@pi-vault/pi-providers

Register MiniMax M3 and M3.1 Flash, StepFun Flash, and Command Code models as custom providers for Pi, and add a typesafe_decide tool for typed decision questions answered by TypeSafe Jev. Requires Pi 1.4.0 or newer.

Install, Upgrade, And Reload

pi install npm:@pi-vault/pi-providers

Then reload Pi:

/reload

Quick Start

Set the API keys for the providers you want to use before launching Pi:

# MiniMax global endpoint (minimax-openai)
export MINIMAX_API_KEY="..."

# MiniMax China endpoint (minimax-openai-cn)
export MINIMAX_CN_API_KEY="..."

# StepFun Step Plan endpoint (stepfun-ai)
export STEP_API_KEY="..."

# Command Code provider (command-code), and the typesafe_decide fallback backend
export CMD_API_KEY="..."

# Optional secondary Command Code key name (CMD_API_KEY takes precedence)
export COMMAND_CODE_API_KEY="..."

# Optional: direct TypeSafe credential for typesafe_decide (preferred)
export TYPESAFE_API_KEY="..."

# Optional: enforce zero data retention for Command Code chat-provider requests only; excludes typesafe_decide
export CMD_ZDR=1

# Optional: override the live catalog endpoint only (inference bases stay unchanged)
export CMD_MODELS_URL="https://catalog.example.test/models"

You can also save a Command Code credential through Pi's login flow:

/login command-code

CMD_API_KEY is the primary environment variable; COMMAND_CODE_API_KEY is a secondary alias. CMD_MODELS_URL changes only where the live model catalog is fetched, never where inference requests are sent.

After reloading, the providers appear in Pi's model picker:

• Select minimax-openai (or minimax-openai-cn) and choose MiniMax-M3.1-Flash or MiniMax-M3.
• Select stepfun-ai and choose one of step-3.7-flash, step-3.5-flash-2603, or step-3.5-flash.
• Select command-code and choose any bundled Command Code model.

Ask Pi normally; there is no provider-specific prompt syntax. Image input works with MiniMax-M3.1-Flash, MiniMax-M3, step-3.7-flash, and command-code models such as minimax/minimax-m3.1-flash.

The Command Code provider uses the Provider API. GOAT membership can use the same API, but access and usage limits follow the current GOAT plan and pricing and limits. This package does not calculate subscription entitlements or included credits.

typesafe_decide is a tool, not a selectable chat model, so it never appears in the model picker and never accepts a chat prompt. Pi calls it like any other tool — see TypeSafe Decisions.

What's New In 0.5.0

• Authenticated MiniMax availability discovery — both direct providers query their regional /models endpoint and keep only supported curated IDs. A successful empty result hides all direct models, the last successful catalog restores offline, and failures retain the prior catalog for retry. Regions use separate credentials and caches.
• Direct MiniMax M3.1 Flash and refreshed M3 metadata — both regional providers now expose MiniMax-M3.1-Flash first and MiniMax-M3 second, with 1,000,000-token context, a 524,288-token output ceiling through max_completion_tokens, standard pricing, and image input. M3 pricing doubles above 512,000 input tokens. MiniMax-M2.7 and MiniMax-M2.7-highspeed are removed from these direct providers only.
• Reasoning continuity across tool rounds — native reasoning_content keeps its exact whitespace and signature, interleaved thinking and text stay in arrival order, and unsigned thinking from saved same-model sessions is replayed as ordered <think> text for the next request.
• Flash effort controls — MiniMax-M3.1-Flash exposes low, medium, high, xhigh, and max. Omitting effort sends no field and uses upstream's default; the simple path clamps minimal to low. M3 keeps automatic thinking.
• Refreshed Command Code catalog — 93 bundled models captured 2026-10-25, with updated capabilities and dated prices, including paid Ling pricing distinct from retired free aliases.
• TypeSafe question compatibility and gateway limits — optional or null instructions, criteria-only Noul questions, nullable Choice descriptions, and one-option Choice or one-level Score questions are accepted. Command Code requests enforce the 100-question limit before sending; direct TypeSafe requests keep their upstream limits.

See CHANGELOG.md for the full release notes, including reasoning replay and catalog refresh fixes.

What's New In 0.4.1

• COMMAND_CODE_API_KEY alias and catalog-only CMD_MODELS_URL override — CMD_API_KEY keeps precedence; CMD_MODELS_URL changes only where the live model catalog is fetched, never where inference requests are sent. Inference bases are unchanged.
• Live pricing, modalities, and reasoning metadata with strict catalog validation — valid live fields override local estimates; missing fields fall back to Pi donor metadata, official family defaults, or the dated bundled pricing snapshot. Malformed refreshes retain the prior catalog.
• Official family and adaptive-thinking fallbacks — retained alongside the bundled snapshot and offline cache. Pricing for temporary offers uses the rates advertised on 2026-09-30.
• Command Code output-cap and cache migration — default max output raised to 32,768 capped by each model's context window. Legacy cache records are migrated in place.
• OpenAI Responses routing — supported alongside Anthropic Messages and OpenAI Chat Completions. Records advertising only /responses no longer get filtered out.
• 86-model bundled snapshot — captured 2026-09-30 (was 82). See CHANGELOG.md for the full release notes.

What's New In 0.4.0

• typesafe_decide tool (new) — answers typed Noul, Choice, and Score questions against a JSON state with TypeSafe Jev and returns structured probabilities, not generated text. Calls TypeSafe System One directly when a TYPESAFE_API_KEY or host-resolved typesafe credential is present; otherwise makes a single fallback request to typesafe/jev through command-code. Direct 400 / 422 responses and cancellation are terminal and never fall back. System One requests never inherit CMD_ZDR, because typesafe/jev has no ZDR-capable Command Code upstream — the tool provides no enforced-ZDR mode.
• Command Code catalog refreshed to 82 models — snapshot captured 2026-09-27 (was 62). Routing is declared per record: /messages takes precedence, /chat/completions is used otherwise, and records advertising only /responses are filtered out because this package does not implement Responses API transport. Pricing estimates were also refreshed against the 2026-09-27 rates. See CHANGELOG.md for the full release notes.

Providers And Models

Provider	Model	Input	Reasoning	Context	Max output	Input / output per 1M tokens	Cache read
minimax-openai / minimax-openai-cn	MiniMax-M3.1-Flash	text, image	low / medium / high / xhigh / max	1,000,000	524,288	$0.15 / $0.60	$0.015
minimax-openai / minimax-openai-cn	MiniMax-M3	text, image	automatic	1,000,000	524,288	$0.30 / $1.20	$0.06
stepfun-ai	step-3.7-flash	text, image	low / medium / high	256,000	256,000	$0.20 / $1.15	$0.04
stepfun-ai	step-3.5-flash-2603	text	low / high	256,000	256,000	$0.10 / $0.30	$0.02
stepfun-ai	step-3.5-flash	text	automatic (shown as high)	256,000	256,000	$0.10 / $0.30	$0.02
command-code	93 bundled snapshot models (captured 2026-10-25)	varies	varies	varies	up to min(32,768, context)	varies	varies
command-code	minimax/minimax-m3.1-flash	text, image	low / medium / high / xhigh / max	1,000,000	up to min(32,768, context)	$0.15 / $0.60	$0.015

API bases:

• minimax-openai: https://api.minimax.io/v1
• minimax-openai-cn: https://api.minimaxi.com/v1
• stepfun-ai: https://api.stepfun.ai/step_plan/v1
• command-code: https://api.commandcode.ai/provider/v1

Direct MiniMax models use standard service pricing. MiniMax-M3 doubles to $0.60 / $2.40 / $0.12 per 1M input / output / cache-read tokens above 512,000 input tokens. Both models cap output at 524,288 tokens within the 1,000,000-token context and send max_completion_tokens. Cache writes are free for both direct MiniMax models and for all stepfun-ai models.

MiniMax-M3.1-Flash exposes the five supported thinking efforts: low, medium, high, xhigh, and max. Omitting effort sends no reasoning_effort field and preserves upstream's maximum default. On the simple provider path, minimal clamps to low. MiniMax-M3 keeps automatic thinking without an effort or disable control. MiniMax-M2.7 and MiniMax-M2.7-highspeed are no longer bundled by the direct providers; historical changelog entries remain.

Availability discovery. Both direct MiniMax providers require the regional API key before querying ${baseUrl}/models. The initial catalog is the curated two-model baseline. A validated response filters availability to curated IDs in curated order: IDs that are not part of the supported catalog are ignored, and a successful response with no supported IDs authoritatively hides every direct MiniMax model. New upstream IDs appear only after this package adds their curated metadata. Pi persists successful results per provider and restores that subset or empty result offline, so global and China caches never cross. Successful checks stay fresh for four hours; failures keep the prior catalog and remain retryable without writing a failed-check timestamp. An SDK caller can force an online check with ModelRuntime.refresh({ providers: ["minimax-openai"], allowNetwork: true, force: true }) (or minimax-openai-cn). China /v1/models parity is assumed but has not been verified against a live authenticated endpoint; a failure in one region leaves the other region unchanged.

TypeSafe Decisions

typesafe_decide is a Pi tool, not a chat model. It takes a JSON state plus one or more named questions and returns structured probabilities, not generated text:

• Noul — a yes/no question. instructions is optional and may be null. When instructions is omitted or null, at least one of criteria.true / criteria.false must be a non-null description.
• Choice — a question over 1–255 named options, each with a string, object, array, or null rubric description.
• Score — a question over 1–10 ordered level descriptions; Score levels cannot be null.

instructions and rubric descriptions accept a string, JSON object, or JSON array, and instructions may also be null. Omitted fields stay omitted rather than becoming empty strings, and explicit null stays null. A one-option Choice or one-level Score is valid; a one-level Score answer uses score 0 and probabilities { "0": 1 }.

See the TypeSafe quick start for how to construct questions, and the TypeSafe API reference for the full upstream contract. A request using all three question types:

{
  "state": { "repo": "pi-providers", "phase": "release review" },
  "questions": {
    "ship_now": {
      "type": "noul",
      "instructions": "Is the release safe to ship today?",
      "criteria": { "true": "All gates pass", "false": "Any gate is open" }
    },
    "risk_level": {
      "type": "score",
      "instructions": "How risky is shipping today?",
      "criteria": ["low", "moderate", "high"]
    },
    "next_step": {
      "type": "choice",
      "instructions": "What should happen next?",
      "criteria": { "ship": "No blockers remain", "fix": "Address open gates" }
    }
  }
}

Add a criteria-only Noul, an explicit-null Noul, a one-option Choice, and a one-level Score:

{
  "state": { "ticket": "payouts failing for three days" },
  "questions": {
    "ready_to_ship": {
      "type": "noul",
      "criteria": { "true": "All release gates pass" }
    },
    "blocked": {
      "type": "noul",
      "instructions": null,
      "criteria": { "false": "No blocker", "true": "Blocked" }
    },
    "next_step": { "type": "choice", "criteria": { "ship": null } },
    "risk": { "type": "score", "criteria": ["Low"] }
  }
}

Routing. The tool prefers direct TypeSafe access: with a configured credential it calls jev-latest on the TypeSafe System One API. If no direct credential is available, or a direct attempt fails in a retryable way, it may make a single fallback request to typesafe/jev through Command Code. Direct 400 / 404 / 413 / 415 / 422 responses mean the request itself is invalid, and a cancelled call is terminal — none fall back. Direct 401 / 403 / 429 / 529, other 5xx, transport failures, timeouts, and malformed success responses are eligible for one backend failover. Credentials are resolved in this order: a host-resolved typesafe credential, then a non-blank TYPESAFE_API_KEY, then command-code fallback auth.

Question limits. Direct TypeSafe requests have no fixed local question cap; upstream token budgets remain authoritative, normally 64K tokens across state plus all questions and 32K across state plus any individual question. Command Code System One accepts at most 100 questions per request. Both gateway paths — an initial request with no direct credential and a retryable fallback after a direct failure — enforce that limit before resolving Command Code credentials or sending. A direct request with 101 or more questions can still succeed. If such an oversized direct request fails retryably, the tool reports the sanitized direct failure together with the gateway limit and never splits, truncates, or forwards the batch.

Billing. The result reports which backend answered: backend is typesafe or command-code, alongside the concrete response model, the answers, and token usage. Direct calls bill your TypeSafe account. Fallback calls consume the Command Code plan's model-specific credits or your pay-as-you-go credit balance. See the Command Code Provider API and the GOAT plan for current rates — this README intentionally does not copy mutable prices.

Privacy. The tool never sends x-cmd-zdr, even when CMD_ZDR=1 is set, because Command Code documents no ZDR-capable upstream for typesafe/jev. It does not provide an enforced-ZDR mode. A direct TypeSafe credential changes the preferred backend but does not disable Command Code fallback after retryable direct failures, and direct TypeSafe handling is governed by your TypeSafe agreement rather than CMD_ZDR. If your policy requires ZDR, do not send that data through this tool unless that agreement independently satisfies the requirement and Command Code fallback authentication is unavailable.

Known Limits

• Key matching. The provider code passes the corresponding environment-variable reference ($MINIMAX_API_KEY, $MINIMAX_CN_API_KEY, $STEP_API_KEY) to Pi. If the variable is unset, the provider cannot authenticate.
• MiniMax tool-call hardening. The minimax-openai and minimax-openai-cn providers wrap their streams with a MiniMax-specific pipeline that folds inline think blocks into a proper thinking content block, repairs empty {} tool-call arguments via second-chance JSON parse, reorders tool-result messages to match the order of preceding tool_use blocks (the MiniMax API rejects mismatched ordering), and fails closed with a retryable error if the upstream ever emits internal tool-call markup instead of proper OpenAI tool calls.
• MiniMax reasoning replay. Native reasoning_content stays a signed thinking block with its exact whitespace, and interleaved thinking/text/tool segments keep their arrival order across a tool round trip. Before a same-model request, unsigned thinking from older saved sessions is replayed as ordered <think>…</think> text without changing the stored transcript; signed and foreign history are left to Pi's serializer. No payload rewriting is added.
• StepFun native stream. StepFun uses Pi's built-in OpenAI-compatible driver directly, without the MiniMax hardening pipeline.
• Command Code catalog. The 93 bundled models captured on 2026-10-25 are the fallback baseline, with a default max output of 32,768 capped by each model's context window. A successful live catalog response is authoritative for exposed model IDs, so bundled models missing from that response stay hidden until a later live refresh includes them. Valid live pricing, modalities, and reasoning metadata overrides local estimates; missing fields fall back to dated local capabilities for the verified additions, then to Pi donor metadata, official family defaults, or the dated bundled pricing snapshot. Pi persists the last successful catalog and restores it offline; source-aware online checks run at most every four hours unless forced, and changing CMD_MODELS_URL triggers a refresh. Refreshing requires Command Code authentication; on the first run, or after a refresh failure, Pi keeps showing the bundled or prior catalog until a successful refresh completes. Routing is declared per record: /messages uses Anthropic Messages, /responses uses OpenAI Responses, and /chat/completions uses OpenAI Chat Completions. Prices are dated estimates in USD per 1M tokens. Temporary offers use the rates advertised on 2026-10-25 and may change or expire; the explicit free model IDs are zero in this snapshot and must be refreshed when their offers expire or the IDs leave the live catalog. Paid inclusionai/ling-3.1-flash is separate from the retired free Ling aliases and keeps its paid rates. DeepSeek V4 uses the displayed off-peak estimate because Pi cannot represent UTC price bands, and its peak input/output rates are higher. Open-model routing and ZDR can also change the actual charge; Command's usage page remains authoritative. CMD_ZDR=1 applies only when making requests and is never persisted in the model cache.
• Command Code Flash effort. minimax/minimax-m3.1-flash accepts text and image input and exposes the five supported thinking efforts: low, medium, high, xhigh, and max. minimal is clamped to low on the simple provider path, and omitting effort sends no reasoning_effort, preserving upstream's default. The gateway still caps output at 32,768 tokens unless the live catalog advertises a smaller or larger explicit supported limit capped by context.
• Deeply nested tool schemas. The MiniMax API may produce collapsed nested arguments on complex JSON schemas; this has been observed with MiniMax-M3. The package emits a diagnostic message instead of retrying.

Development And Verification

Development and peer dependencies use Pi 1.4.0 as the minimum compatibility baseline. Dependency updates are checked against npm publication timestamps and must be at least 48 hours old; this release uses Biome 2.6.2 because 2.6.3 was newer than that cutoff.

pnpm install
pnpm check
pnpm pack:dry-run

Changelog

See CHANGELOG.md for release notes.

License

MIT. See LICENSE.