esoul-sdk 0.20.0 → 0.21.1

package/CHANGELOG.md

@@ -2,6 +2,47 @@
 
 Releases before 0.20.0 are recorded in the repository history only.
 
+## 0.21.1
+
+### Fixed
+
+- **SimGmail people answer an agent's mail.** A persona that is a person (`reply`, `optout`) now reads and
+  answers mail marked `Auto-Submitted` (what an app's autonomous agent sends, RFC 3834) — as people do. Only a
+  machine responder (`ooo`) stays silent to it, and a dead address bounces it. Before, every persona ignored
+  it, so a simulated conversation with an autonomous agent stopped after the first reply.
+
+### Added
+
+- **`esoul-sdk/testing` exports the failed-request list** (`listFailedRequests`, `subscribeFailedRequests`,
+  `resetFailedRequests`) so a test can assert that a click which failed reverted and was named on the banner
+  (`runOptimistic` from `esoul-sdk/react`).
+
+## 0.21.0
+
+### Added
+
+- **Model calls, billed: `llm` in `plugin.json` and `llm(ctx).generate(req)` in `esoul-sdk/server`.**
+  Declare `"llm": { "why", "dailyBudgetUsd"? (default 5, at most 100), "models"? }` and call
+  `llm(ctx).generate({ purpose, model?, system?, messages, tools?, schema?, maxOutputTokens?, temperature? })`
+  from an op or a route — or `ctx.llm.generate(…)` from a task. The app instance's workspace **owner**
+  pays from their AI credits; every call is recorded with the app, the instance and `purpose`; the
+  app's spend per owner per UTC day is capped by `dailyBudgetUsd`. `model` is a tier (`fast`,
+  `balanced`, `best`) or a gateway id on the platform's allowlist (DeepSeek, Claude, GPT mini/nano,
+  Gemini Flash) or in `llm.models`. `tools` are definitions: the model's calls come back in
+  `toolCalls`, never executed by the platform. `schema` (JSON Schema) returns the answer parsed in
+  `object`. A refusal throws `LlmUnavailable` (`isLlmUnavailable`) with `code` `not_declared` |
+  `not_allowed` | `no_credits` | `budget_exceeded` | `model_not_allowed` | `provider_unavailable`.
+  Schema: `PluginLlmSchema`. See `docs/19-model-calls.md`.
+- **`scriptedModel(policy?)` in `esoul-sdk/testing`.** A deterministic stand-in for `llm(ctx)`: scripted
+  texts, objects and tool calls by matching the last user message (or a function), a short echo by
+  default, zero cost, every request in `.calls`. The Forge box answers every model call with one.
+
+### Changed
+
+- **App files may not import a model SDK** (`ai`, `@ai-sdk/*`, `openai`, `@anthropic-ai/*`,
+  `@google/genai`, `@inngest/ai`) — the box check and the install inspection refuse it, tests
+  included. Apps call models through `llm(ctx)`.
+
 ## 0.20.0
 
 ### Added

package/README.md

@@ -573,6 +573,7 @@ Exits 0 when the folder is a well-formed application package; otherwise prints e
 | [14. Your own tables](docs/14-database.md) | `db`, rules, scopes, sealed fields, indexes, migrations |
 | [15. Realtime](docs/15-realtime.md) | Topics, audiences, addressing |
 | [16. Bindings](docs/16-bindings.md) | Slots, contracts, reaching another application |
+| [19. Model calls](docs/19-model-calls.md) | `llm(ctx)`: billed model calls, budgets, `scriptedModel()` for tests |
 
 ## Versions