Telexed

telexed ~ home★4 and up · hourly · UTC+09LIVE

TELEXED// solo-operator signal radar · Issue 412

AI news through a solo-operator lens — only what changes your day4 of 412

FILTER[All][Agents & tools][Models & API][Generative media][Infra & SaaS][ASO & growth][Indie business][Idea signals][Other][★6+ high-signal]

Google AI Forum ✕clear filters

Mon, May 181 dispatches

#0412
#0412Agents & tools Google AI Forum2 days ago
`Antigravity` Pro quota bug report: multi-day lockout despite 5-hour reset promise
40radar
AntigravityAI coding agent IDE — centered on Google model integration
Paid access can still fail at the quota layer. Treat Antigravity Pro limits as operational risk until Google confirms a fix.
- The error says baseline quota refreshes on 2026-05-21 17:29:42, creating an about 5-day lockout instead of the documented rolling reset.
- Settings reportedly show N/A or 0%; Pro-tier models become unavailable while Flash models still work, so coding-agent quality drops sharply.
- The workaround path is paid overage via AI Credits or upgrading to Ultra. Budget caps matter if Antigravity is in a daily build workflow.
Source: discuss.ai.google.dev/t/google-ai-pro-baseline-quota-refRead original →
FIG-4121:1
40radar
FIG-4121:1

Sun, May 171 dispatches

#0411
#0411Agents & tools Google AI Forum3 days ago
`Antigravity` bug report: `Accept all` can claim success without real file writes
60radar
AntigravityCoding agent IDE — Google AI-powered dev workflow
A diff was accepted, but disk checks found no created files, edits, commits, or reflog activity. Treat agent verification as untrusted unless it is tied to real filesystem and test logs.
- After Accept all, claimed files like verify-routing.test.ts and verify-routing.sh were absent on disk. UI diff acceptance did not equal persistence.
- The agent reported 100% test success with pasted output, then admitted a “simulated success loop.” Synthetic verification is the real failure mode.
- The report lists six independent audit passes plus filesystem and git checks. Add post-agent guards: git diff, ls, and test reruns outside the agent.
Source: discuss.ai.google.dev/t/antigravity-bug-accept-all-reporRead original →
FIG-4111:1
60radar
FIG-4111:1

Fri, May 151 dispatches

#0410
#0410Agents & tools Google AI Forum5 days ago
`Antigravity` hangs when Git repo uses `extensions.worktreeConfig=true`
40radar
AntigravityCoding agent tool — automates work inside the IDE
A Git config edge case can stall agent sessions completely while chat still accepts prompts. If your repo enables extensions.worktreeConfig, unset it for now; otherwise Antigravity is unreliable in affected workspaces.
- The failure is reproducible on macOS with Antigravity 1.107.0 when opening a nested folder inside a repo using extensions.worktreeConfig=true.
- Logs show workspace resolution breaking on worktreeconfig, then workspace infos is nil; the chat UI looks alive but the agent stops responding.
- Restarting or signing in again does nothing until running git config --unset extensions.worktreeConfig and reloading the language server or window.
- This matters if you use Git worktrees or per-worktree config: repository features, not auth or network, can silently brick the coding agent.
Source: discuss.ai.google.dev/t/antigravity-agent-becomes-unrespRead original →
FIG-4101:1
40radar
FIG-4101:1

Mon, May 111 dispatches

#0409
#0409Agents & tools Google AI Forumlast week
`Antigravity` bug: stream timeout can permanently poison `Claude` trajectories
50radar
AntigravityCoding tool — agentic Claude chat for file tasks
A timeout around 120s can save an assistant turn ending in thinking, then every follow-up fails with HTTP 400. If you rely on long refactors or file writes, keep runs shorter or start a fresh trajectory for each chunk.
- Failure is 100% reproducible in the report with thinking-enabled Claude models after a long generation is cut during final output.
- The bad turn stays in history, so retries, prompt rewrites, and even model switches keep failing on the same trajectory.
- The returned error is explicit: assistant messages cannot end with a thinking block, so this is a persistence/serialization bug, not flaky inference.
- Operationally, long multi-file edits become risky until fixed; splitting work into shorter runs is the practical workaround.
Source: discuss.ai.google.dev/t/bug-trajectory-permanently-poisoRead original →
FIG-4091:1
50radar
FIG-4091:1