@mindstudio-ai/agent 0.1.170 → 0.1.172

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -4,6 +4,28 @@ What changed in `@mindstudio-ai/agent`, newest first, each version under its own
4
4
 
5
5
  Versions before 0.1.163 aren't recorded here.
6
6
 
7
+ ## 0.1.172
8
+
9
+ ### New
10
+
11
+ - `browser`: a real Chrome an app's backend drives. `browser.open({ url?, viewport?, idleTimeoutSeconds? })` starts a session that stays open across calls, and `browser.session(id)` picks it up from a later one. Its actions (`goto`, `snapshot`, `click`, `type`, `select`, `press`, `scroll`, `waitFor`, `screenshot`, `evaluate`, `close`) each answer with the page afterward, an outline with a ref on everything that can be acted on. `screenshot({ store })` saves straight into a file store, and `screenshot({ cursor: true })` keeps the session's cursor in the shot. `session.tools({ secrets?, screenshots? })` gives `runTask` one tool for a model to drive it, `browser`, which takes a list of steps, runs them in order, stops at the first that fails, and answers with each step's outcome and the page once. With `screenshots`, a store, it adds `browser_screenshot`, which saves a screenshot there and answers with its key and a written description (an `analyzeImage` call). `session.grantView()` is a grant the frontend's `browser.view` (`@mindstudio-ai/interface` 0.1.36) watches it live with. App code isn't affected.
12
+ - `runTask`'s inline tools take `supersedes: '<key>'`, for a result that a newer one makes stale (a page): only the latest under each key stays in full in what the model is resent. `toolCalls`, `onEvent` and the trace still carry every result. App code isn't affected.
13
+
14
+ ## 0.1.171
15
+
16
+ ### New
17
+
18
+ - `evaluateDecision` can run on Nace.AI's Drex, served by DeepInfra: `decisionModelOverride.model` set to `drex-max`, `drex-v2.0` or `drex-v1.5`. Each takes at most 16 questions, and `drex-max` also reads up to 4 PNG, JPEG, WebP or GIF images, 9 MiB together. The usage notes say so. App code isn't affected.
19
+ - `push()`, `update()` and `upsert()` take `created_at`, `updated_at` (unix ms) and `last_updated_by` (a user's id), so an import or a scenario's seed data can carry a record's real history: a case reported three days ago, last touched by someone else. Whichever a write leaves out the platform fills in, as before: now, and the signed-in user (or nobody). An update that leaves them out still advances `updated_at` and stamps `last_updated_by`. `id` is still always the platform's. A database takes the change at its next schema sync in dev, or its next publish in live; until then a given `updated_at` or `last_updated_by` is overwritten as before.
20
+
21
+ ### Changed behavior
22
+
23
+ - A write that carries `created_at`, `updated_at` or `last_updated_by` now writes them, where before they were dropped. The case to check is a row spread from one read back: `push({ ...row, name: 'Copy' })` used to give the copy fresh dates and the current user, and now keeps the original's. Leave those three out of a copy that should look new. An update that spreads a read row back is unaffected, since unchanged values are still advanced and stamped.
24
+
25
+ ### Fixed
26
+
27
+ - `email.send`'s docs say who an app can mail. An app with neither its own subdomain nor a verified sending domain sends from the platform's shared address, which only mails the app's verified users and its workspace's members, in dev and live alike, and refuses the whole send with `unverified_recipients` (400) otherwise. The rule isn't new. An app that mails people who haven't signed in, such as a form's acknowledgement, should check it has a subdomain or a sending domain.
28
+
7
29
  ## 0.1.170
8
30
 
9
31
  ### Changed behavior