feat: store async generated images as WebP

.env.example

@@ -1,5 +1,8 @@
 ADMIN_SECRET=replace-with-a-long-random-secret
 DB_PATH=./data/proxy.db
+OUTPUT_DIR=./data/generated
+OUTPUT_WEBP_QUALITY=90
+OUTPUT_RETENTION_SECONDS=86400
 MAX_UPLOAD_BYTES=52428800
 MAX_IMAGES=8
 MAX_IMAGE_PIXELS=25000000

.gitignore

@@ -5,6 +5,7 @@ __pycache__/
 *.db
 *.db-shm
 *.db-wal
+data/generated/
 .pytest_cache/
 .ruff_cache/
 dist/

PDR.md

@@ -57,7 +57,7 @@ Success is `200` with `provider_used`, `model_used`, `images`, `text`, elapsed t
 
 ### Poll
 
-`GET /v1/jobs/{job_id}?client_id={client_id}` plus the client's secret header. Returns status and, when complete, result; terminal failures include a sanitized error. Unknown and cross-client job IDs return `404`.
+`GET /v1/jobs/{job_id}?client_id={client_id}&include_result=false` plus the client's secret header. Returns small status metadata and, when complete, result metadata with binary image download links. The links are signed, scoped to one image, and expire after one hour by default, so they can be opened directly in a browser without sending a client secret. Treat them as temporary bearer credentials. Authenticated clients may also use `GET /v1/jobs/{job_id}/images/{image_index}?client_id={client_id}` with the secret header. `include_result=true` retains the legacy Base64 response. Terminal failures include a sanitized error. Unknown and cross-client job IDs return `404`.
 
 ### Management
 
@@ -82,12 +82,12 @@ SQLite connections enable foreign keys and busy timeout; WAL is initialized once
 3. A route loads the active client's provider keys from SQLite.
 4. The adapter translates the request to Google or OpenRouter format and sends through shared async `httpx`.
 5. A retryable upstream failure advances to the next configured key; success is normalized into a provider-neutral response.
-6. Polling mode persists a queued job before scheduling the single-process async worker; state and results are persisted.
+6. Polling mode persists a queued job before scheduling the single-process async worker. Compact status and result metadata live in SQLite; generated async images are converted to WebP and written to a private output directory. Image files are deleted after 24 hours while job, timing, and cost metadata remain. Prompt and input-image payloads are cleared from the job record at completion or failure.
 
 ## Operational plan
 
 1. Install Python dependencies and set a unique `ADMIN_SECRET`.
-2. Mount a persistent local volume at the configured database path with restrictive permissions.
+2. Mount persistent local storage for both the SQLite database and generated output directory. Keep database and output backups access-restricted.
 3. Run one Uvicorn worker as `www-data`, bound to `127.0.0.1:8081`. Put TLS and request-size enforcement at Nginx; keep the application port private.
 4. Create clients, add provider credentials, then call `/healthz` and verify one small generation using the intended account/model.
 5. Back up the database securely. Treat backups and job records as secrets and apply retention appropriate to the installation.
@@ -106,4 +106,4 @@ SQLite connections enable foreign keys and busy timeout; WAL is initialized once
 - SQLite WAL assumes local storage and one app process. Multi-host operation needs a different DB and durable queue.
 - Upstream model availability and provider-specific parameter support vary by account and change over time. Configure and validate model IDs per provider.
 - The initial admin portal has no user/session system; protect it behind network access controls and HTTPS.
-- Memory use grows with uploaded image size and serialized base64/job results; the service is asynchronous, not zero-copy or constant-memory.
+- Memory use grows with uploaded image size and provider response size while a generation runs; completed async image payloads are stored as files rather than Base64 in SQLite.

README.md

@@ -51,7 +51,7 @@ curl -X POST http://127.0.0.1:8000/v1/generate \
   -F 'images=@capture.jpg;type=image/jpeg' -F 'images=@style.png;type=image/png'
 ```
 
-Polling mode accepts the same multipart fields with `?async_mode=true`. It returns a `job_id`; poll the returned URL with `client_id` as a query parameter and the same `x-client-secret` header. Jobs and their results are stored in SQLite so they survive a client disconnect. Jobs still processing when the proxy process restarts are marked failed because the in-process worker cannot resume safely.
+Polling mode accepts the same multipart fields with `?async_mode=true`. It returns a `job_id`; poll the returned URL with `client_id` as a query parameter and the same `x-client-secret` header. Job state and compact result metadata are stored in SQLite; generated images are converted to WebP (quality 90 by default) and stored as private files under `OUTPUT_DIR` (default `./data/generated`). Downloaded async results use `image/webp`, so save them with a `.webp` extension or convert them locally to JPEG/PNG if the next step requires it. Output files are automatically deleted after 24 hours (checked at startup and every five minutes); job, timing, and cost records remain. Once a job completes or fails, its uploaded prompt/image request payload is removed from SQLite. Jobs still processing when the proxy process restarts are marked failed because the in-process worker cannot resume safely.
 
 The built-in `nano-banana` alias defaults to Gemini 2.5 Flash Image. Set `GOOGLE_MODEL_NANO_BANANA` and `OPENROUTER_MODEL_NANO_BANANA` to provider model IDs you have confirmed are available for your account. Other model IDs are passed through (with `google/` prefix normalization for the native provider).
 
@@ -61,6 +61,9 @@ The built-in `nano-banana` alias defaults to Gemini 2.5 Flash Image. Set `GOOGLE
 | --- | --- | --- |
 | `ADMIN_SECRET` | required | Protects administration endpoints. Use a long random value. |
 | `DB_PATH` | `./data/proxy.db` | SQLite database location. |
+| `OUTPUT_DIR` | `./data/generated` | Private directory for generated output images. Back it up with the database. |
+| `OUTPUT_WEBP_QUALITY` | `90` | WebP quality used for stored async outputs (1–100). |
+| `OUTPUT_RETENTION_SECONDS` | `86400` | Lifetime of generated image files; cleanup runs every five minutes. |
 | `MAX_UPLOAD_BYTES` | `52428800` | Aggregate raw image upload limit. |
 | `MAX_IMAGES` | `8` | Maximum reference image count. |
 | `MAX_IMAGE_PIXELS` | `25000000` | Maximum width × height for each uploaded image. |
@@ -76,7 +79,8 @@ Only one Uvicorn worker is supported for this initial implementation. SQLite WAL
 See the copyable endpoint guide in [`docs/API.md`](docs/API.md).
 
 - `POST /v1/generate` — multipart `client_id`, `prompt`, `model`, optional `provider`, JSON `params`, repeated `images`; optional `?async_mode=true`.
-- `GET /v1/jobs/{job_id}?client_id=...` — job status/result, tenant-scoped.
+- `GET /v1/jobs/{job_id}?client_id=...&include_result=false` — small, tenant-scoped job status and image download links.
+- `GET /v1/jobs/{job_id}/images/{index}?token=...` — raw image download using a short-lived signed link (or use client-secret authentication).
 - `GET /api/admin/state` — client and redacted key state (`x-admin-secret`).
 - `POST /api/admin/client` — create/update a client (`x-admin-secret`).
 - `POST /api/keys` and `DELETE /api/keys/{id}` — manage keys with `auth-secret` (admin or owning client secret).

app/main.py

@@ -2,29 +2,39 @@ from __future__ import annotations
 
 import asyncio
 import base64
+import binascii
 import hashlib
 import hmac
 import io
 import json
 import logging
 import os
+import re
+import sqlite3
 import time
 import uuid
 from contextlib import asynccontextmanager
+from pathlib import Path
 from typing import Any
+from urllib.parse import quote
 
 import aiosqlite
 import httpx
 from fastapi import FastAPI, File, Form, Header, HTTPException, Query, Request, UploadFile
-from fastapi.responses import HTMLResponse
+from fastapi.responses import FileResponse, HTMLResponse, Response
 from PIL import Image, ImageOps, UnidentifiedImageError
 
 LOG = logging.getLogger("phproxy")
 DB_PATH = os.getenv("DB_PATH", "./data/proxy.db")
+OUTPUT_DIR = Path(os.getenv("OUTPUT_DIR", "./data/generated")).resolve()
 ADMIN_SECRET = os.getenv("ADMIN_SECRET", "")
 MAX_UPLOAD_BYTES = int(os.getenv("MAX_UPLOAD_BYTES", str(50 * 1024 * 1024)))
 MAX_IMAGES = int(os.getenv("MAX_IMAGES", "8"))
 MAX_IMAGE_PIXELS = int(os.getenv("MAX_IMAGE_PIXELS", "25000000"))
+IMAGE_LINK_TTL_SECONDS = int(os.getenv("IMAGE_LINK_TTL_SECONDS", "3600"))
+OUTPUT_WEBP_QUALITY = int(os.getenv("OUTPUT_WEBP_QUALITY", "90"))
+OUTPUT_RETENTION_SECONDS = int(os.getenv("OUTPUT_RETENTION_SECONDS", "86400"))
+OUTPUT_CLEANUP_INTERVAL_SECONDS = 300
 HTTP_TIMEOUT_SECONDS = float(os.getenv("HTTP_TIMEOUT_SECONDS", "300"))
 CONNECT_TIMEOUT_SECONDS = float(os.getenv("CONNECT_TIMEOUT_SECONDS", "15"))
 MODEL_MAP = {
@@ -47,6 +57,175 @@ def secret_matches(value: str | None, stored_hash: str | None) -> bool:
     return bool(value and stored_hash and hmac.compare_digest(secret_hash(value), stored_hash))
 
 
+def make_image_download_token(job_id: str, client_id: str, image_index: int) -> str:
+    claims = json.dumps(
+        {"job_id": job_id, "client_id": client_id, "image_index": image_index,
+         "expires_at": int(time.time()) + IMAGE_LINK_TTL_SECONDS},
+        separators=(",", ":"),
+    ).encode("utf-8")
+    payload = base64.urlsafe_b64encode(claims).rstrip(b"=").decode("ascii")
+    signature = hmac.new(ADMIN_SECRET.encode("utf-8"), payload.encode("ascii"), hashlib.sha256).hexdigest()
+    return f"{payload}.{signature}"
+
+
+def signed_image_download_url(job_id: str, client_id: str, image_index: int) -> str:
+    token = make_image_download_token(job_id, client_id, image_index)
+    return f"/v1/jobs/{job_id}/images/{image_index}?token={quote(token, safe='')}"
+
+
+def verify_image_download_token(token: str, job_id: str, image_index: int) -> str:
+    try:
+        payload, supplied_signature = token.split(".", 1)
+        expected_signature = hmac.new(
+            ADMIN_SECRET.encode("utf-8"), payload.encode("ascii"), hashlib.sha256
+        ).hexdigest()
+        if not hmac.compare_digest(supplied_signature, expected_signature):
+            raise ValueError("invalid signature")
+        raw_claims = base64.urlsafe_b64decode(payload + "=" * (-len(payload) % 4))
+        claims = json.loads(raw_claims)
+        if (claims.get("job_id") != job_id or claims.get("image_index") != image_index
+                or not isinstance(claims.get("client_id"), str)):
+            raise ValueError("token scope mismatch")
+        if not isinstance(claims.get("expires_at"), int) or claims["expires_at"] < int(time.time()):
+            raise HTTPException(status_code=410, detail="Image link has expired")
+        return claims["client_id"]
+    except HTTPException:
+        raise
+    except (ValueError, TypeError, UnicodeDecodeError, json.JSONDecodeError) as exc:
+        raise HTTPException(status_code=401, detail="Invalid image link") from exc
+
+
+def output_image_path(job_id: str, image_index: int, mime_type: str) -> Path:
+    image_format = IMAGE_FORMAT_TYPES.get(mime_type)
+    if image_format is None or image_index < 0:
+        raise ValueError("Unsupported generated image type or index")
+    job_key = uuid.UUID(job_id).hex
+    return OUTPUT_DIR / f"{job_key}-{image_index}.{image_format.lower()}"
+
+
+def convert_image_to_webp(content: bytes, mime_type: str) -> bytes:
+    if not 1 <= OUTPUT_WEBP_QUALITY <= 100:
+        raise ValueError("OUTPUT_WEBP_QUALITY must be between 1 and 100")
+    Image.MAX_IMAGE_PIXELS = MAX_IMAGE_PIXELS
+    try:
+        with Image.open(io.BytesIO(content)) as source:
+            actual_mime = IMAGE_FORMATS.get(source.format or "")
+            if actual_mime is None or actual_mime != mime_type:
+                raise ValueError("Generated image content does not match its MIME type")
+            width, height = source.size
+            if width <= 0 or height <= 0 or width * height > MAX_IMAGE_PIXELS:
+                raise ValueError("Generated image dimensions exceed configured limits")
+            source.load()
+            image = ImageOps.exif_transpose(source)
+            has_alpha = "A" in image.getbands() or "transparency" in image.info
+            image = image.convert("RGBA" if has_alpha else "RGB")
+            output = io.BytesIO()
+            image.save(output, format="WEBP", quality=OUTPUT_WEBP_QUALITY, method=4)
+            return output.getvalue()
+    except (Image.DecompressionBombError, Image.DecompressionBombWarning) as exc:
+        raise ValueError("Generated image dimensions exceed configured limits") from exc
+    except (UnidentifiedImageError, OSError) as exc:
+        raise ValueError("Provider returned an invalid image") from exc
+
+
+def write_output_file(path: Path, content: bytes) -> None:
+    temp_path = OUTPUT_DIR / f".{path.name}.{uuid.uuid4().hex}.tmp"
+    with temp_path.open("xb") as image_file:
+        os.chmod(temp_path, 0o600)
+        image_file.write(content)
+        image_file.flush()
+        os.fsync(image_file.fileno())
+    os.replace(temp_path, path)
+    os.chmod(path, 0o600)
+
+
+def cleanup_expired_output_files(now: float | None = None) -> int:
+    """Delete generated image files once they have been stored for 24 hours."""
+    cutoff = (time.time() if now is None else now) - OUTPUT_RETENTION_SECONDS
+    generated_name = re.compile(r"[0-9a-f]{32}-[0-9]+\.(?:webp|png|jpe?g)")
+    removed = 0
+    try:
+        paths = list(OUTPUT_DIR.iterdir())
+    except FileNotFoundError:
+        return 0
+    for path in paths:
+        if not generated_name.fullmatch(path.name):
+            continue
+        try:
+            if path.is_file() and path.stat().st_mtime <= cutoff:
+                path.unlink()
+                removed += 1
+        except FileNotFoundError:
+            continue
+        except OSError:
+            LOG.exception("Could not remove expired output image %s", path.name)
+    return removed
+
+
+async def output_cleanup_loop() -> None:
+    while True:
+        await asyncio.sleep(OUTPUT_CLEANUP_INTERVAL_SECONDS)
+        try:
+            removed = await asyncio.to_thread(cleanup_expired_output_files)
+            if removed:
+                LOG.info("Removed %s expired generated image(s)", removed)
+        except Exception:
+            LOG.exception("Generated image cleanup failed")
+
+
+def persist_result_images(job_id: str, result: dict[str, Any]) -> tuple[dict[str, Any], bool, list[Path]]:
+    """Store async image outputs as WebP and return compact result metadata."""
+    compact_result = dict(result)
+    compact_images = []
+    written_paths: list[Path] = []
+    superseded_paths: list[Path] = []
+    changed = False
+    try:
+        for image_index, image in enumerate(result.get("images", [])):
+            image_metadata = dict(image)
+            encoded = image_metadata.pop("b64_json", None)
+            if encoded:
+                mime_type = image_metadata.get("mime_type")
+                source_content = base64.b64decode(encoded, validate=True)
+                webp_content = convert_image_to_webp(source_content, mime_type)
+                path = output_image_path(job_id, image_index, "image/webp")
+                write_output_file(path, webp_content)
+                written_paths.append(path)
+                image_metadata["mime_type"] = "image/webp"
+                changed = True
+            elif image_metadata.get("mime_type") in IMAGE_FORMAT_TYPES:
+                old_path = output_image_path(job_id, image_index, image_metadata["mime_type"])
+                if image_metadata["mime_type"] != "image/webp" and old_path.is_file():
+                    source_content = old_path.read_bytes()
+                    webp_content = convert_image_to_webp(source_content, image_metadata["mime_type"])
+                    path = output_image_path(job_id, image_index, "image/webp")
+                    write_output_file(path, webp_content)
+                    written_paths.append(path)
+                    superseded_paths.append(old_path)
+                    image_metadata["mime_type"] = "image/webp"
+                    changed = True
+            compact_images.append(image_metadata)
+        compact_result["images"] = compact_images
+        return compact_result, changed, superseded_paths
+    except Exception:
+        for path in written_paths:
+            path.unlink(missing_ok=True)
+        raise
+
+
+def compact_sqlite_database() -> None:
+    """Checkpoint and compact SQLite without an outstanding async cursor."""
+    conn = sqlite3.connect(DB_PATH, timeout=30)
+    try:
+        checkpoint = conn.execute("PRAGMA wal_checkpoint(TRUNCATE)")
+        checkpoint.fetchall()
+        checkpoint.close()
+        vacuum = conn.execute("VACUUM")
+        vacuum.close()
+    finally:
+        conn.close()
+
+
 @asynccontextmanager
 async def db():
     conn = await aiosqlite.connect(DB_PATH, timeout=10)
@@ -61,6 +240,11 @@ async def db():
 
 async def init_db() -> None:
     os.makedirs(os.path.dirname(os.path.abspath(DB_PATH)), exist_ok=True)
+    OUTPUT_DIR.mkdir(parents=True, exist_ok=True, mode=0o700)
+    os.chmod(OUTPUT_DIR, 0o700)
+    migrated_result_count = 0
+    cleared_request_count = 0
+    superseded_paths: list[Path] = []
     async with db() as conn:
         await conn.execute("PRAGMA journal_mode=WAL")
         await conn.executescript(
@@ -163,19 +347,79 @@ async def init_db() -> None:
                 LOG.exception("Could not backfill generation history for job %s", row["job_id"])
         await conn.commit()
 
+        cur = await conn.execute(
+            "SELECT job_id,result_json,request_json FROM jobs "
+            "WHERE status IN ('completed','failed')"
+        )
+        for row in await cur.fetchall():
+            if row["result_json"]:
+                try:
+                    result = json.loads(row["result_json"])
+                    compact_result, changed, old_paths = await asyncio.to_thread(
+                        persist_result_images, row["job_id"], result
+                    )
+                    superseded_paths.extend(old_paths)
+                    if changed:
+                        await conn.execute(
+                            "UPDATE jobs SET result_json=? WHERE job_id=?",
+                            (json.dumps(compact_result), row["job_id"]),
+                        )
+                        migrated_result_count += 1
+                except Exception:
+                    # Keep the legacy JSON available if an individual historic image cannot be migrated.
+                    LOG.exception("Could not migrate stored output images for job %s", row["job_id"])
+            if row["request_json"] and row["request_json"] != "{}":
+                await conn.execute("UPDATE jobs SET request_json='{}' WHERE job_id=?", (row["job_id"],))
+                cleared_request_count += 1
+        await conn.commit()
+
+    # Remove old PNG/JPEG outputs only after their WebP replacements are committed.
+    for old_path in superseded_paths:
+        try:
+            old_path.unlink(missing_ok=True)
+        except OSError:
+            LOG.exception("Could not remove superseded output image %s", old_path.name)
+
+    async with db() as conn:
+        cur = await conn.execute("PRAGMA freelist_count")
+        free_page_count = (await cur.fetchone())[0]
+    if migrated_result_count or cleared_request_count or free_page_count > 1000:
+        # Reclaim SQLite pages freed by removing Base64 images and completed request payloads.
+        try:
+            await asyncio.to_thread(compact_sqlite_database)
+        except sqlite3.Error:
+            LOG.exception("Could not compact the SQLite database; service will continue")
+    if migrated_result_count or cleared_request_count:
+        LOG.info("Migrated %s job result(s) to files and cleared %s completed request payload(s)",
+                 migrated_result_count, cleared_request_count)
+
 
 @asynccontextmanager
 async def lifespan(_: FastAPI):
     global http_client
+    cleanup_task: asyncio.Task | None = None
     if not ADMIN_SECRET:
         raise RuntimeError("ADMIN_SECRET must be set before starting the service")
+    if OUTPUT_RETENTION_SECONDS <= 0:
+        raise RuntimeError("OUTPUT_RETENTION_SECONDS must be greater than zero")
     await init_db()
+    removed = await asyncio.to_thread(cleanup_expired_output_files)
+    if removed:
+        LOG.info("Removed %s expired generated image(s) during startup", removed)
     http_client = httpx.AsyncClient(
         timeout=httpx.Timeout(HTTP_TIMEOUT_SECONDS, connect=CONNECT_TIMEOUT_SECONDS),
         limits=httpx.Limits(max_keepalive_connections=50, max_connections=200),
         follow_redirects=False,
     )
+    cleanup_task = asyncio.create_task(output_cleanup_loop())
+    try:
         yield
+    finally:
+        cleanup_task.cancel()
+        try:
+            await cleanup_task
+        except asyncio.CancelledError:
+            pass
         await http_client.aclose()
 
 
@@ -511,17 +755,27 @@ async def run_job(job_id: str, client_id: str, request_data: dict[str, Any],
         result.update({"generation_id": job_id,
                        "request_intake_sec": run_metadata["request_intake_sec"],
                        "total_sec": round(completed_at - run_metadata["received_at"], 2)})
+        compact_result, _, superseded_paths = await asyncio.to_thread(
+            persist_result_images, job_id, result
+        )
         async with db() as conn:
-            await conn.execute("UPDATE jobs SET status='completed',result_json=?,updated_at=? WHERE job_id=?",
-                               (json.dumps(result), completed_at, job_id))
+            await conn.execute(
+                "UPDATE jobs SET status='completed',result_json=?,request_json='{}',updated_at=? WHERE job_id=?",
+                (json.dumps(compact_result), completed_at, job_id),
+            )
             await conn.commit()
+        for old_path in superseded_paths:
+            try:
+                old_path.unlink(missing_ok=True)
+            except OSError:
+                LOG.exception("Could not remove superseded output image %s", old_path.name)
         await finish_generation_run(job_id, "completed", completed_at,
                                     result["total_sec"], result=result)
     except HTTPException as exc:
         error = exc.detail
         completed_at = time.time()
         async with db() as conn:
-            await conn.execute("UPDATE jobs SET status='failed',error_json=?,updated_at=? WHERE job_id=?",
+            await conn.execute("UPDATE jobs SET status='failed',request_json='{}',error_json=?,updated_at=? WHERE job_id=?",
                                (json.dumps(error), completed_at, job_id))
             await conn.commit()
         await finish_generation_run(job_id, "failed", completed_at,
@@ -531,7 +785,7 @@ async def run_job(job_id: str, client_id: str, request_data: dict[str, Any],
         completed_at = time.time()
         error = {"message": "Internal generation error"}
         async with db() as conn:
-            await conn.execute("UPDATE jobs SET status='failed',error_json=?,updated_at=? WHERE job_id=?",
+            await conn.execute("UPDATE jobs SET status='failed',request_json='{}',error_json=?,updated_at=? WHERE job_id=?",
                                (json.dumps(error), completed_at, job_id))
             await conn.commit()
         await finish_generation_run(job_id, "failed", completed_at,
@@ -605,6 +859,7 @@ async def generate(request: Request, client_id: str = Form(...), prompt: str = F
 
 @app.get("/v1/jobs/{job_id}")
 async def get_job(job_id: str, client_id: str = Query(...),
+                  include_result: bool = Query(True),
                   x_client_secret: str | None = Header(None)):
     await authorize_client(client_id, x_client_secret)
     async with db() as conn:
@@ -615,12 +870,122 @@ async def get_job(job_id: str, client_id: str = Query(...),
     response: dict[str, Any] = {"job_id": job_id, "status": row["status"],
                                 "created_at": row["created_at"], "updated_at": row["updated_at"]}
     if row["result_json"]:
-        response["result"] = json.loads(row["result_json"])
+        result = json.loads(row["result_json"])
+        if include_result:
+            legacy_result = dict(result)
+            legacy_images = []
+            for index, image in enumerate(result.get("images", [])):
+                image_data = dict(image)
+                if not image_data.get("b64_json") and image_data.get("mime_type") in IMAGE_FORMAT_TYPES:
+                    try:
+                        path = output_image_path(job_id, index, image_data["mime_type"])
+                        if path.is_file():
+                            content = await asyncio.to_thread(path.read_bytes)
+                            image_data["b64_json"] = base64.b64encode(content).decode("ascii")
+                    except (ValueError, OSError):
+                        LOG.exception("Could not read stored output image for job %s, image %s",
+                                      job_id, index)
+                legacy_images.append(image_data)
+            legacy_result["images"] = legacy_images
+            response["result"] = legacy_result
+        else:
+            # Keep status polling small; image bytes are fetched through the binary route.
+            result_metadata = {key: value for key, value in result.items() if key != "images"}
+            result_metadata["images"] = []
+            for index, image in enumerate(result.get("images", [])):
+                image_metadata = {key: value for key, value in image.items()
+                                  if key not in {"b64_json", "url"}}
+                image_is_local = bool(image.get("b64_json"))
+                if not image_is_local and image.get("mime_type") in IMAGE_FORMAT_TYPES:
+                    try:
+                        path = output_image_path(job_id, index, image.get("mime_type"))
+                        image_is_local = path.is_file()
+                    except ValueError:
+                        LOG.warning("Stored output image has unsupported MIME type for job %s, image %s",
+                                    job_id, index)
+                if image_is_local:
+                    image_metadata["download_url"] = signed_image_download_url(
+                        job_id, client_id, index
+                    )
+                elif image.get("url"):
+                    image_metadata["url"] = image["url"]
+                result_metadata["images"].append(image_metadata)
+            response["result"] = result_metadata
     if row["error_json"]:
         response["error"] = json.loads(row["error_json"])
     return response
 
 
+@app.get("/v1/jobs/{job_id}/images/{image_index}")
+async def get_job_image(job_id: str, image_index: int, client_id: str | None = Query(None),
+                        token: str | None = Query(None),
+                        x_client_secret: str | None = Header(None)):
+    """Return raw image bytes via a short-lived signed link or client-secret auth."""
+    if token:
+        token_client_id = verify_image_download_token(token, job_id, image_index)
+        if client_id and client_id != token_client_id:
+            raise HTTPException(status_code=401, detail="Invalid image link")
+        client_id = token_client_id
+    elif client_id:
+        await authorize_client(client_id, x_client_secret)
+    else:
+        raise HTTPException(status_code=401, detail="An image link or client authentication is required")
+
+    async with db() as conn:
+        cur = await conn.execute(
+            "SELECT j.status,j.result_json,c.is_active FROM jobs j "
+            "JOIN clients c ON c.client_id=j.client_id WHERE j.job_id=? AND j.client_id=?",
+            (job_id, client_id),
+        )
+        row = await cur.fetchone()
+    if not row or not row["is_active"]:
+        raise HTTPException(status_code=404, detail="Job not found")
+    if row["status"] != "completed" or not row["result_json"]:
+        raise HTTPException(status_code=409, detail="Job has no completed result")
+    images = json.loads(row["result_json"]).get("images", [])
+    if image_index < 0 or image_index >= len(images):
+        raise HTTPException(status_code=404, detail="Image not found")
+    image = images[image_index]
+    mime_type = image.get("mime_type")
+    if mime_type in IMAGE_FORMAT_TYPES:
+        path = output_image_path(job_id, image_index, mime_type)
+        if path.is_file():
+            extension = IMAGE_FORMAT_TYPES[mime_type].lower()
+            return FileResponse(
+                path,
+                media_type=mime_type,
+                filename=f"generated-{image_index}.{extension}",
+                content_disposition_type="inline",
+                headers={
+                    "Cache-Control": "private, no-store",
+                    "X-Content-Type-Options": "nosniff",
+                    "Referrer-Policy": "no-referrer",
+                },
+            )
+
+    # Keep legacy completed jobs accessible if a file migration was not possible.
+    encoded = image.get("b64_json")
+    if encoded:
+        try:
+            content = base64.b64decode(encoded, validate=True)
+        except (ValueError, TypeError, binascii.Error) as exc:
+            LOG.error("Stored output image is invalid for job %s, image %s", job_id, image_index)
+            raise HTTPException(status_code=500, detail="Stored image is invalid") from exc
+        return Response(
+            content=content,
+            media_type=mime_type if mime_type in IMAGE_FORMAT_TYPES else "application/octet-stream",
+            headers={
+                "Content-Disposition": f'inline; filename="generated-{image_index}"',
+                "Cache-Control": "private, no-store",
+                "X-Content-Type-Options": "nosniff",
+                "Referrer-Policy": "no-referrer",
+            },
+        )
+    if mime_type in IMAGE_FORMAT_TYPES:
+        raise HTTPException(status_code=410, detail="Image is no longer available")
+    raise HTTPException(status_code=409, detail="Image is available only from its upstream URL")
+
+
 @app.get("/api/admin/state")
 async def admin_state(x_admin_secret: str | None = Header(None)):
     auth_admin(x_admin_secret)

docs/API.md

@@ -92,7 +92,7 @@ Success (`200`):
 
 Generation responses also include `generation_id`, `request_intake_sec`, and `total_sec`. `request_intake_sec` covers the request reaching and being parsed by the proxy before the generation is queued; `elapsed_sec` is the upstream call time; `total_sec` runs from request receipt through completion. These measurements help separate slow client uploads from slow model calls.
 
-An image can be returned as `b64_json`, or as a URL when the upstream returns a hosted image. Decode `b64_json` as standard Base64 before saving it. Avoid printing or logging generated image data in an agent script.
+Direct synchronous generation retains its JSON response with image bytes in `b64_json`. For async jobs, request `include_result=false` and use the binary download link described below. The legacy async `include_result=true` response remains supported and reconstructs Base64 from the stored image file.
 
 ## Generate asynchronously and poll
 
@@ -113,10 +113,10 @@ Queued response (`200`):
 {"job_id":"<uuid>","status":"queued","poll_url":"/v1/jobs/<uuid>"}
 ```
 
-Poll every couple of seconds. Include both the `client_id` query parameter and `x-client-secret` header:
+Poll every couple of seconds. To avoid transferring the generated image inside every completed status response, set `include_result=false`. Include both the `client_id` query parameter and `x-client-secret` header:
 
 ```sh
-curl -sS "$BASE_URL/v1/jobs/<uuid>?client_id=booth-01" \
+curl -sS "$BASE_URL/v1/jobs/<uuid>?client_id=booth-01&include_result=false" \
   -H 'x-client-secret: replace-with-a-client-secret'
 ```
 
@@ -126,7 +126,20 @@ While running:
 {"job_id":"<uuid>","status":"processing","created_at":1234567890.0,"updated_at":1234567891.0}
 ```
 
-On success, `status` is `completed` and a `result` field contains the synchronous response object. On failure, `status` is `failed` and `error` contains a sanitized message. Jobs are scoped to the owning `client_id`; an unknown job ID and a job belonging to another client both return `404`.
+On success with `include_result=false`, `status` is `completed` and `result` contains generation metadata plus image descriptors. Each image descriptor includes a `download_url` for binary images, or the upstream `url` when the provider supplies a hosted URL. Download each binary image directly; do not Base64-decode it:
+
+```sh
+curl -sS "$BASE_URL/v1/jobs/<uuid>/images/0?token=<signed-token-from-download_url>" \
+  -o generated.webp
+```
+
+Async generated images are stored as WebP at quality 90 by default (`OUTPUT_WEBP_QUALITY` can be set from 1 to 100). The image endpoint and result descriptor use `image/webp`; save the downloaded bytes with a `.webp` extension. Convert to PNG or JPEG in the client only if a downstream tool requires that format.
+
+`GET /v1/jobs/{job_id}/images/{image_index}?token=...` serves raw image bytes from the proxy's private output directory with the corresponding image MIME type and inline content disposition. Links in the completed-job response are signed, scoped to that job and image, and expire after one hour by default (`IMAGE_LINK_TTL_SECONDS` configures the lifetime). Open the full `download_url` directly in a browser; no client-secret header is needed. Treat the URL like a temporary password: anyone who has it can view that image until it expires. An authenticated client may also call the image route with `client_id` and `x-client-secret` instead of a token. Unknown and cross-client jobs return `404`; an incomplete job returns `409`; expired links or deleted images return `410`.
+
+Async WebP output files are stored under `OUTPUT_DIR` (default `./data/generated`) with directory mode `0700` and file mode `0600`. Image files are automatically deleted after 24 hours by default (`OUTPUT_RETENTION_SECONDS`), with a cleanup pass at startup and every five minutes. Job metadata, timing, and cost history remain in SQLite after image deletion; a deleted image endpoint returns `410`. Completed and failed jobs clear their original prompt and uploaded image payload from SQLite. The database retains only compact result metadata and execution history; startup migrates older completed outputs to WebP and vacuums SQLite to reclaim freed pages.
+
+For compatibility, `include_result` defaults to `true`; that legacy mode includes Base64 image data in the status response. New clients should always request `include_result=false` and download images from their `download_url`. On failure, status is `failed` and `error` contains a sanitized message.
 
 ## Models and parameters