fix(cloudflare): correct catalogue reasoning efforts & base_models

[KastanDay]

Refresh 14 Workers AI catalogue entries using the September 18 Search snapshot. Independent of the sync update; either PR can merge first.

Source priority: Search supplies verified Workers-specific reasoning controls; existing controls remain where Search is silent. Existing canonical model records supply creator metadata, while Workers-specific overrides remain in provider entries. Keep conservative catalogue limits and reduce only the four values above the Workers ceiling of 1,048,576.

Example	Before	After
GLM-4.7-Flash	Toggle plus low/medium/high	Toggle only: chat_template_kwargs.enable_thinking
Gemma 4 / Kimi K2.6	Toggle plus low/medium/high	Effort none/high
GLM-5.3-Flash	No reasoning controls	Effort low/high/max
Kimi K2.7	Single mandatory high effort	No caller control: reasoning_options = []
Four oversized context/output fields	1,310,720	1,048,576

Nine reasoning updates and five links to existing canonical models, using the current schema. Nemotron and Qwen3.8 changes are formatting-only. All 27 models are retained. SEA-LION inherits the canonical description of its Southeast Asian language tuning.

Why Search supersedes the generic input-schema enum: the inspected per-model Search configuration explicitly normalizes low and medium to high for Gemma 4 and Kimi K2.6, with supported efforts high/none. They are aliases, not distinct model effort levels; none already represents off, so the catalogue does not duplicate it with a toggle. GLM-4.7's configuration was corrected to describe only binary thinking, matching the creator's template. Current sync-input.json schemas still show the generic low/medium/high enum; that does not establish model-specific effort controls.

The dated DeepSeek V4 Pro 0813 and Flash 0731 releases retain their model-specific effort sets: the Pro 0813 creator documentation and Workers configuration explicitly support distinct low/high/max levels. Older preview entries are not the baseline for these dated releases. Kimi K2.7's accepted aliases all normalize to mandatory high, so it has no caller control.

Search sources and the input-schema discrepancy are recorded in the changed files. Corresponding Search context corrections are still needed to prevent later syncs from restoring oversized values.

Validation: bun validate passes without the sync PR. Review follow-ups correct SEA-LION's inherited description and remove Kimi K2.7's redundant fixed-effort selector; conservative catalogue limits remain unchanged.

[github-actions]

Action items

• [medium] [possible mistake] providers/cloudflare-workers-ai/models/@cf/google/gemma-4-26b-a4b-it.toml:6 - Check: Workers AI reasoning controls must match this host’s real request surface (provider.toml treats each model’s sync-input.json as authoritative; relay options need host docs/peers without silent drops). Why: These entries replace previously cited native controls (reasoning_effort low|medium|high plus chat_template_kwargs.enable_thinking / .thinking) with Search-only shapes (none|high on Gemma 4 and Kimi K2.6; toggle-only on GLM-4.7-Flash). If sync-input.json still advertises the removed fields, the catalogue will understate callable controls. Action: Re-check current Workers AI sync-input.json (and/or a live probe) for @cf/google/gemma-4-26b-a4b-it, @cf/moonshotai/kimi-k2.6, and @cf/zai-org/glm-4.7-flash; restore any still-supported toggle/effort values, or keep the Search-only sets and document why Search supersedes sync-input for those models.
• [low] [possible mistake] providers/cloudflare-workers-ai/models/@cf/aisingapore/gemma-sea-lion-v4-27b-it.toml:3 - Check: After base_model, keep only real host deltas—not weaker or mismatched metadata that blocks inheritance. Why: Linking base_model = "aisingapore/gemma-sea-lion-v4-27b-it" while keeping description = "Open Gemma instruction model for efficient chat and self-hosted deployments" retains a generic Gemma blurb instead of the lab SEA-LION description (Southeast Asian languages / instruction tuning). Action: Drop description so the provider inherits the lab text, unless Workers AI intentionally uses different public copy (if so, cite that).

[KastanDay]

Evaluated both findings in the review:

1. Reasoning controls — kept the current values and documented the source discrepancy. I rechecked the current Workers AI input schemas for Gemma 4, Kimi K2.6, and GLM-4.7-Flash. They still expose the generic low/medium/high enum. The September 18 per-model Search response is more specific:

   Model	Per-model configuration	Catalogue decision
   Gemma 4	supported_efforts: ["high", "none"]; normalizes_to.low and .medium are "high"	Keep none/high; aliases do not add distinct levels
   Kimi K2.6	supported_efforts: ["high", "none"]; normalizes_to.low and .medium are "high"	Keep none/high; aliases do not add distinct levels
   GLM-4.7-Flash	Corrected per-model configuration: mandatory: false, default_enabled: true, no effort selector	Keep the binary chat_template_kwargs.enable_thinking toggle

   For Gemma/Kimi, none already represents off, so the repository's reasoning-options convention excludes a redundant toggle. GLM's binary control was explicitly confirmed when correcting its Workers configuration and also matches its creator template. Restoring the generic effort enum would reintroduce the inaccurate distinctions this PR fixes. Added the rationale to the three leading TOML headers and PR description.

2. SEA-LION description — fixed. Removed the generic provider override. It now inherits Gemma 3 27B tuned by AI Singapore for Southeast Asian languages and instruction following from the existing canonical record, consistent with creator metadata taking priority.

bun validate passes. A full before/after generated-catalogue comparison confirms that SEA-LION's description is the only resolved-value change from this review follow-up: all reasoning controls, conservative limits, and all 27 Workers models remain unchanged.

[github-actions]

Action items

• [high] [possible mistake] providers/cloudflare-workers-ai/models/@cf/deepseek-ai/deepseek-v4-pro-0813.toml:6 - Check: DeepSeek V4 reasoning baseline for this host (lab + same-surface peers; no invented grades). Why: This changes toggle + high/max to effort none|low|high|max. First-party providers/deepseek/models/deepseek-v4-pro.toml is toggle + high/max, and lab notes map low→high, so publishing low as a distinct grade is the L/M/H anti-pattern called out for DeepSeek V4. A Search enum alone does not show ordered, distinct behavior. Action: Re-verify on Workers AI whether low changes reasoning vs high for this model. If not, drop low (e.g. none|high|max, or restore toggle + high/max if thinking.type is still the off control). Cite the wire path and what was measured.
• [medium] [possible mistake] providers/cloudflare-workers-ai/models/@cf/deepseek-ai/deepseek-v4-flash-0731.toml:8 - Check: Same DeepSeek V4 baseline / distinct-effort rule on Cloudflare. Why: Same none|low|high|max set as Pro. Flash peers vary (toggle+high/max, toggle+low/high/max, or measured none/high only). Without a host-specific check that low is a real grade on Workers AI, this risks cataloging aliases as controls. Action: Confirm whether low is distinct from high on this Workers model; if Search only lists aliases, publish the effective set (and keep none only if off is truly via reasoning_effort=none, not a separate toggle).

[KastanDay]

Both findings compare these dated releases with the older DeepSeek V4 preview. The creator documentation explicitly distinguishes them:

• DeepSeek-V4-Pro-0813 — Chat Template documents low, high, and max as three reasoning levels.
• DeepSeek-V4-Flash-0731 — Chat Template documents the same three levels.

The current Workers AI per-model configuration agrees: supported_efforts contains max/high/low/none; its alias map sends minimal to low and medium to high, rather than mapping low to high. Thus this is not the generic L/M/H enum from the old input schema. none is the host's documented effort-level off setting.

Kept the model values and added exact creator citations beside both entries in 0b12ebe1b. Catalogue validation passes. No new inference measurement is claimed: the evidence is the dated creator contract plus the current host-specific configuration, rather than the preview-model peer entries.

[github-actions]

Action items

• [medium] [violation] providers/cloudflare-workers-ai/models/@cf/moonshotai/kimi-k2.7-code.toml:8 - Check: Always-on / no caller control ⇒ reasoning_options = []. Why: The file states thinking is mandatory and sets effort to only ["high"]. That is not a real graded control; AGENTS.md / the reasoning-options audit guide map always-on reasoners to []. Lab providers/moonshotai/models/kimi-k2.7-code.toml already uses []. Action: Replace with reasoning_options = [] (keep temperature = true if Workers still differs from the lab).
• [medium] [possible mistake] providers/cloudflare-workers-ai/models/@cf/deepseek-ai/deepseek-v4-pro-0813.toml:9 - Check: Effort baseline = lab / same-surface peers; do not publish aliased levels as distinct. Why: First-party DeepSeek Pro is toggle + high/max and maps low→high (providers/deepseek/provider.toml, providers/deepseek/models/deepseek-v4-pro.toml). This PR collapses Gemma/Kimi low/medium as aliases but still lists Pro low as a distinct Workers level. Action: Confirm Workers Search/live behavior treats low as a real Pro grade (not an alias of high); if it aliases, drop low and keep the effective set (e.g. none/high/max or lab-faithful high/max with the correct off control).

[KastanDay]

Addressed the Kimi finding in 62e43e9: reasoning_options = [], with a leading explanation. Its current Workers configuration has only high, mandatory thinking, and all accepted aliases normalize to high; none adds a caller choice. Temperature, pricing, modalities and all limits are unchanged. bun validate passes. The companion sync PR now preserves this fixed-effort meaning too, so it will not recreate the redundant selector.

The DeepSeek Pro finding compares two different releases. This file serves DeepSeek-V4-Pro-0813, while providers/deepseek/models/deepseek-v4-pro.toml describes the older preview. The dated creator release's Chat Template section says reasoning effort “now supports three levels — low, high, and max”. The corresponding inspected Workers configuration lists max/high/low/none; its alias map is minimal → low, medium → high, xhigh → high, and null/default → high. It does not normalize low → high. Together, the exact-release creator documentation and model-specific host configuration support retaining none/low/high/max. This is configuration/documentation evidence, not a claim of a fresh inference benchmark. The generic older preview baseline should not override the dated model's explicit controls.

[github-actions]

No actionable findings.