connect: ai& is a provider of its own, and one key reaches four labs

[fenilmodi00]

What changed

ai& (api.aiand.com) is now a provider of its own. One key reaches models from four labs, so it covers ground that would otherwise take four accounts.

The provider is one row in modelsource.Vendored(). The connect panel, the providers group, the picker group, the Providers tab and codeaf connect aiand all derive from that catalog, so nothing else needed wiring.

internal/config/crew.go grew two tables, because ai& is the first connection here that is itself a router:

 var crewVendors = map[string][]string{
 	"moonshot": {"moonshotai"},
 	"codex":    {"openai"},
+	"aiand": {"deepseek", "z-ai", "moonshotai", "qwen", "openai", "google"},
 }
+
+var crewVendorWire = map[string]map[string]string{
+	"aiand": {"deepseek": "deepseek-ai", "z-ai": "zai-org"},
+}

The first row is what makes one key serve several labs. The second is a spelling, and it is load-bearing: ai& publishes deepseek-ai/deepseek-v4-flash where the catalog says deepseek/deepseek-v4-flash, and asked on 2026-10-02 for the bare names it answered 404 model_not_found for every one. moonshotai is deliberately left alone, because ai& serves moonshotai/kimi-k3 and renaming it would break a send that works.

The row's preferred model moved. https://api.aiand.com/v1/api.json answers 200 without a key and publishes per-model cost, limits and modalities; it says the flagship zai-org/glm-5.3 is attachment: false, so the model codeaf seated a new connection on refused their first screenshot. The preference is now zai-org/glm-5.3-flash: same family, same million tokens, takes image and video, and $0.15/$0.50 against $1.00/$4.00.

The manual gained ai& in the two places that named the provider group and counted it, where leaving the old list would have made the page false. The crew page needed no edit, since it already says every model a connected provider serves is a candidate, and that sentence is now true of ai&.

Merge danger. Two-way door, and the revert is clean. The one thing worth a reviewer's attention is crewVendorWire: it reads close to crewroute.vendorOrgs, which holds the inverse mappings, and the two disagree on moonshot. Folding them together would rename aiand/moonshotai/kimi-k3 into a send that 404s. The commit message says so at the point of the table.

How it was checked

Live against the API on 2026-10-02: the listing, a streamed reply with its usage frame, tool calling, max_tokens, stream_options, the refusal of a bare model name, and the manifest every cost figure above is read from.

The multi-family claim is proved on the seats rather than the catalog, because a row that collapsed to one family would still pass every send assertion:

TestOneMultiFamilyKeyCarriesSeatsOnDifferentLabs
    ai& served z-ai to the worker and planner
    ai& served moonshotai to the checker

ok  	github.com/Agent-Field/codeaf/internal/config	14.797s
ok  	github.com/Agent-Field/codeaf/internal/modelsource	1.318s
ok  	github.com/Agent-Field/codeaf/internal/manual	7.043s
ok  	github.com/Agent-Field/codeaf/internal/namelaw	7.431s

$ make test-laws
laws: 138 files in 30 packages

go build ./..., go vet and gofmt -l are clean. One earlier run showed TestADecisionTakesUnderTwoMilliseconds at 2.11ms against a 2ms budget while five packages ran at once; internal/crewroute is not in this diff and the test passes 3/3 focused and green on its own.

Not in this change

The effort words are per-model and do not line up with codeaf's ladder. Only high is universal across the thirteen: kimi-k2.7-code takes high alone, qwen3.8-27b takes none,low,medium,xhigh and so not high. ai& publishes reasoning_efforts per model, but config.listedOutcome keeps only the id and the adapter's ReasoningProfile seam is fed by the router's catalog, so a direct provider's profile reads unknown and the requested word goes out unclamped. Fixing that means changing internal/catalog, internal/config and internal/provider together and would move effort handling for every direct connection, so it wants its own change and a live key. The change entry records the measurements.

motif-technologies/motif-3 reaches the picker and sends correctly, and the crew cannot weigh it because no catalog row names that vendor. TestAModelTheCatalogCannotWeighStillReachesThePicker holds that split so it is not later "fixed" by dropping the row from the listing.

Checklist

• A change entry — docs/changes/unreleased/1737-aiand-native-provider.md, named for this pull request and carrying the four beliefs it contradicts.
• The manual knows about it — commands.md and services.md corrected where they named the group and counted it. No slash command, key, tool, default, limit or refusal moved, and no gate asks for more.
• No new line in .github/known-red.txt — the file does not exist.
• Only my own paths are staged.

[CLAassistant]

All committers have signed the CLA.

[santoshkumarradha]

Hey, thanks for the PR. Could you sign the CLA when you get a chance so we can review it?