Once I get an error I can't move on

[sai-sy]

Description

I asked the model to get some images and perform some actions, it didn't properly get some images so I got this error:

image formats: ['image/jpeg', 'image/png', 'image/gif', 'image/webp'].```

That's fine I replaced the files in the directory with correct working image. But I can't move past the error:

```The image data you provided does not represent a valid image. Please check your input and try again with one of the supported
image formats: ['image/jpeg', 'image/png', 'image/gif', 'image/webp'].

finish task 3 in l9.md and send it to the server
sai (08:01 AM)

The image data you provided does not represent a valid image. Please check your input and try again with one of the supported
image formats: ['image/jpeg', 'image/png', 'image/gif', 'image/webp'].

okay don't use any image
sai (08:01 AM)

The image data you provided does not represent a valid image. Please check your input and try again with one of the supported
image formats: ['image/jpeg', 'image/png', 'image/gif', 'image/webp'].

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• SVG File Paths Rejected #4052: SVG File Paths Rejected - Similar image format validation error preventing task progression
• fix: Clipboard image paste fails with AI_DownloadError due to binary corruption and encoding mismatch #3816: Clipboard image paste fails with AI_DownloadError - Related issue with invalid image data handling

Feel free to ignore if your specific case involves different circumstances.

[rekram1-node]

Yeah ill fix this

[luisrudge]

having the same issue when OC tries to read a big image

[vaisov]

It happens to me even without any image provided, no idea where it gets this from.

[rosspeoples]

I've hit this twice so far. Both times it was debugging an issue, not even touching images, then just gets stuck:

Now I have to try and figure what it was doing and convince it to continue in a new thread.

[kozak-codes]

Any update on this? @BASIC-BIT ?

My current fix is to open a new opencode thread and tell it to remove the b64 image from the history of that current thread. Willing to push a fix, but I frequently run into this problem when using the chrome MCP.

[kozak-codes]

Additional context + reproduction steps

This is a persistent session-poisoning bug. Once an invalid or oversized image gets stored in the message history, every subsequent request in that session fails with the same error — even if you tell the model to ignore images entirely. The bad image data is already baked into the context sent to the API on every turn.

Root cause: OpenCode stores image attachments (base64-encoded) inside the part records in its local SQLite database (~/.local/share/opencode/opencode.db). When the API rejects an image, the offending data remains in the DB and gets replayed on every new message in that session.

Two distinct error variants I've hit:

1. image exceeds 5 MB maximum: X bytes > 5242880 bytes — full-page screenshots captured via the read tool on binary image files
2. At least one of the image dimensions exceed max allowed size for many-image requests: 2000 pixels — tall/long screenshots (e.g. full-page captures) where height exceeds 2000px
   Steps to reproduce:
3. In a session, ask the model to take a screenshot or read an image file (e.g. via the read tool on a .png)
4. If the image is either >5MB or has a dimension >2000px, the API returns a 400 error
5. Send any follow-up message in the same session — the error repeats identically, regardless of content
6. Telling the model to "ignore images" or "don't use images" has no effect — the bad attachment is already in the stored history

Workaround (manual DB surgery):

python3 << 'EOF'
import sqlite3, json, base64, struct
conn = sqlite3.connect('~/.local/share/opencode/opencode.db')
cursor = conn.cursor()
# Replace SESSION_ID with your session ID
cursor.execute("SELECT id, data FROM part WHERE session_id = 'SESSION_ID'")
rows = cursor.fetchall()
def get_png_dimensions(b64_data):
    try:
        if ',' in b64_data:
            b64_data = b64_data.split(',', 1)[1]
        raw = base64.b64decode(b64_data[:100])
        if raw[:4] == b'\x89PNG':
            w = struct.unpack('>I', raw[16:20])[0]
            h = struct.unpack('>I', raw[20:24])[0]
            return w, h
    except:
        pass
    return None, None
fixed = 0
for part_id, data_str in rows:
    data = json.loads(data_str)
    state = data.get('state', {})
    attachments = state.get('attachments', [])
    bad = False
    for att in attachments:
        if att.get('mime', '').startswith('image/'):
            url = att.get('url', '')
            b64_size = len(url)
            w, h = get_png_dimensions(url)
            if b64_size > 5_000_000 or (w and (w > 2000 or h > 2000)):
                bad = True
                break
    if bad:
        data['state']['attachments'] = []
        cursor.execute("UPDATE part SET data = ? WHERE id = ?", (json.dumps(data), part_id))
        fixed += 1
conn.commit()
print(f"Stripped {fixed} bad image(s)")
conn.close()
EOF

Expected behavior: The session should either (a) not store image attachments that exceed API limits in the first place, or (b) gracefully strip/skip bad attachments from history on retry rather than hard-failing every subsequent turn.