fix(models): refresh model metadata so newer models expose thinking levels

[SamYue1]

Problem

The model metadata shipped with the provider predates several model ids, so those ids fall back to the unknown-model defaults (reasoning: false, no thinkingLevelMap). pi returns ["off"] for non-reasoning models, so deepseek/deepseek-v4.1-flash exposed no selectable thinking level and dropped reasoning_effort from requests.

Change

Refreshed the shipped metadata with the repository's own catalog sync script:

• deepseek/deepseek-v4.1-flash — image + reasoning, efforts low/high/max
• gpt-6-astra — image + reasoning, efforts low…max
• inclusionai/ling-3.0-flash-sante:free — reasoning + output limit
• MiniMaxAI/MiniMax-M3 and the Muse Spark models — effort levels
• dropped the now-redundant manual effort overrides

Verification

• bun test ./tests/test-models.ts ./tests/test-model-metadata-check.ts → 30 pass / 0 fail
• pi --list-models commandcode → 69 models; deepseek/deepseek-v4.1-flash reports thinking: yes
• thinking levels resolve to off/low/high/max, matching the public model reference

[Alice39s]

I reproduced this refresh against the live Provider API and pi's model runtime.

Command Code's gateway rejects minimal for every model I probed, Muse Spark included:

POST /provider/v1/chat/completions  {"model":"meta/muse-spark-1.3", …, "reasoning_effort":"minimal"}
→ 400 {"error":{"message":"Invalid option: expected one of \"low\"|\"medium\"|\"high\"|\"xhigh\"|\"max\""}}

low, medium, high and xhigh return 200. max is validated again upstream and depends on the model. While the override kept minimal selectable, pi offered a level whose request always failed, and tests/test-models.ts refuses to let the sync finish until that override is gone, which is what kept the scheduled job red.

rashidrazak/opencode-cmd-provider builds its MODEL_EFFORTS from the same upstream models.md, and its entries line up with this catalog: deepseek/deepseek-v4.1-flash = low|high|max, MiniMaxAI/MiniMax-M3 = low|medium|high, gpt-6-astra = low…max, and no levels for moonshotai/Kimi-K2.7-Code, Qwen/Qwen3.7-*, thinkingmachines/inkling*, stepfun/*, tencent/hy3-paid, meituan/LongCat-2.0:free, poolside/laguna-s-2.1-free and nvidia/nemotron-3-ultra-550b-a55b. That project calls those "reasoning-capable without published efforts", so ids that still resolve to ["off"] after this PR are expected.

Through pi --mode rpc (set_model, then get_available_thinking_levels):

commandcode/deepseek/deepseek-v4.1-flash   ["off"]                                        → ["off","low","high","max"]
commandcode/MiniMaxAI/MiniMax-M3           ["off"]                                        → ["off","low","medium","high"]
commandcode/meta/muse-spark-1.3            ["off","minimal","low","medium","high","xhigh"] → ["off","low","medium","high","xhigh","max"]
commandcode/meta/muse-spark-1.1            ["off","minimal","low","medium","high","xhigh"] → ["off","low","medium","high","xhigh"]

43 of the 69 live catalog models expose at least one selectable level, and every remaining ["off"] id is one upstream publishes without efforts or marks non-reasoning.

[patlux]

Integrated current main (including repaired CI and sync), regenerated with command-code@1.54.0, and corrected the changelog: Ling has reasoning support but no published effort levels; API-absent IDs are not made selectable by metadata. Documented Muse Spark minimal removal and max addition. Full npm test passes locally with real Pi and OMP 18.1.22, plus formatting and diff checks. Waiting for CI before merge.

[patlux]

Shipped in pi-commandcode-provider@0.7.0: https://github.com/patlux/pi-commandcode-provider/releases/tag/v0.7.0 . Public npm package model discovery and GOAT live smoke verified.