@chrischall/gemini-mcp 1.9.0 → 1.10.0

package/.claude-plugin/marketplace.json

@@ -7,7 +7,7 @@
   },
   "metadata": {
     "description": "MCP server for Google Gemini image generation and editing — generate and edit images via the Gemini API",
-    "version": "1.9.0"
+    "version": "1.10.0"
   },
   "plugins": [
     {
@@ -15,7 +15,7 @@
       "displayName": "Gemini Images",
       "source": "./",
       "description": "Generate and edit images with Google Gemini image models (Nano Banana / Nano Banana Pro) via natural language prompts",
-      "version": "1.9.0",
+      "version": "1.10.0",
       "author": {
         "name": "Chris Hall"
       },

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "gemini-mcp",
   "displayName": "Gemini Images",
-  "version": "1.9.0",
+  "version": "1.10.0",
   "description": "MCP server for Google Gemini image generation and editing — generate and edit images via the Gemini API",
   "author": {
     "name": "Chris Hall",

package/dist/bundle.js

@@ -32868,7 +32868,7 @@ async function loadStdioDotenv() {
 }
 
 // src/version.ts
-var VERSION = "1.9.0";
+var VERSION = "1.10.0";
 
 // src/errors.ts
 init_dist();
@@ -33813,10 +33813,13 @@ function timeoutRiskHint(opts) {
   return lead + `If the host gives up mid-generation the image is still written to disk (and gemini_interact leaves an <image>.json sidecar with the interaction id) \u2014 no rerun needed, just look in the output dir. For a faster in-band result use a Flash model, drop image_size to 1K/2K, or \u2014 in Claude Code \u2014 set MCP_TOOL_TIMEOUT.`;
 }
 function buildMeta(model, seed, opts) {
-  const meta3 = { model, seed };
+  const meta3 = reportShape({ model, seed }, opts);
+  if (opts.image_size) meta3.image_size = opts.image_size;
+  return meta3;
+}
+function reportShape(meta3, opts) {
   if (opts.aspect_ratio) meta3.aspect_ratio = opts.aspect_ratio;
   if (opts.orientation) meta3.orientation = opts.orientation;
-  if (opts.image_size) meta3.image_size = opts.image_size;
   return meta3;
 }
 async function emitMedia(named, kind, opts, meta3, onWritten) {
@@ -36194,7 +36197,7 @@ function registerInteractTools(server, client2) {
           }
         });
         client2.session.lastInteractionId = r.id;
-        const meta3 = { model, interaction_id: r.id };
+        const meta3 = reportShape({ model, interaction_id: r.id }, { ...args, aspect_ratio: aspectRatio });
         if (chainRecovered) meta3.chain_recovered = chainRecovered;
         else if (previousInteractionId) meta3.previous_interaction_id = previousInteractionId;
         if (continuedFromSidecar) meta3.continued_from_sidecar = true;
@@ -36306,7 +36309,7 @@ function registerVideoTools(server, client2) {
           onInteractionStarted: ctx.reportInteraction
         }));
         client2.session.lastVideoInteractionId = r.id;
-        const meta3 = { model, interaction_id: r.id };
+        const meta3 = reportShape({ model, interaction_id: r.id }, { ...args, aspect_ratio: aspectRatio });
         if (previousInteractionId) meta3.previous_interaction_id = previousInteractionId;
         if (r.text) meta3.text = r.text;
         if (report) meta3.image_inputs = report;

package/dist/tools/interact.js

@@ -5,7 +5,7 @@ import { ChainedRequest404Error } from '../client.js';
 import { slugify, baseName, resolveImagePath, writeSidecar, resolveOutputDir, readImageAsInline } from '../images.js';
 import { resolveImageInputs } from '../inputs.js';
 import { findInteractionImages, latestInteractionId } from '../sidecar.js';
-import { emit, resolveAspectRatio, orientationSchema, ASPECT_RATIOS, IMAGE_SIZES, MODEL_CHOICE_GUIDE, resolveVideoInput, videoPathSchema, timeoutMsSchema, timeoutRiskHint, idempotencyKeySchema, asyncSchema, maxWaitMsSchema, withProgressHeartbeat, assertLocalInputsAvailable, imagesUrlSchema, imagesFileUrisSchema } from './shared.js';
+import { emit, reportShape, resolveAspectRatio, orientationSchema, ASPECT_RATIOS, IMAGE_SIZES, MODEL_CHOICE_GUIDE, resolveVideoInput, videoPathSchema, timeoutMsSchema, timeoutRiskHint, idempotencyKeySchema, asyncSchema, maxWaitMsSchema, withProgressHeartbeat, assertLocalInputsAvailable, imagesUrlSchema, imagesFileUrisSchema } from './shared.js';
 import { fingerprintRequest } from '../jobs.js';
 import { previewLocalInputsUnlessConfirmed, schemaConfirm } from './_confirm.js';
 // The last-interaction id and the written-outputs set used to live here, at
@@ -259,7 +259,9 @@ export function registerInteractTools(server, client) {
                 }
             });
             client.session.lastInteractionId = r.id;
-            const meta = { model, interaction_id: r.id };
+            // `aspectRatio` is the RESOLVED value (see resolveAspectRatio above), so
+            // an `orientation: 'portrait'` request can be seen to have become 9:16.
+            const meta = reportShape({ model, interaction_id: r.id }, { ...args, aspect_ratio: aspectRatio });
             // On recovery the prior id was NOT used — reporting it as
             // previous_interaction_id would misrepresent what the model saw.
             if (chainRecovered)

package/dist/tools/shared.js

@@ -446,15 +446,29 @@ export function timeoutRiskHint(opts) {
 }
 /** Build the result metadata echoed to the caller (omitting unset optionals). */
 export function buildMeta(model, seed, opts) {
-    const meta = { model, seed };
-    // The RESOLVED ratio, and the orientation only when one was asked for — so a
-    // caller who said "portrait" can see both what they asked and what it became.
+    const meta = reportShape({ model, seed }, opts);
+    if (opts.image_size)
+        meta.image_size = opts.image_size;
+    return meta;
+}
+/**
+ * Record the shape a generation produced, into any meta object.
+ *
+ * The RESOLVED ratio, and the `orientation` only when one was actually asked
+ * for — so a caller who said "portrait" sees both what they asked and what it
+ * became, while a bare `21:9` is not handed back a word nobody used.
+ *
+ * It lives apart from `buildMeta` because two tools cannot use that function:
+ * `gemini_interact` and `gemini_video_generate` build their metas around
+ * `interaction_id` and chain state rather than a seed, and so reported no shape
+ * at all until #166. Sharing the rule rather than copying it is what keeps
+ * those hand-rolled metas from drifting again.
+ */
+export function reportShape(meta, opts) {
     if (opts.aspect_ratio)
         meta.aspect_ratio = opts.aspect_ratio;
     if (opts.orientation)
         meta.orientation = opts.orientation;
-    if (opts.image_size)
-        meta.image_size = opts.image_size;
     return meta;
 }
 /**

package/dist/tools/video.js

@@ -3,7 +3,7 @@ import { McpToolError } from '@chrischall/mcp-utils';
 import { slugify, baseName } from '../images.js';
 import { resolveImageInputs } from '../inputs.js';
 import { DEFAULT_VIDEO_MODEL } from '../models.js';
-import { emitMedia, resolveAspectRatio, orientationSchema, timeoutMsSchema, idempotencyKeySchema, asyncSchema, maxWaitMsSchema, withProgressHeartbeat, assertLocalInputsAvailable, imagesUrlSchema, imagesFileUrisSchema } from './shared.js';
+import { emitMedia, reportShape, resolveAspectRatio, orientationSchema, timeoutMsSchema, idempotencyKeySchema, asyncSchema, maxWaitMsSchema, withProgressHeartbeat, assertLocalInputsAvailable, imagesUrlSchema, imagesFileUrisSchema } from './shared.js';
 import { fingerprintRequest } from '../jobs.js';
 import { previewLocalInputsUnlessConfirmed, schemaConfirm } from './_confirm.js';
 /** omni's own aspect-ratio enum — NOT the image tools' ASPECT_RATIOS. */
@@ -100,7 +100,9 @@ export function registerVideoTools(server, client) {
                 onInteractionStarted: ctx.reportInteraction,
             }));
             client.session.lastVideoInteractionId = r.id;
-            const meta = { model, interaction_id: r.id };
+            // The RESOLVED ratio: omni takes only 16:9/9:16, so a caller who asked
+            // in words can see which of the two their request became.
+            const meta = reportShape({ model, interaction_id: r.id }, { ...args, aspect_ratio: aspectRatio });
             if (previousInteractionId)
                 meta.previous_interaction_id = previousInteractionId;
             if (r.text)

package/dist/version.js

@@ -1,2 +1,2 @@
 /** Single source of the server version. release-please bumps the literal below. */
-export const VERSION = '1.9.0'; // x-release-please-version
+export const VERSION = '1.10.0'; // x-release-please-version

package/mint.yaml

@@ -0,0 +1,57 @@
+version: 1
+name: Gemini Images
+slug: gemini
+summary: >-
+  Generate and edit images with Google Gemini image models (Nano Banana /
+  Nano Banana Pro) via natural language prompts
+env:
+  - name: GEMINI_API_KEY
+    secret: true
+    required: true
+    help: >-
+      Your Google Gemini API key (aistudio.google.com/apikey)
+  - name: GEMINI_IMAGE_MODEL
+    required: false
+    help: >-
+      Override the default image model (default: gemini-3.1-flash-image)
+  - name: GEMINI_OUTPUT_DIR
+    required: false
+    help: >-
+      Default directory for generated images (default: current working
+      directory)
+  - name: GEMINI_INPUT_DIR
+    required: false
+    help: >-
+      Directory to resolve bare input-image filenames against
+  - name: GEMINI_TIMEOUT_MS
+    required: false
+    help: >-
+      Upstream request timeout in ms (default: 60000, or 120000 for 4K output)
+  - name: GEMINI_CHAIN_RETRY_MS
+    required: false
+    help: >-
+      Retry budget in milliseconds for a chained request that 404s while the
+      upstream result settles (default 120000).
+  - name: GEMINI_HEARTBEAT_MS
+    required: false
+    help: >-
+      How often, in milliseconds, to emit a progress notification while a
+      generation is in flight. Hosts reset their tools/call timeout on progress,
+      so this lets a long 4K/Pro generation outlive the default budget.
+egress:
+  allow:
+    # Only hosts the SERVER process actually fetches. Hosts that appear
+    # solely in a URL this server BUILDS and returns are excluded.
+    - generativelanguage.googleapis.com
+    #
+    # NOTE — this server also reaches at least one host determined at
+    # RUNTIME, which cannot appear as a literal here:
+    #   - a signed upload URL returned by the API (often storage.googleapis.com / s3)
+    #   - a media/CDN host supplied at runtime
+    # Confirm the concrete host(s) against a live run and add them, or
+    # the isolated tier will block the call.
+    #
+    # Also: the images_url argument makes this server fetch an ARBITRARY
+    # caller-supplied https host. That cannot be enumerated in advance, so on
+    # the isolated tier either allow the hosts you intend to pass, or expect
+    # image-by-URL to be blocked there.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chrischall/gemini-mcp",
-  "version": "1.9.0",
+  "version": "1.10.0",
   "mcpName": "io.github.chrischall/gemini-mcp",
   "description": "Gemini image-generation MCP server for Claude — developed and maintained by AI (Claude Code)",
   "author": "Claude Code (AI) <https://www.anthropic.com/claude>",
@@ -33,7 +33,8 @@
     ".claude-plugin",
     "skills/",
     ".mcp.json",
-    "server.json"
+    "server.json",
+    "mint.yaml"
   ],
   "scripts": {
     "build": "tsc && npm run bundle",

package/server.json

@@ -6,12 +6,12 @@
     "url": "https://github.com/chrischall/gemini-mcp",
     "source": "github"
   },
-  "version": "1.9.0",
+  "version": "1.10.0",
   "packages": [
     {
       "registryType": "npm",
       "identifier": "@chrischall/gemini-mcp",
-      "version": "1.9.0",
+      "version": "1.10.0",
       "transport": {
         "type": "stdio"
       },