | Code complete + reviewed; full benchmark Δ +0.00pp. Awaits manual EDH check of the two UI bits (confidence filter, sources badge). |
-| **2** | Traceability (accuracy) | #81, #113 | ⬜ | Corpus expansion + dual locations. |
+| **5** | Housekeeping (accuracy) | #118, #119 | 🟢 | Both closed 2026-05-27 (benchmark runner sort + D1 CI gate verified). |
+| **3** | Resolver follow-ups (accuracy) | #114, #115, #116 | 🟢 | Shipped via PR [#126](https://github.com/recost-dev/extension/pull/126). Follow-ups #127 (detection threading) + #128 (dashboard badge) landed on `feat/wave3-followups`. |
+| **1** | Findings quality (accuracy) | #84, #85, #112 | 🟢 | All closed 2026-05-28 (C2 dedupe, C3 confidence, CACHE/BATCH_GUARD tightening). |
+| **2** | Traceability (accuracy) | #81, #113 | ⬜ | Corpus expansion + dual locations. Next wave. |
 | **4** | Recall recovery (accuracy) | #117 | ⬜ | Risky — depended on Wave 3 (#116), now unblocked. |
 
 ## Wave 9 — Local-mode IPC architecture (CLOSED)
@@ -232,6 +232,8 @@ Post-foundation work organized into 10 waves + 2 standalones, tracked via `wave/
 
 > Append `YYYY-MM-DD HH:MM — <one-line update>`. Newest at top.
 
+- 2026-05-29 — **Wave 3 follow-ups #127 + #128 landed** on `feat/wave3-followups`. #127 (`a1972bf`): threaded `inlineParallelCapable` through the regex pattern path (`openai-compatible.ts` emit, registry-only), the regex-only `local-waste-detector.ts` (new inline-parallel finding with the n/count suggestion, distinct `inline_parallel` id), the pattern dedup key (`utils.ts`), and the intelligence graph (`ApiCallNode` + `builder.ts`). #128 (`e9d7d7e`): `EndpointRecord` in both UI type files + `Endpoints.tsx` inline-parallel chip + `scan-results.ts` population at all 3 endpoint-construction sites — restores the indicator DALL·E lost when #116 reclassified `images.generate` off `batchCapable`. Gates: full `test:scanner` green (3 new tests), `build` (webview+ext) clean, benchmark Δ +0.00pp on all 5 metrics. **Caveat:** the web dashboard reads endpoints from the API, whose schema has no `inline_parallel` column — dashboard chip needs an api-repo migration to light up; VS Code webview badge works now. Next: accuracy Wave 2 (#113 → #81).
+- 2026-05-29 — **Tracker reconciliation.** Waves 5 (#118/#119, closed 05-27) and 1 (#84/#85/#112, closed 05-28) marked 🟢 — both had merged but were never reflected here. Remaining accuracy work is now Wave 2 → 4.
 - 2026-05-28 — **Wave 1 (findings quality) code-complete** on `claude/superpowers-plugins-skills-1Yaij` (plan `docs/superpowers/plans/2026-05-28-wave1-findings-quality.md`). #112: comment-stripped guard window stops the `CACHE_GUARD`/`BATCH_GUARD` literal-word leak (URL-safe `//` lookbehind). #85: detectors carry a structural `riskScore`; single `deriveSeverity()` (hybrid floor+amplifier, not pure confidence×cost) + `computeCostImpact()` applied at all 5 `Suggestion`-construction sites; `costImpactUsd` internal-only; confidence filter in the sidebar. #84: `collapseSuggestions()` dedupes by `type::file::endpoint|line-bucket`, unions `sources`, max confidence, AI-preferred description; `mergeAiSuggestions` collapses instead of dropping; "detected by N sources" badge. Subagent-driven (impl + spec + code-quality review per unit). Gates: full `test:scanner` green, full benchmark Δ +0.00pp on all 5 metrics. **Pending:** (1) manual EDH check of the two UI bits; (2) follow-ups — pre-existing `scope === "internal"` guard divergence between the two `buildAggressiveSuggestions` copies (scan-results.ts lacks it), and dead `SourceBadge` + duplicated badge inline-style in `ResultsPage.tsx` (extract a shared secondary-badge).
 - 2026-05-27 — **Wave 3 shipped (PR #126)** — #114/#115/#116 closed. Subagent-driven across 2 parallel worktree tracks; gates green (test:scanner, build:ext, benchmark Δ +0.00pp). Follow-ups filed: #127 (detection threading) + #128 (dashboard badge). **Tracker reconciliation:** also corrected stale statuses — Waves 7 (PR #122), 8 (PR #123), and 10 (PR #124) had merged earlier but were never marked 🟢 here. All platform waves (6–10) now confirmed shipped; remaining work is accuracy waves 5→1→2→4 + standalones.
 - 2026-05-21 — **Wave 7 PR opened (#122).** Closes #92 + #93 as design-resolved via labeling, not reconciliation. 7 commits on `wave7/cost-simulator-labeling` (worktree `../extension-wave7`): shared `EstimateDisclaimer` component, render on Simulate tab, code comments on `LOCAL_PRICING` + `FREQUENCY_CLASS_MULTIPLIERS`, CLAUDE.md "Cost numbers: heuristic vs authoritative" section. All gates green (build, 4/4 tests, D1 Δ +0.00pp). Spec + plan committed to main. Wave 8 (#46/#94/#100) is next per severity order.

src/intelligence/__tests__/builder.test.ts

@@ -25,6 +25,7 @@ run("buildRepoIntelligenceSnapshot normalizes paths, stores null providers, and
         library: "openai",
         cacheCapable: false,
         batchCapable: false,
+        inlineParallelCapable: true,
       },
       {
         file: "./src/payments.ts",
@@ -127,11 +128,14 @@ run("buildRepoIntelligenceSnapshot normalizes paths, stores null providers, and
   assert.equal(workerCall.provider, "openai");
   assert.equal(workerCall.library, "openai");
   assert.equal(workerCall.filePath, "src/worker.ts");
+  assert.equal(workerCall.inlineParallelCapable, true);
 
   const stripeCall = Object.values(snapshot.apiCalls).find((apiCall) => apiCall.filePath === "src/payments.ts" && apiCall.line === 15);
   assert.ok(stripeCall);
   assert.equal(stripeCall.provider, "stripe");
   assert.equal(stripeCall.library, "stripe");
+  // Absent on input → coerced to false (parallels batchCapable).
+  assert.equal(stripeCall.inlineParallelCapable, false);
 
   const regexOnlyCall = Object.values(snapshot.apiCalls).find((apiCall) => apiCall.filePath === "src/regex-only.ts" && apiCall.line === 22);
   assert.ok(regexOnlyCall);

src/intelligence/builder.ts

@@ -197,6 +197,7 @@ export function buildRepoIntelligenceSnapshot(
         costModel: call.costModel ?? null,
         frequencyClass: call.frequencyClass ?? null,
         batchCapable: Boolean(call.batchCapable),
+        inlineParallelCapable: Boolean(call.inlineParallelCapable),
         cacheCapable: Boolean(call.cacheCapable),
         streaming: Boolean(call.streaming),
         isMiddleware: Boolean(call.isMiddleware),

src/intelligence/types.ts

@@ -22,6 +22,7 @@ export interface ApiCallNode {
   costModel: ApiCallInput["costModel"] | null;
   frequencyClass: ApiCallInput["frequencyClass"] | null;
   batchCapable: boolean;
+  inlineParallelCapable: boolean;
   cacheCapable: boolean;
   streaming: boolean;
   isMiddleware: boolean;

src/scan-results.ts

@@ -492,6 +492,7 @@ export function mergeRemoteAndLocalEndpoints(
       if (!endpoint.costModel && call.costModel) endpoint.costModel = call.costModel;
       endpoint.frequencyClass = pickMostSevereFrequency(endpoint.frequencyClass, call.frequencyClass);
       if (call.batchCapable) endpoint.batchCapable = true;
+      if (call.inlineParallelCapable) endpoint.inlineParallelCapable = true;
       if (call.cacheCapable) endpoint.cacheCapable = true;
       if (call.streaming) endpoint.streaming = true;
       if (call.isMiddleware) endpoint.isMiddleware = true;
@@ -547,6 +548,7 @@ export function mergeRemoteAndLocalEndpoints(
         costModel: call.costModel,
         frequencyClass: call.frequencyClass,
         batchCapable: call.batchCapable,
+        inlineParallelCapable: call.inlineParallelCapable,
         cacheCapable: call.cacheCapable,
         streaming: call.streaming,
         isMiddleware: call.isMiddleware,
@@ -580,6 +582,7 @@ export function mergeRemoteAndLocalEndpoints(
     if (!synthetic.costModel && call.costModel) synthetic.costModel = call.costModel;
     synthetic.frequencyClass = pickMostSevereFrequency(synthetic.frequencyClass, call.frequencyClass);
     if (call.batchCapable) synthetic.batchCapable = true;
+    if (call.inlineParallelCapable) synthetic.inlineParallelCapable = true;
     if (call.cacheCapable) synthetic.cacheCapable = true;
     if (call.streaming) synthetic.streaming = true;
     if (call.isMiddleware) synthetic.isMiddleware = true;

src/scanner/local-waste-detector.ts

@@ -308,6 +308,43 @@ function detectBatchFinding(relativePath: string, site: MatchedCallSite): LocalW
   );
 }
 
+// Inline-parallel: the endpoint itself accepts an n/count parameter (a
+// registry-only signal, e.g. DALL·E `images.generate`), so fanning out one
+// request per item is wasteful. Mirrors the AST inline-parallel detector for
+// the regex-only path (RECOST_DISABLE_AST=1). Distinct from `detectBatchFinding`
+// — uses an `inline_parallel` id so it never collapses with a batch finding.
+function detectInlineParallelFinding(relativePath: string, site: MatchedCallSite): LocalWasteFinding | null {
+  const ext = path.extname(relativePath).toLowerCase();
+  if (!JS_TS_EXTENSIONS.has(ext)) return null;
+  if (!site.match.inlineParallelCapable) return null;
+  if (site.batchGuard) return null;
+  if (!(site.loopDepth > 0 || site.promiseAll || site.mapFanout || site.arrayFanout)) return null;
+
+  const evidence: string[] = [];
+  pushEvidence(evidence, site.loopDepth > 0, "Call executes inside a loop.");
+  pushEvidence(
+    evidence,
+    site.promiseAll || site.mapFanout || site.arrayFanout,
+    "Parallel collection fanout issues one request per item."
+  );
+  evidence.push("This endpoint accepts an n/count parameter that returns multiple results from a single request.");
+  const score = baseScore(site) + 1 - (site.smallBounded ? 1 : 0);
+  const confidence = scoreToConfidence(score, site, true);
+  if (confidence < 0.35) return null;
+  return {
+    id: `local-inline_parallel-${relativePath}:${site.line}`,
+    type: "batch" as SuggestionType,
+    severity: scoreToSeverity(score),
+    confidence,
+    riskScore: score,
+    description:
+      "This endpoint accepts an n/count parameter — request multiple results in a single call instead of issuing one request per item.",
+    affectedFile: relativePath,
+    line: site.line,
+    evidence,
+  };
+}
+
 function detectRedundancyFinding(relativePath: string, site: MatchedCallSite, clientInitCount: number): LocalWasteFinding | null {
   const ext = path.extname(relativePath).toLowerCase();
   if (!JS_TS_EXTENSIONS.has(ext)) return null;
@@ -415,7 +452,11 @@ function dedupeFindings(findings: Array<LocalWasteFinding | null>): LocalWasteFi
   const deduped = new Map<string, LocalWasteFinding>();
   for (const finding of findings) {
     if (!finding) continue;
-    const key = `${finding.type}:${finding.affectedFile}:${finding.line ?? 0}`;
+    // Key on the detector-specific id so distinct detectors that share a
+    // SuggestionType at the same site stay separate — e.g. batch
+    // (`local-batch-…`) vs inline-parallel (`local-inline_parallel-…`), which
+    // both emit type "batch". Fall back to type:file:line if an id is missing.
+    const key = finding.id || `${finding.type}:${finding.affectedFile}:${finding.line ?? 0}`;
     const existing = deduped.get(key);
     if (!existing || finding.confidence > existing.confidence) {
       deduped.set(key, finding);
@@ -433,6 +474,7 @@ export function detectLocalWasteFindingsInText(relativePath: string, text: strin
   for (const site of callSites) {
     findings.push(detectCacheFinding(relativePath, site));
     findings.push(detectBatchFinding(relativePath, site));
+    findings.push(detectInlineParallelFinding(relativePath, site));
     findings.push(detectRedundancyFinding(relativePath, site, clientInitCount));
     findings.push(detectNPlusOneFinding(relativePath, site));
     findings.push(detectRateLimitFinding(relativePath, site));

src/scanner/patterns/openai-compatible.ts

@@ -173,6 +173,9 @@ export const openAiCompatibleMatcher: LineMatcher = {
 
       const streaming = reg?.streaming ?? /stream/i.test(action);
       const batchCapable = reg?.batchCapable ?? /batches|batch/.test(path);
+      // No hint-regex fallback: inline-parallel capability (an n/count param)
+      // is a registry-only signal — there is no reliable lexical hint for it.
+      const inlineParallelCapable = reg?.inlineParallelCapable ?? false;
       const cacheCapable = reg?.cacheCapable ?? /responses|chat|assistants|threads/.test(path);
 
       results.push({
@@ -187,6 +190,7 @@ export const openAiCompatibleMatcher: LineMatcher = {
         host,
         streaming,
         batchCapable,
+        inlineParallelCapable,
         cacheCapable,
         rawMatch: match[0],
       });

src/scanner/patterns/types.ts

@@ -22,6 +22,7 @@ export interface ApiCallMatch {
   loopContext?: boolean;
   streaming?: boolean;
   batchCapable?: boolean;
+  inlineParallelCapable?: boolean;
   cacheCapable?: boolean;
   inferredCostRisk?: string[];
   rawMatch?: string;

src/scanner/patterns/utils.ts

@@ -39,6 +39,7 @@ export function uniqueMatches(matches: ApiCallMatch[]): ApiCallMatch[] {
       match.host ?? "",
       match.streaming ? "1" : "0",
       match.batchCapable ? "1" : "0",
+      match.inlineParallelCapable ? "1" : "0",
     ].join("|");
 
     if (seen.has(key)) continue;

src/test/local-waste-detector.test.ts

@@ -59,6 +59,18 @@ run("reduces confidence when polling already has backoff and concurrency guards"
   assert.ok((rateLimit?.confidence ?? 1) < 0.75);
 });
 
+run("flags inline-parallel fanout for an n/count-capable endpoint (regex-only path)", () => {
+  const text = [
+    "async function makeThumbnails(prompts) {",
+    "  return Promise.all(prompts.map((prompt) => client.images.generate({ prompt })));",
+    "}",
+  ].join("\n");
+  const findings = detectLocalWasteFindingsInText("src/lib/thumbnails.ts", text);
+  const inlineParallel = findings.find((finding) => finding.id.includes("inline_parallel"));
+  assert.ok(inlineParallel, "expected an inline-parallel finding for images.generate fanout");
+  assert.match(inlineParallel?.description ?? "", /n\/count parameter/);
+});
+
 run("#112: bare 'cache' in a comment does not suppress a cache finding", () => {
   const text = [
     "// we should cache this someday but do not yet",

src/test/scanner-patterns.test.ts

@@ -99,4 +99,18 @@ run("keeps route definition detection", () => {
   assert.ok(routes.some((r) => r.method === "POST" && r.url === "/webhook/stripe"));
 });
 
+run("emits inlineParallelCapable from the registry for images.generate", () => {
+  const match = pickProvider("await client.images.generate({ prompt: 'a cat', n: 4 });", "openai");
+  assert.ok(match);
+  assert.equal(match?.inlineParallelCapable, true);
+  // images.generate is inline-parallel, not batch-capable (reclassified in #116).
+  assert.notEqual(match?.batchCapable, true);
+});
+
+run("leaves inlineParallelCapable unset for non-inline-parallel methods", () => {
+  const match = pickProvider("await client.chat.completions.create({ model: 'gpt-4.1' });", "openai");
+  assert.ok(match);
+  assert.notEqual(match?.inlineParallelCapable, true);
+});
+
 console.log("All scanner matcher tests passed");

webview/src/types.ts

@@ -43,6 +43,7 @@ export interface EndpointRecord {
   costModel?: "per_token" | "per_transaction" | "per_request" | "free";
   frequencyClass?: string;
   batchCapable?: boolean;
+  inlineParallelCapable?: boolean;
   cacheCapable?: boolean;
   streaming?: boolean;
   isMiddleware?: boolean;