docs: both api keys, example model gpt-5.4-nano

This commit is contained in:
pj committed 2026-06-12 10:52:14 +05:30
1 parent 3ca0e9a5b9
commit ede495d935
3 files changed
+24 -19

No files matched your search

+10 -8
View File
@@ -81,19 +81,21 @@ Traces land in `./sanderling/runs/<timestamp>/`.
## Run with the LLM action backend
`sanderling/spec-llm.ts` reuses the same properties and login setup but swaps the
seeded fuzzer for `llm({ model })`: an OpenRouter vision model chooses which of
the already-enumerated candidates to act on each step, from the screenshot.
seeded fuzzer for `llm({ model })`: a vision model chooses which of the
already-enumerated candidates to act on each step, from the screenshot.
```sh
export OPENROUTER_API_KEY=sk-or-...
export OPENROUTER_API_KEY=sk-or-... # or OPENAI_API_KEY=sk-... for OpenAI direct
sanderling test --spec sanderling/spec-llm.ts --bundle-id app.folio --duration 2m
```
The model must support image input **and** strict `json_schema` structured
outputs (e.g. `openai/gpt-4o`); edit the `model` string in `spec-llm.ts` to pick
another. Each step is one multimodal call, so keep the duration / step budget
modest. The trace records the model's reasoning and `source: "llm"` on each
chosen action, so the replay UI shows why each pick was made.
OpenRouter wins when both keys are set. With a plain OpenAI key, drop the vendor
prefix from the model id in `spec-llm.ts` (`gpt-5.4-nano`, not
`openai/gpt-5.4-nano`). The model must support image input **and** strict
`json_schema` structured outputs. Each step is one multimodal call, so keep the
duration / step budget modest. The trace records the model's reasoning and
`source: "llm"` on each chosen action, so the replay UI shows why each pick was
made.
## Run a sanderling test (iOS)
+9 -7
View File
@@ -2,17 +2,19 @@
//
// The properties and the login `setup` are reused verbatim; only the action
// generator changes. Instead of the seeded fuzzer drawing a random candidate,
// `llm({ model })` hands selection to an OpenRouter vision model: each step it
// sees the screenshot plus the candidate list the system already enumerates and
// returns which candidate to act on. The candidate set, the typed input values,
// action execution, and the trace are all identical to the seeded run.
// `llm({ model })` hands selection to a vision model: each step it sees the
// screenshot plus the candidate list the system already enumerates and returns
// which candidate to act on. The candidate set, the typed input values, action
// execution, and the trace are all identical to the seeded run.
//
// Requirements:
// - OPENROUTER_API_KEY in the environment.
// - OPENROUTER_API_KEY (OpenRouter) or OPENAI_API_KEY (OpenAI) in the
// environment; OpenRouter wins when both are set. With a plain OpenAI key,
// drop the vendor prefix from the model id ("gpt-5.4-nano").
// - A model that supports image input AND strict json_schema structured
// outputs (e.g. "openai/gpt-4o"). A model lacking either fails clearly.
// outputs. A model lacking either fails clearly.
import { llm } from "@sanderling/spec";
export { properties, setup } from "./spec";
export const actionsRoot = llm({ model: "openai/gpt-4o" });
export const actionsRoot = llm({ model: "openai/gpt-5.4-nano" });
+5 -4
View File
@@ -35,10 +35,11 @@ export function actions(generator: () => Action[]): GeneratorNode {
}
// llm selects the LLM action backend: instead of the seeded picker drawing a
// random candidate, Go drives an OpenRouter model that chooses which candidate
// to act on from the screenshot + the candidate list. The returned marker is
// inert on the JS picker (pick.ts walks it to null); Go reads its config.model
// off globalThis.actions. API key comes from OPENROUTER_API_KEY.
// random candidate, Go drives an OpenAI-compatible model that chooses which
// candidate to act on from the screenshot + the candidate list. The returned
// marker is inert on the JS picker (pick.ts walks it to null); Go reads its
// config.model off globalThis.actions. API key comes from OPENROUTER_API_KEY
// (OpenRouter) or OPENAI_API_KEY (OpenAI); OpenRouter wins when both are set.
export function llm(config: { model: string }): GeneratorNode {
return { kind: "llm", config };
}