feat(subconscious): add GLM-5.2

[github-actions]

I've made the fix. Here's the review-ready summary:

Summary

Subconscious serves GLM-5.2 as subconscious/glm-5.2 but the model was missing from the catalog, so tools that read models.dev data (e.g. opencode) could not resolve the model on the Subconscious provider. This adds a provider model entry that inherits provider-agnostic facts from the existing zhipuai/glm-5.2 base-model metadata and declares only the provider-specific pricing, reasoning controls, and reasoning delivery format.

Changes

New file providers/subconscious/models/subconscious/glm-5.2.toml (model id resolves to subconscious/glm-5.2):

• base_model = "zhipuai/glm-5.2" — inherits name, description, family, attachment=false, reasoning=true, temperature=true, tool_call=true, structured_output=true, open_weights=true, release_date/last_updated, limit (context 1,000,000; output 131,072), and modalities (text/text).
• [cost] — input = 1.4, output = 4.4, cache_read = 0.26 (USD per 1M tokens).
• interleaved = true — reasoning streams inline within the content field for the OpenAI-compatible endpoint.
• reasoning_options:
  • type = "toggle" — chat_template_kwargs.enable_thinking (true | false).
  • type = "budget_tokens" — thinking.budget_tokens via the Anthropic Messages format (bounds omitted; not documented).

Evidence

• Subconscious Pricing — lists GLM-5.2 with model id subconscious/glm-5.2 and per-1M-token rates: cached $0.26, input $1.40, output $4.40. (TIM-Qwen3.6 27B also listed there.)
• Subconscious Thinking Mode docs — documents the OpenAI-format toggle chat_template_kwargs.enable_thinking and the Anthropic-format thinking config (type: "enabled", budget_tokens) that "lets you cap reasoning with budget_tokens"; states reasoning streams inline in content wrapped in think tags, with streaming code reading delta.content.
• Subconscious API reference (OpenAI-compatible) — confirms the OpenAI-compatible endpoint and chat_template_kwargs extension surface.
• Base model models/zhipuai/glm-5.2.toml — confirms the provider-agnostic metadata (1M context, 131,072 output, text-only, reasoning, tool/structured-output support) inherited via base_model.

Validation

• Verified base_model = "zhipuai/glm-5.2" resolves to the existing models/zhipuai/glm-5.2.toml (model id zhipuai/glm-5.2).
• Manually checked the merged object against AuthoredModelShape: all required fields (name, description, attachment, reasoning, tool_call, release_date, last_updated, modalities, open_weights, limit with output) are inherited from the base model; reasoning_options is declared (satisfies the "must set reasoning_options when reasoning is true" refine); cost uses valid schema fields (input, output, cache_read).
• Confirmed there is no sync module for subconscious, so the leading source-comment block and inline reasoning-option API comments are preserved (not stripped by re-serialization).
• Did not run bun validate or any shell/API commands.

Review notes

• The pre-existing Subconscious model subconscious/tim-qwen3.6-27b declares interleaved = { field = "reasoning_content" } and cost = { input = 0, output = 0 }. The Thinking Mode docs show reasoning inline in content (think tags) for the OpenAI format, so I used interleaved = true for GLM-5.2 to match the documented behavior. The TIM-Qwen pricing on the pricing page ($0.15/$0.30/$3.00) also differs from its catalog cost = 0. Both are pre-existing discrepancies and were intentionally left out of scope.
• budget_tokens is declared without min/max because Subconscious does not document bounds (the docs only show an example value of 2000); per the reasoning-options audit guidance, unverified bounds are omitted.
• Context/output limits are inherited from the base model (1M / 131,072) since Subconscious does not publish separate per-model limits for GLM-5.2 and markets long-context inference; if Subconscious documents tighter limits, limit can be overridden later.

Closes #3059

Automated by the issue fixer: https://github.com/anomalyco/models.dev/actions/runs/28793474685

[kraplegit]

add glm-5.2

[kraplegit]

wrong code. reference https://github.com/anomalyco/models.dev/blob/dev/providers/subconscious/models/subconscious/tim-qwen3.6-27b.toml

[rekram1-node]

Follow-up in 645321823 removes the unsupported interleaved claim, aligns reasoning controls with the Anthropic Messages adapter from #3062, and adds Subconscious’s GLM-5.2 launch post as model-specific evidence for Messages support. Thinking is disabled by omitting thinking and enabled with thinking.type = "enabled" plus budget_tokens. bun validate and git diff --check pass.