@chrischall/gemini-mcp 0.7.0 → 0.7.2

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": "0.7.0"
+    "version": "0.7.2"
   },
   "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": "0.7.0",
+      "version": "0.7.2",
       "author": {
         "name": "Chris Hall"
       },

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "gemini-mcp",
   "displayName": "Gemini Images",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "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

@@ -32510,7 +32510,7 @@ var init_clipboard = __esm({
 init_dist();
 
 // src/version.ts
-var VERSION = "0.7.0";
+var VERSION = "0.7.2";
 
 // src/tools/models.ts
 init_dist();
@@ -32544,6 +32544,8 @@ var UPLOAD_BASE_URL = "https://generativelanguage.googleapis.com/upload/v1beta";
 var FILE_MAX_BYTES = 2 * 1024 ** 3;
 var FILE_POLL_INTERVAL_MS = 2e3;
 var FILE_POLL_MAX_ATTEMPTS = 150;
+var CHAIN_404_RETRIES = 2;
+var CHAIN_404_RETRY_MS = 2e3;
 var GeminiClient = class {
   apiKey;
   configError;
@@ -32756,9 +32758,28 @@ var GeminiClient = class {
       if (opts.searchTypes?.length) searchTool.search_types = opts.searchTypes;
       body.tools = [searchTool];
     }
-    const data = await this.interactionsApi.fetchJson("POST", "/interactions", {
-      body
-    });
+    let data;
+    for (let attempt = 0; ; attempt++) {
+      try {
+        data = await this.interactionsApi.fetchJson("POST", "/interactions", {
+          body
+        });
+        break;
+      } catch (err) {
+        if (!(opts.previousInteractionId && err instanceof ApiError && err.status === 404)) throw err;
+        if (attempt < CHAIN_404_RETRIES) {
+          await this.sleep(CHAIN_404_RETRY_MS * (attempt + 1));
+          continue;
+        }
+        throw new McpToolError(
+          `Previous interaction "${opts.previousInteractionId}" was not found (retried ${CHAIN_404_RETRIES}\xD7 \u2014 the interactions store can lag briefly after a turn completes). If the id is old: interactions are retained 55 days on the paid tier (1 day on the free tier) and are scoped to the API key that created them.`,
+          {
+            hint: "Retry once more in a few seconds; if it keeps failing, start a new chain: call gemini_interact again with the last output image passed via `images` (or `images_base64`) plus the same instruction, then chain the NEW interaction id.",
+            cause: err
+          }
+        );
+      }
+    }
     const images = [];
     const textParts = [];
     const searchQueries = [];
@@ -32871,7 +32892,7 @@ function resolveInputPath(p, kind) {
   const inputDir = readEnvVar("GEMINI_INPUT_DIR");
   if (inputDir) {
     const candidate = join3(inputDir, p);
-    if (existsSync(candidate)) return candidate;
+    if (existsSync(candidate)) return resolve(candidate);
   }
   const cwd = resolve(p);
   if (existsSync(cwd)) return cwd;
@@ -32990,7 +33011,7 @@ function buildMeta(model, seed, opts) {
   if (opts.image_size) meta3.image_size = opts.image_size;
   return meta3;
 }
-async function emit(named, opts, meta3) {
+async function emit(named, opts, meta3, onWritten) {
   if (opts.inline) {
     const imageBlocks = named.map((n) => ({ type: "image", data: n.image.base64, mimeType: n.image.mimeType }));
     if (meta3 && Object.keys(meta3).length > 0) {
@@ -33003,6 +33024,7 @@ async function emit(named, opts, meta3) {
   const dir = resolveOutputDir(opts.output_dir);
   const images = [];
   for (const n of named) images.push(await writeImage(dir, n.base, n.image.base64, n.image.mimeType));
+  onWritten?.(images);
   return textResult({ images, ...meta3 });
 }
 
@@ -33168,6 +33190,21 @@ function registerSetTools(server) {
 init_zod();
 init_dist();
 var lastInteractionId;
+var writtenOutputs = /* @__PURE__ */ new Set();
+var NEW_REFERENCES_ONLY = "NEW reference images only (e.g. a style or target photo). When chaining with previous_interaction_id, do NOT re-attach the prior turn's output \u2014 the interaction already contains it, and re-sending it anchors the model against your edit.";
+function splitReattachedOutputs(images) {
+  const kept = [];
+  const dropped = [];
+  for (const p of images) {
+    let resolved;
+    try {
+      resolved = resolveImagePath(p);
+    } catch {
+    }
+    (resolved && writtenOutputs.has(resolved) ? dropped : kept).push(p);
+  }
+  return { kept, dropped };
+}
 function registerInteractTools(server) {
   server.registerTool(
     "gemini_interact",
@@ -33178,8 +33215,8 @@ function registerInteractTools(server) {
         input: external_exports.string().min(1).describe("Text prompt or editing instruction"),
         previous_interaction_id: external_exports.string().optional().describe("ID from a prior gemini_interact call \u2014 continues that multi-turn conversation"),
         continue_last: external_exports.boolean().optional().describe("Continue from the most recent interaction this server created (convenience for previous_interaction_id; an explicit id wins)"),
-        images: external_exports.array(external_exports.string().min(1)).optional().describe("Paths to reference input images"),
-        images_base64: external_exports.array(external_exports.string().min(1)).optional().describe("Reference images as base64 strings or data URIs"),
+        images: external_exports.array(external_exports.string().min(1)).optional().describe(`Paths to reference input images. ${NEW_REFERENCES_ONLY}`),
+        images_base64: external_exports.array(external_exports.string().min(1)).optional().describe(`Reference images as base64 strings or data URIs. ${NEW_REFERENCES_ONLY}`),
         model: external_exports.string().optional().describe(`Model id override (default: server default; see gemini_list_models). ${MODEL_CHOICE_GUIDE}`),
         aspect_ratio: external_exports.enum(ASPECT_RATIOS).optional().describe("Output aspect ratio"),
         image_size: external_exports.enum(IMAGE_SIZES).optional().describe("Output resolution (512 = 0.5K, Flash-only)"),
@@ -33206,7 +33243,12 @@ function registerInteractTools(server) {
         }
         previousInteractionId = lastInteractionId;
       }
-      const inputs = await gatherImageInputs({ images: args.images, images_base64: args.images_base64, from_clipboard: args.from_clipboard });
+      let images = args.images;
+      let dropped = [];
+      if (previousInteractionId && images?.length) {
+        ({ kept: images, dropped } = splitReattachedOutputs(images));
+      }
+      const inputs = await gatherImageInputs({ images, images_base64: args.images_base64, from_clipboard: args.from_clipboard });
       const video = await resolveVideoInput(args);
       const r = await client.interact({
         input: args.input,
@@ -33225,7 +33267,8 @@ function registerInteractTools(server) {
       const model = resolveModel(args.model, readEnvVar("GEMINI_IMAGE_MODEL"));
       const meta3 = { model, interaction_id: r.id };
       if (previousInteractionId) meta3.previous_interaction_id = previousInteractionId;
-      meta3.hint = `To refine this image, call gemini_interact again with previous_interaction_id: "${r.id}" (or continue_last: true).`;
+      if (dropped.length) meta3.dropped_previous_output = dropped;
+      meta3.hint = `To refine this image, call gemini_interact again with previous_interaction_id: "${r.id}" (or continue_last: true) \u2014 do NOT re-attach the returned image as an input; the interaction already contains it.`;
       if (r.text) meta3.text = r.text;
       if (r.grounding) meta3.grounding = r.grounding;
       if (video.videoFileMeta) meta3.video_file = video.videoFileMeta;
@@ -33234,7 +33277,9 @@ function registerInteractTools(server) {
         image,
         base: r.images.length > 1 ? `${slug}-${String(i + 1).padStart(2, "0")}` : slug
       }));
-      return emit(named, args, meta3);
+      return emit(named, args, meta3, (paths) => {
+        for (const p of paths) writtenOutputs.add(p);
+      });
     }
   );
 }

package/dist/client.js

@@ -1,6 +1,6 @@
 import { basename, dirname, join } from 'node:path';
 import { fileURLToPath } from 'node:url';
-import { loadDotenvSafely, readEnvVar, McpToolError, createApiClient, formatApiError, fileBlob } from '@chrischall/mcp-utils';
+import { loadDotenvSafely, readEnvVar, McpToolError, ApiError, createApiClient, formatApiError, fileBlob } from '@chrischall/mcp-utils';
 import { resolveModel, filterImageModels } from './models.js';
 const __dirname = dirname(fileURLToPath(import.meta.url));
 await loadDotenvSafely({ path: join(__dirname, '..', '.env'), override: false });
@@ -21,6 +21,13 @@ const FILE_MAX_BYTES = 2 * 1024 ** 3;
 // in ~1s (verified); longer videos take proportionally longer to process.
 const FILE_POLL_INTERVAL_MS = 2_000;
 const FILE_POLL_MAX_ATTEMPTS = 150;
+// The interactions store is eventually consistent: a freshly returned id can
+// 404 for a short window right after its turn completes (observed live
+// 2026-07-06 — an id that 404'd on an immediate chain resolved fine minutes
+// later). A chained 404 generated nothing (no cost), so retry before giving up:
+// attempts wait 2s then 4s (~6s total).
+const CHAIN_404_RETRIES = 2;
+const CHAIN_404_RETRY_MS = 2_000;
 export class GeminiClient {
     apiKey;
     configError;
@@ -256,9 +263,33 @@ export class GeminiClient {
                 searchTool.search_types = opts.searchTypes;
             body.tools = [searchTool];
         }
-        const data = await this.interactionsApi.fetchJson('POST', '/interactions', {
-            body,
-        });
+        let data;
+        for (let attempt = 0;; attempt++) {
+            try {
+                data = await this.interactionsApi.fetchJson('POST', '/interactions', {
+                    body,
+                });
+                break;
+            }
+            catch (err) {
+                // A 404 on a chained call means the referenced interaction wasn't
+                // found. Two causes: (a) store lag right after the id was created —
+                // retry with backoff; (b) genuinely gone — interactions are retained
+                // 55 days (paid tier) / 1 day (free tier) and are scoped to the API
+                // key's project. The chain is recoverable from the last output file,
+                // so the exhausted-retries error says how.
+                if (!(opts.previousInteractionId && err instanceof ApiError && err.status === 404))
+                    throw err;
+                if (attempt < CHAIN_404_RETRIES) {
+                    await this.sleep(CHAIN_404_RETRY_MS * (attempt + 1));
+                    continue;
+                }
+                throw new McpToolError(`Previous interaction "${opts.previousInteractionId}" was not found (retried ${CHAIN_404_RETRIES}× — the interactions store can lag briefly after a turn completes). If the id is old: interactions are retained 55 days on the paid tier (1 day on the free tier) and are scoped to the API key that created them.`, {
+                    hint: 'Retry once more in a few seconds; if it keeps failing, start a new chain: call gemini_interact again with the last output image passed via `images` (or `images_base64`) plus the same instruction, then chain the NEW interaction id.',
+                    cause: err,
+                });
+            }
+        }
         // Only surface the `model_output` step — that's the caller-facing result.
         // `thought` steps hold internal reasoning (and, with includeThoughts, draft
         // "thinking" images); collecting those would leak reasoning into `text` and

package/dist/images.js

@@ -62,7 +62,9 @@ export async function writeImage(dir, base, base64, mimeType) {
  *  1. Absolute path that exists → returned as-is.
  *  2. `GEMINI_INPUT_DIR` env var is set → look for `join(inputDir, p)`.
  *  3. Relative to cwd → `resolve(p)`.
- * Throws a helpful `McpToolError` (labelled `kind`) if none are found.
+ * Always returns an absolute path (callers compare resolved paths — e.g. the
+ * interact re-attach guard — so a relative GEMINI_INPUT_DIR must not leak
+ * through). Throws a helpful `McpToolError` (labelled `kind`) if none are found.
  */
 function resolveInputPath(p, kind) {
     if (isAbsolute(p) && existsSync(p))
@@ -73,7 +75,7 @@ function resolveInputPath(p, kind) {
         // single-user MCP — the caller already has the user's own filesystem access.
         const candidate = join(inputDir, p);
         if (existsSync(candidate))
-            return candidate;
+            return resolve(candidate);
     }
     const cwd = resolve(p);
     if (existsSync(cwd))

package/dist/tools/interact.js

@@ -2,12 +2,40 @@ import { z } from 'zod';
 import { McpToolError, readEnvVar } from '@chrischall/mcp-utils';
 import { resolveModel } from '../models.js';
 import { client } from '../client.js';
-import { slugify, baseName, gatherImageInputs } from '../images.js';
+import { slugify, baseName, gatherImageInputs, resolveImagePath } from '../images.js';
 import { emit, ASPECT_RATIOS, IMAGE_SIZES, MODEL_CHOICE_GUIDE, resolveVideoInput, videoPathSchema } from './shared.js';
 // The most recent interaction id this server process created — what
 // `continue_last: true` resumes. In-memory only: an MCP server lives for the
 // host session, so "last" means "last in this session".
 let lastInteractionId;
+// Absolute paths of every image this tool has written this session. Used to
+// drop a chained call's re-attached prior output: the interaction state
+// already contains that image, and re-sending it as a fresh reference anchors
+// the model against the requested edit (observed in the wild — callers pass
+// the last result path back via `images` despite the description saying not to).
+const writtenOutputs = new Set();
+/** Shared warning for the reference-image params. */
+const NEW_REFERENCES_ONLY = 'NEW reference images only (e.g. a style or target photo). When chaining with previous_interaction_id, ' +
+    "do NOT re-attach the prior turn's output — the interaction already contains it, and re-sending it " +
+    'anchors the model against your edit.';
+/**
+ * On a chained call, split `images` into genuinely new references vs paths
+ * this server itself generated (which the interaction already contains).
+ * Unresolvable paths are kept so `gatherImageInputs` surfaces its usual error.
+ */
+function splitReattachedOutputs(images) {
+    const kept = [];
+    const dropped = [];
+    for (const p of images) {
+        let resolved;
+        try {
+            resolved = resolveImagePath(p);
+        }
+        catch { /* kept; loading will error actionably */ }
+        (resolved && writtenOutputs.has(resolved) ? dropped : kept).push(p);
+    }
+    return { kept, dropped };
+}
 export function registerInteractTools(server) {
     server.registerTool('gemini_interact', {
         description: 'Preferred tool for iterative or multi-step refinement of a single image — ' +
@@ -30,11 +58,11 @@ export function registerInteractTools(server) {
             images: z
                 .array(z.string().min(1))
                 .optional()
-                .describe('Paths to reference input images'),
+                .describe(`Paths to reference input images. ${NEW_REFERENCES_ONLY}`),
             images_base64: z
                 .array(z.string().min(1))
                 .optional()
-                .describe('Reference images as base64 strings or data URIs'),
+                .describe(`Reference images as base64 strings or data URIs. ${NEW_REFERENCES_ONLY}`),
             model: z
                 .string()
                 .optional()
@@ -90,7 +118,12 @@ export function registerInteractTools(server) {
             }
             previousInteractionId = lastInteractionId;
         }
-        const inputs = await gatherImageInputs({ images: args.images, images_base64: args.images_base64, from_clipboard: args.from_clipboard });
+        let images = args.images;
+        let dropped = [];
+        if (previousInteractionId && images?.length) {
+            ({ kept: images, dropped } = splitReattachedOutputs(images));
+        }
+        const inputs = await gatherImageInputs({ images, images_base64: args.images_base64, from_clipboard: args.from_clipboard });
         const video = await resolveVideoInput(args);
         const r = await client.interact({
             input: args.input,
@@ -110,7 +143,11 @@ export function registerInteractTools(server) {
         const meta = { model, interaction_id: r.id };
         if (previousInteractionId)
             meta.previous_interaction_id = previousInteractionId;
-        meta.hint = `To refine this image, call gemini_interact again with previous_interaction_id: "${r.id}" (or continue_last: true).`;
+        if (dropped.length)
+            meta.dropped_previous_output = dropped;
+        meta.hint =
+            `To refine this image, call gemini_interact again with previous_interaction_id: "${r.id}" (or continue_last: true) — ` +
+                'do NOT re-attach the returned image as an input; the interaction already contains it.';
         if (r.text)
             meta.text = r.text;
         if (r.grounding)
@@ -122,6 +159,8 @@ export function registerInteractTools(server) {
             image,
             base: r.images.length > 1 ? `${slug}-${String(i + 1).padStart(2, '0')}` : slug,
         }));
-        return emit(named, args, meta);
+        // writeImage returns already-resolved absolute paths.
+        return emit(named, args, meta, (paths) => { for (const p of paths)
+            writtenOutputs.add(p); });
     });
 }

package/dist/tools/shared.js

@@ -85,8 +85,10 @@ export function buildMeta(model, seed, opts) {
  * return their paths as a text result. `inline` wins when true.
  * Optional `meta` is merged into the disk-path JSON result, or prepended as a
  * text block in inline mode (only if meta has keys).
+ * Optional `onWritten` receives the absolute paths written in disk mode (never
+ * called inline) — lets a tool remember its own outputs.
  */
-export async function emit(named, opts, meta) {
+export async function emit(named, opts, meta, onWritten) {
     if (opts.inline) {
         const imageBlocks = named.map((n) => ({ type: 'image', data: n.image.base64, mimeType: n.image.mimeType }));
         if (meta && Object.keys(meta).length > 0) {
@@ -100,5 +102,6 @@ export async function emit(named, opts, meta) {
     const images = [];
     for (const n of named)
         images.push(await writeImage(dir, n.base, n.image.base64, n.image.mimeType));
+    onWritten?.(images);
     return textResult({ images, ...meta });
 }

package/dist/version.js

@@ -1,2 +1,2 @@
 /** Single source of the server version. release-please bumps the literal below. */
-export const VERSION = '0.7.0'; // x-release-please-version
+export const VERSION = '0.7.2'; // x-release-please-version

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chrischall/gemini-mcp",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "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>",

package/server.json

@@ -6,12 +6,12 @@
     "url": "https://github.com/chrischall/gemini-mcp",
     "source": "github"
   },
-  "version": "0.7.0",
+  "version": "0.7.2",
   "packages": [
     {
       "registryType": "npm",
       "identifier": "@chrischall/gemini-mcp",
-      "version": "0.7.0",
+      "version": "0.7.2",
       "transport": {
         "type": "stdio"
       },