acking-you/codex-bridge/qq-quoted-image-recovery

QQ Quoted Image Recovery

Overview

Use this skill when a QQ reply chain points at an earlier image message but the bridge did not hand you the image itself.

The core rule is simple:

• a quote marker is only a handle,
• the flattened quote text is not the image,
• you have not "seen the image" until you recover the original media payload.

When to Use

• The inbound text contains [quote<msg:12345> ...] or [quote<msg:12345>].
• The user is asking what a quoted screenshot/image says or means.
• The quote block is empty or nearly empty, but the surrounding context suggests

the original message was visual.

• You need the quoted image pixels before answering.

Do not use this skill for ordinary text-only quote context. Use qq-current-history for that.

Workflow

1. Read the quoted message_id from [quote<msg:...>].
2. Recover the raw quoted message via OneBot get_msg.
3. Inspect the raw message JSON for image segments.
4. Recover the actual image URL or local path from that raw payload.
5. Download the image into .run/artifacts/.
6. Inspect locally first.
7. Only if local inspection is insufficient, send the recovered artifact to a

vision-capable model.

Rules

• Stay inside the current conversation and the quoted message_id.
• Do not pretend the bridge already showed you the image when it only showed a

quote marker.

• Do not hallucinate missing text from a screenshot you have not recovered.
• Prefer a bounded, deterministic recovery path over broad history scanning.
• If recovery fails, say that clearly and explain that only the quote marker was

available.

Practical Notes

• Today the bridge flattens quoted message text with extract_text, so image

segments can disappear from the visible quote preamble.

• The raw get_msg payload is the source of truth when you need to recover

quoted media.

• If the raw payload contains multiple segments, do not assume the first one is

text; inspect the segment list directly.

• Save recovered media under .run/artifacts/quoted_<message_id>.<ext>.

Local-First Inspection

• Check file type, dimensions, and whether it is obviously a screenshot, photo,

or meme.

• Use local OCR / preview / terminal-friendly inspection first when available.
• Escalate to a vision model only when the answer really depends on pixels that

local inspection cannot reliably recover.

Common Mistakes

• Treating [quote<msg:...>] as if it already contained the image body.
• Summarizing a screenshot before recovering the actual image payload.
• Jumping straight to a vision model without first reconstructing the quoted

artifact.

• Broadly scraping conversation history instead of targeting the quoted

message_id.