thinking levels: `max` is never offered (shift+tab stops at `high`), and `minimal` is offered but 400s

[km-tr]

Versions — pi 0.87.1 (macOS), @commandcode/pi-commandcode-provider 0.3.0.

Problem

pi's thinking levels for every Command Code model stop at high. max never shows up when cycling with shift+tab or in /thinking, and asking for it on the command line is silently clamped back to high:

$ pi -p --model command-code/deepseek/deepseek-v4.1-flash:max "reply with exactly: ok"
ok
$ # the session log records the level that was actually used:
$ grep thinking_level_change ~/.pi/agent/sessions/**/*.jsonl
{"type":"thinking_level_change","thinkingLevel":"high"}

PI_REASONING_LEVEL=max behaves the same way, and nothing warns that the request was downgraded. This applies to all 81 models in the catalog, not just DeepSeek, and no configuration reaches max — modelThinkingLevels and defaultThinkingLevel are clamped the same way.

max is a value the gateway accepts (POST /provider/v1/chat/completions, deepseek/deepseek-v4.1-flash, max_tokens: 16):

reasoning_effort	gateway
low / medium / high / max	200
minimal	400 (see below)

Why

getSupportedThinkingLevels() in @earendil-works/pi-ai:

const EXTENDED_THINKING_LEVELS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"];
export function getSupportedThinkingLevels(model) {
    if (!model.reasoning) return ["off"];
    return EXTENDED_THINKING_LEVELS.filter((level) => {
        const mapped = model.thinkingLevelMap?.[level];
        if (mapped === null) return false;
        if (level === "xhigh" || level === "max") return mapped !== undefined;  // opt-in
        return true;
    });
}

xhigh and max are opt-in — pi only offers them when thinkingLevelMap names them, and docs/rpc-commands.md:278 describes them as levels "exposed only when supported by the selected model". toModel never sets thinkingLevelMap, so the offered set is off | minimal | low | medium | high, and clampThinkingLevel() pulls anything above high back down to it.

The mirror image: minimal is offered but rejected

The same missing map makes the fallback "every level that is not opt-in is supported", so pi offers minimal and the gateway answers 400:

$ pi -p --model command-code/deepseek/deepseek-v4.1-flash:minimal "reply with exactly: ok"
400: {"message":"Invalid option: expected one of \"low\"|\"medium\"|\"high\"|\"xhigh\"|\"max\"","type":"invalid_request_error","param":"reasoning_effort"}

Reproduced against 12 models on the /chat/completions route — deepseek/deepseek-v4.1-flash, deepseek/deepseek-v4-pro, Qwen/Qwen3.8-Max, MiniMaxAI/MiniMax-M3, moonshotai/Kimi-K3, zai-org/GLM-5.3, google/gemini-3.8-flash, xai/grok-4.7, meta/muse-spark-1.3, xiaomi/mimo-v2.6-pro, sakana/fugu-ultra, gpt-6-astra — all return that same message. off is unaffected: pi omits the field entirely and the request returns 200. The /messages route (Claude) is untested here, plan-limited on this key.

Fix

Only minimal is a clear-cut bug — the gateway rejects it, so it should not be offered:

model.thinkingLevelMap = { minimal: null };

Adding max: "max" on top of that puts max back into shift+tab for the /chat/completions models, which is what this report is about:

model.thinkingLevelMap = { minimal: null, max: "max" };

Whether max deserves a place on every open model, and whether xhigh belongs anywhere here, is a call for the maintainer — xhigh is an OpenAI-style value that pi treats as a per-model opt-in, and nothing in the catalog currently claims it. /provider/v1/models returns no capability fields today (id, object, created, owned_by, name, context_length, supported_endpoints), so a per-model reasoning_efforts list there would let toModel derive the whole map instead of hardcoding one, the same way pricing, max_output_tokens, modalities and reasoning are already read when the API sends them.

[km-tr]

Workaround without waiting for a release

No fork or local checkout needed. models.json modelOverrides is applied on top of extension-provided models, and it merges thinkingLevelMap key by key, so the missing map can be supplied from user config:

~/.pi/agent/models.json

{
  "providers": {
    "command-code": {
      "modelOverrides": {
        "deepseek/deepseek-v4.1-flash": {
          "thinkingLevelMap": { "minimal": null, "max": "max" }
        }
      }
    }
  }
}

Keys are the bare model ids, so any other model in the catalog can be added the same way. Opening /model reloads the file.

Verified on pi 0.87.1 by reading the thinking_level_change entry the session writes:

request	before	after
--model command-code/deepseek/deepseek-v4.1-flash:max	clamped to high	max
--model command-code/deepseek/deepseek-v4.1-flash:minimal	400 reasoning_effort	clamped to low, no error

max also appears in shift+tab / /thinking once the map names it, since the offered levels are derived from the same map.

This is a stopgap — once the provider ships a map of its own, the models.json entry can be dropped. It is only a user-side patch; it does not fix the other 80 models for anyone who does not write the file.