@jaypie/mcp 0.8.181 → 0.8.183

package/dist/assets.js

@@ -55,11 +55,11 @@ function getMcpAssetPaths() {
  */
 function getMcpBuildInfo() {
     return {
-        commit: "9b837db4"
+        commit: "77ae81b6"
             ,
-        version: "0.8.181"
+        version: "0.8.183"
             ,
-        versionString: "@jaypie/mcp@0.8.181#9b837db4"
+        versionString: "@jaypie/mcp@0.8.183#77ae81b6"
             ,
     };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jaypie/mcp",
-  "version": "0.8.181",
+  "version": "0.8.183",
   "description": "Jaypie MCP",
   "repository": {
     "type": "git",

package/release-notes/jaypie/1.2.117.md

@@ -0,0 +1,9 @@
+---
+version: 1.2.117
+date: 2026-10-05
+summary: Pulls @jaypie/llm 1.3.58
+---
+
+## Changes
+
+- **`@jaypie/llm` 1.3.58**: a `format` request answered with no content fails over instead of resolving with `content: undefined`, and throws `LlmIncompleteError` when every model comes back empty; Gemini block reasons and `OTHER` finishes count as incomplete.

package/release-notes/jaypie/1.2.118.md

@@ -0,0 +1,9 @@
+---
+version: 1.2.118
+date: 2026-10-05
+summary: Pulls @jaypie/llm 1.3.59
+---
+
+## Changes
+
+- **`@jaypie/llm` 1.3.59**: the structured-output corrective turn is taken once per run; an empty Gemini reply makes 2 requests before failing over instead of 24.

package/release-notes/llm/1.3.58.md

@@ -0,0 +1,28 @@
+---
+version: 1.3.58
+date: 2026-10-05
+summary: A format request answered with no content fails over instead of resolving with undefined content, and throws LlmIncompleteError when every model comes back empty
+---
+
+## Changes
+
+### Empty structured output (issue #608)
+
+`operate` with `format` resolved successfully when the provider returned no content (observed: Gemini, 0 output tokens, `fallbackUsed: false`, no warning). The caller received `content: undefined`.
+
+- Empty content (`undefined`, `null`, or whitespace) on a `format` request settles `status: "incomplete"` with `error.reason: "incomplete"`, `error.detail` `Model returned no content: <stop reason>`, and `stopReason: "other"`. It logs at `warn` with the provider's stop reason.
+- Adapters with a corrective turn (Fireworks, Google) try it once before settling empty. Salvage and fresh-context conversion are skipped: there is no text to read.
+- A fallback chain moves past an empty response like any cut-off. An earlier partial is kept over a later empty response.
+- When only empty responses come back, `operate` throws `LlmIncompleteError` (status 502) whatever `incomplete` says: there is no partial to return.
+- Requests without `format` are unchanged.
+
+### Gemini stop reasons
+
+- A blocked prompt returns no candidate; `promptFeedback.blockReason` is now reported as the stop reason, so it fails over and names its cause.
+- `OTHER`, `LANGUAGE`, `IMAGE_SAFETY`, and `MALFORMED_FUNCTION_CALL` are incomplete finishes. `IMAGE_SAFETY` normalizes to `content_filter`.
+
+## Testing
+
+- `OperateLoop` specs cover empty and whitespace content with `format`, the corrective turn on empty content, settling empty once turns run out, no conversion of empty content, and unchanged behavior without `format`.
+- `Llm` class specs cover failover on empty content, throwing for a lone model and a whole chain, and a partial kept over a later empty response.
+- `runWithFallback`, `GoogleAdapter`, and `incompleteReason` specs cover `isEmptyResult`, the block reason, and the new Gemini finishes.

package/release-notes/llm/1.3.59.md

@@ -0,0 +1,18 @@
+---
+version: 1.3.59
+date: 2026-10-05
+summary: The structured-output corrective turn is taken once per run instead of repeating until the turn budget runs out
+---
+
+## Changes
+
+### One corrective turn (issue #610)
+
+A `format` request answered with prose or with nothing took the corrective turn (only the `structured_output` tool offered) on every turn until `turns` ran out. With the default budget of 24, an empty Gemini reply cost 24 requests, each resending the full prompt, before the chain failed over.
+
+- The corrective turn is taken at most once per run. A second miss settles: empty content fails over as in 1.3.58, and prose completes as it did when turns ran out.
+- Applies to every adapter with a corrective turn: Fireworks, Google, and Mistral.
+
+## Testing
+
+- `OperateLoop` specs cover one corrective turn with the default budget for empty content and for prose.

package/release-notes/mcp/0.8.182.md

@@ -0,0 +1,10 @@
+---
+version: 0.8.182
+date: 2026-10-05
+summary: Release notes for @jaypie/llm 1.3.58 and jaypie 1.2.117; llm skill documents empty format responses
+---
+
+## Changes
+
+- **Release notes**: `@jaypie/llm` 1.3.58 and `jaypie` 1.2.117
+- **`llm` skill**: documents how an empty `format` response fails over and throws

package/release-notes/mcp/0.8.183.md

@@ -0,0 +1,10 @@
+---
+version: 0.8.183
+date: 2026-10-05
+summary: Release notes for @jaypie/llm 1.3.59 and jaypie 1.2.118; llm skill documents the single corrective turn
+---
+
+## Changes
+
+- **Release notes**: `@jaypie/llm` 1.3.59 and `jaypie` 1.2.118
+- **`llm` skill**: the corrective turn is taken once per run

package/skills/llm.md

@@ -384,7 +384,7 @@ string. The operate loop first parses the text as JSON, stripping a Markdown
 code fence — this runs for every provider. If that fails, adapters that opt in
 (`supportsStructuredOutputRetry`: **Fireworks** and **Google**) take a
 corrective turn offering only the `structured_output` tool and demanding it be
-called, bounded by the `turns` budget. That turn withholds the caller's own
+called. The corrective turn is taken once per run, within the `turns` budget. That turn withholds the caller's own
 tools and does not send the schema natively, leaving one unambiguous way to
 answer.
 
@@ -857,6 +857,8 @@ try {
 
 An incomplete `content` is the raw partial text, a string even when `format` was requested: a truncated answer is not the object the format promises. Check `status` (or use `incomplete: "throw"`) before reading `content` as the formatted object.
 
+A `format` request answered with **no content** (zero output tokens, an empty candidate, a blocked prompt) settles incomplete with `error.detail` `Model returned no content: <stop reason>`, logged at `warn`, and `stopReason: "other"`. An adapter that supports the corrective turn tries it once first. A chain moves past an empty response like any cut-off, and an earlier partial is kept over a later empty one. When nothing but empty responses come back, `operate` throws `LlmIncompleteError` whatever `incomplete` says: there is no partial to return (issue #608).
+
 ### Provider Options in a Chain
 
 `providerOptions` belongs to the model it was written for. Per-call `providerOptions` reach the primary only (including its linger pass); a fallback receives only the `providerOptions` on its own entry, or none. Options shaped for one provider (OpenAI `reasoning`, Gemini `thinkingConfig`) would otherwise break every other provider in the chain. Prefer first-class `effort` and `temperature`, which translate per provider.
@@ -1068,7 +1070,7 @@ exactly like a provider rate limit.
 | ------------- | ------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `max_turns`   | 429    | The model asked for another tool call after `turns` ran out. Nothing failed; the run did not converge.                                                                       |
 | `tool_errors` | 502    | Tool execution failed six times in a row and the loop stopped.                                                                                                               |
-| `incomplete`  | 502    | The provider cut the model off before it finished (an output token ceiling, a content filter). `content` holds the partial text; `error.detail` names the provider's reason and `stopReason` normalizes it. A fallback chain moves past it. |
+| `incomplete`  | 502    | The provider cut the model off before it finished (an output token ceiling, a content filter). `content` holds the partial text; `error.detail` names the provider's reason and `stopReason` normalizes it. A fallback chain moves past it. With `format`, empty content settles here too and throws once the chain is spent. |
 
 ```typescript
 import { LlmResponseErrorReason } from "@jaypie/llm";