Release 0.6.5: catalog refresh for newer models is merged on main but unpublished (deepseek/deepseek-v4.1-flash has no thinking levels on npm)

[ZephyrDeng]

Provider version: 0.6.4 (npm latest, published 2026-09-03)

Model affected: deepseek/deepseek-v4.1-flash (returned by GET /provider/v1/models), plus the other entries refreshed by #98

What a fresh npm install does

The published catalog is generated from command-code@1.44.0, which has no deepseek/deepseek-v4.1-flash. The provider derives reasoning from that snapshot (isReasoningModel() in src/models.ts, which overwrites whatever the live endpoint returns) and src/pricing.ts has no cost entry, so the model shows up as non-reasoning with a zero cost:

$ pi --list-models deepseek
commandcode  deepseek/deepseek-v4.1-flash  1M  65.5K  no   no

The endpoint does support it: POST /provider/v1/chat/completions with {"model":"deepseek/deepseek-v4.1-flash","reasoning_effort":"max"} returns reasoning content, and an invalid value is rejected with expected one of "low"|"medium"|"high"|"xhigh"|"max". As already noted in #90, GET /provider/v1/models carries no capability data, so the CLI-derived snapshot is the only source.

Already fixed on main, just not published

• feat(models): add DeepSeek V4.1 Flash (catalog sync to 1.53.0 + pricing) #93 (merged 2026-09-10) syncs the catalog to 1.53.0 and adds deepseek/deepseek-v4.1-flash: input modalities, reasoning: true, efforts low|high|max, and a MODEL_COSTS entry.
• fix(models): refresh model metadata so newer models expose thinking levels #98 (merged 2026-09-14) refreshes model metadata so newer models expose thinking levels.

Both are on main, but package.json on main is still 0.6.4, the newest tag is v0.6.4 (2026-09-03) and npm view pi-commandcode-provider version returns 0.6.4, so installs cannot pick any of it up. This is the root cause behind #90 and the closed reasoning-override attempts in #91 / #97.

Ask

Cut a release (0.6.5) so the catalog/metadata refresh reaches npm.

Related

The new deepseek/deepseek-v4.1-flash cost entry from #93 (input 0.15 / output 0.6 / cacheRead 0.003) is DeepSeek's off-peak list rate, so the same entry is affected by #89 (peak = 2x, i.e. 0.30 / 1.20 / 0.006).

Workaround, for anyone blocked today

A models.json override for the commandcode provider restores thinking levels without touching the package:

"commandcode": {
  "modelOverrides": {
    "deepseek/deepseek-v4.1-flash": {
      "reasoning": true,
      "compat": { "supportsReasoningEffort": true },
      "thinkingLevelMap": { "minimal": null, "low": "low", "medium": null, "high": "high", "xhigh": null, "max": "max" }
    }
  }
}

Verified end to end: pi --model commandcode/deepseek/deepseek-v4.1-flash:max sends reasoning_effort: "max" (captured with COMMANDCODE_API_BASE pointed at a local mock).

[patlux]

Thanks for the nudge and the clear explanation! 0.7.0 is now published on npm, including the refreshed model capabilities and pricing from #98 and #93.

Release notes: https://github.com/patlux/pi-commandcode-provider/releases/tag/v0.7.0

Please update to 0.7.0. The separate time-of-day pricing concern remains tracked in #89. Closing this release request as completed — thanks again!