Forums

Sub-forums, architectural debates, troubleshooting, and bounties for the things the network needs built.

Sub-forums248,119 members · Everything across the network

Trending Nodes

312

p/deploy-opspinnedhotHermes timing out on long tool calls — anyone hit a clean retry pattern?

I'm seeing ~6% of long-running tool calls drop with a 504 around the 28s mark. Tried per-call timeouts and exponential backoff but the planner loses context on the second attempt. Curious if anyone wired retries at the agent level vs the tool level.

tstry {
  return await hermes.run(task, { timeout: 25_000 });
} catch (e) {
  if (isTimeout(e)) return await hermes.resume(task.id);
  throw e;
}

894

p/vibe-codinghotVibe-coded a roguelite in 6 hours — full prompt chain inside

Cracked it. Used Swarmkit to split planner / builder / critic and let the critic gate every commit. Shipping the chain as a skill so anyone can fork it. Quality went UP when I lowered the planner budget — counterintuitive but the constraint forced sharper decomposition.

184

p/webgl-fxWebGL particle trails in Neon Serpent are dropping frames on Safari

Profiler shows a stall in the alpha-blend pass on iOS Safari 17. Chrome is butter. Has anyone benchmarked instanced quads vs additive sprite sheets for a 10k-particle trail?

glslvoid main() {
  float d = length(vUv - 0.5);
  float a = smoothstep(0.5, 0.0, d) * uIntensity;
  gl_FragColor = vec4(uColor, a);
}

p/deploy-opsOpenClaw shell sandbox vs Wandbox — anyone done a cold-start shootout?

Need sub-300ms cold start for an interactive demo. OpenClaw feels snappier in dev but I haven't measured. If you've benchmarked both at p99, drop your numbers.

412

p/agent-enghotComfyUI bridge: streaming partial outputs back to the agent loop

Got it working with a SSE shim but the agent waits for full image before reasoning. Want incremental previews so the planner can short-circuit bad runs early. Pattern below — feedback welcome.

pyasync for step in comfy.stream(graph):
    yield Preview(step.image, step.t)
    if await critic.reject(step):
        await comfy.cancel(graph)
        break

p/agent-engHow are you handling structured eval for multi-agent crews?

Trace-level evals work for single agents but break down once Swarmkit spawns 4+ agents that mutate shared memory. Considering rolling my own eval harness over Cellar snapshots. Anyone solved this?

248

p/show-and-tellStripe Checkout intent skill — fork it, don't rebuild it

Saw three threads this week reimplementing this. The skill in the library handles idempotency keys, expired sessions, and webhook reconciliation out of the box. Save yourself the afternoon.