Hermes NoneType Fix, React Doctor, Seedance TTS — AI Daily May 27

484 messages · 87 active members

484
messages
87
active members
@arielletolome, @samb69, @roki1122
top contributors

Overview

Hermes infrastructure dominated the day. A bad Codex response shape (response.output = None) crashed the OpenAI Python SDK parser with 'NoneType object is not iterable,' taking down Hermes community-wide. @rmktg shipped a custom wrapper with a Google Doc writeup and @weslindquist had his agent document the root cause — update Hermes to the latest version to apply the fix. Compounding the pain: @sav310 and @fuckyesiwannatalkbusiness discovered the post-update Hermes config has TWO fallback sections (main chat + auxiliary/system), and the system one silently overrode their fallback_providers, routing auxiliary tasks to paid OpenRouter while Codex OAuth was down. On the build side, @Wootbro's `npx react-doctor@latest` one-shot agent prompt was the day's most actionable share — builders ran it on staging code and saw scores jump from 66 to 100, with @archimorty finding the root cause of a long-standing bug (warning: running overnight on /fast burned 20% of his weekly Claude usage). @samb69 cracked Seedance pronunciation issues by feeding TTS-generated audio directly instead of script + voice reference, and @swh800 shared a technique of stacking reference images over text prompts. @Max_Bernstein published ad pipeline notes including a 'brief at each stage' pattern to stop agents skipping gates. Rounding out the day: @mattkowald confirmed OpenAI Ads (ads.openai.com) is accepting signups with a rumored $200k minimum spend, @navuud wired an old Android phone to his agent to auto-lock his Tacoma and Lexus at 10pm, and a long productivity thread collapsed (as always) to the consensus that calendar time blocking is what actually ships work regardless of upstream tool.

Topics

A Codex response with output = None crashed the OpenAI Python SDK parser, taking Hermes down community-wide; @rmktg and @weslindquist published the fix. Separately, the post-update Hermes has two fallback sections — @sav310 found the auxiliary/system one silently overrode fallback_providers and routed to paid OpenRouter during the OAuth outage.

@Wootbro shared `/goal run npx react-doctor@latest and fix issues until you get a score of 100` as a drop-in prompt. Multiple builders reported scores jumping from 66 to 100 and long-standing bugs surfaced, but @archimorty warned overnight runs on /fast can burn 20% of weekly Claude quota.

@samb69 fixed persistent Seedance pronunciation issues by feeding TTS-generated audio directly instead of script + voice reference, and called it the most comprehensive video gen suite (12+ reference elements, no morphing). @swh800 separately demonstrated that stacking reference images outperforms text-only prompting, with Omni's compliance filters as the main friction point.

@mattkowald confirmed approval on ads.openai.com, with @samarh90 noting access is still gated. @rmktg referenced a rumored $200k minimum spend, while @justingacina reported running ads with only a handful of impressions so far. @Max_Bernstein also shared a public doc of ad pipeline lessons including a per-stage brief pattern to stop agents skipping gates.

Members compared Notion, ClickUp, Asana, EOS, Workflowy, and Apple Notes, with @realcrischico, @benjifenji, and @seekersight converging on calendar time blocking as the real shipping mechanism. On the automation side, @navuud rigged an old Android phone so his agent auto-locks his Tacoma and Lexus at 10pm, and @fmill1's AI reply agent is booking so many calls his VA wants it disabled.

Key Takeaways

  • If Hermes throws 'NoneType object is not iterable,' update to the latest version — and check BOTH fallback sections (main + system/auxiliary), since the system one supersedes and can silently route to paid OpenRouter.
  • `npx react-doctor@latest` is a worthwhile one-shot agent audit for React apps — but don't run it overnight on /fast unless you want to torch your weekly Claude quota.
  • Seedance accepts TTS-generated audio as direct input (fixing pronunciation issues script+voice-reference can't) and stacking reference images beats text-only prompting.
  • OpenAI's ads platform (ads.openai.com) is accepting signups but is gated, with a rumored $200k minimum spend — early access is no guarantee of impressions yet.
  • For multi-stage agent pipelines, force the agent to write its own brief at each gate to prevent it from skipping ahead despite strict instructions.

Hot Threads

@roki1122started

Hermes / GPT-5.5 down — NoneType crash and the community fix

18 replies8 participants
@launchmanmathiasstarted

Planning and time management systems for AI builders

22 replies7 participants
@Wootbrostarted

React Doctor one-shot agent prompt and score results

15 replies8 participants

Linked Items