* docs: tighten agent/review policy for reasoning_options and base_model
Stop agents defaulting OpenAI gateways to empty reasoning_options from
uncertainty; baseline effort is low/medium/high from upstream/peers.
Clarify budget_tokens as narrow/legacy and require override-only base_model.
* docs: rewrite AGENTS.md as catalog-only guide
Drop JS/code-style noise. Focus on lab models vs providers, base_model
(create models/ when missing), override-only hosts, logos, costs, and
reasoning_options.
* docs: fix model field required/optional guidance in AGENTS.md
description is required; prefer cost.tiers over legacy context_over_200k;
split strongly recommended (family, knowledge) from truly optional (status).
* docs: clarify none-vs-toggle and require toggle wire comments
Effort with none plus graded levels must not also claim toggle. Binary
off may use toggle with a leading top-of-file wire-path comment.
* docs: align reviewer/fixer with create-models-if-missing base_model rule
Subagent review: bots still used the weak 'base_model only if models/
exists' wording. Bind create-lab-entry + override-only; fix stale
section refs, README effort example, and required logo label.
* docs: fix toggle+effort coexistence and lab inheritance requirements
Allow toggle beside graded effort when off is a separate wire control;
forbid only toggle+effort when none is already an effort value. Require
complete lab models/ files for base_model inheritance; mark interleaved
as provider-only.
* docs: resolve reasoning policy contradictions in one pass
Classify hosts by lab vs multi-model relay (not npm). Baseline is the
underlying model's native/peer option set, not fixed L/M/H. Fix examples
to match DeepSeek and Alibaba wire paths; align skill, reviewer, fixer.
* docs: fix opus-4.6 example options and OpenRouter path
README base_model snippet matches lab effort+budget; AGENTS table uses
real openrouter claude-opus-4.6.toml filename.
The daily model sync rewrites synced provider TOMLs by parsing and
re-serializing them, which discards every comment except a leading header
block (the preservation added in f07ac11f only re-attaches the top-of-file
comment run). Any comment authored between sections was silently deleted on
every rewrite that touched the file — e.g. sync PR #3004 stripped Venice's
"# Live /models ..." citation from before [[reasoning_options]].
Rather than teach the serializer to track inline comments, move every
existing inline comment to the top of its file (a one-time data migration)
and document the rule in AGENTS.md. The existing leading-header
preservation then keeps them across rewrites.
Migrated 82 model files across Venice, Baseten, OVHcloud, and Cloudflare
Workers AI. Verified: every file parses to identical data before and after
(0 data mismatches), and round-tripping all 1000 synced files through the
existing formatToml + leadingComments preservation loses 0 comments.
No code change.
Document review requirements for provider/model PRs:
- New providers must ship a compliant logo (blocker) and should add a
sync module when the source is context-rich (recommended)
- New models must use base_model when a models/ metadata entry exists
- Reasoning models must declare reasoning_options, with API request
syntax as TOML comments for niche providers
- PRs that change data should cite their sources (recommended)
- Include logo guidelines inline
Aiden Cline