feat(sync): add Cortecs model sync

[ASMAE20]

Summary

Add Cortecs model sync

• sync the Cortecs model catalog from https://api.cortecs.ai/v1/models
• convert API prices from EUR to USD per million tokens
• add newly available models and update existing pricing, limits, modalities, and capabilities
• use canonical base_model metadata where available

Test plan

• bun models:sync cortecs
• bun models:sync cortecs --dry-run
• bun validate

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/cortecs.ts:72 - Check: Do not invent limit.output when the API only exposes context. Why: output is set to model.context_size, which then overrides correct base_model completion caps (e.g. Gemini Flash base 65_536 → 1_048_576, GPT-5 family base 128_000 → 400_000/1_050_000, Claude Opus 5 base 128_000 → 1_000_000). Action: When max output is unknown, omit limit.output so factored models inherit metadata (and only write output when existing/hand-authored or the API provides a real max).
• [high] [violation] providers/cortecs/models/claude-4-5-sonnet.toml:1 - Check: Provider models must use base_model when models/ metadata exists. Why: Several Claude IDs still ship full inline defs despite canonical metadata (claude-4-5-sonnet → anthropic/claude-sonnet-4-5, claude-4-6-sonnet → anthropic/claude-sonnet-4-6, claude-opus4-5/6/7 → anthropic/claude-opus-4-5/6/7, claude-sonnet-4 → anthropic/claude-sonnet-4-0). Exact-slug resolution cannot find these aliases. Action: Add Cortecs ID→canonical aliases in the sync (and re-sync) so these files factor like claude-haiku-4-5 / claude-opus-5.
• [high] [possible mistake] providers/cortecs/models/gpt-4.1.toml:2 - Check: reasoning must match the model, not a mis-tagged gateway flag. Why: New/updated GPT-4.1/GPT-4o entries set reasoning = true (and empty reasoning_options) while models/openai/gpt-4*.toml and native OpenAI entries are reasoning = false. Same pattern likely affects other non-reasoning models if supported_features is trusted blindly. Action: Verify Cortecs feature tags; for non-reasoning bases, do not override reasoning to true (prefer base metadata when the API flag conflicts).
• [high] [possible mistake] packages/core/src/sync/providers/cortecs.ts:92 - Check: reasoning_options must reflect Cortecs’ documented OpenAI-compatible controls for the same surface. Why: New reasoning models default to existing?.reasoning_options ?? [], so claude-opus-5, claude-sonnet-5, claude-sonnet-4, GPT-5 / GPT-5.6 / gpt-oss-20b, etc. get [] while same-provider peers keep effort (and Claude peers also budget_tokens), matching providers/cortecs/provider.toml. Action: Stop blank-defaulting new models; seed options from provider-documented controls / same-family Cortecs peers (e.g. Claude effort+budget, GPT/DeepSeek/gpt-oss effort) or parse API metadata when present.
• [medium] [possible mistake] providers/cortecs/models/glm-4.7-flash.toml:2 - Check: reasoning, interleaved, and base metadata must be consistent. Why: Diff sets reasoning = false while keeping [interleaved] field = "reasoning_content", and zhipuai/glm-4.7-flash is reasoning = true. Action: Verify Cortecs’ feature flag; either restore reasoning (and options) or drop interleaved when reasoning is truly unavailable.

[github-actions]

Action items

• [high] [violation] providers/cortecs/models/gemini-2.5-flash.toml:11 - Check: limit.output must reflect the provider max completion size, not the context window, especially when base_model already defines output. Why: Several new factored models set limit.output equal to Cortecs context_size (e.g. Gemini Flash 1_048_576 vs base 65_536; GPT-5 family 400_000/1_050_000 vs base 128_000; Claude Opus/Sonnet 5 1_000_000 vs base 128_000; GPT-4.1 mini/nano 1_047_576 vs base 32_768; GPT-4o/4o-mini 128_000 vs base 16_384). That inflates usable max tokens for callers. Same pattern appears in gemini-3.1-flash-lite.toml, gemini-3.5-flash.toml, gpt-5*.toml, gpt-5.6-*.toml, gpt-4.1-*.toml, gpt-4o*.toml, claude-opus-5.toml, and claude-sonnet-5.toml. Action: Drop those incorrect limit.output overrides (or set verified Cortecs completion caps) so models inherit base output, then re-sync and confirm the bad values do not return.
• [high] [possible mistake] providers/cortecs/models/claude-opus4-7.toml:7 - Check: reasoning_options must match Cortecs’ verified request surface, not be copied across the whole Claude family. Why: Patch 2 adds budget_tokens to Opus 4.7/4.8, Opus 5, and Sonnet 5. Prior Cortecs notes for Opus 4.7/4.8 said only reasoning_effort / output_config.effort is exposed and “no explicit budget.” Native Anthropic and OpenRouter entries for these IDs are effort-only (no budget_tokens). Action: Remove budget_tokens from claude-opus4-7, claude-opus4-8, claude-opus-5, and claude-sonnet-5 unless you can cite Cortecs model/API evidence that thinking.budget_tokens is accepted for those IDs; keep effort-only (or empty) until verified.
• [medium] [violation] packages/core/src/sync/providers/cortecs.ts:88 - Check: Sync must not invent completion limits from context_size, and must not keep bad authored outputs forever. Why: The first sync used output: existing?.limit?.output ?? model.context_size, which created the bad overrides above. The follow-up output: authored?.limit?.output stops new inventions only when no authored output exists; once a bad limit.output is on disk, later syncs preserve it. Action: For factored models, omit unknown output (inherit base). For standalone models, only write output when Cortecs exposes a real completion cap; never default to context_size. After fixing the code, clear residual bad outputs and re-run sync.
• [medium] [violation] providers/cortecs/models/nvidia-nemotron-3-nano-30b-a3b.toml:1 - Check: Provider models must use base_model when matching models/ metadata exists. Why: Full inline definitions duplicate provider-agnostic facts for models that already have metadata: nvidia/nemotron-3-nano-30b-a3b and nvidia/llama-3.1-nemotron-ultra-253b (Cortecs IDs nvidia-nemotron-3-nano-30b-a3b, llama-3.1-nemotron-ultra-253b-v1). Auto-resolve misses them because of the nvidia- prefix / -v1 suffix. Action: Add Cortecs base-model aliases (like claude-sonnet-4) and emit base_model with only provider-specific overrides (cost, limits, modalities, reasoning_options).

[bodo6778]

Thank you for opening this PR! Hope it gets merged soon!

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/cortecs.ts:70 - Check: Provider sync must not invent completion limits from context size; API-authoritative fields only. Why: limit.output is taken only from authored files (and previously defaulted to context_size). Cortecs exposes context_size, not max completion tokens, so the sync wrote/kept output equal to the full context window and will omit output on new standalone models. That permanently overrides correct lab completion caps after base_model merge. Action: Stop authoring limit.output unless Cortecs documents a real completion cap; inherit lab/base output. Re-sync/fix every model whose output was set to context (e.g. Gemini 1_048_576, Claude 5 1_000_000, GPT-5.x 400_000/1_050_000, Haiku 200_000).
• [high] [violation] providers/cortecs/models/gemini-2.5-flash.toml:10 - Check: Relay reasoning_options must follow lab + same-surface peers; [] means no caller control, not uncertainty (AGENTS.md / audit skill). Why: New/updated Gemini entries use reasoning_options = [] while Google and OpenRouter expose real controls (2.5: toggle + budget_tokens; 3.x: effort minimal|low|medium|high). Same pattern on Qwen3.5/3.6 ([] vs lab/peer toggle±budget), Gemma-4, and other controlled reasoners. Action: Author host-accurate options from lab/peers (or Cortecs wire docs), not empty arrays; fix Gemini, Qwen3.5/3.6, Gemma-4, and other reasoning = true models defaulted to [].
• [high] [violation] providers/cortecs/models/claude-opus4-7.toml:7 - Check: budget_tokens only for real reasoning-budget APIs; not Claude 4.7+ adaptive effort. Why: Patch 2 added budget_tokens min 1024 to Opus 4.7/4.8, Opus 5, and Sonnet 5. First-party Anthropic entries for these are effort-only (4.7/4.8/Opus 5: low|medium|high|xhigh|max; Sonnet 5: toggle + those efforts)—no budget. Prior Cortecs comments for 4.7/4.8 said no budget is exposed. Action: Remove budget_tokens from Claude 4.7+ / 5 Cortecs files; align effort/toggle with lab/peers and what Cortecs actually maps.
• [high] [violation] providers/cortecs/models/pixtral-12b-2409.toml:1 - Check: Non-lab hosts must use base_model; add complete models/<lab>/… when missing. Why: Cortecs is a multi-model relay, but many new entries are full inline definitions for nameable lab models (e.g. Pixtral vs models/mistral/pixtral-12b.toml, Nemotron Nano 30B vs models/nvidia/nemotron-3-nano-30b-a3b.toml, Llama 3.1 Nemotron Ultra vs models/nvidia/llama-3.1-nemotron-ultra-253b.toml, plus other Mistral/Meta/Amazon/etc. IDs). Sync only factors when resolveModelMetadataBaseModel matches, so most third-party IDs stay full definitions. Action: For every nameable third-party model, add/complete lab metadata and emit override-only provider files with base_model; extend ID→canonical mapping where Cortecs slugs differ.
• [high] [violation] packages/core/src/sync/providers/cortecs.ts:54 - Check: reasoning = true ⇒ non-empty policy-compliant reasoning_options; relays must not default []. Why: Standalone path uses existing?.reasoning_options ?? []; base_model path only keeps options when existing.reasoning === true, so new base_model reasoners get no options and new standalones get empty arrays. That reintroduces the empty-options violations on every sync. Action: Derive options from known Cortecs/lab/peer baselines (or require hand-authored options and fail closed); never default relay reasoners to [].
• [medium] [violation] providers/cortecs/models/kimi-k2.5.toml:1 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: Kimi K2.5/K2.6 keep toggle but the PR dropped the prior thinking.type = enabled|disabled comments; sync only preserves a leading header, so mid-file notes are lost. Action: Restore a leading header on each toggle model documenting the exact Cortecs request field.
• [medium] [possible mistake] providers/cortecs/models/deepseek-v4-pro.toml:7 - Check: DeepSeek V4 baseline is toggle + high/max, not invented L/M/H. Why: Cortecs still authors effort = low|medium|high with no toggle; lab is toggle + high|max, OpenRouter peers use high|xhigh. Supporting comments were removed. Action: Verify Cortecs wire behavior; if it proxies DeepSeek controls, switch to toggle + high/max (with wire comment); if it only accepts L/M/H, document that host-specific mapping in a leading comment.
• [low] [possible mistake] providers/cortecs/models/claude-opus-5.toml:4 - Check: Relay effort sets should match lab/peers unless the host documents a narrower surface. Why: Claude 4.7+/5 lab entries include xhigh/max (Sonnet 5 also has toggle); Cortecs only lists low|medium|high. Action: Confirm Cortecs reasoning_effort values for these IDs and expand (or document the narrower map) accordingly.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/cortecs.ts:78 - Check: Relay reasoning_options must follow lab/same-surface peers; [] means no caller control, not uncertainty (AGENTS.md Reasoning options). Why: For standalone reasoners the sync does existing?.reasoning_options ?? [], and many synced reasoners are left as [] even when lab/peers expose controls (e.g. Gemini toggle/budget, Qwen3.5 toggle/budget, GPT effort). Re-running sync will keep baking false empties. Action: Stop defaulting to []. Preserve only verified host options; for new reasoners require lab/peer-derived options (or skip auto-create until authored). Fix catalog entries that currently claim [] without evidence of no control.

• [high] [violation] providers/cortecs/models/gemini-2.5-flash.toml:2 - Check: Multi-model relay of controlled reasoners must not author empty reasoning_options. Why: Final files use reasoning_options = [] for google/gemini-2.5-flash, gemini-2.5-pro, gemini-3.1-flash-lite, gemini-3.5-flash, and Alibaba qwen3.5-* / qwen3.6-*, while first-party/OpenRouter peers expose toggle and/or budget_tokens (or effort). Action: Author Cortecs-real options from lab + peers (and providers/cortecs/provider.toml wire mapping); remove unjustified [].

• [high] [violation] providers/cortecs/models/gemini-2.5-flash.toml:11 - Check: Provider limit.output must be a real completion cap, not context size. Why: Sync wrote output = context_size (e.g. Gemini Flash 1_048_576 vs lab 65_536; GPT-5 family 400_000 / GPT-5.6 1_050_000 vs lab 128_000). Patch 2 only keeps authored?.limit?.output, so these wrong values stick. Action: Drop bogus limit.output overrides so base limits inherit (or set verified caps). Change sync so unknown output is omitted for base_model rows, never copied from context_size.

• [high] [violation] providers/cortecs/models/claude-opus4-7.toml:7 - Check: budget_tokens only when this host has a real reasoning-budget field; not for Claude 4.7+ adaptive effort. Why: Patch 2 adds budget_tokens min 1024 on claude-opus4-7 / claude-opus4-8 (and related). Lab + OpenRouter peers are effort-only (low…max/xhigh); prior Cortecs notes said no budget for 4.7+. Action: Remove budget_tokens on 4.7/4.8+. Align effort with lab/peers (xhigh/max if Cortecs forwards them). Do not auto-copy older Claude budget onto adaptive models.

• [high] [violation] providers/cortecs/models/pixtral-12b-2409.toml:1 - Check: Non-lab hosts must use base_model; add complete models/<lab>/… when missing. Why: Many third-party models are full inline definitions despite nameable labs (e.g. pixtral-* → mistral/pixtral-*, DeepSeek/Mistral/Meta/Qwen/Nemotron/Nova stacks, gemma-3-27b-it, etc.). Cortecs is a relay. Action: For each nameable lab model, add/complete lab metadata if needed and factor provider files to override-only (cost, real deltas, reasoning_options). Keep full inline only for true host-unique IDs.

• [medium] [violation] providers/cortecs/models/kimi-k2.5.toml:1 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: Sync dropped the prior Kimi toggle comments; final kimi-k2.5 / kimi-k2.6 still have type = "toggle" with no leading wire comment (thinking.type). Action: Restore a leading header such as # Toggle: thinking.type = enabled|disabled (and ensure sync preserves leading headers).

• [medium] [possible mistake] providers/cortecs/models/gpt-5.toml:4 - Check: GPT-style relay effort should match lab/peer sets for that model, not a fixed L/M/H. Why: New/updated GPT and gpt-oss rows use only low|medium|high, while OpenAI lab includes none / minimal / xhigh / max by model. Provider notes say Cortecs maps OpenAI reasoning_effort. Action: Verify Cortecs accepted values per model; expand to the lab set Cortecs actually forwards, or document host-narrowing with evidence.

• [medium] [possible mistake] providers/cortecs/models/deepseek-v4-pro.toml:5 - Check: DeepSeek V4 baseline is lab/peer controls (toggle + high/max), not invented L/M/H. Why: Files keep effort = low|medium|high without toggle; lab is toggle + high|max, OpenRouter peers use high/xhigh. Action: Confirm Cortecs wire mapping for V4; if it proxies DeepSeek thinking controls, match lab (toggle + high/max). Do not keep L/M/H unless Cortecs docs/tests show those exact values.

• [low] [possible mistake] providers/cortecs/models/deepseek-v4-pro.toml:12 - Check: EUR→USD cache prices should stay plausible vs prior Cortecs/lab cache rates. Why: cache_read jumps from 0.003625 to 0.432 on V4 Pro (Flash 0.0028 → 0.037) under EUR_TO_USD = 1.114. That is a large discontinuity vs lab cache pricing. Action: Verify API unit/field for cache_read_cost and the FX conversion; correct if the endpoint is per-token or a different fee.

[github-actions]

Action items

• [high] [violation] providers/cortecs/models/nvidia-nemotron-3-nano-30b-a3b.toml:1 - Check: Non-lab hosts must use base_model (and add complete models/<lab>/… when missing). Why: Cortecs is a multi-model relay, but many new/updated entries are full inline definitions for nameable lab models. Clear existing targets include nvidia/nemotron-3-nano-30b-a3b, nvidia/llama-3.1-nemotron-ultra-253b (llama-3.1-nemotron-ultra-253b-v1), mistral/pixtral-12b (pixtral-12b-2409), mistral/mistral-nemo (mistral-nemo-instruct-2407), plus other third-party IDs still authored as standalone (gemma-3-27b-it, ministral-*, nova-*, qwen3-vl-235b-a22b, qwen2.5-vl-72b-instruct, deepseek-r1-0528, hermes-*, etc.). Action: Point each third-party entry at canonical lab metadata (add complete lab files when absent); keep provider files override-only (cost, host reasoning_options, real deltas). Extend resolveCortecsBaseModel aliases where Cortecs IDs differ from lab slugs.
• [high] [violation] providers/cortecs/models/claude-opus4-7.toml:10 - Check: budget_tokens only for real reasoning-budget surfaces; Claude 4.7+ is adaptive effort-only. Why: Patch 2 adds budget_tokens min 1024 on claude-opus4-7, claude-opus4-8, claude-opus-5, and claude-sonnet-5. First-party/OpenRouter peers for 4.7/4.8/Opus 5 are effort-only (no budget); prior Cortecs comments also said 4.7/4.8 expose no budget. This invents a control this generation does not have. Action: Drop budget_tokens on Claude 4.7+ / Claude 5 entries; keep effort only. Align effort values with lab/same-surface peers (lab uses low|medium|high|xhigh|max for these, not just L/M/H) if Cortecs actually forwards them.
• [high] [violation] packages/core/src/sync/providers/cortecs.ts:75 - Check: Do not set limit.output from context size; preserve only real max-output. Why: Sync used output = context_size (then later freezes authored output). That wrote wrong provider overrides such as Gemini output = 1_048_576 (lab 65_536), GPT-5 output = 400_000 (lab 128_000), GPT-5.6-* output = 1_050_000 (lab 128_000), and many standalone models with output == context. Consumers will treat full context as max completion. Action: Stop defaulting output to context_size. For factored models, omit limit.output unless Cortecs differs from lab; for standalone models, author verified max-output (or inherit via base_model). Fix already-written bad limit.output overrides.
• [high] [violation] providers/cortecs/models/deepseek-v4-flash-0731.toml:2 - Check: On relays, reasoning_options = [] means no caller control—not uncertainty—and must match lab/same-host peers when controls exist. Why: New V4 Flash 0731 is [] while same-provider deepseek-v4-flash / deepseek-v4-pro publish effort, OpenRouter peers publish effort, and DeepSeek first-party is toggle + high|max. Empty understates controls. Action: Set host-accurate options (at minimum match other Cortecs DeepSeek V4 entries; prefer lab toggle + high|max if Cortecs exposes those). Do not leave [] because the control set was not verified.
• [medium] [violation] packages/core/src/sync/providers/cortecs.ts:88 - Check: Relay sync must not invent empty reasoning_options for every reasoner. Why: existing?.reasoning_options ?? [] (and runner fallback to []) stamps “no control” on new reasoning models without evidence. That produced empty options across Gemini, Qwen3.5/3.6, GLM, Nemotron, and other reasoners even when lab/peers expose controls and providers/cortecs/provider.toml documents reasoning_effort / Kimi thinking.type mappings. Action: Preserve hand-authored options; for new models, seed from lab/same-surface peers or leave for hand-authorship—never default to [] from uncertainty. Re-audit empty reasoners after the fix.
• [medium] [violation] providers/cortecs/models/kimi-k2.5.toml:1 - Check: Every toggle needs a leading top-of-file wire comment. Why: Sync re-serialization drops the prior # Cortecs maps thinking.type = enabled|disabled headers on kimi-k2.5 / kimi-k2.6 while keeping type = "toggle". AGENTS requires the exact wire path above the first key because mid-file comments are stripped. Action: Restore a leading header such as # Toggle: thinking.type = enabled|disabled on each toggle model, and teach the sync to re-emit that header (or stop rewriting files that only need comment preservation).
• [medium] [possible mistake] providers/cortecs/models/pixtral-12b-2409.toml:6 - Check: API supported_features must not force reasoning = true on non-reasoners when using full inline definitions. Why: Several standalone entries mark reasoning = true (e.g. pixtral-12b-2409, llama-3.1-8b-instruct, older Llama/Mistral instruct IDs) while lab metadata for the same families is non-reasoning. Without base_model, the feature flag becomes catalog truth and requires bogus reasoning_options. Action: Prefer base_model so lab reasoning wins; only override when Cortecs truly exposes a reasoner control for that ID.

[ASMAE20]

@rekram1-node