@jaypie/mcp 0.8.48 → 0.8.51

package/dist/suites/docs/index.js

@@ -9,7 +9,7 @@ import { gt } from 'semver';
 /**
  * Docs Suite - Documentation services (skill, version, release_notes)
  */
-const BUILD_VERSION_STRING = "@jaypie/mcp@0.8.48#ce72b43d"
+const BUILD_VERSION_STRING = "@jaypie/mcp@0.8.51#4c00d3ed"
     ;
 const __filename$1 = fileURLToPath(import.meta.url);
 const __dirname$1 = path.dirname(__filename$1);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jaypie/mcp",
-  "version": "0.8.48",
+  "version": "0.8.51",
   "description": "Jaypie MCP",
   "repository": {
     "type": "git",

package/release-notes/jaypie/1.2.43.md

@@ -0,0 +1,12 @@
+---
+version: 1.2.43
+date: 2026-04-29
+summary: Pull in @jaypie/llm 1.2.31 (native structured outputs across all providers)
+---
+
+## Changes
+
+- Bumped `@jaypie/llm` peer-dep range to `^1.2.31`, exposing:
+  - Anthropic, OpenAI, Gemini 3, and xAI all use native structured outputs by default (no more `structured_output` fake-tool emulation in the happy path).
+  - Per-provider runtime fallback that engages the legacy fake-tool path transparently when a model rejects native structured output, with a `log.warn` so silent degradations don't hide.
+  - Anthropic specifically: native field is `output_config.format` (the earlier `output_format` was deprecated by the API).

package/release-notes/jaypie/1.2.44.md

@@ -0,0 +1,11 @@
+---
+version: 1.2.44
+date: 2026-04-29
+summary: Pull in @jaypie/llm 1.2.32 (OpenAI temperature-deprecation handling)
+---
+
+## Changes
+
+- Bumped `@jaypie/llm` peer-dep range to `^1.2.32`, exposing:
+  - The OpenAI adapter strips `temperature` for denylisted models (`^gpt-5\.5`, `^o\d`) and transparently retries without the param when the API returns `400 Unsupported parameter: 'temperature' …`.
+  - Capability matrix now exercises `temperature` across every configured model.

package/release-notes/llm/1.2.30.md

@@ -0,0 +1,29 @@
+---
+version: 1.2.30
+date: 2026-04-28
+summary: Anthropic provider uses native structured outputs in place of the structured_output fake-tool emulation
+---
+
+## Changes
+
+- `AnthropicAdapter` (`operate()`/`stream()`):
+  - When `format` is set, the request body now includes `output_config: { format: { type: "json_schema", schema } }` instead of injecting a synthetic `structured_output` tool with `tool_choice: "any"`.
+  - JSON schemas are sanitized for Anthropic's grammar compiler: objects always get `additionalProperties: false`, unsupported numeric/string/array constraints (`minimum`, `maxLength`, `multipleOf`, …) are appended to `description` rather than sent to the API, unsupported string formats are stripped, and `minItems` is preserved only when 0 or 1.
+  - Responses to structured-output requests are parsed from the first `text` content block (guaranteed valid JSON by the Anthropic grammar). `stop_reason: "refusal"` and `stop_reason: "max_tokens"` are recognized and surface as text rather than triggering a parse error.
+  - Streaming structured outputs now arrive as `LlmStreamChunkType.Text` deltas (consumers concat to assemble JSON), instead of a single `ToolCall` chunk.
+- `AnthropicProvider.send(message, { response })`: the system prompt is no longer stuffed with the JSON schema, and the regex JSON extraction is gone. The native `output_config.format` field handles the contract.
+- A runtime fallback re-engages the legacy fake-tool emulation transparently when a model rejects `output_config` (cached per model in `runtimeNoStructuredOutputModels`). Citations-related 400s and `output_format`-deprecation 400s are explicitly excluded so they propagate instead of being masked by a retry.
+
+## Why
+
+Anthropic launched native structured outputs publicly on 2025-11-14 (now GA on Claude 4.5+). The previous tool-emulation hack predates the feature. Native:
+- Removes a fragile regex extraction in `provider.send`.
+- Removes the synthetic `structured_output` tool from the operate loop.
+- Returns guaranteed-valid JSON via grammar-constrained generation rather than relying on the model to comply with a system-prompt instruction.
+- Keeps a fallback path for any model that does not (yet) support `output_config`.
+
+## Notes
+
+- Streaming consumers that previously branched on a `tool_call` chunk to receive structured output will now see `text` chunks. Concatenate to assemble the final JSON, or use `operate()` for the parsed object.
+- Schema constraints not supported by Anthropic's grammar (regex `pattern`, `minLength`, numeric ranges, recursion) are stripped at request time. The original constraints are still enforced client-side by the caller's Zod schema during result validation.
+- Gemini and OpenRouter still use the fake-tool emulation pattern; native migration for those adapters is a separate task.

package/release-notes/llm/1.2.31.md

@@ -0,0 +1,33 @@
+---
+version: 1.2.31
+date: 2026-04-29
+summary: OpenRouter and Gemini adapters drop the structured_output fake-tool hack in favor of native response_format / responseJsonSchema (with runtime fallback for unsupported models)
+---
+
+## Changes
+
+- `OpenRouterAdapter` (`operate()`/`stream()`):
+  - When `format` is set, the request body now includes `response_format: { type: "json_schema", json_schema: { name, schema, strict: true } }` instead of injecting a synthetic `structured_output` tool.
+  - `formatTools` no longer injects the synthetic tool. The legacy fake tool is added in `buildRequest` only as a runtime fallback for models cached as not supporting native `response_format`.
+  - `formatOutputSchema` now forces `additionalProperties: false` on every object (required for `strict: true`).
+  - `executeRequest` annotates the response with a `__jaypieStructuredOutput` flag so `hasStructuredOutput` / `extractStructuredOutput` can detect intent statelessly. The native path parses `choices[0].message.content` as JSON; the fallback path keeps the existing tool-call extraction.
+  - 400/422 errors mentioning `response_format`/`json_schema`/`structured_output`/`require_parameters` cache the model in `runtimeNoStructuredOutputModels` and trigger a one-shot retry via the legacy fake-tool path. Unrelated 4xx propagate.
+  - `log.warn` at both fallback entry points (initial cached-model build and 4xx retry).
+- `GeminiAdapter` (`operate()`/`stream()`):
+  - Format **and** tools combined now uses native `responseJsonSchema` + tools when the model id matches `^gemini-3` (Gemini 3 preview). Gemini 2.5 (including thinking) and earlier keep the `structured_output` fake-tool emulation with the system-prompt nudge.
+  - `formatTools` no longer injects the synthetic structured_output tool; the fake tool is added in `buildRequest` only when the runtime falls back.
+  - 400 errors mentioning `responseJsonSchema`/`responseSchema`/`responseMime`/`function_call`/`tools` cache the model in `runtimeNoStructuredOutputComboModels` and trigger a one-shot retry via the legacy fake-tool path.
+  - `log.warn` at both fallback entry points.
+
+## Why
+
+OpenRouter exposes the OpenAI-style structured-outputs shape end-to-end, and Gemini 3 launched as a preview supporting tools + `responseJsonSchema` together. The `structured_output` fake-tool emulation predates both. Native:
+- Returns guaranteed-valid JSON via grammar-constrained generation rather than relying on the model to comply with a system-prompt instruction.
+- Removes the synthetic tool from the operate-loop tool list (cleaner traces, no risk of the model ignoring the nudge).
+- Keeps a fallback path that re-engages the fake-tool emulation transparently when a model rejects the native field.
+
+## Notes
+
+- Anthropic was migrated in 1.2.30; with this release, all four LLM providers (Anthropic, OpenAI, OpenRouter, Gemini) use native structured outputs by default. xAI inherits from `OpenAiAdapter` and was already native.
+- Gemini 2.5 thinking models stay on the fake-tool path because the combo is unreliable on 2.5; this matches Google's documentation that the combo is a Gemini 3 preview feature.
+- Streaming consumers that previously branched on a `tool_call` chunk to receive structured output from OpenRouter/Gemini will now see `text` chunks. Concatenate to assemble the final JSON, or use `operate()` for the parsed object.

package/release-notes/llm/1.2.32.md

@@ -0,0 +1,23 @@
+---
+version: 1.2.32
+date: 2026-04-29
+summary: OpenAI adapter strips/retries on temperature deprecation; matrix gains a temperature capability
+---
+
+## Changes
+
+- `OpenAiAdapter` (`operate()`/`stream()`) now mirrors the Anthropic adapter's temperature-deprecation handling:
+  - `MODELS_WITHOUT_TEMPERATURE` regex denylist (`^gpt-5\.5`, `^o\d`) is checked in `buildRequest` and `temperature` is stripped before the request is sent.
+  - Session-level `runtimeNoTemperatureModels` Set populated by `executeRequest` when the API returns `400 Unsupported parameter: 'temperature' …`. Subsequent calls for the same model strip the param up-front via `supportsTemperature`.
+  - `executeRequest` catches the 400, caches the model, and transparently retries without `temperature`. Unrelated 400s and 400s on requests without `temperature` still propagate.
+  - `clearRuntimeNoTemperatureModels()` exposed for tests.
+- Capability matrix (`test/matrix.ts`, `test/models.ts`) gains a `temperature` capability that runs a basic prompt with `temperature: 0` and verifies a non-empty response. Runs across every model in the matrix.
+
+## Why
+
+`gpt-5.5` rejects `temperature` outright (`400 Unsupported parameter`). Every prior adapter that hit this returned an unrecoverable error to the caller. Matching the Anthropic pattern lets the OpenAI adapter handle the deprecation transparently — first call for an unknown model pays one round-trip to learn; denylisted models pay nothing — so callers can pass `temperature` portably across the model line.
+
+## Notes
+
+- The denylist regex is forward-compatible: any future `gpt-5.5.*` or o-series variant strips automatically. Other models that newly deprecate the param fall into the runtime-cache path on first 400.
+- Streaming (`executeStreamRequest`) does not retry on temperature errors; it relies on the denylist + warmed runtime cache from prior `operate()` calls. Models not yet in the cache will surface the 400 to the streaming consumer.

package/release-notes/mcp/0.8.49.md

@@ -0,0 +1,9 @@
+---
+version: 0.8.49
+date: 2026-04-28
+summary: Add release notes for @jaypie/llm 1.2.30 (Anthropic native structured outputs)
+---
+
+## Changes
+
+- Added `release-notes/llm/1.2.30.md` covering the Anthropic provider's switch from the `structured_output` fake-tool emulation to Anthropic's native `output_format` structured outputs (with a runtime fallback to the legacy path for unsupported models).

package/release-notes/mcp/0.8.50.md

@@ -0,0 +1,9 @@
+---
+version: 0.8.50
+date: 2026-04-29
+summary: Add release notes for @jaypie/llm 1.2.31 (OpenRouter and Gemini native structured outputs)
+---
+
+## Changes
+
+- Added `release-notes/llm/1.2.31.md` covering the OpenRouter and Gemini adapters' migration from the `structured_output` fake-tool emulation to native `response_format: json_schema` (OpenRouter) and `responseJsonSchema` + tools (Gemini 3), with a runtime fallback for unsupported models.

package/release-notes/mcp/0.8.51.md

@@ -0,0 +1,10 @@
+---
+version: 0.8.51
+date: 2026-04-29
+summary: Add release notes for @jaypie/llm 1.2.32 and jaypie 1.2.44 (OpenAI temperature-deprecation handling)
+---
+
+## Changes
+
+- Added `release-notes/llm/1.2.32.md` covering the OpenAI adapter's temperature-deprecation strip-and-retry path and the new `temperature` capability in the matrix harness.
+- Added `release-notes/jaypie/1.2.44.md` for the corresponding `jaypie` peer-dep bump to `@jaypie/llm` `^1.2.32`.