@jaypie/mcp 0.8.186 → 0.8.188

package/dist/assets.js

@@ -55,11 +55,11 @@ function getMcpAssetPaths() {
  */
 function getMcpBuildInfo() {
     return {
-        commit: "5f9c6bfe"
+        commit: "b2041bca"
             ,
-        version: "0.8.186"
+        version: "0.8.188"
             ,
-        versionString: "@jaypie/mcp@0.8.186#5f9c6bfe"
+        versionString: "@jaypie/mcp@0.8.188#b2041bca"
             ,
     };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jaypie/mcp",
-  "version": "0.8.186",
+  "version": "0.8.188",
   "description": "Jaypie MCP",
   "repository": {
     "type": "git",

package/release-notes/llm/1.3.63.md

@@ -0,0 +1,26 @@
+---
+version: 1.3.63
+date: 2026-10-08
+summary: MODEL.HAIKU points at Claude Haiku 5.5
+---
+
+## Changes
+
+### Claude Haiku 5.5
+
+- `MODEL.HAIKU` is now `claude-haiku-5-5`, superseding `claude-haiku-4-5`. Every `MODEL.HAIKU` consumer follows, including the live matrix.
+- `COST` prices `claude-haiku-5-5` at $0.10 input and $0.50 output per million tokens, with cache reads at $0.01 and cache writes at $0.125 (5m) and $0.20 (1h). This is the rate card for prompts up to 100K tokens; the $0.50 / $2.50 long-prompt rate is not modeled. The `claude-haiku-4-5` entry stays so historic usage remains replayable.
+- Streaming requests resolve Haiku 5.5's 128,000 output-token maximum. Haiku 4.5 keeps 64,000.
+- `AnthropicAdapter` strips `temperature` for Haiku 5 and later, which reject non-default sampling parameters with a 400.
+- The deprecated `PROVIDER.ANTHROPIC.MODEL.TINY` tier stays frozen on `claude-haiku-4-5`.
+
+### Behavior to expect from the new model
+
+- Thinking is adaptive and on by default, and it counts toward `max_tokens`. A caller passing a small explicit output limit may see a response stop before any text.
+- Cost per token drops tenfold against Haiku 4.5, though the newer tokenizer counts the same text as about 30% more tokens.
+
+## Testing
+
+- `maxOutputTokens.spec.ts` covers the 128,000 ceiling for `claude-haiku-5-5` alongside 64,000 for `claude-haiku-4-5`.
+- `AnthropicAdapter.spec.ts` covers the temperature strip for `claude-haiku-5-5`.
+- Verified live: `claude-haiku-5-5` passes all seven capability matrix cells.

package/release-notes/llm/1.3.64.md

@@ -0,0 +1,43 @@
+---
+version: 1.3.64
+date: 2026-10-09
+summary: MODEL.MISTRAL.LARGE points at Mistral Large 4, MODEL.MISTRAL.MEDIUM returns, and models that cannot combine format with tools are marked
+---
+
+## Changes
+
+### Models that cannot combine `format` with tools
+
+- `LLM.FORMAT_WITH_TOOLS_UNSUPPORTED` lists, by literal id, the models known not to settle when one call carries both `format` and `tools`: `mistral-large-4`, `mistral-large-4-0`, and `mistral-medium-3-5`. Either option alone works on them.
+- `operate()` and `stream()` log a warning when such a model is asked for the pair, then run the call anyway.
+- In a fallback chain the attempt throws `LlmUnrecoverableError` before any request goes out, so the next model takes over at once. The linger pass on the primary still runs the call.
+- The live matrix skips the `both` cell for every listed id, cataloged or passed through `APP_MODELS`. `APP_FORCE=1` still runs it.
+
+### Mistral Large 4
+
+- `MODEL.MISTRAL.LARGE` is now `mistral-large-4-0` (public preview since 2026-10-06), superseding `mistral-large-2512`. `PROVIDER.MISTRAL.DEFAULT` follows. Mistral's own `mistral-large-latest` alias still names Large 3.
+- `COST` prices `mistral-large-4-0` at its list rate: $1.36 input, $4.18 output, and $0.14 cached input per million tokens. Mistral currently bills half that under an open-ended "sale price"; promotional rates are not modeled. The `mistral-large-2512` entry stays so historic usage remains replayable.
+- `MistralAdapter` sends `reasoning_effort` to Large 4 and later. Large 4 accepts `none` and `high` like the other reasoning models. Large 3 and `mistral-large-latest` still omit the field.
+
+### Mistral Medium
+
+- `MODEL.MISTRAL.MEDIUM` is back in the catalog as `mistral-medium-3-5`. It was removed on 2026-08-02 because it cannot combine `format` with tools; the marker above now carries that limit.
+
+### Greedy sampling
+
+- `MistralAdapter` sends `top_p: 1` alongside `temperature: 0` unless the caller set `top_p`. Large 4 rejects a bare `temperature: 0` with "top_p must be 1 when using greedy sampling" while it reasons. 1 is the API default, so other models are unaffected.
+
+### Behavior to expect from Large 4
+
+- A call that passes both `format` and `tools` on `MODEL.MISTRAL.LARGE` or the Mistral default now warns, and usually does not finish. It worked on Large 3.
+- PDF and other `document_url` input fails with "The model `mistral-ocr-2512` does not exist for router `ocr`". This is on Mistral's side and worked on Large 3. `Llm.ocr` is unaffected.
+- Reasoning is on by default and counts toward output tokens. Pass `effort: "lowest"` to turn it off.
+- List price is about 2.7 times Large 3's.
+
+## Testing
+
+- `MistralAdapter.spec.ts` covers the effort gate for Large 4 against Large 3, and the `top_p` pin.
+- `constants.spec.ts` covers the Mistral catalog, its prices, and the marked ids.
+- `guardFormatWithTools.spec.ts` covers the warning, the fail-fast throw, and the silent cases; `Llm.class.spec.ts` covers `failFast` reaching chain attempts and not the linger pass or a lone model.
+- Verified live on 2026-10-09: a `format` plus `tools` call on `["mistral-large-4-0", MODEL.MISTRAL.SMALL]` failed over and answered in 1.1 seconds.
+- Verified live on 2026-10-09 with `APP_GROUP=mistral`: `mistral-large-4-0` passes five cells with `both` and `pdf` skipped, `mistral-medium-3-5` passes six with `both` skipped, and `mistral-small-2603` passes all seven.

package/skills/llm.md

@@ -59,7 +59,7 @@ LlamaCloud serves LlamaParse, which extracts documents and generates no text. Th
 ### Model Constants
 
 - **`PROVIDER.<name>.DEFAULT`** — the single default model per provider (above), used when no `model` is given.
-- **`LLM.MODEL.*`** — the named model catalog (e.g. `MODEL.SONNET`, `MODEL.ASTRA`, `MODEL.SOL`, `MODEL.GEMINI_FLASH`, `MODEL.GROK`, `MODEL.MUSE_SPARK`, `MODEL.MUSE_SPARK_CONTRIBUTOR`, `MODEL.NOVA_PRO`, `MODEL.NOVA_LITE`), plus three nested subtrees: `MODEL.FIREWORKS.*` for Fireworks serverless models (`DEEPSEEK`, `DEEPSEEK_FLASH`, `GLM`, `GLM_FLASH`, `GPT_OSS`, `INKLING`, `KIMI`, `MINIMAX`, `NEMOTRON`, `QWEN`), `MODEL.MISTRAL.*` (`LARGE`, `OCR`, `SMALL`), `MODEL.LLAMAPARSE.*` for LlamaParse tiers (`FAST`, `COST_EFFECTIVE`, `AGENTIC`, `AGENTIC_PLUS`), and `MODEL.OPENROUTER.*` for provider-prefixed routes (`GLM`, `LUNA`, `MIMO`, `SONNET`). Pick specific models from here. `MODEL.MISTRAL.OCR` and `MODEL.LLAMAPARSE.*` are document-extraction engines reached through `Llm.ocr`, not chat completions: they are priced per page in `LLM.PAGE_COST` (USD per 1,000 pages, keyed by literal id) and carry no `COST` entry. Amazon's Nova models are first-class ids served over Bedrock; Bedrock's third-party routes are not catalogued — pass the literal id (e.g. `us.anthropic.claude-sonnet-4-6`) and `determineModelProvider` resolves it to `bedrock`.
+- **`LLM.MODEL.*`** — the named model catalog (e.g. `MODEL.SONNET`, `MODEL.ASTRA`, `MODEL.SOL`, `MODEL.GEMINI_FLASH`, `MODEL.GROK`, `MODEL.MUSE_SPARK`, `MODEL.MUSE_SPARK_CONTRIBUTOR`, `MODEL.NOVA_PRO`, `MODEL.NOVA_LITE`), plus three nested subtrees: `MODEL.FIREWORKS.*` for Fireworks serverless models (`DEEPSEEK`, `DEEPSEEK_FLASH`, `GLM`, `GLM_FLASH`, `GPT_OSS`, `INKLING`, `KIMI`, `MINIMAX`, `NEMOTRON`, `QWEN`), `MODEL.MISTRAL.*` (`LARGE`, `MEDIUM`, `OCR`, `SMALL`), `MODEL.LLAMAPARSE.*` for LlamaParse tiers (`FAST`, `COST_EFFECTIVE`, `AGENTIC`, `AGENTIC_PLUS`), and `MODEL.OPENROUTER.*` for provider-prefixed routes (`GLM`, `LUNA`, `MIMO`, `SONNET`). Pick specific models from here. `MODEL.MISTRAL.OCR` and `MODEL.LLAMAPARSE.*` are document-extraction engines reached through `Llm.ocr`, not chat completions: they are priced per page in `LLM.PAGE_COST` (USD per 1,000 pages, keyed by literal id) and carry no `COST` entry. Amazon's Nova models are first-class ids served over Bedrock; Bedrock's third-party routes are not catalogued — pass the literal id (e.g. `us.anthropic.claude-sonnet-4-6`) and `determineModelProvider` resolves it to `bedrock`.
 - `MODEL.MUSE_SPARK` and `MODEL.MUSE_SPARK_CONTRIBUTOR` are the same Meta model at two tiers. The contributor tier is roughly a tenth of the price because prompts and completions may train Meta models; it is limited to 100 RPM and does not accept `reasoning.effort: "max"`. It is an explicit opt-in: `PROVIDER.META.DEFAULT` is `MODEL.MUSE_SPARK`.
 - The catalog is the **single source of truth for CI coverage**: `packages/llm/test/models.ts` derives the live capability matrix from `MODEL.*` plus each `PROVIDER.*.DEFAULT`, and the workflow shards it by provider. Adding a model to `MODEL.*` puts it under test; no id list exists anywhere else.
 - **Deprecated:** the size-tier map `PROVIDER.<name>.MODEL.{DEFAULT,LARGE,SMALL,TINY}`, the `DEFAULT.MODEL` bundle, and `ALL` are `@deprecated` and retired in 2.0 — use `PROVIDER.*.DEFAULT` for defaults and `MODEL.*` for named models.
@@ -1240,10 +1240,30 @@ erroring):
   answers in prose.
 - **Mistral** — only models that reason accept the field, and they accept only
   `none` and `high`, so the neutral scale collapses to a binary: `lowest` maps
-  to `none`, everything else to `high`. Mistral Large 3 rejects the field
-  outright and is skipped; a model that rejects it at runtime is cached and the
+  to `none`, everything else to `high`. Mistral Large 3 (`mistral-large-2512`)
+  rejects the field outright and is skipped, while Large 4
+  (`MODEL.MISTRAL.LARGE`) accepts it and reasons by default when `effort` is
+  unset; a model that rejects it at runtime is cached and the
   request transparently retried without it. Structured output combines with
   tools natively — no emulation.
+
+### Models that cannot combine `format` with tools
+
+`LLM.FORMAT_WITH_TOOLS_UNSUPPORTED` lists, by literal id, the models known not
+to settle when one call carries both `format` and `tools` (currently
+`mistral-large-4`, `mistral-large-4-0`, and `mistral-medium-3-5`, which are
+`MODEL.MISTRAL.LARGE` and `MODEL.MISTRAL.MEDIUM`). They restate
+the answer or re-call the tool, and a request can run for minutes. Either
+option alone works.
+
+- `operate()` and `stream()` log a warning and run the call anyway when such a
+  model is the only one.
+- In a fallback chain (a `model` array or `fallback`), the attempt throws
+  `LlmUnrecoverableError` before any request goes out and the next model takes
+  over. If every model fails, the primary still runs once more as usual.
+
+Jaypie 2.0 splits `operate` into `ask` (structured) and `loop` (tools).
+
 - **Bedrock** — not yet wired; `effort` is ignored.
 
 ### Per-model effort
@@ -1346,7 +1366,7 @@ truncate:
 - **Non-streaming** (`operate()`, `send()`): capped at 16,384 tokens —
   larger non-streaming responses risk HTTP timeouts (stream instead)
 - **Streaming** (`stream()`): the model maximum — e.g., 128,000 for current
-  Claude models (64,000 for Haiku), 65,536 for Gemini 2.5/3.x
+  Claude models (64,000 for Haiku 4.5), 65,536 for Gemini 2.5/3.x
 
 Override per model with `modelOptions`: