feat(ambient): sync model catalog from API

[ambient-gregory]

Ambient now syncs from /v1/models in the hourly workflow instead of a hand-maintained allowlist. Replaces the ambient:generate script with a SyncProvider and picks up the full model list (15 models).

Ambient's API reports an openrouter.slug, per-token pricing, and supported_features, so each model is adapted to the OpenRouter shape and run through buildOpenRouterModel (same as cloudflare-workers-ai). Models with a canonical definition inherit it via base_model and override only cost, limit, and modalities; the two gpt-oss models have no canonical entry and are emitted in full.

Models

Model	base_model	Input $/M	Output $/M
zai-org/GLM-5.1-FP8	zhipuai/glm-5.1	1.40	4.40
ambient/large	zhipuai/glm-5.1	1.40	4.40
z-ai/glm-4.5-air	zhipuai/glm-4.5-air	0.125	0.85
moonshotai/kimi-k2.6	moonshotai/kimi-k2.6	0.67	3.39
moonshotai/kimi-k2.7-code	moonshotai/kimi-k2.7-code	0.75	3.50
google/gemma-4-26b-a4b-it	google/gemma-4-26b-a4b-it	0.10	0.30
google/gemma-4-31b-it	google/gemma-4-31b-it	0.20	0.80
nvidia/nemotron-3-nano-30b-a3b	nvidia/nemotron-3-nano-30b-a3b	0.05	0.20
qwen/qwen3-coder-30b-a3b-instruct	alibaba/qwen3-coder-30b-a3b-instruct	0.07	0.27
qwen/qwen3-next-80b-a3b-instruct	alibaba/qwen3-next-80b-a3b-instruct	0.09	1.10
qwen/qwen3.5-35b-a3b	alibaba/qwen3.5-35b-a3b	0.14	1.00
qwen/qwen3.6-27b	alibaba/qwen3.6-27b	0.32	3.20
qwen/qwen3.6-35b-a3b	alibaba/qwen3.6-35b-a3b	0.15	1.00
openai/gpt-oss-120b	—	0.15	0.60
openai/gpt-oss-20b	—	0.029	0.14

[ambient-gregory]

Hi @rekram1-node, would love to get this reviewed 🙏

[rekram1-node]

AUTOMATED REVIEW: Blocking merge: generated reasoning models lack explicit provider options, transient API omissions can delete local files, and the sync ignores source is_ready state while most returned models are not ready. Add conservative deletion/readiness handling, options or [], regenerate, and resolve conflicts.

[ambient-gregory]

Thank you for the review @rekram1-node, this is ready for a final look 🙏

[ambient-gregory]

@rekram1-node just following up here, thank you 🙏

[github-actions]

Action items

• [medium] [violation] providers/ambient/models/zai-org/GLM-5.2-FP8.toml:1 and sync completeness — Check: Sync idempotence and full-output commit (sync.md steps 8–10). Why: The PR body lists 15 synced models, but the diff only creates 1 new model file (zai-org/GLM-5.2-FP8.toml) and modifies 1 (moonshotai/kimi-k2.7-code.toml); the remaining ~13 models (e.g., ambient/large, openai/gpt-oss-120b, qwen/qwen3-coder-30b-a3b-instruct) are absent from providers/ambient/models/. The one new file that is committed does not match what bun models:sync ambient would produce: buildOpenRouterModel → factorBaseModel generates a description via describeModel (e.g., "Flagship GLM model for hybrid reasoning, coding, and agentic engineering") that differs from the base model's hand-authored description, so the sync would write it as an override, but the committed file omits it. The next CI sync run would both create the missing models and rewrite this file, violating the clean dry-run expectation. The PR body also contradicts the diff — it lists zai-org/GLM-5.1-FP8 and kimi prices 0.75/3.50, while the diff creates GLM-5.2-FP8 and sets kimi to 0.84/3.99. Action: Run bun models:sync ambient, commit the full generated output, then run bun models:sync ambient --dry-run and bun validate; reconcile the PR body with the actual diff.

• [low] [possible mistake] providers/ambient/models/zai-org/GLM-5.2-FP8.toml:1 — Check: Provider-specific interleaved field. Why: The sibling providers/ambient/models/zai-org/GLM-5.1-FP8.toml:7 declares [interleaved] field = "reasoning_content" for Ambient's GLM-5.1 serving. GLM-5.2 is also a reasoning model from the same family, but the new file omits [interleaved]. The sync code sets interleaved: existing?.interleaved (undefined for new files), so it would not add this field automatically. Action: Verify whether Ambient's GLM-5.2-FP8 returns interleaved reasoning in reasoning_content and add [interleaved] if so.

• [low] [possible mistake] Pricing data in moonshotai/kimi-k2.7-code.toml:5-6 and zai-org/GLM-5.2-FP8.toml:5-7 — Check: Source citations for material pricing changes. Why: The diff updates kimi-k2.7-code prices (0.75→0.84 input, 3.5→3.99 output) and introduces GLM-5.2-FP8 prices (1.2/4.2, cache_read 0.26, limit.output 202_752). The PR body references the /v1/models endpoint but does not link to it or to Ambient's pricing documentation, so the values cannot be reviewed against evidence. Action: Add a direct citation to Ambient's /v1/models endpoint or pricing page stating what each cited value supports.