feat(bash): 在终端标题栏展示最近一次 LLM call 输出速度 (tok/s)

.github/workflows/ci.yml

@@ -92,16 +92,15 @@ jobs:
           sudo apt-get update
           sudo apt-get install -y gcc-aarch64-linux-gnu
 
-      - name: Build goagent (linux, Debian 11 container)
+      - name: Build goagent (linux, Ubuntu 20.04 container)
         if: matrix.goos == 'linux'
         run: |
-          docker run --rm -v /opt/hostedtoolcache:/opt/hostedtoolcache -v "$PWD":/github/workspace -w /github/workspace debian:11 \
+          docker run --rm -v /opt/hostedtoolcache:/opt/hostedtoolcache -v "$PWD":/github/workspace -w /github/workspace ubuntu:20.04 \
             bash -euxo pipefail -c '
               GO_BIN="$(ls -d /opt/hostedtoolcache/go/*/x64/bin | sort -V | tail -1)/go"
               export PATH="$(dirname "$GO_BIN"):$PATH"
               export GOCACHE="$PWD/go/.gocache" GOMODCACHE="$PWD/go/.gomodcache"
               go version
-              sed -i "s|deb.debian.org|deb.debian.org|g; s|security.debian.org/debian-security|security.debian.org/debian-security|g" /etc/apt/sources.list 2>/dev/null || true
               apt-get update
               apt-get install -y --no-install-recommends gcc make libc6-dev
               mkdir -p dist
@@ -181,12 +180,11 @@ jobs:
           cd rust && cargo build --release --target ${{ matrix.target }}
           cp target/${{ matrix.target }}/release/${{ matrix.bin }} ../dist/rustagent-${{ matrix.suffix }}${{ matrix.ext }}
 
-      - name: Build rustagent (linux, Debian 11 container)
+      - name: Build rustagent (linux, Ubuntu 20.04 container)
         if: matrix.runner == 'ubuntu-latest'
         run: |
-          docker run --rm -v "$PWD":/github/workspace -w /github/workspace debian:11 \
+          docker run --rm -v "$PWD":/github/workspace -w /github/workspace ubuntu:20.04 \
             bash -euxo pipefail -c '
-              sed -i "s|deb.debian.org|deb.debian.org|g" /etc/apt/sources.list 2>/dev/null || true
               apt-get update
               apt-get install -y --no-install-recommends curl ca-certificates build-essential pkg-config
               if [ "${{ matrix.target }}" = "aarch64-unknown-linux-gnu" ]; then
@@ -257,12 +255,11 @@ jobs:
           cd webagent && cargo build --release --target ${{ matrix.target }}
           cp target/${{ matrix.target }}/release/${{ matrix.bin }} ../dist/webagent-${{ matrix.suffix }}${{ matrix.ext }}
 
-      - name: Build webagent (linux, Debian 11 container)
+      - name: Build webagent (linux, Ubuntu 20.04 container)
         if: matrix.runner == 'ubuntu-latest'
         run: |
-          docker run --rm -v "$PWD":/github/workspace -w /github/workspace debian:11 \
+          docker run --rm -v "$PWD":/github/workspace -w /github/workspace ubuntu:20.04 \
             bash -euxo pipefail -c '
-              sed -i "s|deb.debian.org|deb.debian.org|g" /etc/apt/sources.list 2>/dev/null || true
               apt-get update
               apt-get install -y --no-install-recommends curl ca-certificates build-essential pkg-config
               if [ "${{ matrix.target }}" = "aarch64-unknown-linux-gnu" ]; then
@@ -313,12 +310,11 @@ jobs:
           sudo apt-get update
           sudo apt-get install -y ripgrep gcc-aarch64-linux-gnu
 
-      - name: Build cagent (linux, Debian 11 container)
+      - name: Build cagent (linux, Ubuntu 20.04 container)
         if: runner.os == 'Linux'
         run: |
-          docker run --rm -v "$PWD":/github/workspace -w /github/workspace debian:11 \
+          docker run --rm -v "$PWD":/github/workspace -w /github/workspace ubuntu:20.04 \
             bash -euxo pipefail -c '
-              sed -i "s|deb.debian.org|deb.debian.org|g" /etc/apt/sources.list 2>/dev/null || true
               apt-get update
               apt-get install -y --no-install-recommends gcc make xxd libc6-dev libcurl4-openssl-dev pkg-config
               mkdir -p dist

c/agent.c

@@ -171,6 +171,9 @@ static void stream_display_callback(void *ctx, const SseEvent *evt) {
         if (evt->out_tokens > 0) sctx->accum.out_tokens = evt->out_tokens;
         if (evt->cache_read_tokens > 0) sctx->accum.cache_read_tokens = evt->cache_read_tokens;
         if (evt->cache_creation_tokens > 0) sctx->accum.cache_creation_tokens = evt->cache_creation_tokens;
+        sctx->accum.start_ms = evt->start_ms;
+        sctx->accum.end_ms = evt->end_ms;
+        sctx->accum.speed_ready = evt->speed_ready;
         break;
 
     case SSE_TOOL_CALL_START: {
@@ -222,6 +225,7 @@ static void stream_display_callback(void *ctx, const SseEvent *evt) {
         }
         sctx->accum.tool_count = 0;
         sctx->accum.stopped = 0;
+        sctx->accum.speed_ready = 0;
         if (sctx->accum.stop_reason) { free(sctx->accum.stop_reason); sctx->accum.stop_reason = NULL; }
         if (sctx->accum.error) { free(sctx->accum.error); sctx->accum.error = NULL; }
         sctx->accum.in_tokens = 0;
@@ -1001,6 +1005,8 @@ int agent_loop(Agent *agent, const char *user_input, const char *turn_kind) {
         agent->last_cache_creation_tokens = accum->cache_creation_tokens;
         agent->last_context_tokens = accum->in_tokens + accum->out_tokens +
             accum->cache_read_tokens + accum->cache_creation_tokens;
+        agent->last_start_ms = accum->start_ms;
+        agent->last_end_ms = accum->end_ms;
 
         if (accum->in_tokens > 0 || accum->out_tokens > 0) {
             DisplayMessage *dm = malloc(sizeof(DisplayMessage));
@@ -1027,9 +1033,17 @@ int agent_loop(Agent *agent, const char *user_input, const char *turn_kind) {
                 store_stats_set_int_file(agent->paths.stats, "current_context_tokens",
                     agent->last_context_tokens);
             }
-            /* 对齐 bash 版 store_stats_update 末尾的 display_term_title */
-            agent_update_title(agent);
         }
+        /* 仅完整协议终结且 HTTP 成功后的最终 USAGE 才更新速度。
+         * 独立于正 token 守卫：零用量成功需要清零，失败和中断保留旧值。 */
+        if (accum->speed_ready && !accum->error && !agent->interrupted) {
+            long long dur = agent->last_end_ms - agent->last_start_ms;
+            int speed = (dur > 0) ? (int)((long long)agent->last_output_tokens * 1000LL / dur) : 0;
+            store_stats_set_int_file(agent->paths.stats, "last_call_speed_tok_per_sec", speed);
+        }
+        /* 对齐 bash 版 store_stats_update 末尾的 display_term_title */
+        if (accum->in_tokens > 0 || accum->out_tokens > 0 || accum->speed_ready)
+            agent_update_title(agent);
 
         /* ---- 执行工具调用 ---- */
         const char **result_ids = NULL;
@@ -3019,27 +3033,29 @@ void agent_update_title_status(Agent *agent, const char *status) {
     int ao = json_get_int(jp.val, "total_output_tokens");
     long long ll_cr = json_get_ll(jp.val, "total_cache_read_tokens");
     int ct = json_get_int(jp.val, "current_context_tokens");
+    int speed = json_get_int(jp.val, "last_call_speed_tok_per_sec");
 
-    /* 对齐 bash 版 term_title.awk: model T:turn R:req I:in+cr(pct) O:out C:ctx */
+    /* 对齐 bash 版 term_title.awk: model T:turn R:req I:in+cr(pct) O:out C:ctx S:speedtok/s */
     long long total_i = ll_ai + ll_cr;
     int pct = (total_i > 0) ? (int)((ll_cr * 100.0 / total_i) + 0.5) : 0;
     char pct_s[32];
     if (total_i > 0) snprintf(pct_s, sizeof(pct_s), "%d%%", pct);
     else snprintf(pct_s, sizeof(pct_s), "—");
 
-    char tc_s[32], ar_s[32], total_i_s[32], ao_s[32], ct_s[32];
+    char tc_s[32], ar_s[32], total_i_s[32], ao_s[32], ct_s[32], speed_s[32];
     format_int_commas(tc, tc_s, sizeof(tc_s));
     format_int_commas(ar, ar_s, sizeof(ar_s));
     format_ll_commas(total_i, total_i_s, sizeof(total_i_s));
     format_int_commas(ao, ao_s, sizeof(ao_s));
     format_int_commas(ct, ct_s, sizeof(ct_s));
+    format_int_commas(speed, speed_s, sizeof(speed_s));
 
     int idle = (status && strcmp(status, "idle") == 0 && agent->active_task_count <= 0);
     const char *prefix = idle ? "" : "\xe2\x8f\xb3 ";
     int progress = idle ? 0 : 3;
     FILE *err = agent->err ? agent->err : stderr;
-    fprintf(err, "\x1b]0;%s%s T:%s R:%s I:%s(%s) O:%s C:%s\x07\x1b]9;4;%d\x07",
-            prefix, agent->model, tc_s, ar_s, total_i_s, pct_s, ao_s, ct_s, progress);
+    fprintf(err, "\x1b]0;%s%s T:%s R:%s I:%s(%s) O:%s C:%s S:%stok/s\x07\x1b]9;4;%d\x07",
+            prefix, agent->model, tc_s, ar_s, total_i_s, pct_s, ao_s, ct_s, speed_s, progress);
     fflush(err);
 
     free(stats_content);

c/agent.h

@@ -77,6 +77,8 @@ typedef struct {
     int last_output_tokens;
     int last_cache_read_tokens;
     int last_cache_creation_tokens;
+    long long last_start_ms;
+    long long last_end_ms;
     int active_task_count;
     int sub_agent_depth;      /* 主代理为 0，第一层子代理为 1 */
 

c/store.c

@@ -152,7 +152,7 @@ int store_session_init(const SessionPaths *p, int is_new) {
                    "\"total_input_tokens\":0,"
                    "\"total_output_tokens\":0,\"total_cache_read_tokens\":0,"
                    "\"total_cache_creation_tokens\":0,\"current_context_tokens\":0,"
-                   "\"last_updated\":\"\"}\n");
+                   "\"last_call_speed_tok_per_sec\":0,\"last_updated\":\"\"}\n");
         fclose(f);
 
         /* 写入 session_start 事件（与 bash 版对齐） */
@@ -543,18 +543,19 @@ static void stats_write_canonical(const char *path,
                                   int total_cache_read_tokens,
                                   int total_cache_creation_tokens,
                                   int current_context_tokens,
+                                  int last_call_speed_tok_per_sec,
                                   const char *last_updated) {
     StrBuf buf;
     sb_init(&buf);
     sb_appendf(&buf, "{\"current_turn_count\":%d,\"agent_request_count\":%d,"
                "\"compact_request_count\":%d,\"sub_agent_request_count\":%d,"
                "\"total_input_tokens\":%d,\"total_output_tokens\":%d,"
                "\"total_cache_read_tokens\":%d,\"total_cache_creation_tokens\":%d,"
-               "\"current_context_tokens\":%d,\"last_updated\":",
+               "\"current_context_tokens\":%d,\"last_call_speed_tok_per_sec\":%d,\"last_updated\":",
                current_turn_count, agent_request_count, compact_request_count,
                sub_agent_request_count, total_input_tokens, total_output_tokens,
                total_cache_read_tokens, total_cache_creation_tokens,
-               current_context_tokens);
+               current_context_tokens, last_call_speed_tok_per_sec);
     sb_append_json_string(&buf, last_updated ? last_updated : "");
     sb_append(&buf, "}\n");
     util_write_file(path, buf.data);
@@ -617,6 +618,7 @@ void store_stats_set_int_file(const char *path, const char *key, int value) {
     int total_input_tokens = 0, total_output_tokens = 0;
     int total_cache_read_tokens = 0, total_cache_creation_tokens = 0;
     int current_context_tokens = 0;
+    int last_call_speed_tok_per_sec = 0;
 
     char *content = store_stats_read(path);
     if (content && content[0]) {
@@ -631,6 +633,7 @@ void store_stats_set_int_file(const char *path, const char *key, int value) {
             total_cache_read_tokens = json_get_int(jp.val, "total_cache_read_tokens");
             total_cache_creation_tokens = json_get_int(jp.val, "total_cache_creation_tokens");
             current_context_tokens = json_get_int(jp.val, "current_context_tokens");
+            last_call_speed_tok_per_sec = json_get_int(jp.val, "last_call_speed_tok_per_sec");
         }
     }
 
@@ -643,6 +646,7 @@ void store_stats_set_int_file(const char *path, const char *key, int value) {
     else if (strcmp(key, "total_cache_read_tokens") == 0) total_cache_read_tokens = value;
     else if (strcmp(key, "total_cache_creation_tokens") == 0) total_cache_creation_tokens = value;
     else if (strcmp(key, "current_context_tokens") == 0) current_context_tokens = value;
+    else if (strcmp(key, "last_call_speed_tok_per_sec") == 0) last_call_speed_tok_per_sec = value;
 
     time_t now = time(NULL);
     struct tm tm_buf;
@@ -653,7 +657,7 @@ void store_stats_set_int_file(const char *path, const char *key, int value) {
                           compact_request_count, sub_agent_request_count,
                           total_input_tokens, total_output_tokens,
                           total_cache_read_tokens, total_cache_creation_tokens,
-                          current_context_tokens, ts);
+                          current_context_tokens, last_call_speed_tok_per_sec, ts);
     if (content) {
         free(content);
     }

c/test_transport.c

@@ -42,7 +42,80 @@ static void check(int condition, const char *label) {
     }
 }
 
+/* 走实际逐行解析器，分别检查提前 STOP、协议终结和最终 HTTP 门槛。 */
+static void feed_stream(StreamCtx *stream, const char *wire) {
+    check(stream_cb((char *)wire, 1, strlen(wire), stream) == strlen(wire), "流数据完整接收");
+}
+
+static void test_speed_completion(void) {
+    SseAccumulator acc;
+    sse_accum_init(&acc);
+    StreamCtx stream = {0};
+    stream.callback = sse_accum_callback;
+    stream.ctx = &acc;
+    stream.provider = "claude";
+    sb_init(&stream.line_buf);
+
+    feed_stream(&stream, "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\"},\"usage\":{\"output_tokens\":10}}\n\n");
+    check(acc.stopped && !acc.speed_ready, "提前停止原因不是最终用量");
+    check(!stream_speed_ready(&stream, CURLE_OK, 200), "Claude 有停止原因但缺少 message_stop 的提前 EOF 保留速度");
+    feed_stream(&stream, "data: {\"type\":\"message_stop\"}\n\n");
+    check(stream_speed_ready(&stream, CURLE_OK, 200), "Claude 完整终结且 HTTP 成功允许更新");
+    check(!acc.speed_ready, "仅协议终结尚未发最终用量时不更新");
+    check(!stream_speed_ready(&stream, CURLE_PARTIAL_FILE, 200), "协议完整但 HTTP 截断保留速度");
+    check(!stream_speed_ready(&stream, CURLE_OK, 401), "HTTP 失败保留速度");
+    volatile int cancelled = 1;
+    stream.cancelled = &cancelled;
+    check(!stream_speed_ready(&stream, CURLE_OK, 200), "协议终结后取消仍保留速度");
+    check(!stream_speed_ready(&stream, CURLE_WRITE_ERROR, 200), "取消引发写入错误保留速度");
+    emit_simple_event(stream.callback, stream.ctx, SSE_STOP, "interrupted");
+    check(acc.stopped && !acc.speed_ready, "中断 STOP 不允许更新速度");
+    stream.cancelled = NULL;
+
+    SseEvent final_usage = {0};
+    final_usage.type = SSE_USAGE;
+    final_usage.speed_ready = stream_speed_ready(&stream, CURLE_OK, 200);
+    final_usage.end_ms = 100;
+    sse_accum_callback(&acc, &final_usage);
+    check(acc.speed_ready && acc.out_tokens == 10, "最终用量允许更新且保留累计 token");
+    emit_simple_event(stream.callback, stream.ctx, SSE_RETRY, NULL);
+    check(!acc.speed_ready && acc.out_tokens == 0, "重试清除速度门槛及用量");
+    sb_free(&stream.line_buf);
+
+    /* 与传输重试一样，每次尝试使用全新的上下文。 */
+    memset(&stream, 0, sizeof(stream));
+    stream.callback = sse_accum_callback;
+    stream.ctx = &acc;
+    stream.provider = "openai";
+    sb_init(&stream.line_buf);
+    check(!stream_speed_ready(&stream, CURLE_OK, 200), "重试不继承协议终结状态");
+    feed_stream(&stream, "data: {\"object\":\"chat.completion.chunk\",\"choices\":[{\"finish_reason\":\"stop\"}]}\n\n");
+    check(acc.stopped && !stream_speed_ready(&stream, CURLE_OK, 200), "OpenAI 有 finish_reason 但缺少 DONE 的提前 EOF 保留速度");
+    feed_stream(&stream, "data: [DONE]\n\n");
+    check(stream_speed_ready(&stream, CURLE_OK, 200), "OpenAI 完整终结允许更新");
+    final_usage.speed_ready = stream_speed_ready(&stream, CURLE_OK, 200);
+    sse_accum_callback(&acc, &final_usage);
+    check(acc.speed_ready && acc.out_tokens == 0, "零用量成功仍允许清零速度");
+    stream.protocol_complete = stream.openai_finished = 0;
+    feed_stream(&stream, "data: [DONE]\n\n");
+    check(!stream_speed_ready(&stream, CURLE_OK, 200), "仅 DONE 无 finish_reason 不更新");
+    sb_free(&stream.line_buf);
+
+    const char *failed_events[] = {"response.failed", "response.incomplete"};
+    for (int i = 0; i < 2; i++) {
+        memset(&stream, 0, sizeof(stream));
+        stream.callback = sse_accum_callback;
+        stream.ctx = &acc;
+        parse_responses_sse_event(&stream, failed_events[i], "{}", 2);
+        check(stream.responses_terminal && !stream_speed_ready(&stream, CURLE_OK, 200),
+            "Responses 失败或不完整终结保留速度");
+        streamctx_free_openai_tools(&stream);
+    }
+    sse_accum_free(&acc);
+}
+
 int main(void) {
+    test_speed_completion();
     Capture capture = {0};
     StreamCtx stream = {0};
     stream.callback = capture_event;
@@ -55,6 +128,7 @@ int main(void) {
     parse_responses_sse_event(&stream, "response.function_call_arguments.delta", delta, strlen(delta));
     parse_responses_sse_event(&stream, "response.completed", completed, strlen(completed));
 
+    check(stream_speed_ready(&stream, CURLE_OK, 200), "Responses completed 允许最终用量更新速度");
     check(capture.count == 3, "Responses completed emits tool, usage, stop");
     check(capture.events[0].type == SSE_TOOL_CALL && strcmp(capture.events[0].tool_name, "Read") == 0 && strcmp(capture.events[0].tool_id, "call_1") == 0 && strcmp(capture.events[0].tool_input, "{\"path\":\"/tmp/a\"}") == 0, "Responses maps item id and argument delta");
     check(capture.events[1].type == SSE_USAGE && capture.events[1].in_tokens == 9 && capture.events[1].out_tokens == 8 && capture.events[1].cache_read_tokens == 6, "Responses prefers nested cached tokens");
@@ -68,6 +142,7 @@ int main(void) {
     stream.ctx = &capture;
     const char *error = "{\"reason\":\"upstream failed\"}";
     parse_responses_sse_event(&stream, "error", error, strlen(error));
+    check(!stream_speed_ready(&stream, CURLE_OK, 200), "Responses 错误终结不更新速度");
     check(capture.count == 3, "Responses error emits error, usage, stop");
     check(capture.events[0].type == SSE_ERROR && strcmp(capture.events[0].content, "upstream failed") == 0, "Responses error uses reason fallback");
     check(capture.events[1].type == SSE_USAGE, "Responses error emits usage event");