🔄 Weekly pulse report: 2026-07-31

reports/2026-07-31.md

@@ -0,0 +1,75 @@
+# 📊 GitHub Pulse Scout — areal-project/AReaL
+
+**Window:** 2026-07-24T05:02:54.772Z → 2026-07-31T05:02:54.772Z **Issues fetched:** 8
+
+## Activity
+
+| Metric            | Count |
+| :---------------- | ----: |
+| 🆕 New issues     |     1 |
+| ✅ Closed issues  |     4 |
+| 📂 Currently open |     4 |
+
+## Response efficiency
+
+| Metric                 | Value |
+| :--------------------- | ----: |
+| Avg close time (hours) | 629.3 |
+| Avg close time (days)  | 26.22 |
+| Closed sample size     |     4 |
+| Unique issue authors   |     8 |
+
+## Top labels
+
+| Label                   | Count |
+| :---------------------- | ----: |
+| `stale`                 |     5 |
+| `bug`                   |     3 |
+| `question`              |     1 |
+| `call-for-contribution` |     1 |
+| `roadmap`               |     1 |
+
+## 🔥 Hottest issues
+
+- [#1381](https://github.com/areal-project/AReaL/issues/1381) — [Roadmap] 2026 H2
+  Milestones (💬 **10**)
+- [#1560](https://github.com/areal-project/AReaL/issues/1560) — [BUG] Megatron–vLLM
+  weight sync incorrectly requires matching pipeline-parallel sizes (💬 **2**)
+- [#1451](https://github.com/areal-project/AReaL/issues/1451) — [Feature] Support
+  cross-tokenizer on-policy distillation (💬 **2**)
+- [#1442](https://github.com/areal-project/AReaL/issues/1442) — [BUG] attn_impl=sdpa
+  silently produces wrong logp with packed sequences on FSDP+HF backend (💬 **2**)
+- [#1419](https://github.com/areal-project/AReaL/issues/1419) — [BUG] Partial rollout
+  groups silently corrupt GRPO group normalization (💬 **2**)
+
+## 🤖 AI Summary
+
+# Pulse summary — last 7 days
+
+### Overall
+
+Momentum is **steady but slow**: 4 issues closed against 1 new, leaving 4 open amid
+**~26-day average resolution**. Stale flags and deferred bugs suggest bandwidth
+constraints and a backlog of correctness fixes.
+
+### Trends & focus
+
+- Activity is light; **bug and stale** dominate labels, with roadmap planning (#1381)
+  driving most discussion.
+- Response efficiency lags; deep technical bugs (e.g., #1560, #1442) linger weeks before
+  closure or deferral.
+- Participation is broad (**8 unique authors**) but thin per issue; many reports stall
+  after initial triage.
+- Recurring themes: silent correctness bugs in training/rollout logic, backend coupling,
+  and teardown reliability.
+
+### Recommendations
+
+- Triage stale bugs for quick patches or explicit defer plans to stop silent
+  regressions.
+- Add targeted tests for SDPA/packed-sequence and GRPO normalization edge cases.
+- Clarify ownership and SLAs for critical-path bugs to accelerate turnaround.
+
+<small>Generated on 2026-07-31T05:02:54.775Z by
+[GH Pulse Scout](https://github.com/oss-infra/gh-pulse-scout), a tool that analyzes
+GitHub issue data to provide insights into project health and trends with AI 💖.</small>