feat(web): add follow-up suggestion chips and ghost text (#1156)

* feat(web): add follow-up suggestion chips and ghost text to chat UI

The LLM now always generates 1-3 follow-up command suggestions via
<suggestions> tags in its response. These are extracted server-side,
broadcast as SSE events, and rendered as clickable chips above the
chat input. The first suggestion also appears as ghost text in the
input field (Tab to accept). Includes debug logging for LLM responses
in the agentic loop and removes noisy NEAR AI status logging.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: resolve deferred review items from PR #1156 [skip-regression-check]

- Remove literal backslashes from raw string prompt (reasoning.rs)
- Make WASM channels skip Suggestions status (no-op instead of empty callback)
- Add !e.shiftKey guard to Tab-to-accept ghost text handler
- Cap extracted suggestions at 3 and trim whitespace-only entries
- Extract suggestions in approval-resume path (prevents tag leaking)
- Remove stale .has-ghost class during showSuggestionChips reset

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>
This commit is contained in:
Illia Polosukhin
2026-03-14 18:57:24 +00:00
committed by GitHub
co-authored by Claude Opus 4.6
parent f9b880c2e9
commit 757d24bd90
14 changed files with 368 additions and 35 deletions
+24
View File
@@ -152,6 +152,30 @@ pub async fn run_agentic_loop(
// Call LLM
let output = delegate.call_llm(reasoning, reason_ctx, iteration).await?;
match &output.result {
RespondResult::Text(text) => {
tracing::debug!(
iteration,
len = text.len(),
has_suggestions = text.contains("<suggestions>"),
response = %text,
"LLM text response"
);
}
RespondResult::ToolCalls {
tool_calls,
content,
} => {
let names: Vec<&str> = tool_calls.iter().map(|tc| tc.name.as_str()).collect();
tracing::debug!(
iteration,
tools = ?names,
has_content = content.is_some(),
"LLM tool_calls response"
);
}
}
match output.result {
RespondResult::Text(text) => {
// Tool intent nudge: if the LLM says "let me search..." without
+97
View File
@@ -1051,6 +1051,54 @@ fn strip_internal_tool_call_text(text: &str) -> String {
}
}
/// Extract `<suggestions>["...","..."]</suggestions>` from a response string.
///
/// Returns `(cleaned_text, suggestions)`. The `<suggestions>` block is stripped
/// from the text regardless of whether the JSON inside parses successfully.
/// Only the **last** `<suggestions>` block is used (closest to end of response).
/// Blocks inside markdown code fences are ignored.
pub(crate) fn extract_suggestions(text: &str) -> (String, Vec<String>) {
use regex::Regex;
use std::sync::LazyLock;
static RE: LazyLock<Regex> = LazyLock::new(|| {
Regex::new(r"(?s)<suggestions>\s*(.*?)\s*</suggestions>").expect("valid regex") // safety: constant pattern
});
// Find the position of the last closing code fence to avoid matching inside code blocks
let last_code_fence = text.rfind("```").unwrap_or(0);
// Find all matches, take the last one that's after the last code fence
let mut best_match: Option<regex::Match<'_>> = None;
let mut best_capture: Option<String> = None;
for caps in RE.captures_iter(text) {
if let (Some(full), Some(inner)) = (caps.get(0), caps.get(1))
&& full.start() >= last_code_fence
{
best_match = Some(full);
best_capture = Some(inner.as_str().to_string());
}
}
let Some(full) = best_match else {
return (text.to_string(), Vec::new());
};
let cleaned = format!("{}{}", &text[..full.start()], &text[full.end()..]); // safety: regex match boundaries are valid UTF-8
let cleaned = cleaned.trim().to_string();
// Parse the JSON array
let suggestions = best_capture
.and_then(|json| serde_json::from_str::<Vec<String>>(&json).ok())
.unwrap_or_default()
.into_iter()
.filter(|s| !s.trim().is_empty() && s.len() <= 80)
.take(3)
.collect();
(cleaned, suggestions)
}
#[cfg(test)]
mod tests {
use std::sync::Arc;
@@ -2197,6 +2245,55 @@ mod tests {
assert_eq!(result, input);
}
#[test]
fn test_extract_suggestions_basic() {
let input = "Here is my answer.\n<suggestions>[\"Check logs\", \"Deploy\"]</suggestions>";
let (text, suggestions) = super::extract_suggestions(input);
assert_eq!(text, "Here is my answer."); // safety: test
assert_eq!(suggestions, vec!["Check logs", "Deploy"]); // safety: test
}
#[test]
fn test_extract_suggestions_no_tag() {
let input = "Just a plain response.";
let (text, suggestions) = super::extract_suggestions(input);
assert_eq!(text, "Just a plain response."); // safety: test
assert!(suggestions.is_empty()); // safety: test
}
#[test]
fn test_extract_suggestions_malformed_json() {
let input = "Answer.\n<suggestions>not json</suggestions>";
let (text, suggestions) = super::extract_suggestions(input);
assert_eq!(text, "Answer."); // safety: test
assert!(suggestions.is_empty()); // safety: test
}
#[test]
fn test_extract_suggestions_inside_code_fence() {
let input = "```\n<suggestions>[\"foo\"]</suggestions>\n```";
let (text, suggestions) = super::extract_suggestions(input);
// The tag is inside a code fence, so it should not be extracted
assert_eq!(text, input); // safety: test
assert!(suggestions.is_empty()); // safety: test
}
#[test]
fn test_extract_suggestions_after_code_fence() {
let input = "```\ncode\n```\nAnswer.\n<suggestions>[\"foo\"]</suggestions>";
let (text, suggestions) = super::extract_suggestions(input);
assert_eq!(text, "```\ncode\n```\nAnswer."); // safety: test
assert_eq!(suggestions, vec!["foo"]); // safety: test
}
#[test]
fn test_extract_suggestions_filters_long() {
let long = "x".repeat(81);
let input = format!("Answer.\n<suggestions>[\"{}\", \"ok\"]</suggestions>", long);
let (_, suggestions) = super::extract_suggestions(&input);
assert_eq!(suggestions, vec!["ok"]); // safety: test
}
#[test]
fn test_tool_error_format_includes_tool_name() {
// Regression test for issue #487: tool errors sent to the LLM should
+28
View File
@@ -420,6 +420,10 @@ impl Agent {
// Complete, fail, or request approval
match result {
Ok(AgenticLoopResult::Response(response)) => {
// Extract <suggestions> from response text before user sees it
let (response, suggestions) =
crate::agent::dispatcher::extract_suggestions(&response);
// Hook: TransformResponse — allow hooks to modify or reject the final response
let response = {
let event = crate::hooks::HookEvent::ResponseTransform {
@@ -473,6 +477,18 @@ impl Agent {
)
.await;
// Send suggestions after response (best-effort, rendered by web gateway)
if !suggestions.is_empty() {
let _ = self
.channels
.send_status(
&message.channel,
StatusUpdate::Suggestions { suggestions },
&message.metadata,
)
.await;
}
Ok(SubmissionResult::response(response))
}
Ok(AgenticLoopResult::NeedApproval { pending }) => {
@@ -1334,6 +1350,8 @@ impl Agent {
match result {
Ok(AgenticLoopResult::Response(response)) => {
let (response, suggestions) =
crate::agent::dispatcher::extract_suggestions(&response);
thread.complete_turn(&response);
let (turn_number, tool_calls) = thread
.turns
@@ -1364,6 +1382,16 @@ impl Agent {
&message.metadata,
)
.await;
if !suggestions.is_empty() {
let _ = self
.channels
.send_status(
&message.channel,
StatusUpdate::Suggestions { suggestions },
&message.metadata,
)
.await;
}
Ok(SubmissionResult::response(response))
}
Ok(AgenticLoopResult::NeedApproval {