@tonyclaw/llm-inspector 1.14.8 → 1.14.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -3,61 +3,39 @@ import { cn } from "../../lib/utils";
3
3
  import type { StopReason } from "../../lib/stopReason";
4
4
 
5
5
  export type ThreadConnectorProps = {
6
- /** The stop reason extracted from this log entry's response. */
7
6
  stopReason: StopReason;
8
- /** True when the response is still in-flight (responseStatus === null). */
9
7
  isPending: boolean;
10
- /** True when this is the first entry in the group. */
11
8
  isFirst: boolean;
12
- /** True when this is the last entry in the group. */
13
9
  isLast: boolean;
10
+ /** True when this entry starts a new turn (first overall, or after end_turn/stop). */
11
+ isTurnStart: boolean;
14
12
  };
15
13
 
16
14
  /**
17
- * Renders the vertical timeline connector on the left side of a log entry
18
- * in thread view. A continuous line runs through intermediate entries, with:
19
- * - Line starts from the top for the first entry
20
- * - Line continues through intermediate entries
21
- * - Line ends at a turn-boundary dot/cap when stop_reason is end_turn/stop
22
- * - Dimmed/dashed line for pending (in-flight) entries
23
- * - A solid dot marks turn boundaries
15
+ * Vertical timeline connector for thread view. Uses flexbox layout (no
16
+ * absolute positioning) so the connector naturally tracks its sibling
17
+ * LogEntry height no scroll jitter.
24
18
  */
25
19
  export function ThreadConnector({
26
20
  stopReason,
27
21
  isPending,
28
22
  isFirst,
29
- isLast,
23
+ isLast: _isLast,
24
+ isTurnStart,
30
25
  }: ThreadConnectorProps): JSX.Element {
31
- // Turn boundary: "end_turn" for Anthropic, "stop" for OpenAI
32
26
  const isBoundary = stopReason === "end_turn" || stopReason === "stop";
33
- // Tool use: the turn continues
34
27
  const isToolUse = stopReason === "tool_use";
35
28
 
36
- // Compute classes for each segment of the connector:
37
- // - top half line (from previous entry down to this entry's top)
38
- // - center dot/circle (this entry's response marker)
39
- // - bottom half line (from this entry's bottom down to the next entry)
40
-
41
- const lineClass = cn(
42
- "w-0.5 bg-muted-foreground/30",
43
- isPending && "border-dashed bg-transparent border-l-2 border-muted-foreground/20",
44
- );
45
-
46
29
  return (
47
- <div className="flex items-stretch h-full w-6 shrink-0 relative">
48
- {/* Top line segment */}
49
- <div className="absolute left-1/2 -translate-x-1/2 w-0.5 top-0 flex flex-col items-center">
50
- {!isFirst && (
51
- <div
52
- className={cn("w-0.5 grow", "bg-muted-foreground/30", "h-4")}
53
- style={{ minHeight: "0.5rem" }}
54
- />
55
- )}
56
- {isFirst && <div className="h-4" />}
30
+ <div className="flex flex-col items-center w-6 shrink-0">
31
+ {/* Top: incoming line from previous entry, or empty spacer for first.
32
+ Fixed height so the marker stays near the LogEntry header row. */}
33
+ <div className="flex justify-center h-4">
34
+ {!isFirst && <div className="w-0.5 bg-muted-foreground/30" />}
57
35
  </div>
58
36
 
59
- {/* Center marker */}
60
- <div className="absolute left-1/2 -translate-x-1/2 top-4 flex items-center justify-center z-10">
37
+ {/* Center marker — aligned with the LogEntry header row */}
38
+ <div className="flex items-center justify-center py-0.5">
61
39
  {isBoundary ? (
62
40
  <div
63
41
  className={cn(
@@ -69,8 +47,13 @@ export function ThreadConnector({
69
47
  />
70
48
  ) : isToolUse ? (
71
49
  <div
72
- className="size-2 rounded-full bg-muted-foreground/25"
73
- title="Tool Use — turn continues"
50
+ className={cn(
51
+ "size-2 rounded-full",
52
+ isTurnStart
53
+ ? "bg-emerald-400 shadow-[0_0_6px_rgba(52,211,153,0.5)]"
54
+ : "bg-muted-foreground/25",
55
+ )}
56
+ title={isTurnStart ? "Tool Use — start of turn" : "Tool Use — turn continues"}
74
57
  />
75
58
  ) : isPending ? (
76
59
  <div
@@ -78,26 +61,32 @@ export function ThreadConnector({
78
61
  title="Response pending"
79
62
  />
80
63
  ) : (
81
- <div className="size-1.5 rounded-full bg-muted-foreground/30" />
64
+ <div
65
+ className={cn(
66
+ "size-1.5 rounded-full",
67
+ isTurnStart
68
+ ? "bg-emerald-400 shadow-[0_0_6px_rgba(52,211,153,0.5)]"
69
+ : "bg-muted-foreground/30",
70
+ )}
71
+ />
82
72
  )}
83
73
  </div>
84
74
 
85
- {/* Bottom line segment */}
86
- <div
87
- className="absolute left-1/2 -translate-x-1/2 w-0.5 top-4 flex flex-col items-center"
88
- style={{ bottom: 0 }}
89
- >
90
- {!isBoundary && (
75
+ {/* Bottom: outgoing line to next entry, or short terminator at boundaries.
76
+ flex-1 fills the remaining height of the LogEntry card. */}
77
+ <div className="flex-1 flex justify-center min-h-1">
78
+ {isBoundary ? (
79
+ <div className="w-0.5 bg-muted-foreground/10 h-4" />
80
+ ) : (
91
81
  <div
92
82
  className={cn(
93
- "w-0.5 flex-1",
83
+ "w-0.5 h-full",
94
84
  isPending
95
85
  ? "border-dashed bg-transparent border-l-2 border-muted-foreground/20"
96
86
  : "bg-muted-foreground/30",
97
87
  )}
98
88
  />
99
89
  )}
100
- {isBoundary && <div className="w-0.5 h-4 bg-muted-foreground/10" />}
101
90
  </div>
102
91
  </div>
103
92
  );
@@ -60,11 +60,8 @@ export const AnthropicFormatHandler: FormatHandler = {
60
60
  const json: unknown = JSON.parse(rawBody);
61
61
  if (typeof json === "object" && json !== null && !Array.isArray(json)) {
62
62
  const keys = Object.keys(json);
63
- if (keys.includes("model") && keys.includes("messages")) {
64
- if (keys.includes("system") || keys.includes("tools")) {
65
- return true;
66
- }
67
- }
63
+ // Anthropic puts `system` as a top-level key alongside `model` and `messages`
64
+ return keys.includes("model") && keys.includes("messages") && keys.includes("system");
68
65
  }
69
66
  return false;
70
67
  } catch {
@@ -25,11 +25,41 @@ export const OpenAIFormatHandler: FormatHandler = {
25
25
  extractTokens(responseBody: string): TokenUsage {
26
26
  const parsed = parseOpenAIResponse(responseBody);
27
27
  if (parsed) {
28
+ // OpenAI puts cached_tokens in usage.prompt_tokens_details (passthrough field)
29
+ let cacheReadInputTokens: number | null = null;
30
+ try {
31
+ const raw: unknown = JSON.parse(responseBody);
32
+ if (raw !== null && typeof raw === "object" && !Array.isArray(raw)) {
33
+ const usageDesc = Object.getOwnPropertyDescriptor(raw, "usage");
34
+ if (
35
+ usageDesc !== undefined &&
36
+ typeof usageDesc.value === "object" &&
37
+ usageDesc.value !== null
38
+ ) {
39
+ const detailsDesc = Object.getOwnPropertyDescriptor(
40
+ usageDesc.value,
41
+ "prompt_tokens_details",
42
+ );
43
+ if (
44
+ detailsDesc !== undefined &&
45
+ typeof detailsDesc.value === "object" &&
46
+ detailsDesc.value !== null
47
+ ) {
48
+ const cacheDesc = Object.getOwnPropertyDescriptor(detailsDesc.value, "cached_tokens");
49
+ if (cacheDesc !== undefined && typeof cacheDesc.value === "number") {
50
+ cacheReadInputTokens = cacheDesc.value;
51
+ }
52
+ }
53
+ }
54
+ }
55
+ } catch {
56
+ // ignore parse errors
57
+ }
28
58
  return {
29
59
  inputTokens: parsed.usage.prompt_tokens ?? null,
30
60
  outputTokens: parsed.usage.completion_tokens ?? null,
31
61
  cacheCreationInputTokens: null,
32
- cacheReadInputTokens: null,
62
+ cacheReadInputTokens,
33
63
  };
34
64
  }
35
65
  return {
@@ -55,12 +85,8 @@ export const OpenAIFormatHandler: FormatHandler = {
55
85
  const json: unknown = JSON.parse(rawBody);
56
86
  if (typeof json === "object" && json !== null && !Array.isArray(json)) {
57
87
  const keys = Object.keys(json);
58
- if (keys.includes("model") && keys.includes("messages")) {
59
- // OpenAI doesn't use "system" or "tools" keys at root level
60
- if (!keys.includes("system") && !keys.includes("tools")) {
61
- return true;
62
- }
63
- }
88
+ // OpenAI has `model` and `messages` at the top level, but NOT `system`
89
+ return keys.includes("model") && keys.includes("messages") && !keys.includes("system");
64
90
  }
65
91
  return false;
66
92
  } catch {
@@ -73,6 +73,7 @@ export const OpenAIRequestSchema = z.object({
73
73
  tools: z.array(OpenAIToolDefinition).optional(),
74
74
  tool_choice: z
75
75
  .union([
76
+ z.enum(["auto", "none", "required"]),
76
77
  z.object({ type: z.literal("auto") }),
77
78
  z.object({ type: z.literal("none") }),
78
79
  z.object({ type: z.literal("function"), function: z.object({ name: z.string() }) }),
@@ -93,6 +93,30 @@ export function extractOpenAIStream(
93
93
  promptTokens = chunk.usage.prompt_tokens ?? 0;
94
94
  completionTokens = chunk.usage.completion_tokens ?? 0;
95
95
  log.inputTokens = promptTokens;
96
+ // Extract cached_tokens from raw parsed object (passthrough field in usage)
97
+ if (typeof parsed === "object" && parsed !== null && !Array.isArray(parsed)) {
98
+ const usageDesc = Object.getOwnPropertyDescriptor(parsed, "usage");
99
+ if (
100
+ usageDesc !== undefined &&
101
+ typeof usageDesc.value === "object" &&
102
+ usageDesc.value !== null
103
+ ) {
104
+ const detailsDesc = Object.getOwnPropertyDescriptor(
105
+ usageDesc.value,
106
+ "prompt_tokens_details",
107
+ );
108
+ if (
109
+ detailsDesc !== undefined &&
110
+ typeof detailsDesc.value === "object" &&
111
+ detailsDesc.value !== null
112
+ ) {
113
+ const cacheDesc = Object.getOwnPropertyDescriptor(detailsDesc.value, "cached_tokens");
114
+ if (cacheDesc !== undefined && typeof cacheDesc.value === "number") {
115
+ log.cacheReadInputTokens = cacheDesc.value;
116
+ }
117
+ }
118
+ }
119
+ }
96
120
  usageCaptured = true;
97
121
  }
98
122
 
@@ -5,7 +5,7 @@ import { extractRequestMetadata } from "./schemas";
5
5
  import { registry } from "./formats";
6
6
  import { findProviderByModel } from "./providers";
7
7
  import { getClientInfo } from "./socketTracker";
8
- import { formatForPath, type FormatHandler } from "./formats";
8
+ import { formatForPath, formatRegistry, type FormatHandler } from "./formats";
9
9
  import {
10
10
  PROXY_IDENTITY,
11
11
  PRESERVE_HEADERS,
@@ -308,6 +308,12 @@ export async function handleProxy(req: Request): Promise<Response> {
308
308
  upstreamHeaders.forEach((value, key) => {
309
309
  upstreamHeadersObj[key.toLowerCase()] = value;
310
310
  });
311
+ // Detect the true format from the request body for accurate UI display.
312
+ // The path-based format (formatHandler.format) drives routing, but the body
313
+ // structure determines whether it's actually Anthropic or OpenAI.
314
+ const bodyFormat = formatRegistry.detectFormat(requestBody);
315
+ const displayApiFormat = bodyFormat !== "unknown" ? bodyFormat : formatHandler.format;
316
+
311
317
  const log = await createLog(
312
318
  req.method,
313
319
  parsed.apiPath,
@@ -316,7 +322,7 @@ export async function handleProxy(req: Request): Promise<Response> {
316
322
  clientInfo,
317
323
  rawHeaders,
318
324
  upstreamHeadersObj,
319
- formatHandler.format,
325
+ displayApiFormat,
320
326
  model,
321
327
  sessionId,
322
328
  preAcquiredId,
@@ -111,9 +111,12 @@ function detectFormat(rawBody: string | null): RequestFormat {
111
111
  try {
112
112
  const json: unknown = JSON.parse(rawBody);
113
113
  if (typeof json === "object" && json !== null && !Array.isArray(json)) {
114
- const keys = Object.keys(json);
115
- if (keys.includes("model") && keys.includes("messages")) {
116
- if (keys.includes("system") || keys.includes("tools")) {
114
+ const hasModel = safeGetProperty(json, "model") !== undefined;
115
+ const hasMessages = safeGetProperty(json, "messages") !== undefined;
116
+ if (hasModel && hasMessages) {
117
+ // Anthropic requests put `system` as a top-level key; OpenAI does not.
118
+ // Both formats can have `tools`, so we check `system` as the discriminator.
119
+ if (safeGetProperty(json, "system") !== undefined) {
117
120
  return "anthropic";
118
121
  }
119
122
  const msgVal = safeGetProperty(json, "messages");