fix(bridge): recover inline-prefixed metadata in user message body
OpenClaw's canonical convention is to emit metadata envelopes (chat_id, sender, reply target, …) as SEPARATE user-role messages folded into the openai-completions request right after the real one — `extractLatestUserMessage` skips those whole. Fabric.OpenclawPlugin's dispatch does not split: it passes metadata blocks and the real user content as ONE merged user-message body, separated by blank lines. With the prior filter that meant the entire turn was dropped with "no user message found" (HTTP 400) because the first line matched a sentinel — the actual prompt sitting after the metadata blocks never reached the bridge. When the whole-body check fails for a single-message body, walk past leading sentinel-prefixed blocks (sentinel header + optional ```json code fence + blank-line separator) and use whatever non-metadata block follows. Falls back to the previous "skip entirely" semantics when the body is metadata-only. End-user symptom that surfaced this: every contractor agent (Claude / Gemini) subscribed to a Fabric channel silently failed to reply to sub-discussion messages during recruitment — fabric dispatch said "completed" in 1.6s but trajectory had `assistantTexts: []`, `terminalError: non_deliverable_terminal_turn`, `errorMessage: "400 \"no user message found\""`. Surfaced recruiting developer1 on prod-t2 2026-05-31.
This commit is contained in:
@@ -61,6 +61,51 @@ function isRuntimeContextMessage(text: string): boolean {
|
||||
return INBOUND_META_SENTINELS.includes(firstLine);
|
||||
}
|
||||
|
||||
/**
|
||||
* Strip *prefixed* metadata blocks from a single user message body.
|
||||
*
|
||||
* Background: OpenClaw's older / canonical convention is to emit metadata
|
||||
* envelopes (chat_id, sender, reply target, …) as SEPARATE user-role
|
||||
* messages folded into the openai-completions request right after the real
|
||||
* one. `extractLatestUserMessage` skips those separate messages whole.
|
||||
*
|
||||
* The Fabric channel plugin (Fabric.OpenclawPlugin) does not split: it
|
||||
* passes the metadata blocks and the actual user content as ONE merged
|
||||
* user-message body, separated by blank lines. Result: a Fabric inbound
|
||||
* turn has a single user message whose first line matches a sentinel, so
|
||||
* `isRuntimeContextMessage` returned true and the whole turn was dropped
|
||||
* with "no user message found" (HTTP 400). Symptom: every contractor agent
|
||||
* subscribed to a Fabric channel silently fails to reply.
|
||||
*
|
||||
* Strip leading sentinel-prefixed blocks (sentinel header + optional code
|
||||
* fence + blank-line separator) until we hit a block whose first line is
|
||||
* NOT a sentinel — that's the real prompt. Returns "" if the entire body
|
||||
* is metadata-only (still a no-op turn, same as before).
|
||||
*/
|
||||
function stripPrefixedMetadataBlocks(raw: string): string {
|
||||
// Split on blank-line block boundaries. Within a metadata block the JSON
|
||||
// code fence has only single newlines, so it stays in the same chunk as
|
||||
// its sentinel header.
|
||||
const blocks = raw.split(/\n\n+/);
|
||||
const kept: string[] = [];
|
||||
let stillStripping = true;
|
||||
for (const block of blocks) {
|
||||
if (stillStripping) {
|
||||
const trimmed = block.trimStart();
|
||||
const firstLine = trimmed.split("\n", 1)[0].trim();
|
||||
if (
|
||||
INBOUND_META_SENTINELS.includes(firstLine) ||
|
||||
trimmed.startsWith(LEGACY_RUNTIME_CONTEXT_MARKER)
|
||||
) {
|
||||
continue; // drop this prefix block
|
||||
}
|
||||
stillStripping = false;
|
||||
}
|
||||
kept.push(block);
|
||||
}
|
||||
return kept.join("\n\n").trim();
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract the latest user-authored message from the OpenClaw request.
|
||||
*
|
||||
@@ -85,7 +130,18 @@ export function extractLatestUserMessage(req: BridgeInboundRequest): string {
|
||||
for (let i = userMessages.length - 1; i >= 0; i -= 1) {
|
||||
const raw = messageText(userMessages[i]);
|
||||
if (!raw) continue;
|
||||
if (isRuntimeContextMessage(raw)) continue;
|
||||
// First try the separate-message convention (Discord/Telegram path): a
|
||||
// user message whose entire body is one metadata envelope — skip whole.
|
||||
if (isRuntimeContextMessage(raw)) {
|
||||
// ...but also try inline-prefixed-metadata recovery (Fabric path):
|
||||
// some channels splice metadata + real content into one body. Walk
|
||||
// past leading sentinel blocks; if any non-metadata block follows,
|
||||
// that's the prompt. If nothing follows, this turn really is pure
|
||||
// metadata and we keep skipping.
|
||||
const stripped = stripPrefixedMetadataBlocks(raw);
|
||||
if (stripped) return stripOpenClawTimestampPrefix(stripped);
|
||||
continue;
|
||||
}
|
||||
return stripOpenClawTimestampPrefix(raw);
|
||||
}
|
||||
return "";
|
||||
|
||||
Reference in New Issue
Block a user