pandemicsyn · pandemicsyn · Jul 18, 2026 · Jul 18, 2026 · Jul 18, 2026
@@ -0,0 +1,5 @@
+---
+'neondeck': patch
+---
+
+Load PR review threads through a lean cancellable GitHub query and a short-lived mutation-aware cache so warm review surfaces open within the performance budget.
@@ -871,6 +871,9 @@ <h2 id="summary-title">Executive summary</h2>
               long tasks. That baseline remains retained. Stable thread identity, local metadata
               reuse, and active-patch-first scheduling now produce one thread request, zero
               abandoned patches, and a 798 ms first-patch median that passes all three samples.
+              The final thread-surface pass keeps the full Flue action intact while using an
+              84.7% smaller web query plus a bounded 15-second cache; warm thread visibility now
+              measures 459 ms median and passes its target. The original results remain retained.
             </p>
           </div>
 
@@ -920,10 +923,10 @@ <h3>Correctness risks</h3>
             <article class="panel-block">
               <h3>Performance risks</h3>
               <p>
-                The large-patch and chat render bottlenecks are now mitigated. Remaining risks are
-                duplicate review-thread reads, repeated full local metadata work for each patch,
-                Runtime endpoint cost and invalidation bursts, and the eager plugin registry
-                pending explicit budgets.
+                The large-patch, chat render, repeated metadata, and warm review-thread
+                bottlenecks are now mitigated. Remaining risks are the 642 ms tree median,
+                one-time cold object/thread reads, Runtime endpoint cost and invalidation bursts,
+                and the eager plugin registry pending explicit budgets.
               </p>
             </article>
           </div>
@@ -993,7 +996,9 @@ <h2 id="performance-title">Performance first: discussion report</h2>
             complete. The two approved request-path fixes are now implemented and remeasured:
             review-thread identity is stable and immutable local PR metadata is reused across
             patch reads. Active-patch prioritization is also implemented and remeasured: the
-            selected patch now passes the one-second target in all three retained samples. Keep
+            selected patch now passes the one-second target in all three retained samples. The
+            production-only thread pass is complete: a lean cancellable query and short-lived
+            mutation-aware cache reduce warm thread visibility to 459 ms median. Keep tree and
             cold-fetch changes, Runtime aggregation, review-subtree consolidation, and registry
             splitting behind workload and budget evidence.
           </div>
@@ -1022,6 +1027,8 @@ <h2 id="performance-title">Performance first: discussion report</h2>
                 <tr><td>Real PR · request stability</td><td>1 thread request · 0 aborted patches</td><td>381 kB thread · 743 kB total API</td><td>Last API response: 2,782 ms</td><td>Duplicate and settlement churn eliminated</td></tr>
                 <tr><td>Real PR · active-patch priority</td><td>612 ms tree · 798 ms first patch</td><td>620 ms LCP · 0.0012 CLS</td><td>1,724 ms threads · zero long tasks</td><td>Completed · first patch passes all samples</td></tr>
                 <tr><td>Real PR · prioritized requests</td><td>1 thread request · 4 patch requests</td><td>381 kB thread · 724 kB total API</td><td>0 aborted · last API response 2,021 ms</td><td>Active read settles before background fanout</td></tr>
+                <tr><td>Real PR · lean threads, uncached</td><td>639 ms tree · 924 ms first patch</td><td>58 kB thread · 401 kB total API</td><td>1,511 ms threads · 16.8 ms post-response render</td><td>Payload fixed; GitHub latency remains</td></tr>
+                <tr><td>Real PR · lean threads, warm</td><td>642 ms tree · 934 ms first patch</td><td>530 ms traced LCP · 0.00 CLS</td><td>459 ms threads · 6.8 ms backend cache hit</td><td>Completed · warm thread target passes</td></tr>
                 <tr><td>Real PR · cold local / GitHub</td><td>4,978 / 3,894 ms</td><td>295 / 765 kB metadata</td><td>0 / 1 truncated files</td><td>Both miss cold target; local is complete</td></tr>
                 <tr><td>Production initial preload</td><td>194.6 kB gzip</td><td>157.7 kB entry · 21.6 kB ccount</td><td>Focused-review wrapper 0.77 kB gzip</td><td>Track; no budget miss declared</td></tr>
               </tbody>
@@ -1047,7 +1054,10 @@ <h2 id="performance-title">Performance first: discussion report</h2>
             measured 551 ms LCP, 0.00 CLS, and the same 39 ms total Pierre forced reflow with no
             estimated savings. After active-patch prioritization, the browser median reached 798
             ms to first patch and a direct trace measured 498 ms LCP, 0.00 CLS, and 36 ms total
-            Pierre forced reflow with no estimated savings.
+            Pierre forced reflow with no estimated savings. The final production-only thread
+            pass measured 459 ms median thread visibility, 58 kB per thread response, 6.5 ms
+            post-response rendering, zero aborted/failed API requests, 530 ms traced LCP, and
+            0.00 CLS. Its initial uncached GitHub read remains outside the warm target.
           </div>
 
           <div class="finding-list">
@@ -1352,6 +1362,7 @@ <h2 id="actions-title">Recommended sequence</h2>
             <li class="completed-action"><span class="completion">COMPLETED · MEASUREMENT + FIRST WAVE</span> <strong>Performance measurement gate:</strong> the five-part fixture pass, chat isolation, thresholded Pierre CodeView, and real registered-PR trace are complete. Runtime topology, duplicate review DOM, and lazy registry remain deferred behind workload/budget evidence.</li>
             <li class="completed-action"><span class="completion">COMPLETED · REQUEST-PATH REMEDIATION</span> <strong>P1 · Real PR request path:</strong> discuss the measured duplicate review-thread identity and repeated local metadata work, then implement only the approved fixes and repeat <code>npm run bench:pr-review</code>. The stable thread key and bounded revision-metadata cache are implemented and verified on the retained target.</li>
             <li class="completed-action"><span class="completion">COMPLETED · ACTIVE PATCH PRIORITY</span> <strong>P1 · Active patch priority:</strong> discuss whether neighbor and unresolved-path prefetch should wait for the active patch. The retained remediated trace was 195 ms over target; the implemented active-first schedule now reaches a 798 ms median and passes all three samples while preserving zero abandoned reads.</li>
+            <li class="completed-action"><span class="completion">COMPLETED · WARM THREAD PATH</span> <strong>P1 · Review-thread latency:</strong> correct the harness's production-mode detection, measure request versus render cost, slim the web-only GitHub query, and add only a bounded 15-second mutation-aware cache after the lean query remained too slow. Warm thread visibility now reaches 459 ms median with an 84.7% smaller response; cold GitHub latency remains explicit.</li>
             <li class="completed-action"><span class="completion">COMPLETED · COPY/DISCLOSURE</span> <strong>P2 · Copy/disclosure:</strong> row-specific action names and full-value affordances for operational text. Suggested command: <code>$impeccable clarify</code>.</li>
             <li class="completed-action"><span class="completion">COMPLETED · PR #140</span> <strong>P2 · Editor state ownership:</strong> keep new-comment, draft-edit, and thread-reply bodies scoped to their editor generation so older async completions cannot clear newer work.</li>
             <li class="completed-action"><span class="completion">COMPLETED · FINAL PASS</span> <strong>P3 · Final pass:</strong> tabular metrics, timer cleanup, operation-scoped status, stable finding IDs, and config tightening. Suggested command: <code>$impeccable polish</code>.</li>

@@ -1,6 +1,6 @@
 # PR Review / File Tree Performance Plan
 
-Status: phases 1–5 implemented; real-PR verification, request-path remediation, and active-patch prioritization complete; cold-fetch decision remains
+Status: phases 1–5 implemented; real-PR verification, request-path remediation, active-patch prioritization, and warm review-thread remediation complete; tree and cold-fetch decisions remain
 Prior art: `.plans/archived/DIFF_UI_PLAN.md`, `.plans/archived/DIFF_REVIEW.md`
 
 ## 2026-07-17 reconciliation
@@ -266,6 +266,59 @@ but before the custom element painted. A direct DevTools trace measured 498 ms
 LCP, 0.00 CLS, and a 36 ms Pierre forced-reflow total with no estimated
 user-visible savings.
 
+## 2026-07-18 review-thread surface result
+
+A production-only follow-up isolated the remaining thread delay. The previous
+1,724 ms result is retained above, but the rerun found that its browser origin
+included the Vite development client despite being described as production.
+React Strict Mode could start and abort an initial request in that mode. The
+benchmark now detects Vite development mode and refuses it unless
+`--allow-development` is explicit. It also records failed and aborted API
+requests plus thread request start, response end, duration, bytes, and
+post-response render time.
+
+The clean production baseline still missed: one 381 KB thread response took
+about 810 ms directly and thread visibility measured 1,644 ms median. Three
+instrumented reloads showed a roughly 362 ms request start, 1,011 ms request
+duration, and only 11 ms of React work after the response. This confirmed the
+GitHub GraphQL path, not TanStack Query, React, or Pierre rendering, as the
+bottleneck.
+
+The review surface now uses a dedicated GitHub query that omits diff hunks,
+review/database ids, and pull-request backreferences that the UI does not read.
+The full Flue action continues using the full-fidelity query. The HTTP response
+also omits duplicate unresolved-thread and unresolved-comment collections; the
+client derives unresolved threads from the canonical list. The lean uncached
+run reduced thread transfer from 381,244 B to 58,240 B, but still measured
+1,511 ms thread visibility because the GitHub request remained 906–1,119 ms.
+
+Because the lean query could not reach the warm budget alone, it is backed by a
+small in-process cache: 15-second TTL, 16 entries, token-scoped keys, and
+explicit invalidation after review submission, thread reply, and
+resolve/unresolve. Reads invalidated while in flight are not stored. In-flight
+requests are not shared, so one browser cancellation cannot cancel another
+caller's work. Browser cancellation is propagated through the GraphQL request
+to GitHub.
+
+| Path                           | Active priority | Lean, no cache | Lean + warm cache | Change / verdict                      |
+| ------------------------------ | --------------: | -------------: | ----------------: | ------------------------------------- |
+| Production tree visible        |          612 ms |         639 ms |            642 ms | Still misses; separate tree follow-up |
+| Production first patch visible |          798 ms |         924 ms |            934 ms | Median passes                         |
+| Production threads visible     |        1,724 ms |       1,511 ms |            459 ms | 73.4% faster; median passes           |
+| Initial backend thread read    |          917 ms |         684 ms |            655 ms | Cold GitHub round trip remains        |
+| Warm backend thread read       |    not recorded | not applicable |            6.8 ms | Short-lived in-process reuse          |
+| Thread transfer per sample     |       381,244 B |       58,240 B |          58,240 B | 84.7% smaller                         |
+| Total API transfer per sample  |       742,850 B |      400,821 B |         400,821 B | 46.0% smaller                         |
+| Thread render after response   |    not recorded |        16.8 ms |            6.5 ms | Rendering remains negligible          |
+| Aborted or failed API requests |               0 |              0 |                 0 | Stable production path                |
+
+The final three thread samples were 608, 452, and 459 ms. The first sample's
+604 ms FCP made a sub-500 ms thread paint impossible even though its cached
+thread request took 23 ms; the median satisfies the warm harness target. A
+separate final DevTools trace measured 530 ms LCP, 0.00 CLS, one successful
+thread request, and zero estimated FCP/LCP savings from render-blocking
+resources. The cold first read remains explicitly outside the warm-cache pass.
+
 ## Confirmed follow-up candidates and remediation status
 
 1. **Completed — stabilize review-thread identity.** `reviewThreads(pr)`
@@ -283,15 +336,21 @@ user-visible savings.
    isolated from adjacent, draft, and unresolved background reads until it
    settles. The next run reached a 798 ms median and passed the target in all
    three samples without abandoned reads.
-4. **Discuss later — revisit cold fetch.** The 4.98-second object fetch misses the target,
+4. **Completed — slim and briefly reuse review-surface threads.** The web path
+   now uses an 84.7% smaller query response plus a bounded 15-second cache with
+   mutation invalidation and race protection. Production thread visibility is
+   459 ms median, while the full-fidelity Flue action is unchanged.
+5. **Discuss later — revisit cold fetch.** The 4.98-second object fetch misses the target,
    but it is a one-time revision cost. Separate network fetch time from local
    metadata time before changing refspecs or the `<3s` budget.
 
 Acceptance is partial on the same real target: duplicate thread requests and
 settlement-driven abandoned patch reads are eliminated, the first-patch browser
-budget now passes, backend targets pass, and fallback code is unchanged. The
-tree, threads, and one-time cold-object budgets still miss and remain separate
-follow-ups. Raw baseline, remediation, and active-priority results are
+budget now passes, warm thread visibility passes on the median, backend targets
+pass, and fallback code is unchanged. The tree and one-time cold-object budgets
+still miss and remain separate follow-ups; a cold GitHub thread read also
+remains slower than the warm UI budget. Raw baseline, remediation, and
+active-priority results are
 gitignored at
 `benchmarks/results/pr-12204-real-local.json` and
 `benchmarks/results/pr-12204-remediation-local.json`, and