includes check can produce false-positive replacements for true deltas

nextText.includes(streamText) treats nextText as a full cumulative payload whenever streamText appears as a substring anywhere inside it — including as part of an unrelated word. This will misfire for true delta chunks if the short accumulated streamText appears as a substring inside the incoming delta.

For example:

• streamText = "is" (early in the stream, only two characters accumulated)
• nextText = "This result is important" (a genuine delta to be appended)
• "This result is important".includes("is") → true ("is" appears in "Th**is**" and standalone)
• Result: streamText is silently replaced with the delta text, permanently dropping everything before "is"

The same false-positive risk applies to any common short word or phrase that may appear as a substring in subsequent delta chunks. The risk is highest early in a streaming session when streamText is still short.

A more conservative guard — e.g. only triggering when nextText is meaningfully longer than streamText, or requiring the match to be in the second half of nextText — would reduce false positives. For instance:

// Require nextText to end with streamText, or that streamText occupies
// at least half of nextText, to reduce false-positive substring matches.
if (
  nextText.includes(streamText) &&
  nextText.length >= streamText.length * 2
) {
  streamText = nextText;
  return;
}

Or alternatively, since the bug is triggered by leading whitespace/directive stripping, trimming both sides before comparing is a more targeted fix:

if (nextText.trimStart().startsWith(streamText.trimStart())) {
  streamText = nextText;
  return;
}

This is a comment left during a code review.
Path: extensions/feishu/src/reply-dispatcher.ts
Line: 163-166

Comment:
**`includes` check can produce false-positive replacements for true deltas**

`nextText.includes(streamText)` treats `nextText` as a full cumulative payload whenever `streamText` appears as a substring anywhere inside it — including as part of an unrelated word. This will misfire for true delta chunks if the short accumulated `streamText` appears as a substring inside the incoming delta.

For example:
- `streamText = "is"` (early in the stream, only two characters accumulated)
- `nextText = "This result is important"` (a genuine delta to be appended)
- `"This result is important".includes("is")` → `true` (`"is"` appears in `"Th**is**"` and standalone)
- Result: `streamText` is silently **replaced** with the delta text, permanently dropping everything before `"is"`

The same false-positive risk applies to any common short word or phrase that may appear as a substring in subsequent delta chunks. The risk is highest early in a streaming session when `streamText` is still short.

A more conservative guard — e.g. only triggering when `nextText` is meaningfully longer than `streamText`, or requiring the match to be in the second half of `nextText` — would reduce false positives. For instance:

```ts
// Require nextText to end with streamText, or that streamText occupies
// at least half of nextText, to reduce false-positive substring matches.
if (
  nextText.includes(streamText) &&
  nextText.length >= streamText.length * 2
) {
  streamText = nextText;
  return;
}
```

Or alternatively, since the bug is triggered by leading whitespace/directive stripping, trimming both sides before comparing is a more targeted fix:

```ts
if (nextText.trimStart().startsWith(streamText.trimStart())) {
  streamText = nextText;
  return;
}
```

How can I resolve this? If you propose a fix, please make it concise.

Avoid replacing stream text on arbitrary substring matches

Using nextText.includes(streamText) to classify cumulative partials can drop already-streamed content when the upstream sends true deltas that happen to repeat prior output. In that case a delta chunk like "abc def" after "abc" matches this condition, so streamText is replaced instead of appended and the earlier prefix is lost in the card update/final close path. This regression appears whenever partial payloads are delta-style (which this function still claims to support via the append fallback), so the replacement check should be constrained to a safer cumulative-only signal.

Useful? React with 👍 / 👎.

Tighten reverse-prefix dedupe before dropping partial chunks

The new streamText.trimStart().startsWith(nextText.trimStart()) fast-path can suppress valid delta chunks whenever a new token happens to match the beginning of the response (or is whitespace-only after trimming). For example, after streamText = "Hello there. ", an incoming delta "Hello" is dropped because "Hello there. ".trimStart().startsWith("Hello") is true, so the repeated word is never appended; the same happens for "\n" because trimming yields an empty string. This regresses content accuracy in streaming updates and can leak into the final card text when the flow closes from partials without a final payload.

Useful? React with 👍 / 👎.

Remove reverse-prefix suppression for streamed deltas

The new reverse-prefix guard drops valid delta chunks whenever a chunk starts with text that already appeared at the beginning of the response. In mergeStreamingText, streamText.trimStart().startsWith(nextText.trimStart()) returns true for inputs like existing "Hello there. " plus delta "Hello" (or whitespace-only chunks where trimStart() becomes empty), so the function returns early and never appends text; if the stream closes from partials, the final card content is missing those tokens.

Useful? React with 👍 / 👎.