Symptom

A channel bot (bot-A) with a message-coalesce window configured (message_coalesce_window_ms = 3000), routed to an agent (agent-A) on a text-only model (DeepSeek, supports_vision = false), never receives a description of inbound photos. The agent only sees the redacted placeholder [image omitted: model has no vision support] (the #6010 text-only redaction) and replies that it cannot see the image.

An otherwise identical bot without a coalesce window, same model, does get a Groq description and can answer about the image. The only differing variable is message_coalesce_window_ms.

[media] image_description = true is set globally; the Groq provider/key work (verified by calling the vision model directly with the same image — it returns a correct description).

Root cause (read from current source, tag v2026.6.22-beta.22)

Inbound dispatch forks on the debounce window in BridgeManager::start_adapter (crates/librefang-channels/src/bridge.rs:1435-1438):

• Fast path (debounce_ms == 0) → dispatch_message(). Its image branch downloads the photo, builds an ImageFile block, and calls the description step:

// crates/librefang-channels/src/bridge.rs:3959
if let Some(desc_block) = maybe_describe_inbound_image(handle, Some(saved)).await {
    final_blocks.insert(0, desc_block);   // prepend Groq description
}
// ...
// crates/librefang-channels/src/bridge.rs:3966
dispatch_with_blocks(final_blocks, message, /* ... */).await;   // shared tail

• Debounce path (debounce_ms > 0) → pre-downloads images into image_blocks and dispatches via flush_debounced() → dispatch_with_blocks() (bridge.rs:1223). This branch builds the image blocks itself and never calls maybe_describe_inbound_image. dispatch_with_blocks() (bridge.rs:6245) contains no description call in its entire body.

So maybe_describe_inbound_image (defined bridge.rs:5453) is reachable only from dispatch_message. The [media] image_description flag is consequently only ever consulted on the fast path (the flag check lives in describe_inbound_image at crates/librefang-api/src/channel_bridge.rs, reached only via maybe_describe_inbound_image).

Why it surfaces as it does

For a text-only model the un-described image block is then redacted to a placeholder by the #6010 vision-gate (redact_images_for_text_only, crates/librefang-runtime/src/agent_loop/mod.rs), so there is no HTTP 400 / crash — the failure is silent: the agent simply never receives any image content or description. Vision-capable models are unaffected (they would see the raw image either way). The bug is therefore invisible except as "the bot ignores photos" on text-only + coalesce setups.

Proposed fix

The two paths already converge on the shared tail dispatch_with_blocks(). The divergence is only that the content→blocks conversion (which includes description) is duplicated, and the debounced copy omits the description step.

• Minimal: in flush_debounced (where the image blocks are assembled, around bridge.rs:1073-1083), mirror the ImageFile-extraction + maybe_describe_inbound_image prepend from bridge.rs:3950-3963 before calling dispatch_with_blocks.
• Proper (recommended): factor the content→blocks conversion including maybe_describe_inbound_image into a single shared helper that both the immediate and the debounced paths call before the common dispatch_with_blocks. The fork should be only "coalesce or not" at ingest; all downstream media handling (description, future transcription/OCR) should live in one path so a parallel branch cannot drift out of parity again.

Setup (for reference)

• v2026.6.22-beta.22
• Pattern: sidecar/channel bot with message_coalesce_window_ms > 0, default agent on a text-only model (supports_vision = false), [media] image_description = true.
• Repro: send a photo to a coalesced bot vs. a non-coalesced bot on the same model — only the non-coalesced one is described.
• Workaround: remove message_coalesce_window_ms for affected bots (forces the fast path), at the cost of losing message coalescing.

Happy to test a patch.

Refs: #6010 (text-only image redaction — closed; this is the adjacent description gap on the debounce path)