okami.mom/albedo
git clone https://git.nkp.pet/okami.mom/albedo
b82006c
main
web search
the model gets one search, await web_search(query, limit=8), whatever is signed in. it answers with a record: answer (written prose, empty from a plain engine) and sources, each with title, url, snippet, and published; printing it shows the answer and numbered sources. the model is told only that, not which provider answered or why another was passed over.
providers
a provider is a SearchPlugin(web_search.Provider(name, label, search)) on the extension that owns its sign-in, so credentials, account rotation, and wire formats stay where they already live. search(Query(text, limit, session)) answers Result(Answer, String); the error is a short reason the fallthrough can list. shared types and helpers are in harness/web_search.gleam: distinct dedupes sources by url and cuts them to the limit, failure words a transport error, and listen folds a streamed exchange's payloads into a state for providers whose answer arrives as SSE.
| extension | name | how it searches |
|---|---|---|
codex | codex | the ChatGPT backend's Responses with the hosted web_search tool forced; sources are the answer's url_citations, then what the search call turned up. utm_source=openai is stripped. |
claude | claude | Messages with the server tool web_search_20250305 (max_uses 5) appended through wire.encode_with; sources are citations_deltas, then web_search_tool_result pages. |
antigravity | antigravity | Cloud Code streamGenerateContent with googleSearch grounding through wire.encode_with, at the model's lowest effort. its own inline [label](http…) links are cut to their labels (search.unlink; code like f[T any](v T) is left alone), since some point at pages it never grounded on; grounding chunks are Google redirect urls; albedo_http:locations resolves each with one HEAD, all at once, 5s, keeping the redirect when it fails. |
exa | exa | Exa's /search with highlights as snippets. no prose. |
the subscription providers refuse an answer that never searched (no search call, no grounding): it would be the model's memory passed off as a search result. each one checks its sign-in before resolving a model, so a signed-out provider fails at once without reaching the network. searchModel in that extension's section of extensions.json picks the model; otherwise codex uses the account's first listed model, claude the newest listed model, and antigravity the first listed Gemini.
exa reads its key from creds.json as {"exa": {"apiKey": "..."}}; {"exa": {"endpoint": "..."}} in extensions.json points it at another host (the e2e suite points it at a fake).
transcript
each search notes a web activity in the cell's trace: the query as its target, and as its detail the answer and a list of linked sources as markdown, at most 16000 characters. the TUI names it in the burst's summary (looked up “query”) and shows the detail under it, cut to its first four rows with a click to expand row that opens the rest (burst.go foldLookup, rowactions.go toggleLookup). the model never sees the detail; it reads what the cell printed.
order
Session searches use the SearchPlugin contributions of the loaded active extensions. Saving a different selection takes effect after reload. The daemon's /web-search page lists globally enabled providers. Session overrides can select a different provider set. Quarantined or inactive extensions cannot supply providers.
The order is a preference within that eligible set. The top provider searches alone; the query falls back to the next one only when it fails or finds nothing. Providers below one that answered are never called. When every eligible provider that is on fails, the call raises with each reason in order.
the order is global, in extensions.json:
{ "web-search" : { "order" : ["claude" , "codex" , "antigravity" , "exa" ], "off" : ["exa" ] } }
names order leaves out follow it in registry order; names in off are never tried. /web-search is the page for it: u and d move the selected provider, t turns it on or off, and each change writes both lists under one settings lock. Providers hidden by extension selection keep their saved positions and exclusions. the service is GET /extensions/web-search/providers and POST /extensions/web-search/providers/{name} with {"change": "up" | "down" | "toggle"}; a move past either end changes nothing. see OpenAPI.
tests: test/e2e/web_search_test.py.
1# web search 2 3the model gets one search, `await web_search(query, limit=8)`, whatever is signed in. it answers with a record: `answer` (written prose, empty from a plain engine) and `sources`, each with `title`, `url`, `snippet`, and `published`; printing it shows the answer and numbered sources. the model is told only that, not which provider answered or why another was passed over. 4 5## providers 6 7a provider is a `SearchPlugin(web_search.Provider(name, label, search))` on the extension that owns its sign-in, so credentials, account rotation, and wire formats stay where they already live. `search(Query(text, limit, session))` answers `Result(Answer, String)`; the error is a short reason the fallthrough can list. shared types and helpers are in [`harness/web_search.gleam`](../src/albedo/harness/web_search.gleam): `distinct` dedupes sources by url and cuts them to the limit, `failure` words a transport error, and `listen` folds a streamed exchange's payloads into a state for providers whose answer arrives as SSE. 8 9| extension | name | how it searches | 10| --- | --- | --- | 11| `codex` | `codex` | the ChatGPT backend's Responses with the hosted `web_search` tool forced; sources are the answer's `url_citation`s, then what the search call turned up. `utm_source=openai` is stripped. | 12| `claude` | `claude` | Messages with the server tool `web_search_20250305` (`max_uses` 5) appended through `wire.encode_with`; sources are `citations_delta`s, then `web_search_tool_result` pages. | 13| `antigravity` | `antigravity` | Cloud Code `streamGenerateContent` with `googleSearch` grounding through `wire.encode_with`, at the model's lowest effort. its own inline `[label](http…)` links are cut to their labels (`search.unlink`; code like `f[T any](v T)` is left alone), since some point at pages it never grounded on; grounding chunks are Google redirect urls; `albedo_http:locations` resolves each with one HEAD, all at once, 5s, keeping the redirect when it fails. | 14| `exa` | `exa` | Exa's `/search` with highlights as snippets. no prose. | 15 16the subscription providers refuse an answer that never searched (no search call, no grounding): it would be the model's memory passed off as a search result. each one checks its sign-in before resolving a model, so a signed-out provider fails at once without reaching the network. `searchModel` in that extension's section of `extensions.json` picks the model; otherwise codex uses the account's first listed model, claude the newest listed model, and antigravity the first listed Gemini. 17 18exa reads its key from `creds.json` as `{"exa": {"apiKey": "..."}}`; `{"exa": {"endpoint": "..."}}` in `extensions.json` points it at another host (the e2e suite points it at a fake). 19 20## transcript 21 22each search notes a `web` activity in the cell's trace: the query as its `target`, and as its `detail` the answer and a list of linked sources as markdown, at most 16000 characters. the TUI names it in the burst's summary (`looked up “query”`) and shows the detail under it, cut to its first four rows with a `click to expand` row that opens the rest (`burst.go` `foldLookup`, `rowactions.go` `toggleLookup`). the model never sees the detail; it reads what the cell printed. 23 24## order 25 26Session searches use the `SearchPlugin` contributions of the loaded active extensions. Saving a different selection takes effect after reload. The daemon's `/web-search` page lists globally enabled providers. Session overrides can select a different provider set. Quarantined or inactive extensions cannot supply providers. 27 28The order is a preference within that eligible set. The top provider searches alone; the query falls back to the next one only when it fails or finds nothing. Providers below one that answered are never called. When every eligible provider that is on fails, the call raises with each reason in order. 29 30the order is global, in `extensions.json`: 31 32``` json 33{ "web-search" : { "order" : [ "claude" , "codex" , "antigravity" , "exa" ], "off" : [ "exa" ] } } 34``` 35 36names `order` leaves out follow it in registry order; names in `off` are never tried. `/web-search` is the page for it: `u` and `d` move the selected provider, `t` turns it on or off, and each change writes both lists under one settings lock. Providers hidden by extension selection keep their saved positions and exclusions. the service is `GET /extensions/web-search/providers` and `POST /extensions/web-search/providers/{name}` with `{"change": "up" | "down" | "toggle"}`; a move past either end changes nothing. see [OpenAPI](../docs/openapi.yaml). 37 38tests: `test/e2e/web_search_test.py`.