`DeepSWE` Flags Leakage and False Negatives in AI Coding Benchmarks

// related

• #0001

  #0001Agents & toolsGoogle AI Forum3 hours ago

  `Gemini Code Assist` Individual IDE/CLI Access May End June 18

  80radar

  Gemini Code AssistAI coding assistant — Gemini in IDE extensions and CLI

  Individual Pro, Ultra, and free access stops on June 18, 2026. VS Code users should line up alternate auth or tools before agent workflows break.

  • —The notice names Gemini CLI and Gemini Code Assist IDE extensions. Individual Pro, Ultra, and free accounts are the affected paths.
  • —The cutoff date is June 18, 2026. Any VS Code workflow wired to personal Gemini access needs a fallback before then.
  • —The forum post does not confirm whether an Antigravity-specific VS Code extension continues. Treat it as unresolved until Google posts a clearer migration path.

  Source: discuss.ai.google.dev/t/will-there-no-longer-be-a-geminiRead original →
• #0002

  #0002Agents & toolsHacker News · Show HN AI3 hours ago

  `AISlop`, a CLI that catches AI-generated code smells

  60radar

  AISlopAI code-quality CLI — local code-smell scanning

  Scans for patterns tests often miss: empty catches, dead code, duplicate helpers, and useless comments. Local-only hook integration makes it worth trying in agent-heavy repos.

  • —npx aislop scan runs locally, so source code is not uploaded. That lowers the barrier for private app repos.
  • —Targets AI-agent residue rather than syntax errors: empty catch, dead code, duplicated helpers, and low-value comments.
  • —Hook-based use after each tool call fits Claude Code, Codex, and opencode workflows where small slop accumulates fast.

  Source: github.com/scanaislop/aislopRead original →
• #0003

  #0003Agents & toolsGoogle AI Forum5 hours ago

  `Antigravity Ultra` users hit persistent `MODEL_CAPACITY_EXHAUSTED` on Claude Opus thinking

  60radar

  AntigravityCoding agent IDE — uses Google's backend model routing

  Paid agent sessions can fail before real work starts. Treat Antigravity Ultra as capacity-risky for Claude-heavy workflows until routing improves.

  • —The backend returns 503 `MODEL_CAPACITY_EXHAUSTED` from cloudcode-pa.googleapis.com, so this is server capacity, not local quota.
  • —Failures recur after 1-29 seconds on multi-step agent tasks like file edits, bash runs, and code generation. Long tasks are unreliable.
  • —Server-Timing shows about 13s spent waiting for model capacity before failure. Retry loops burn time without changing the result.
  • —Multiple similar forum threads over nearly a month point to account or routing-level fragility, not a one-off client setup problem.

  Source: discuss.ai.google.dev/t/persistent-model-capacity-exhausRead original →