@jaypie/mcp 0.8.144 → 0.8.146

package/dist/suites/docs/index.js

@@ -10,7 +10,7 @@ import { gt } from 'semver';
 /**
  * Docs Suite - Documentation services (skill, version, release_notes)
  */
-const BUILD_VERSION_STRING = "@jaypie/mcp@0.8.144#c397172d"
+const BUILD_VERSION_STRING = "@jaypie/mcp@0.8.146#5d742935"
     ;
 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": "0.8.144",
+  "version": "0.8.146",
   "description": "Jaypie MCP",
   "repository": {
     "type": "git",

package/release-notes/jaypie/1.2.96.md

@@ -0,0 +1,9 @@
+---
+version: 1.2.96
+date: 2026-09-04
+summary: Pick up @jaypie/llm 1.3.38
+---
+
+## Changes
+
+- `@jaypie/llm` 1.3.38: catalog `MODEL.ASTRA` as `gpt-6-astra`

package/release-notes/jaypie/1.2.97.md

@@ -0,0 +1,10 @@
+---
+version: 1.2.97
+date: 2026-09-05
+summary: Pick up @jaypie/llm 1.3.39
+---
+
+## Changes
+
+- `@jaypie/llm` 1.3.39: Meta Model API provider with `MODEL.MUSE_SPARK` and
+  `MODEL.MUSE_SPARK_CONTRIBUTOR`

package/release-notes/llm/1.3.38.md

@@ -0,0 +1,43 @@
+---
+version: 1.3.38
+date: 2026-09-04
+summary: Catalog MODEL.ASTRA as gpt-6-astra
+---
+
+## Changes
+
+- `MODEL.ASTRA` is `gpt-6-astra`, added alongside `MODEL.SOL` rather than
+  replacing it. Astra is a new-generation flagship at twice Sol's rate, so the
+  two coexist as distinct offerings and `PROVIDER.OPENAI.DEFAULT` stays on
+  `MODEL.SOL`. The catalog addition puts Astra in the live capability matrix and
+  the OpenAI hot tests.
+- `COST` gains `gpt-6-astra` at $10.00 input / $50.00 output per million tokens
+  with $1.00 cache reads.
+- `reasoning.effort` follows the gpt-6 ladder: `highest` reaches `max`, the rung
+  above `xhigh`, and `lowest` clamps to `low` because gpt-6 drops `minimal`. The
+  gpt-5 line is unchanged.
+- `PROVIDER.OPENAI.MODEL_MATCH_WORDS` gains `astra`, matching how `sol`, `terra`,
+  and `luna` are carried.
+
+- CI no longer shards Mistral. The provider's availability does not survive a
+  per-push run — a `Rate limit exceeded` 429 across all seven cells, and an
+  entitlement refusal before that — so `npm-check.yml` and `npm-deploy.yml` drop
+  the `mistral` group. Mistral models stay cataloged, priced, and adapter-tested;
+  `APP_GROUP=mistral npm run test:llm:matrix` still reaches the live cells.
+
+- `AnthropicAdapter` now logs a warning when a response carries no text block,
+  naming the `stop_reason` and the block types actually returned. The call still
+  settles with `undefined` content, but a refusal, a thinking-only turn, and a
+  truncated one are no longer indistinguishable from each other. Tool-use turns
+  are unaffected: they legitimately carry no text.
+
+## Testing
+
+- `determineModelProvider` resolves the id on the existing `gpt` match word;
+  `astra` is added for the bare name.
+- `maxOutputTokens` leaves OpenAI unset by policy: Astra's 128K ceiling is not a
+  low ceiling, and streaming requests carry no limit.
+- The matrix takes no `MATRIX_EXPECT` entry. Astra advertises the same input
+  modalities as Sol (text and image), and Sol passes every capability cell.
+- Deprecated `PROVIDER.OPENAI.MODEL` size tiers and the `MODEL.GPT*` aliases are
+  unchanged by policy.

package/release-notes/llm/1.3.39.md

@@ -0,0 +1,70 @@
+---
+version: 1.3.39
+date: 2026-09-05
+summary: Meta Model API provider with MODEL.MUSE_SPARK and MODEL.MUSE_SPARK_CONTRIBUTOR
+---
+
+## Changes
+
+- `MetaProvider` (`PROVIDER.META`, name `meta`) calls Meta's Model API at
+  `https://api.meta.ai/v1`. The API serves the OpenAI Responses protocol, so
+  `MetaAdapter` extends `OpenAiAdapter` the way `XaiAdapter` does and inherits
+  request building, tool handling, structured output, streaming, and `input_file`
+  data URIs. `Llm` resolves the provider from the model id, the bare name
+  `meta`, or a `meta:` prefix.
+- `MODEL.MUSE_SPARK` is `muse-spark-1.3` and `PROVIDER.META.DEFAULT`.
+  `MODEL.MUSE_SPARK_CONTRIBUTOR` is `muse-spark-1.3-contributor`: the same
+  model at roughly a tenth of the price in exchange for data use (prompts and
+  completions may train Meta models), limited to 100 RPM. The contributor tier
+  is an explicit opt-in and is never the default.
+- `COST` prices all five Muse Spark ids: `muse-spark-1.3` and `muse-spark-1.2`
+  at $1.25 input / $4.25 output with $0.15 cache reads, the `-contributor` ids
+  at $0.10 / $0.20 with $0.002 cache reads, and `muse-spark-1.1` at the
+  standard rate. Reasoning bills as output; no write premium is published.
+- The API key is read from `META_API_KEY`, then from `MODEL_API_KEY` (the name
+  Meta's own docs and SDK examples use), then an explicit `apiKey` option wins
+  over both. `ConfigurationError` when none resolves.
+- `reasoning.effort` walks Meta's `minimal | low | medium | high | xhigh | max`
+  ladder. `max` is limited to the Standard-tier `muse-spark-1.3`, so `highest`
+  on the contributor id or an older release clamps to `xhigh` and logs at
+  debug. Muse Spark always reasons: `none` is rejected and never emitted.
+  `reasoning.summary: "auto"` is requested on every Muse Spark id.
+- The `user` option is sent as `safety_identifier`, the field Meta names as the
+  successor to the deprecated `user`.
+- A 402 from Meta is classified as a terminal billing failure (`LlmQuotaError`,
+  `reason: "billing"`) rather than retried.
+- `PROVIDER.META.MODEL_MATCH_WORDS` is `["muse"]`. `meta` is deliberately not a
+  match word: `meta.llama-*` stays on Bedrock and `meta-llama/*` on OpenRouter.
+- `OpenAiAdapter` gains a `protected supportsReasoningSummary(model)` hook so
+  OpenAI-compatible subclasses can request reasoning summaries for models
+  outside OpenAI's `gpt-5+` / o-series naming.
+
+## Testing
+
+- Unit coverage for the adapter (field renaming, reasoning summary, effort
+  ladder and the `max` gate, `prompt_cache_key`, usage with `cached_tokens` and
+  `reasoning_tokens`, 402 classification), the provider (key resolution order),
+  routing (every cataloged and priced id, the `meta:` prefix, Bedrock and
+  OpenRouter ids left alone), and constants (default never the contributor
+  tier, every Muse Spark price within the `COST` invariants).
+- A `meta` group joins the CI capability matrix in `npm-check.yml` and
+  `npm-deploy.yml`, keyed on `CICD_META_API_KEY`.
+- Live verification on 2026-09-05: the capability matrix passed on both
+  cataloged models. The `pdf` cell on `muse-spark-1.3` overran the 180 second
+  cell deadline on the first run and passed on a rerun, so no `MATRIX_EXPECT`
+  entry is pinned for it. Live probes confirmed `max` is accepted on
+  `muse-spark-1.3` and rejected with a 400 on the contributor tier, the
+  response echoes the literal model id, a repeat call reports `cached_tokens`,
+  and a two-turn tool call replays the reasoning item without error. Before
+  billing was configured every call returned 402 `billing_not_configured`,
+  which exercised the billing classification.
+- `test/matrix.ts` now exits the process explicitly after a passing run. An
+  abandoned cell keeps retrying in the background, and its backoff timers and
+  open socket held the event loop open after the verdict printed.
+- `test/matrix.ts` treats a provider refusal (`stopReason: "refusal"`) as an
+  inconclusive cell, alongside the cell deadline and an exhausted turn budget.
+  The `claude-opus-5` pdf cell refused on five of seven CI runs across two
+  branches and answered on every local run, so failing on it reported a
+  classifier flake as a missing capability. Every cell now requests the
+  exchange envelope so the stop reason is observable; the refusal still prints
+  in ISSUES.

package/release-notes/mcp/0.8.145.md

@@ -0,0 +1,10 @@
+---
+version: 0.8.145
+date: 2026-09-04
+summary: Release notes for @jaypie/llm 1.3.38 and jaypie 1.2.96
+---
+
+## Changes
+
+- Added release notes for `@jaypie/llm` 1.3.38 and `jaypie` 1.2.96
+- `skills/llm.md` documents `MODEL.ASTRA` and the gpt-6 reasoning effort ladder

package/release-notes/mcp/0.8.146.md

@@ -0,0 +1,14 @@
+---
+version: 0.8.146
+date: 2026-09-05
+summary: Release notes for @jaypie/llm 1.3.39, @jaypie/testkit 1.2.68, and jaypie 1.2.97
+---
+
+## Changes
+
+- Added release notes for `@jaypie/llm` 1.3.39, `@jaypie/testkit` 1.2.68, and
+  `jaypie` 1.2.97
+- `skills/llm.md` documents the Meta provider: the `muse` match word and
+  `meta:` prefix, `MODEL.MUSE_SPARK` and `MODEL.MUSE_SPARK_CONTRIBUTOR` with the
+  contributor tier's data-use terms, `META_API_KEY` with the `MODEL_API_KEY`
+  fallback, the Meta column of the reasoning effort table, and prompt caching

package/release-notes/testkit/1.2.68.md

@@ -0,0 +1,10 @@
+---
+version: 1.2.68
+date: 2026-09-05
+summary: Mock MetaProvider from @jaypie/llm
+---
+
+## Changes
+
+- `@jaypie/testkit/mock` wraps `MetaProvider` alongside the other `@jaypie/llm`
+  providers, matching `@jaypie/llm` 1.3.39.

package/skills/llm.md

@@ -32,23 +32,27 @@ the alias rather than the id it currently resolves to. Read the id off
 
 | Provider | Match Keywords | Default Model (`PROVIDER.*.DEFAULT`) |
 |----------|----------------|---------------|
-| OpenAI | "openai", "gpt", "sol", "terra", "luna", /^o\d/ | `MODEL.SOL` |
+| OpenAI | "openai", "gpt", "astra", "sol", "terra", "luna", /^o\d/ | `MODEL.SOL` |
 | Anthropic | "anthropic", "claude", "fable", "haiku", "mythos", "opus", "sonnet" | `MODEL.SONNET` |
 | Google | "google", "gemini" | `MODEL.GEMINI_FLASH` |
 | Fireworks | "fireworks" (also matched inside ids like `accounts/fireworks/models/...`) | `MODEL.FIREWORKS.GLM` |
 | Mistral | "mistral", "ministral", "codestral", "devstral", "magistral", "pixtral", "voxtral" | `MODEL.MISTRAL.LARGE` |
 | OpenRouter | "openrouter" | `MODEL.OPENROUTER.SONNET` |
 | xAI | "xai", "grok" | `MODEL.GROK` |
+| Meta | "muse" (plus the exact name `meta` and a `meta:` prefix) | `MODEL.MUSE_SPARK` |
 | Bedrock | "amazon.nova", "anthropic.claude", "meta.llama", "deepseek.", "google.gemma", "moonshotai.", "openai.gpt-oss", … | `MODEL.NOVA_PRO` |
 
 The provider name for Gemini models is `"google"` — `"gemini"` is accepted as a deprecated alias.
 
 Mistral's family names mostly do not contain the substring "mistral" (`ministral` is m-i-n-i-s-t-r-a-l; `codestral`/`devstral`/`pixtral`/`voxtral` share only the `-tral` suffix), so each family carries its own match word. Bedrock-hosted `mistral.mistral-*` ids still resolve to `bedrock`, and `mistralai/*` routes still resolve to `openrouter`; use the `mistral:` prefix to force the direct API.
 
+"meta" is deliberately not a Meta match word: `meta.llama-*` is a Bedrock id and `meta-llama/*` an OpenRouter route. Muse Spark ids resolve on "muse"; use the `meta:` prefix to force the Meta Model API for any other id.
+
 ### 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.SOL`, `MODEL.GEMINI_FLASH`, `MODEL.GROK`, `MODEL.NOVA_PRO`, `MODEL.NOVA_LITE`), plus three nested subtrees: `MODEL.FIREWORKS.*` for Fireworks serverless models (`DEEPSEEK`, `GLM`, `GPT_OSS`, `INKLING`, `KIMI`, `MINIMAX`, `NEMOTRON`, `QWEN`), `MODEL.MISTRAL.*` (`LARGE`, `OCR`, `SMALL`), and `MODEL.OPENROUTER.*` for provider-prefixed routes (`GLM`, `LUNA`, `SONNET`). Pick specific models from here. `MODEL.MISTRAL.OCR` serves `POST /v1/ocr` rather than chat completions: it is priced per page, carries no `COST` entry, and is reached through `MistralProvider.ocr()`. 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`, `GLM`, `GPT_OSS`, `INKLING`, `KIMI`, `MINIMAX`, `NEMOTRON`, `QWEN`), `MODEL.MISTRAL.*` (`LARGE`, `OCR`, `SMALL`), and `MODEL.OPENROUTER.*` for provider-prefixed routes (`GLM`, `LUNA`, `SONNET`). Pick specific models from here. `MODEL.MISTRAL.OCR` serves `POST /v1/ocr` rather than chat completions: it is priced per page, carries no `COST` entry, and is reached through `MistralProvider.ocr()`. 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.
 
@@ -644,6 +648,7 @@ MISTRAL_API_KEY     # Required for Mistral
 OPENAI_API_KEY      # Required for OpenAI
 OPENROUTER_API_KEY  # Required for OpenRouter
 XAI_API_KEY         # Required for xAI (Grok)
+META_API_KEY        # Required for Meta (Muse Spark); MODEL_API_KEY is read as a fallback
 ```
 
 Keys are resolved via `getEnvSecret()` which supports AWS Secrets Manager.
@@ -858,13 +863,13 @@ await Llm.operate("Solve this step by step", {
 Per-provider translation (`medium`/`high` stay aligned across providers; ends
 collapse where a provider has fewer rungs):
 
-| `effort`  | OpenAI `reasoning.effort` | Anthropic `output_config.effort` | Gemini 3 `thinkingLevel` | Gemini 2.5 `thinkingBudget` | Grok `reasoning_effort` | OpenRouter `reasoning.effort` | Fireworks `reasoning_effort` | Mistral `reasoning_effort` |
-|-----------|---------------------------|----------------------------------|--------------------------|-----------------------------|-------------------------|-------------------------------|------------------------------|----------------------------|
-| `lowest`  | minimal | low    | MINIMAL | 512   | low    | minimal | low    | none |
-| `low`     | low     | low    | LOW     | 4096  | low    | low     | low    | high |
-| `medium`  | medium  | medium | MEDIUM  | 8192  | medium | medium  | medium | high |
-| `high`    | high    | high   | HIGH    | 16384 | high   | high    | high   | high |
-| `highest` | xhigh   | max    | HIGH    | 24576 | high   | xhigh   | high   | high |
+| `effort`  | OpenAI `reasoning.effort` | Anthropic `output_config.effort` | Gemini 3 `thinkingLevel` | Gemini 2.5 `thinkingBudget` | Grok `reasoning_effort` | Meta `reasoning.effort` | OpenRouter `reasoning.effort` | Fireworks `reasoning_effort` | Mistral `reasoning_effort` |
+|-----------|---------------------------|----------------------------------|--------------------------|-----------------------------|-------------------------|-------------------------|-------------------------------|------------------------------|----------------------------|
+| `lowest`  | minimal | low    | MINIMAL | 512   | low    | minimal | minimal | low    | none |
+| `low`     | low     | low    | LOW     | 4096  | low    | low     | low     | low    | high |
+| `medium`  | medium  | medium | MEDIUM  | 8192  | medium | medium  | medium  | medium | high |
+| `high`    | high    | high   | HIGH    | 16384 | high   | high    | high    | high   | high |
+| `highest` | xhigh   | max    | HIGH    | 24576 | high   | max     | xhigh   | high   | high |
 
 When a neutral level has no distinct native rung and collapses onto a neighbor
 it is "papered over." Papering that follows from the provider's native scale
@@ -873,21 +878,27 @@ Anthropic `lowest`) is the only possible outcome for that provider and is not
 logged. Papering that depends on the specific model, such as OpenAI clamping
 `highest` on a model predating `xhigh`, logs at `log.debug`. OpenAI's extremes
 are also version-gated: `xhigh` (`highest`) applies
-only to gpt-5.2+ and `minimal` (`lowest`) only to gpt-5.4+ (its history is
-non-monotonic); older gpt-5 / o-series clamp the end to `high`/`low`.
+only to gpt-5.2+ and `minimal` (`lowest`) only to gpt-5.4 through gpt-5 (its
+history is non-monotonic and gpt-6 drops it again); older gpt-5 / o-series
+clamp the end to `high`/`low`. From gpt-6 onward `highest` reaches `max`, the
+rung above `xhigh`, and `lowest` clamps to `low`.
 
 Gating (effort is silently ignored where reasoning is unavailable, never
 erroring):
 
 - **OpenAI** — only reasoning models (`gpt-5+`, `o`-series); merges with the
   auto `reasoning.summary`. `xhigh` requires gpt-5.2+, `minimal` requires
-  gpt-5.4+ (else clamped).
+  gpt-5.4 through gpt-5, and gpt-6+ takes `max` for `highest` (else clamped).
 - **Anthropic** — only Claude 4.5+ / 5 models (`output_config.effort`); merges
   alongside a structured-output `format` config.
 - **Google** — `thinkingLevel` for Gemini 3.x, `thinkingBudget` for Gemini 2.5;
   other models get nothing.
 - **xAI** — only models whose name advertises reasoning (a `*-reasoning`
   suffix); a bare version name and `*-non-reasoning` are skipped.
+- **Meta** — every `muse-spark` id; merges with the auto `reasoning.summary`.
+  `max` is limited to the Standard-tier `muse-spark-1.3`, so `highest` on the
+  contributor id or an older release clamps to `xhigh` (logged at debug).
+  Muse Spark always reasons: `none` is rejected by the API and never emitted.
 - **OpenRouter** — always forwarded; OpenRouter maps to the routed model's
   nearest supported level.
 - **Fireworks** — always forwarded (`reasoning_effort`); the API accepts it on
@@ -925,7 +936,7 @@ elsewhere); `false`/`0` = disabled; `"5m"`/`"1h"` = that TTL.
 
 Per provider: Anthropic `cache_control` on system + last tool; Bedrock
 `cachePoint` blocks (model-gated, auto-denylisted and retried without on a 400;
-5m only); OpenAI/xAI automatic caching + a stable `prompt_cache_key`; OpenRouter
+5m only); OpenAI/xAI/Meta automatic caching + a stable `prompt_cache_key`; OpenRouter
 `cache_control` on the system message (forwarded to Anthropic/Gemini backends);
 Google implicit caching only (Gemini 2.5+). TTL applies to Anthropic/OpenRouter;
 others use provider defaults. Sub-threshold prefixes silently no-op. Cache
@@ -965,7 +976,7 @@ await Llm.operate(input, {
 });
 ```
 
-OpenAI and xAI leave the limit unset (their defaults do not truncate early).
+OpenAI, xAI, and Meta leave the limit unset (their defaults do not truncate early).
 Mistral is capped (32,768 streaming / 16,384 non-streaming) despite publishing
 no low ceiling: a Mistral model can degenerate into restating its answer when
 `format` and tools are combined, and an uncapped completion turns that into a