bun validate accepts limit.output > limit.context (64 published models, incl. a dropped zero on jiekou/claude-haiku-4-5)

bun validate requires limit.context and limit.output on every resolved provider model (AGENTS.md -> "Required on resolved provider models"), but nothing checks the relation between them. limit.output > limit.context is currently publishable, and 64 text-generation models across 18 providers ship that way in models.dev/api.json right now.

For a text model that's a contradiction: output tokens come out of the context window, so output can't exceed context.

Worked example - the check would have caught a real dropped zero

providers/jiekou/models/claude-haiku-4-5-20251001.toml:

[limit]
context = 20000
output = 64000

Claude Haiku 4.5 has a 200,000-token context window. The catalog already knows this in two other places:

• models/anthropic/claude-haiku-4-5.toml -> context = 200_000
• providers/anthropic/models/claude-haiku-4-5-20251001.toml -> context = 200_000

Of the 59 provider entries for this model in api.json, 57 say 200000, one says 192000, and this one says 20000.

The value is imported faithfully from upstream rather than mistyped here: https://api.jiekou.ai/openai/models currently returns "context_size": 20000 for claude-haiku-4-5-20251001, while the same response returns 200000 for its own sibling SKUs claude-haiku-4-5-20251001-dd, -r and -cc. So it's an upstream typo - but it reaches api.json as published fact because nothing downstream contradicts it, and output (64000) > context (20000) was the signal sitting right there in the same file.

Suggested check

In bun validate, per resolved provider model:

if (limit.output > limit.context) fail(`limit.output (${limit.output}) exceeds limit.context (${limit.context})`)

Two caveats, from actually reading the hits rather than just counting them:

1. Restrict to text generation (modalities.output === ["text"], non-audio input). Embedding entries put the vector dimension in limit.output - azure/cohere-embed-v3-english is 512 -> 1024, digitalocean/e5-large-v2 512 -> 1024 - and ASR entries put decoder frames there (whisper-large-v3, 448 -> 4096). Those trip the rule for an unrelated reason and are probably a separate modelling question, so I've excluded them from the count above.
2. About 20 of the 64 are near-misses that look like a decimal-vs-binary mismatch: context = 128000 with output = 131072, or 256000/262144. In those the context is likely the wrong field, not the output - but they're still contradictions, and the same check surfaces them.

The 64 entries

Method: https://models.dev/api.json fetched 2026-09-02, filtered to limit.output > limit.context with limit.context > 0, modalities.output == ["text"], no audio input, excluding embedding/rerank/TTS/ASR entries by name. Happy to open a PR for the validator rule if that's useful - I didn't want to guess at where you'd want it in the schema vs. a separate lint step.

Start at the bun validate entry point and read AGENTS.md for the resolved provider model requirements. Inspect the mentioned model TOML files and the generated models.dev/api.json data to understand the existing validation path. Done means text-generation models with limit.output > limit.context are rejected while embedding and audio-related entries remain excluded.