[sh1pt-actions-fleet]

Anthony: "summarizing is failing right now" (on pit.world/search?q=saas).

What was wrong

1. Same summary for different results. Three results summarized at once (Wikipedia, Salesforce, nichedb) all came back with one summary. The TronBrowser relay handed concurrent reads each other's page: one Obscura tab, navigate then read, interleaved. That is fixed at the source in fix(automate): concurrent fetch_page calls returned another call's page profullstack/tronbrowser.dev#176. Here, summaries written since 2026-11-02 11:00 UTC, when pages started being read through the relay, are dropped once at startup and rewritten on the next request.
2. "The summarizer did not answer". Infernet answers 503 No live node serves "qwen2.5:7b", so jobs fell to any node. One node (dev-vienna) took a job and never answered, which used up the whole 90 s.
   • On the 503, pit.world now asks for the best general model Infernet lists as live (FALLBACK_MODELS, never a fine-tune or an abliterated model), then any node.
   • The first node gets about 55% of the budget, then there is one retry on another node. Refused keys (4xx) are not retried.

Tests

bun test: 413 pass, 0 fail. tsc is clean. New tests cover:

• the model pick
• the 503 → listed model → any node chain
• a stuck node retried once, in under 6 s
• a 401 not retried
• the one-time cleanup drops only rows in the window and survives a restart

🤖 Generated with Claude Code