@plumbus/core 0.6.9 → 0.6.12

package/CHANGELOG.md

@@ -1,5 +1,40 @@
 # @plumbus/core changelog
 
+## 0.6.12 — 2026-07-26 — agent wiring version 10
+
+### Changed
+
+- **`AGENT_WIRING_VERSION` bumped to 10.** The `@plumbus/auth` instruction pointers generated by `plumbus init` now describe **invitation-only admission via `loginContext`** — so a coding agent asked "the invite flow needs user context, but an invite has no user yet" is routed to `node_modules/@plumbus/auth/instructions/resolvers.md` instead of inventing a workaround. The `configure-runtime.md` pointer mentions the optional `loginContext` step. Requires `@plumbus/auth` **0.1.1+** for the referenced sections; the pointers are inert when auth is not installed.
+- Run **`plumbus init --patch`** on existing projects to refresh the managed wiring block. `plumbus doctor` reports files below version 10 as stale and recommends the safest follow-up command.
+
+### Notes
+
+No runtime, API, or wire-format change — the bump only affects generated agent wiring files and the freshness check that reports them. Projects that skip the patch keep working; their agent wiring simply describes the auth resolver pointers as before.
+
+## 0.6.11 — 2026-07-26 — provider-native tool calling
+
+### Added
+
+- **Provider tool protocol** — `AITool`, `AIToolChoice`, `AIToolCall` (discriminated on `argumentsStatus`), `AIToolExecutionOptions`, and `AIProviderCapabilities`. `ChatMessage` is now a union (`user` | `assistant` with optional `toolCalls` | `tool`); the existing `user`/`assistant` construction forms are unchanged.
+- **Adapter caller tools** — `createOpenAIAdapter` and `createAnthropicAdapter` implement caller tools natively (OpenAI `tools`/`tool_calls`; Anthropic `tool_use`/`tool_result` + `input_schema`). `AIProviderAdapter.capabilities` is optional; an adapter that omits it is treated as declaring every capability `false`.
+- **`generateWithUsage` result overloads** — no-tool callers keep the flat `AIFinalGenerateResult<T>` with unconditional `.data`; tool-enabled config returns the discriminated `AIToolEnabledGenerateResult<T>` (keyed on `finishReason`). `AIGenerateResult` is preserved as a back-compat alias of the flat result; `finishReason` is an additive field existing consumers may ignore.
+- **`runToolLoop`** (`src/ai/tool-loop.ts`) — bounded request→tool→observe loop for capability authors. Default `maxRounds` 8, hard maximum 20; the exhausted-loop final request omits **both** `tools` and `toolChoice`. Invalid-argument calls are never executed and surface as a `tool_arguments_invalid` observation.
+- **`EntityIndexDefinition.unique`** — `indexes` accepts the legacy `string[][]` form or `{ columns, unique? }`. Additive; existing entity index declarations are unchanged.
+- **`FlowService.describe()`** — returns a `FlowDescription` (flow name, input JSON schema, input-schema hash) so callers such as `@plumbus/chat` can bind flows as provider-native tools without reaching into flow internals.
+- **`safeJsonStringify`** (exported) — best-effort JSON serialization that never throws (handles circular refs / non-serializable values), used for tool observations and structured logging. **`AIInvalidRequestError`** (exported) — structured `PlumbusError` subclass for provider-rejected/invalid AI requests; existing `catch` blocks are unaffected.
+- **`normalizeFinishReason`** — TIER-1 export that maps provider-specific finish reasons onto the framework's canonical set, so tool-loop callers can branch on one vocabulary across adapters.
+- **`Repository.updateWhere` (optional conditional/CAS update) + transaction-capable write path** — `Repository` gains the optional `updateWhere` method returning `ConditionalUpdateResult`, surfaced to `@plumbus/chat` for atomic lease-based session mutations. `updateWhere` is **optional** on the `Repository` interface, so pre-existing custom `Repository` implementations still compile; stores that lack it fail closed with `chat.storage_unsupported` — raised on the first turn or confirm that needs a conditional write, not at process start — rather than silently losing conditional semantics.
+
+### Notes
+
+All changes are additive and stay within the `0.6.x` line. Declared add-on peers remain satisfied (`@plumbus/chat` / `@plumbus/chat-ui` keep `0.5.x || 0.6.x`; `@plumbus/voice` keeps `0.6.x`). **`@plumbus/chat` 0.1.11 requires this core line in practice** (≥ 0.6.11) for the tool protocol and `updateWhere` path. No install-time or wire-format break for callers that stay on older chat.
+
+## 0.6.10
+
+### Fixed
+
+- **OpenAI `gpt-5.5+` temperature** — the OpenAI adapter no longer sends `temperature` for `gpt-5.5` and later models (including `-pro` and dated snapshots). Those models only accept the API default (`1`); sending Plumbus's `0.7` default (or any other value) caused HTTP 400 `unsupported_value`. Earlier `gpt-5` lines such as `gpt-5.4-mini` still receive the configured temperature. `max_completion_tokens` mapping for the broader `gpt-5*` / `o*` family is unchanged.
+
 ## 0.6.9
 
 ### Added (M5 — authentication contract)