Pi Mono: Reliability Sweep Across Retry Logic and Thinking Models
A wave of fixes hardened retry behavior for flaky provider errors and repaired how thinking and reasoning tokens survive model renames, replays, and gateway quirks. Two coding-agent fixes also corrected how prompts and shell errors are recorded, closing gaps that had been silently degrading eval accuracy.
Duration: PT2M42S
Episode overview
This episode is a short developer briefing from Pi Mono.
It explains recent repository work in plain language.
- Show: Pi Mono
- Published: 2026-09-18T06:05:29Z
- Audio duration: PT2M42S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's September 18th, and this is Pi Mono.
The clear signal today: retry logic and thinking-model handling both got a serious hardening pass, touching nearly every major provider integration.
On retries, four separate fixes landed for distinct failure modes. Azure peak-load capacity errors and Cloudflare 520 responses now get retried instead of failing immediately, from commits e98f287 and e5d1838. PR 9724 fixes a case where malformed Retry-After dates produced invalid delays, causing 429 retries to fire…
The second theme is thinking and reasoning token integrity across providers. David Brailovsky's commits fixed Gemini's unsupported thinking levels, preserved thinking replay through renamed Anthropic models by separating the requested and provider-reported model IDs, and enabled unsigned thinking replay for Vercel…
On the coding-agent side, two fixes matter for anyone relying on evals or transcripts. Commit 1629239 changed how forced system prompts are recorded, ensuring the transcript reflects the actual prompt sent to providers rather than a stale replacement. PR 9706 then aligned eval validation to read from that corrected…
Also worth noting: PR 9630 added event…
Nearby episodes from Pi Mono
- Rewriting the Rules of the Conversation Transcript
- Extension Exports and Provider Cleanup
- Provider Sprawl and Small Fixes with Big Payoffs
- Weekly Recap - Model Provider Cleanup & Session Reliability
- Hardening the Model Provider Layer
- Smoothing the Chat Experience
- Cross-Platform Fixes and Token Accounting Cleanup
- Model Catalog Cleanup and Agent Reliability Fixes