feat(pioneer): add 26 models (Kimi K3, Claude Opus 5, GPT-5.6)

[samrathreddy]

Summary

Adds 26 new Pioneer models (from the live https://api.pioneer.ai/v1/models), each authored as a base_model reference so provider-agnostic facts are inherited rather than duplicated inline. Includes 9 new models/ metadata specs for models that lacked one, plus three correctness fixes to the Pioneer sync.

Sync fixes (packages/core/src/sync/providers/pioneer.ts)

1. Filter alias duplicates. Pioneer's API returns every model twice — once under its real id and once under an anthropic/pioneer/<id> alias. Aliases are now dropped in parseModels so the sync no longer authors phantom anthropic/pioneer/* TOMLs.
2. Derive cost from the API. New models previously got no pricing (cost was only preserved from an existing file). Cost is now derived from the API's per-1M-token prices, preferring any hand-authored value.
3. Honor base-model reasoning. Pioneer reports identical boilerplate low/medium/high reasoning levels for every model, which was flipping non-reasoning models (Pixtral, Mistral Large 3, Devstral 2) to reasoning = true. Reasoning is now driven by the base model's authored metadata.

New model metadata entries + sources

Costs and context/output limits for all Pioneer listings come from Pioneer /v1/models (https://api.pioneer.ai/v1/models). Model facts (name, release date, license, modalities) in each new models/ spec are sourced from the vendor's model card:

models/ entry	Source
deepseek/deepseek-v3	https://huggingface.co/deepseek-ai/DeepSeek-V3
deepseek/deepseek-v3.1	https://huggingface.co/deepseek-ai/DeepSeek-V3.1
meta/llama-3.2-1b	https://huggingface.co/meta-llama/Llama-3.2-1B
meta/llama-3.2-3b	https://huggingface.co/meta-llama/Llama-3.2-3B
mistral/codestral-22b-v0.1	https://huggingface.co/mistralai/Codestral-22B-v0.1
mistral/magistral-small-2506	https://huggingface.co/mistralai/Magistral-Small-2506
mistral/ministral-8b-instruct-2410	https://huggingface.co/mistralai/Ministral-8B-Instruct-2410
alibaba/qwen2.5-coder-0.5b	https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B
alibaba/qwen3-235b-a22b-instruct-2507	https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507

Pioneer API models with no authoritative metadata (and a non-existent ministral-14b) were intentionally not added rather than shipping placeholder data.

Review feedback addressed

• [high] Dropped the added models/anthropic/claude-opus-5.toml — the canonical entry already exists on dev (feat: add Claude Opus 5 #3706). Rebased onto latest dev; Pioneer's claude-opus-5 listing now inherits the existing anthropic/claude-opus-5 spec.
• [low] Added the source table above.

Compliance

• All new models use base_model (a models/ entry exists for each).
• Reasoning models declare reasoning_options (bun validate passes).
• No new provider; openrouter.ts untouched.
• bun validate and SDK typecheck pass; catalog (generated.ts) regenerated.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] models/anthropic/claude-opus-5.toml:1 - Check: Do not add or overwrite provider-agnostic models/ metadata that already exists; inherit it with base_model. Why: Current base already has models/anthropic/claude-opus-5.toml (pdf input, established description). This PR adds the path as a new file with a weaker definition (modalities.input drops pdf, different description), which will conflict on rebase/merge and can regress canonical Opus 5 metadata used across providers. Action: Rebase onto current dev, drop the models/anthropic/claude-opus-5.toml addition, and keep only providers/pioneer/models/claude-opus-5.toml pointing at the existing base_model = "anthropic/claude-opus-5".
• [low] [possible mistake] models/alibaba/qwen2.5-coder-0.5b.toml:1 - Check: PRs that add material model facts should cite first-party sources mapped to claims. Why: Nine new models/ entries introduce release dates, limits, licenses, and capabilities (plus Pioneer costs/limits in provider TOMLs) with only a high-level “live API” mention in the PR body, so reviewers cannot corroborate those facts from the submission. Action: In the PR body, link Pioneer /v1/models or pricing docs for costs/limits and model cards/HF pages for each new models/ entry, stating what each URL supports.

[github-actions]

Action items

• [medium] [possible mistake] providers/pioneer/models/deepseek-ai/DeepSeek-V3.toml:10 - Check: Provider limit.output should reflect Pioneer’s served max completion tokens. Why: This entry only overrides context = 163_840 and inherits output = 8_192 from the new base metadata, while the sibling DeepSeek-V3.1 on Pioneer sets output = 131_072 and other hosts commonly advertise ~128K–164K output for DeepSeek-V3. An 8K cap is a large under-statement if Pioneer’s max_tokens is higher. Action: Verify Pioneer /v1/models max_tokens for deepseek-ai/DeepSeek-V3 and set limit.output (or correct the base default) if the API allows more than 8,192.
• [medium] [violation] packages/core/src/sync/providers/pioneer.ts:39 - Check: Sync behavior changes need focused regression coverage when they alter catalog correctness. Why: This PR changes alias filtering, cost derivation, and base-model reasoning handling, but packages/core/test/sync.test.ts still only asserts Pioneer skipCreates / trackMissingModels. Those paths can silently reintroduce phantom anthropic/pioneer/* models, missing costs, or false reasoning flags. Action: Add unit tests for (1) dropping anthropic/pioneer/ IDs in parseModels, (2) pioneerCost preferring existing cost then API prices, and (3) base-model path keeping non-reasoning bases free of reasoning_options / interleaved while reasoning bases retain options.
• [low] [possible mistake] models/alibaba/qwen2.5-coder-0.5b.toml:12 - Check: Provider-agnostic tool_call should match the actual checkpoint. Why: The weights URL is the base Qwen/Qwen2.5-Coder-0.5B checkpoint, not the Instruct variant, yet tool_call = true. Peer Llama base entries in this PR correctly use tool_call = false. Action: Confirm whether Pioneer’s Qwen/Qwen2.5-Coder-0.5B is the base or Instruct build; set tool_call = false (or point metadata/weights at the Instruct ID) if it is the base model.
• [low] [violation] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite first-party sources mapped to claims. Why: 9 new models/ entries introduce release dates, context/output limits, licenses, and capabilities, and 26 Pioneer rows add pricing, but the PR body only briefly notes the live Pioneer models API without links that state what each source supports for the new metadata. Action: Add PR-body citations (Pioneer models/pricing API or docs, plus model cards/HF pages) and state which claim each link supports—especially for the new models/* limits, dates, and licenses.

[samrathreddy]

Thanks for the review. Status on the remaining items:

Fixed

• [low] qwen2.5-coder-0.5b tool_call - Confirmed against the vendor card: Qwen/Qwen2.5-Coder-0.5B is the base (pretrained) checkpoint ("we do not recommend using base language models for conversations") and documents no tool/function calling. Set tool_call = false, which also matches the Llama base entries in this PR. Pushed; bun validate passes.

No change needed (with rationale)

• [medium] DeepSeek-V3 limit.output = 8192 - This is correct for Pioneer, not an under-statement. Pioneer serves deepseek-ai/DeepSeek-V3 with max_output_tokens = 8192 (verified against /v1/models); only DeepSeek-V3.1 is 131072, which this PR sets explicitly. The inherited 8192 reflects the actual served cap.
• [medium] sync regression tests - Happy to add Pioneer coverage if you'd like it, but flagging that it isn't required by the contribution checklist (the gate is bun validate), and by current precedent only vercel.parseModels has dedicated tests - openrouter (the reference rich sync) has none. So this reads as a nice-to-have rather than a blocker. Let me know if you'd prefer I add parseModels/cost/reasoning tests for Pioneer to match the vercel pattern.

The two hard-blocker checklist items (base_model inheritance, reasoning_options on reasoning models) and the citation recommendation are all satisfied.

[github-actions]

No actionable findings.