pi-auto-compat

A Pi coding agent extension that automatically fixes missing compat flags in models.json — so warnings like 💡 pi-cache-optimizer: ... merged compat lacks ... and the ⚠️ compat footer marker never appear.

Detection mirrors pi-cache-optimizer exactly (same priority chain), and the source of truth is the merged model list from ctx.modelRegistry — not a static file scan. Providers without a models.json entry are patched too (a minimal compat-only entry is created; credentials are never touched).

What it fixes

0. Universal long cache retention — every model/provider gets supportsLongCacheRetention: true when not explicitly set (pi already defaults it to true; declaring it surfaces the "Caching" badge and silences the cache-optimizer warning). Explicit false is respected; Pi's built-in llama.cpp is excluded. New models added via /better-custom are covered automatically — no per-model injection needed.
1. Adaptive generation (anthropic-messages + Opus/Sonnet ≥ 4.6, Fable ≥ 5, or Kimi Coding K3) → forceAdaptiveThinking: true (+ allowEmptySignature for K3 empty-signature models)
2. DeepSeek-like (openai-completions / openai-responses) → supportsLongCacheRetention, requiresReasoningContentOnAssistantMessages, thinkingFormat: "deepseek" (+ sendSessionAffinityHeaders on completions)
3. Claude-like on OpenAI-compatible proxies → cacheControlFormat: "anthropic"
4. Non-official OpenAI-compatible proxies (openai-completions) → sendSessionAffinityHeaders: true — only when undefined; explicit false is a valid anti-403 opt-out and is never overwritten
5. Reasoning models without thinkingLevelMap (static rule) → default { low, medium, high, xhigh } map

How it works

• Fix placement mirrors pi-cache-optimizer /fix: channel keys (affinity/retention) go provider-level; model-behavior keys go model-level (models[].compat or modelOverrides) unless every sibling model is compatible.
• Extension-owned providers are special-cased: when an extension registers a provider with its own models list (or refreshModels), Pi applies that list after models.json, so provider-level compat and models[].compat never reach the merged model — only modelOverrides do. All fixes for such providers are therefore written to modelOverrides (detected via ctx.modelRegistry.getRegisteredProviderConfig).
• After writing, the registry is refreshed in-process (modelRegistry.refresh({ allowNetwork: false })) — no /reload needed.
• Triggers: session_start, model_select, the models.json file-watcher, and the manual /auto-compat command.
• A timestamped backup is written before each save (max 3 kept).

Install

pi install npm:@bismawy/pi-auto-compat

Then run /reload in Pi (or restart it). Optionally run /auto-compat once to fix everything immediately.

Manual install (from source): copy index.ts into %USERPROFILE%\.pi\agent\extensions\.

Safety

• Never copies, moves, or writes credentials (apiKey, OAuth) — new provider entries are compat-only.
• Never deletes existing values; only fills/repairs the specific compat keys listed above.
• Skips official OpenAI endpoints and Pi's built-in llama.cpp provider.