X Platform November 15 AI Brief | SpaceX Starship Completes First Tower Catch, Grok Bot Team Shares Code Quality Practices, Alibaba Open Sources OpenCodeReview
01 Starship Flight 15 Completes First Tower Catch, Booster Loss Cause Awaits Telemetry Review
SpaceX officially announced Mechazilla caught the Starship, with Musk celebrating and stating he hugged the returning Ship 42; this is described as the first Starship tower catch. Musk further explained that the booster loss was related to slight underperformance of the engines, with green flames resulting from copper burning after the regenerative cooling loop exhausted its fuel and the main combustion chamber being in an oxygen-rich state; improving Raptor V3.2 performance consistency and moving MECO1 forward by about one second were mentioned as adjustment directions. However, he cautioned that the initial assessment might be biased and requires full flight telemetry analysis, so the cause should be considered preliminary.
- @SpaceX: https://x.com/SpaceX/status/2121715216597827756
- @elonmusk: https://x.com/elonmusk/status/2121648438144024943
- @elonmusk: https://x.com/elonmusk/status/2121666488214204441
02 Engineering Practices Shared by Grok Bot Team: Solidifying Corrections into Checks and Having Agents Submit Reviewable Runtime Evidence
A translated summary from an AI Engineer conference talk focuses on code quality governance: turning repeatedly corrected issues into automated rules like lint, cleaning up bad patterns that subsequent Agents might imitate, and directly disabling error-prone features; the speaker also introduced having Agents run programs like users, submitting screen recordings and screenshots for human review. The summary states the team merged about 11,000 PRs last month, with a rollback rate of about 0.5%, and application speed has improved compared to the initial release. These figures and practices are from the blogger’s recounting of the talk, not independent audit data.
- @dotey: https://x.com/dotey/status/2121835140850602200
- @poteto: https://x.com/poteto/status/2121673563489140863
03 Local and Web Test Times Drop Significantly After Cleaning Up AI-Generated Invalid Tests
A developer shared their investigation into test and CI bottlenecks: issues included treating grep source code as tests, serially cold-starting Node and git in loops, and uncleaned update-check requests causing server.close() to wait an extra 3.47 seconds. After removing useless tests and stray requests, and splitting fast/slow tests, they reported local unit tests dropping from 39 seconds to 8.8 seconds, and web tests from 200 seconds to 68 seconds, and released the open-source test-value skill to organize methods for handling slow tests and invalid assertions. This is a single retrospective from the author’s project, not a universal benchmark.
- @t20000622yy: https://x.com/t20000622yy/status/2121894954050163153
04 Darwin.skill 3.0 Switches to Evaluating Skills Based on Actual Task Output, Not Just Comparing Documentation
The author explains that the new Darwin.skill version uses two skill versions on actual tasks, then has judges blind-review the outputs, with scripts making the final call, keeping the shorter version if tied. The motivation for the change: a documentation review selected the new version 3-0, but actual runtime resulted in 33 more tool calls; the author also cites NVIDIA research indicating low correlation between reading documentation scores and actual runtime effectiveness. This is the author’s explanation of the tool design and a single iteration experience.
- @AlchainHust: https://x.com/AlchainHust/status/2121744774877389261
05 Alibaba Open Sources OpenCodeReview, Providing Review Comments Per Code Line and Integrable with Programming Agents
The blogger introduces that Alibaba has open-sourced OpenCodeReview, internally refined for two years, usable via API and integrable with Claude Code or Codex; the tool provides modification suggestions for specific code lines and can check for security and logic issues in unfamiliar project directories. The post also claims it has served tens of thousands of Alibaba developers and intercepted millions of defects; these internal scale and effectiveness figures are from the blogger’s introduction, with no independent verification in the daily materials.
- @Gorden_Sun: https://x.com/Gorden_Sun/status/2121934388363768062
06 Alma 0.4.672 Integrates Poppy, Tests Shopify UCP Checkout Capability Negotiation
The Alma author explains that version 0.4.672 can automatically discover a merchant’s `/.well-known/poppy.json`, carrying a signed Poppy-Session header when accessing APIs, pages, or MCP servers to identify the personal Agent acting on behalf of the user. The author also states that using Alma’s UCP profile to test Allbirds’ Shopify UCP endpoint negotiated capabilities like checkout, cart, and order; this is a single endpoint test reported by the version’s author and does not mean all merchants already support it.