pi-minimax-m3

A standalone pi extension that fixes two issues with the built-in MiniMax-M3 integration:

1. Silent over-billing on the Anthropic-compatible endpoint. M3's /anthropic/v1/messages endpoint ignores cache_control markers, so every turn was billed at the full input price ($0.60/Mtok) instead of the cache-read price ($0.12/Mtok). M3 does support passive/automatic prompt caching on its OpenAI-compatible endpoint (/v1/chat/completions).
2. Duplicated thinking in the response. M3 emits thinking content twice: once in reasoning_content (consumed by pi as a thinking block) and once in content wrapped in <think>…</think> markers (which would otherwise appear inside the visible text).

This extension registers two new providers — minimax-m3-clean and minimax-cn-m3-clean — that route MiniMax-M3 to the OpenAI-compatible endpoint so passive caching works. The thinking cleanup is performed during the stream by wrapping the built-in openai-completions streamSimple driver and rewriting the event stream in flight: duplicated thinking from M3's reasoning_content / reasoning field alternation is suppressed, <think>…</think> spans are filtered out of text deltas (and their inner content is routed to a real thinking block when no reasoning fields were streamed), and text_start is deferred until the first non-whitespace character.

It mirrors the upstream fix in pi-mono@b85b91c9 ("route MiniMax-M3 to openai-completions for passive caching") so users can get the fix on any pi version without waiting for an upstream release.

Install

From npm:

pi install npm:pi-minimax-m3-caching-fix

From a git checkout (latest, or pinned):

pi install git:github.com/rwese/pi-minimax-m3-caching-fix
pi install git:github.com/rwese/pi-minimax-m3-caching-fix@v0.2.0

For local development from a clone:

git clone https://github.com/rwese/pi-minimax-m3-caching-fix
pi install ./pi-minimax-m3-caching-fix

The extension reuses the env vars you already have for the built-in minimax provider — no new credentials required:

Provider	Env var	Endpoint
minimax-m3-clean	MINIMAX_API_KEY	https://api.minimax.io/v1
minimax-cn-m3-clean	MINIMAX_CN_API_KEY	https://api.minimaxi.com/v1

Quickstart (for the impatient)

# 1. Make sure your MiniMax API key is exported
export MINIMAX_API_KEY="sk-..."

# 2. Install the extension
pi install npm:pi-minimax-m3-caching-fix

# 3. Restart any running pi session, then start one
pi

# 4. Inside pi, switch the model
/model
#   pick:  minimax-m3-clean / MiniMax-M3 (clean)

# 5. Verify caching — look at the footer or session log
#    Turn 1: ~99% cache miss (system prompt being written to cache)
#    Turn 2+: ~99% cache read (system prompt being reused)

That's it. No new credentials, no config file, no restart of the upstream minimax provider. Just pick the right model in /model and the rest happens automatically.

Use

1. Run pi.
2. Open the model picker with /model.
3. Pick minimax-m3-clean / MiniMax-M3 (clean) for the global endpoint or minimax-cn-m3-clean / MiniMax-M3 (clean — CN) for the China endpoint.
4. Send a prompt. The first turn is a cache miss; subsequent turns of the same session show a CH (cache hit rate) in the footer as the system prompt gets reused.

In the session log, the usage object on each assistant message shows the cache reads. For example, a 3-turn session looks like:

Turn	input	cacheRead	Hit rate
1	8932	114	1%
2	128	8946	99%
3	128	8946	99%

Why a separate provider (not overriding the built-in)

pi.registerProvider(name, { models }) replaces every model registered for that provider. There are two ways that breaks the built-in integration:

• Override minimax with baseUrl only — this lumps M2.x onto the OpenAI-compatible endpoint too, breaking M2.x.
• Override minimax with new models — this wipes M2.x from the registry.

So this extension registers new provider names (minimax-m3-clean, minimax-cn-m3-clean) that don't collide with minimax or minimax-cn. Users opt in by switching the model in /model. The built-in minimax / MiniMax-M3 model is still listed — pick the one with "(clean)" in the name.

Limitations

• Two MiniMax-M3 entries in /model. The built-in (broken, billing at full input price) and the extension's (clean) both appear. Pick the one with (clean) in the name.
• Requires both env vars for both providers to show. pi only lists providers that have auth configured. If you only have MINIMAX_API_KEY, only minimax-m3-clean shows up; set MINIMAX_CN_API_KEY (even to a dummy value) to also see minimax-cn-m3-clean.

How the fix works

The extension does two things:

1. Routes M3 to /v1/chat/completions by registering the two new providers under a custom api id (the provider name) so the wrapper below only intercepts these models. The model metadata mirrors packages/ai/src/models.generated.ts from the upstream fix: input: ["text", "image"], reasoning: true, cost $0.6 / $2.4 / $0.12 per million tokens, 1M-token context window, 512K max output.

2. Cleans M3's thinking in the stream wrapper. The wrapper sits in front of the built-in openai-completions streamSimple driver and rewrites events as they arrive:

   • All driver thinking blocks are merged into ONE thinking block. M3 re-streams the same reasoning when it switches between reasoning_content and reasoning fields, which would otherwise start a new (truncated) thinking block on every field switch. The wrapper dedupes by prefix and emits only the new portion of reasoning.
   • A ThinkScanner filters <think>…</think> spans from text deltas in real time and holds back bytes that look like the start of a tag so markers split across deltas are classified correctly. If the model never streamed reasoning fields, the captured inner content is routed to a real thinking block instead of being dropped; otherwise it's a duplicate of the reasoning fields and is discarded.
   • text_start is deferred until the first non-whitespace character so empty / whitespace-only text blocks are not rendered.

   This is the same effect as the upstream compat.skipThinkingBlock flag, but applied in the stream wrapper because the user's installed @earendil-works/pi-ai (0.79.1) predates that compat field. When a future pi-ai release includes skipThinkingBlock, the wrapper becomes a thin pass-through and can be deleted.

Removing the extension (when upstream ships the fix)

When pi-mono ships a release that includes b85b91c9 (or any release whose models.generated.ts lists MiniMax-M3 with api: "openai-completions" and skipThinkingBlock: true), retire the extension:

pi remove npm:pi-minimax-m3-caching-fix

The built-in minimax / MiniMax-M3 model will then route correctly out of the box.

License

MIT — see LICENSE.

Credits

The in-flight thinking-cleanup wrapper introduced in v0.2.0 (the ThinkScanner, the merged-thinking block, and the deferred text_start) was contributed by Thunder Guardian (Discord: @Thunder Guardian).

Development

npm run check    # tsc --noEmit using the bundled tsconfig.json

The tsconfig.json configures --skipLibCheck and --moduleResolution bundler so the type check is reproducible without depending on transitive type packages of the user's installed pi.