mirror of
https://github.com/block/buzz.git
synced 2026-08-18 06:50:31 +02:00
feat(agent): emit friendly tool titles with async fast-pass summaries
Port the Goose two-stage tool-title pattern for individual tool calls. emit_pending now sends a deterministic friendly title (server: tool plus a detail excerpt from path/command/query-style args, 60-char first-line cap, multibyte-safe) alongside an explicit toolName so the desktop never loses tool identity. For runnable tools a fire-and-forget fast pass asks the LLM (64 output tokens, 300-char args excerpt, one retry) for a friendlier title and publishes it as a title-only tool_call_update tagged _meta.buzz.toolSummary — deliberately without a status field so it can never clobber pending/executing/terminal state. Cancel-aware via the session watch channel; silent on failure. Kill switch BUZZ_AGENT_NO_TOOL_SUMMARY=1 disables the fast pass; model override via BUZZ_AGENT_TOOL_SUMMARY_MODEL. The regression harness sets the kill switch since those tests queue exact LLM response sequences and count requests; golden_transcripts.rs carries the explicit summary-path coverage (pending title + toolName, title-only async update, kill switch). No chain summaries in this slice. Co-authored-by: Taylor Ho <taylorkmho@gmail.com> Signed-off-by: Taylor Ho <taylorkmho@gmail.com>
This commit is contained in:
npub1223z34hd7vtwc6qj4s7flsxkj644nlre2nthu7lrrmkumhu3xddsrx9r6w
co-authored by
Taylor Ho
Taylor Ho
parent
cc42a49799
commit
b4b66f8ecb
@@ -28,7 +28,7 @@ pub struct RunCtx<'a> {
|
||||
pub effective_model: &'a str,
|
||||
pub session_id: &'a str,
|
||||
pub system_prompt: &'a str,
|
||||
pub llm: &'a Llm,
|
||||
pub llm: &'a Arc<Llm>,
|
||||
pub mcp: &'a Arc<McpRegistry>,
|
||||
/// Skills discovered at session creation; used by the built-in `load_skill` tool.
|
||||
pub skills: &'a [SkillEntry],
|
||||
@@ -333,6 +333,20 @@ impl RunCtx<'_> {
|
||||
results[idx] = Some(synthetic_tool_result(call, err));
|
||||
continue;
|
||||
}
|
||||
// Second stage of the friendly-title flow: fire-and-forget fast
|
||||
// summarization for real (runnable) tool calls only — fast-failed
|
||||
// and built-in calls keep their deterministic title.
|
||||
if self.cfg.tool_summary_enabled {
|
||||
spawn_tool_summary(
|
||||
Arc::clone(self.llm),
|
||||
self.cfg.clone(),
|
||||
self.effective_model.to_owned(),
|
||||
self.wire.clone(),
|
||||
self.session_id.to_owned(),
|
||||
call.clone(),
|
||||
self.cancel.clone(),
|
||||
);
|
||||
}
|
||||
runnable.push(idx);
|
||||
}
|
||||
|
||||
@@ -557,7 +571,11 @@ async fn emit_pending(wire: &WireSender, sid: &str, call: &ToolCall) {
|
||||
json!({
|
||||
"sessionUpdate": "tool_call",
|
||||
"toolCallId": call.provider_id,
|
||||
"title": call.name,
|
||||
// Deterministic friendly title, available immediately. The
|
||||
// exact tool identity rides alongside in `toolName` so
|
||||
// clients never lose it to the friendlier phrasing.
|
||||
"title": friendly_tool_title(&call.name, &call.arguments),
|
||||
"toolName": call.name,
|
||||
"kind": "other",
|
||||
"status": "pending",
|
||||
"rawInput": call.arguments,
|
||||
@@ -567,6 +585,175 @@ async fn emit_pending(wire: &WireSender, sid: &str, call: &ToolCall) {
|
||||
.await;
|
||||
}
|
||||
|
||||
/// Format a qualified tool name (`server__tool`) as a human-readable base
|
||||
/// label. Mirrors goose's `format_tool_name`.
|
||||
fn format_tool_name(tool_name: &str) -> String {
|
||||
if let Some((server, tool)) = tool_name.split_once("__") {
|
||||
format!("{}: {}", server.replace('_', " "), tool.replace('_', " "))
|
||||
} else {
|
||||
tool_name.replace('_', " ")
|
||||
}
|
||||
}
|
||||
|
||||
/// Build a short deterministic title from the tool name plus the most useful
|
||||
/// argument value (file path, command, query, url, etc.). Port of goose's
|
||||
/// `summarize_tool_call` fallback-title builder.
|
||||
fn friendly_tool_title(tool_name: &str, arguments: &serde_json::Value) -> String {
|
||||
const DETAIL_KEYS: [&str; 9] = [
|
||||
"path", "file", "command", "query", "url", "uri", "name", "pattern", "source",
|
||||
];
|
||||
const MAX_DETAIL_CHARS: usize = 60;
|
||||
|
||||
let base = format_tool_name(tool_name);
|
||||
let detail = arguments.as_object().and_then(|obj| {
|
||||
for key in &DETAIL_KEYS {
|
||||
if let Some(v) = obj.get(*key) {
|
||||
let s = match v {
|
||||
serde_json::Value::String(s) => s.clone(),
|
||||
other => other.to_string(),
|
||||
};
|
||||
if !s.is_empty() {
|
||||
let first_line = s.lines().next().unwrap_or(&s);
|
||||
let mut out: String = first_line.chars().take(MAX_DETAIL_CHARS).collect();
|
||||
if first_line.chars().count() > MAX_DETAIL_CHARS {
|
||||
out.push('…');
|
||||
}
|
||||
return Some(out);
|
||||
}
|
||||
}
|
||||
}
|
||||
None
|
||||
});
|
||||
match detail {
|
||||
Some(d) => format!("{base} · {d}"),
|
||||
None => base,
|
||||
}
|
||||
}
|
||||
|
||||
const TOOL_SUMMARY_SYSTEM_PROMPT: &str =
|
||||
"Summarize this tool call in a short lowercase phrase (3-8 words). \
|
||||
No punctuation. No quotes. Examples: reading project configuration, \
|
||||
checking network connectivity, listing files in src directory";
|
||||
|
||||
/// Max output tokens for a tool-title summary — a short phrase, not prose.
|
||||
const TOOL_SUMMARY_MAX_OUTPUT_TOKENS: u32 = 64;
|
||||
|
||||
/// Cap on the serialized-arguments excerpt sent to the fast model.
|
||||
const TOOL_SUMMARY_MAX_ARGS_CHARS: usize = 300;
|
||||
|
||||
/// Cap on the accepted summary phrase; longer responses are rejected as
|
||||
/// non-conforming rather than truncated mid-thought.
|
||||
const TOOL_SUMMARY_MAX_TITLE_CHARS: usize = 80;
|
||||
|
||||
/// Normalize a fast-model response into a row-label phrase, or `None` when
|
||||
/// the response is unusable (empty, multi-paragraph rambling, oversized).
|
||||
fn sanitize_tool_summary(raw: &str) -> Option<String> {
|
||||
let first_line = raw.trim().lines().next()?.trim();
|
||||
let cleaned = first_line.trim_matches(|c| c == '"' || c == '\'' || c == '`');
|
||||
if cleaned.is_empty() || cleaned.chars().count() > TOOL_SUMMARY_MAX_TITLE_CHARS {
|
||||
return None;
|
||||
}
|
||||
Some(cleaned.to_string())
|
||||
}
|
||||
|
||||
/// Fire-and-forget fast summarization pass for one tool call (goose's
|
||||
/// two-stage title pattern). Publishes a title-only `tool_call_update`
|
||||
/// tagged `_meta.buzz.toolSummary` — deliberately no `status` field, so a
|
||||
/// late summary can never clobber pending/executing/terminal state. Failures
|
||||
/// are silent: the deterministic title from `emit_pending` stays.
|
||||
fn spawn_tool_summary(
|
||||
llm: Arc<Llm>,
|
||||
cfg: Config,
|
||||
model: String,
|
||||
wire: WireSender,
|
||||
sid: String,
|
||||
call: ToolCall,
|
||||
cancel: watch::Receiver<bool>,
|
||||
) {
|
||||
tokio::spawn(async move {
|
||||
let args_json = {
|
||||
let s = call.arguments.to_string();
|
||||
if s.chars().count() > TOOL_SUMMARY_MAX_ARGS_CHARS {
|
||||
let mut t: String = s.chars().take(TOOL_SUMMARY_MAX_ARGS_CHARS).collect();
|
||||
t.push('…');
|
||||
t
|
||||
} else {
|
||||
s
|
||||
}
|
||||
};
|
||||
let user_prompt = format!("Tool: {}\nArguments: {args_json}", call.name);
|
||||
let effective_model = cfg.tool_summary_model.as_deref().unwrap_or(&model);
|
||||
|
||||
// The fast model occasionally returns an empty/errored response under
|
||||
// load. One retry with a short backoff recovers the common cases.
|
||||
let mut title: Option<String> = None;
|
||||
for attempt in 0..2 {
|
||||
if *cancel.borrow() {
|
||||
return;
|
||||
}
|
||||
match llm
|
||||
.summarize(
|
||||
&cfg,
|
||||
TOOL_SUMMARY_SYSTEM_PROMPT,
|
||||
&user_prompt,
|
||||
TOOL_SUMMARY_MAX_OUTPUT_TOKENS,
|
||||
effective_model,
|
||||
)
|
||||
.await
|
||||
{
|
||||
Ok(s) => {
|
||||
if let Some(clean) = sanitize_tool_summary(&s) {
|
||||
title = Some(clean);
|
||||
break;
|
||||
}
|
||||
if attempt == 0 {
|
||||
tracing::debug!(
|
||||
"tool summary: empty/unusable response for {} ({}), retrying once",
|
||||
call.provider_id,
|
||||
call.name
|
||||
);
|
||||
tokio::time::sleep(std::time::Duration::from_millis(150)).await;
|
||||
}
|
||||
}
|
||||
Err(e) => {
|
||||
if attempt == 0 {
|
||||
tracing::debug!(
|
||||
"tool summary: fast pass errored for {} ({}): {e}, retrying once",
|
||||
call.provider_id,
|
||||
call.name
|
||||
);
|
||||
tokio::time::sleep(std::time::Duration::from_millis(150)).await;
|
||||
} else {
|
||||
tracing::debug!(
|
||||
"tool summary: fast pass errored for {} ({}) after retry: {e}",
|
||||
call.provider_id,
|
||||
call.name
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
let Some(title) = title else { return };
|
||||
if *cancel.borrow() {
|
||||
return;
|
||||
}
|
||||
wire::send(
|
||||
&wire,
|
||||
wire::session_update(
|
||||
&sid,
|
||||
json!({
|
||||
"sessionUpdate": "tool_call_update",
|
||||
"toolCallId": call.provider_id,
|
||||
"title": title,
|
||||
"toolName": call.name,
|
||||
"_meta": { "buzz": { "toolSummary": true } },
|
||||
}),
|
||||
),
|
||||
)
|
||||
.await;
|
||||
});
|
||||
}
|
||||
|
||||
async fn emit_in_progress(wire: &WireSender, sid: &str, call: &ToolCall) {
|
||||
wire::send(
|
||||
wire,
|
||||
@@ -744,3 +931,72 @@ fn map_stop(p: ProviderStop) -> StopReason {
|
||||
ProviderStop::Refusal => StopReason::Refusal,
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::{friendly_tool_title, sanitize_tool_summary};
|
||||
use serde_json::json;
|
||||
|
||||
#[test]
|
||||
fn friendly_title_formats_qualified_name_with_detail() {
|
||||
assert_eq!(
|
||||
friendly_tool_title("developer__shell", &json!({ "command": "git status" })),
|
||||
"developer: shell · git status"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn friendly_title_prefers_path_over_later_keys() {
|
||||
assert_eq!(
|
||||
friendly_tool_title(
|
||||
"fs__read_file",
|
||||
&json!({ "name": "x", "path": "/tmp/a.txt" })
|
||||
),
|
||||
"fs: read file · /tmp/a.txt"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn friendly_title_without_useful_args_is_just_the_name() {
|
||||
assert_eq!(friendly_tool_title("do_thing", &json!({})), "do thing");
|
||||
assert_eq!(friendly_tool_title("do_thing", &json!(null)), "do thing");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn friendly_title_truncates_long_first_line_only() {
|
||||
let long = "x".repeat(200);
|
||||
let title = friendly_tool_title("t", &json!({ "command": format!("{long}\nsecond") }));
|
||||
assert!(title.ends_with('…'));
|
||||
// "t · " + 60 chars + ellipsis.
|
||||
assert_eq!(title.chars().count(), 4 + 60 + 1);
|
||||
assert!(!title.contains("second"));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn friendly_title_handles_multibyte_without_panicking() {
|
||||
let title = friendly_tool_title("t", &json!({ "query": "héllo wörld 🚀".repeat(20) }));
|
||||
assert!(title.starts_with("t · "));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn sanitize_accepts_short_phrase_and_strips_quotes() {
|
||||
assert_eq!(
|
||||
sanitize_tool_summary("\"checking repository state\"\n"),
|
||||
Some("checking repository state".to_string())
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn sanitize_rejects_empty_and_oversized() {
|
||||
assert_eq!(sanitize_tool_summary(" \n "), None);
|
||||
assert_eq!(sanitize_tool_summary(&"x".repeat(200)), None);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn sanitize_keeps_only_first_line() {
|
||||
assert_eq!(
|
||||
sanitize_tool_summary("reading config\nextra rambling"),
|
||||
Some("reading config".to_string())
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -637,6 +637,15 @@ pub struct Config {
|
||||
/// Thinking/reasoning effort level. `None` = use provider default (no
|
||||
/// thinking config sent). Set via `BUZZ_AGENT_THINKING_EFFORT`.
|
||||
pub thinking_effort: Option<ThinkingEffort>,
|
||||
/// Per-tool friendly title summarization: after emitting a tool call,
|
||||
/// spawn a fast LLM pass that publishes a short human phrase as a
|
||||
/// title-only `tool_call_update`. Default on; disable via
|
||||
/// `BUZZ_AGENT_NO_TOOL_SUMMARY=1`.
|
||||
pub tool_summary_enabled: bool,
|
||||
/// Optional model override for tool title summaries (a cheaper/faster
|
||||
/// model than the session model). Set via `BUZZ_AGENT_TOOL_SUMMARY_MODEL`;
|
||||
/// `None` falls back to the session's effective model.
|
||||
pub tool_summary_model: Option<String>,
|
||||
}
|
||||
|
||||
impl Config {
|
||||
@@ -731,6 +740,8 @@ impl Config {
|
||||
hook_servers: parse_hook_servers_env("MCP_HOOK_SERVERS"),
|
||||
hints_enabled: parse_env("BUZZ_AGENT_NO_HINTS", 0u8)? == 0,
|
||||
thinking_effort: parse_thinking_effort(env("BUZZ_AGENT_THINKING_EFFORT").as_deref())?,
|
||||
tool_summary_enabled: parse_env("BUZZ_AGENT_NO_TOOL_SUMMARY", 0u8)? == 0,
|
||||
tool_summary_model: env("BUZZ_AGENT_TOOL_SUMMARY_MODEL"),
|
||||
};
|
||||
cfg.validate()?;
|
||||
Ok(cfg)
|
||||
@@ -771,6 +782,8 @@ impl Config {
|
||||
hook_servers: HookServers::None,
|
||||
hints_enabled: false,
|
||||
thinking_effort: None,
|
||||
tool_summary_enabled: false,
|
||||
tool_summary_model: None,
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -1193,6 +1193,8 @@ mod tests {
|
||||
openai_api: OpenAiApi::Chat,
|
||||
hints_enabled: true,
|
||||
thinking_effort: None,
|
||||
tool_summary_enabled: false,
|
||||
tool_summary_model: None,
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -808,3 +808,198 @@ async fn test_cancel_notification_no_reply() {
|
||||
|
||||
h.shutdown().await;
|
||||
}
|
||||
|
||||
/// Per-tool friendly titles, stage 1: the immediate `tool_call` (pending)
|
||||
/// update must carry the deterministic friendly title plus the exact tool
|
||||
/// identity in `toolName`, and receipts in `rawInput`.
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn test_pending_tool_call_carries_friendly_title_and_tool_name() {
|
||||
let url = spawn_fake_llm(vec![
|
||||
openai_tool_call(
|
||||
"call_t1",
|
||||
"fake__do_thing",
|
||||
json!({ "command": "git status" }),
|
||||
),
|
||||
openai_text("done"),
|
||||
])
|
||||
.await;
|
||||
let mut h = Harness::spawn(&[("OPENAI_COMPAT_BASE_URL", &url)]).await;
|
||||
|
||||
let sid = handshake(&mut h).await;
|
||||
let p = h
|
||||
.send(
|
||||
"session/prompt",
|
||||
json!({
|
||||
"sessionId": sid,
|
||||
"prompt": [{ "type": "text", "text": "use the tool" }],
|
||||
}),
|
||||
)
|
||||
.await;
|
||||
|
||||
let pending = h
|
||||
.recv_until(|v| {
|
||||
v.get("method") == Some(&json!("session/update"))
|
||||
&& v["params"]["update"]["sessionUpdate"] == "tool_call"
|
||||
})
|
||||
.await;
|
||||
let update = &pending["params"]["update"];
|
||||
assert_eq!(update["title"], "fake: do thing · git status");
|
||||
assert_eq!(update["toolName"], "fake__do_thing");
|
||||
assert_eq!(update["status"], "pending");
|
||||
assert_eq!(update["rawInput"]["command"], "git status");
|
||||
|
||||
let final_resp = h.recv_for_id(p).await;
|
||||
assert_eq!(final_resp["result"]["stopReason"], "end_turn");
|
||||
h.shutdown().await;
|
||||
}
|
||||
|
||||
/// Per-tool friendly titles, stage 2: the async fast pass publishes a
|
||||
/// title-only `tool_call_update` tagged `_meta.buzz.toolSummary` with NO
|
||||
/// status field, so it can never regress pending/executing/terminal state.
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn test_async_tool_summary_publishes_title_only_update() {
|
||||
// Queue: round-1 tool call, then the fast summary response (the tool is
|
||||
// delayed 1s so the summary request reaches the fake LLM first), then the
|
||||
// round-2 final text.
|
||||
let url = spawn_fake_llm(vec![
|
||||
openai_tool_call(
|
||||
"call_s1",
|
||||
"fake__tool_0",
|
||||
json!({ "command": "git status" }),
|
||||
),
|
||||
openai_text("checking repository state"),
|
||||
openai_text("done"),
|
||||
])
|
||||
.await;
|
||||
let mut h = Harness::spawn(&[("OPENAI_COMPAT_BASE_URL", &url)]).await;
|
||||
|
||||
let init_id = h
|
||||
.send(
|
||||
"initialize",
|
||||
json!({ "protocolVersion": 2, "clientCapabilities": {} }),
|
||||
)
|
||||
.await;
|
||||
let _ = h.recv_for_id(init_id).await;
|
||||
let fake_mcp = env!("CARGO_BIN_EXE_fake-mcp");
|
||||
let new_id = h
|
||||
.send(
|
||||
"session/new",
|
||||
json!({
|
||||
"cwd": "/tmp",
|
||||
"mcpServers": [{
|
||||
"name": "fake",
|
||||
"command": fake_mcp,
|
||||
"args": [],
|
||||
"env": [
|
||||
{ "name": "FAKE_MCP_TOOL_COUNT", "value": "1" },
|
||||
{ "name": "FAKE_MCP_TOOL_DELAY", "value": "1" },
|
||||
],
|
||||
}],
|
||||
}),
|
||||
)
|
||||
.await;
|
||||
let new = h.recv_for_id(new_id).await;
|
||||
let sid = new["result"]["sessionId"].as_str().unwrap().to_owned();
|
||||
|
||||
let p = h
|
||||
.send(
|
||||
"session/prompt",
|
||||
json!({
|
||||
"sessionId": sid,
|
||||
"prompt": [{ "type": "text", "text": "use the tool" }],
|
||||
}),
|
||||
)
|
||||
.await;
|
||||
|
||||
let summary = h
|
||||
.recv_until(|v| {
|
||||
v.get("method") == Some(&json!("session/update"))
|
||||
&& v["params"]["update"]["sessionUpdate"] == "tool_call_update"
|
||||
&& v["params"]["update"]["_meta"]["buzz"]["toolSummary"] == json!(true)
|
||||
})
|
||||
.await;
|
||||
let update = &summary["params"]["update"];
|
||||
assert_eq!(update["toolCallId"], "call_s1");
|
||||
assert_eq!(update["title"], "checking repository state");
|
||||
assert_eq!(update["toolName"], "fake__tool_0");
|
||||
assert!(
|
||||
update.get("status").is_none(),
|
||||
"summary update must not carry a status: {update}"
|
||||
);
|
||||
|
||||
let final_resp = h.recv_for_id(p).await;
|
||||
assert_eq!(final_resp["result"]["stopReason"], "end_turn");
|
||||
h.shutdown().await;
|
||||
}
|
||||
|
||||
/// Kill switch: BUZZ_AGENT_NO_TOOL_SUMMARY=1 must suppress the fast pass
|
||||
/// entirely — no summary update on the wire, and no extra LLM request
|
||||
/// (the canned-queue accounting would break the turn if one fired).
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn test_tool_summary_kill_switch() {
|
||||
let url = spawn_fake_llm(vec![
|
||||
openai_tool_call("call_k1", "fake__tool_0", json!({ "command": "ls" })),
|
||||
openai_text("done"),
|
||||
])
|
||||
.await;
|
||||
let mut h = Harness::spawn(&[
|
||||
("OPENAI_COMPAT_BASE_URL", &url),
|
||||
("BUZZ_AGENT_NO_TOOL_SUMMARY", "1"),
|
||||
])
|
||||
.await;
|
||||
|
||||
let init_id = h
|
||||
.send(
|
||||
"initialize",
|
||||
json!({ "protocolVersion": 2, "clientCapabilities": {} }),
|
||||
)
|
||||
.await;
|
||||
let _ = h.recv_for_id(init_id).await;
|
||||
let fake_mcp = env!("CARGO_BIN_EXE_fake-mcp");
|
||||
let new_id = h
|
||||
.send(
|
||||
"session/new",
|
||||
json!({
|
||||
"cwd": "/tmp",
|
||||
"mcpServers": [{
|
||||
"name": "fake",
|
||||
"command": fake_mcp,
|
||||
"args": [],
|
||||
"env": [{ "name": "FAKE_MCP_TOOL_COUNT", "value": "1" }],
|
||||
}],
|
||||
}),
|
||||
)
|
||||
.await;
|
||||
let new = h.recv_for_id(new_id).await;
|
||||
let sid = new["result"]["sessionId"].as_str().unwrap().to_owned();
|
||||
|
||||
let p = h
|
||||
.send(
|
||||
"session/prompt",
|
||||
json!({
|
||||
"sessionId": sid,
|
||||
"prompt": [{ "type": "text", "text": "use the tool" }],
|
||||
}),
|
||||
)
|
||||
.await;
|
||||
|
||||
// Drain everything up to the final response; no notification along the
|
||||
// way may carry the toolSummary marker. If the fast pass had fired it
|
||||
// would also have consumed the "done" response and broken the turn.
|
||||
let mut saw_summary = false;
|
||||
let final_resp = loop {
|
||||
let v = h.recv().await;
|
||||
if v["params"]["update"]["_meta"]["buzz"]["toolSummary"] == json!(true) {
|
||||
saw_summary = true;
|
||||
}
|
||||
if v["id"] == json!(p) {
|
||||
break v;
|
||||
}
|
||||
};
|
||||
assert!(
|
||||
!saw_summary,
|
||||
"kill switch did not suppress the summary pass"
|
||||
);
|
||||
assert_eq!(final_resp["result"]["stopReason"], "end_turn");
|
||||
h.shutdown().await;
|
||||
}
|
||||
|
||||
@@ -102,7 +102,12 @@ impl Harness {
|
||||
.env("BUZZ_AGENT_LLM_TIMEOUT_SECS", "5")
|
||||
.env("BUZZ_AGENT_TOOL_TIMEOUT_SECS", "5")
|
||||
.env("BUZZ_AGENT_MAX_ROUNDS", "8")
|
||||
.env("BUZZ_AGENT_MCP_INIT_TIMEOUT_SECS", "2");
|
||||
.env("BUZZ_AGENT_MCP_INIT_TIMEOUT_SECS", "2")
|
||||
// These tests queue exact LLM response sequences and count
|
||||
// requests; the async tool-summary fast pass would consume
|
||||
// queued responses nondeterministically. Covered explicitly
|
||||
// in golden_transcripts.rs instead.
|
||||
.env("BUZZ_AGENT_NO_TOOL_SUMMARY", "1");
|
||||
for (k, v) in extra {
|
||||
cmd.env(k, v);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user