1 00:00:00,100 --> 00:00:05,580 Picture this: your autonomous coding agent just completely mangled a multi-file 2 00:00:05,640 --> 00:00:06,180 refactor. 3 00:00:07,080 --> 00:00:10,240 You pull up your trace waterfall in Jaeger or Honeycomb to see what happened, 4 00:00:10,940 --> 00:00:13,240 and you see the tool executed with a green checkmark. 5 00:00:13,920 --> 00:00:17,360 But the returned payload is totally, completely blank. 6 00:00:18,438 --> 00:00:20,018 Mate, the classic ghost run. 7 00:00:20,938 --> 00:00:24,458 It is like opening the hood of a car and finding someone replaced the engine with a 8 00:00:24,518 --> 00:00:26,958 handwritten note saying, yep, it worked. 9 00:00:27,203 --> 00:00:27,903 Exactly! 10 00:00:28,243 --> 00:00:32,443 You have no idea what the model actually read before it decided to nuke your code. 11 00:00:33,063 --> 00:00:39,263 But in Claude Code version 2.1.283, Anthropic finally addressed that black box. 12 00:00:40,163 --> 00:00:47,063 Under the changelog for 2.1.283, they expanded the OTEL LOG TOOL CONTENT flag so 13 00:00:47,163 --> 00:00:53,583 MCP tool, WebFetch, and WebSearch outputs are now explicitly recorded inside the 14 00:00:53,643 --> 00:00:55,303 tool dot output span event. 15 00:00:55,333 --> 00:00:57,573 Which was a massive blind spot before, right? 16 00:00:57,673 --> 00:01:01,893 Because earlier versions only captured local Read and Bash calls. 17 00:01:02,186 --> 00:01:07,653 If Claude was fetching context from a remote MCP database server or hitting the web 18 00:01:07,713 --> 00:01:10,613 via WebSearch, you were basically flying blind. 19 00:01:10,725 --> 00:01:11,425 Totally blind. 20 00:01:12,405 --> 00:01:16,645 Now, getting this flowing into something like Jaeger, Dash0, 21 00:01:16,745 --> 00:01:20,185 or Honeycomb takes a few specific environment variables. 22 00:01:20,825 --> 00:01:24,585 First, you set CLAUDE CODE ENABLE TELEMETRY to one. 23 00:01:25,205 --> 00:01:28,685 Then, because distributed tracing is still technically in beta, 24 00:01:29,145 --> 00:01:33,565 you have to export CLAUDE CODE ENHANCED TELEMETRY BETA equals one. 25 00:01:33,625 --> 00:01:38,105 Right, and then your standard OpenTelemetry plumbing: OTEL TRACES EXPORTER set to 26 00:01:38,225 --> 00:01:44,025 otlp, and your OTEL EXPORTER OTLP ENDPOINT pointing to your local collector, 27 00:01:44,125 --> 00:01:47,945 usually port 4317 over gRPC. 28 00:01:47,958 --> 00:01:54,198 And the key switch: OTEL LOG TOOL CONTENT set to one, ideally paired with OTEL LOG 29 00:01:54,246 --> 00:01:57,638 TOOL DETAILS equals one so you get the parameters too. 30 00:01:57,807 --> 00:02:01,467 You know, watching the actual trace context flow makes it click. 31 00:02:02,447 --> 00:02:07,467 You get that root span, claude code dot interaction, which spawns the child span, 32 00:02:07,767 --> 00:02:08,927 claude code dot tool. 33 00:02:09,587 --> 00:02:14,967 And then Claude Code injects the standard W3C TRACEPARENT into the subprocess 34 00:02:15,007 --> 00:02:15,587 environment. 35 00:02:15,947 --> 00:02:19,307 It reminds me of an old midnight database migration I botched years ago. 36 00:02:20,107 --> 00:02:23,807 If you cannot see what the query actually returned, you are just guessing in the 37 00:02:23,827 --> 00:02:24,167 dark. 38 00:02:24,350 --> 00:02:24,610 Right. 39 00:02:25,190 --> 00:02:29,590 But, before anyone goes throwing OTEL LOG TOOL CONTENT into their base company 40 00:02:29,690 --> 00:02:32,550 image, there are some serious security traps here. 41 00:02:32,923 --> 00:02:34,283 Oh, huge traps! 42 00:02:34,883 --> 00:02:37,083 Think about what lives in tool output. 43 00:02:37,943 --> 00:02:42,723 If your MCP server queries an internal Postgres database, or your agent uses 44 00:02:42,763 --> 00:02:47,223 WebFetch on internal wiki pages, you are now dumping raw SQL rows, 45 00:02:47,623 --> 00:02:51,783 corporate documentation, and possibly customer data straight into your telemetry 46 00:02:51,883 --> 00:02:52,383 pipeline. 47 00:02:52,542 --> 00:02:56,942 Yeah, unless your OpenTelemetry collector has aggressive scrubbers or redaction 48 00:02:57,007 --> 00:03:01,102 processors, that stuff is sitting in plaintext in your logging backend. 49 00:03:01,246 --> 00:03:03,422 Plus, there is the truncation ceiling. 50 00:03:03,513 --> 00:03:08,782 Claude Code enforces a default limit of sixty kilobytes, sixty one thousand four 51 00:03:08,822 --> 00:03:13,102 hundred and forty UTF sixteen code units, on content attributes. 52 00:03:13,125 --> 00:03:18,405 So if your WebFetch grabs a massive page, it gets chopped unless you dial up CLAUDE 53 00:03:18,501 --> 00:03:20,725 CODE OTEL CONTENT MAX LENGTH. 54 00:03:20,945 --> 00:03:25,925 And, importantly, if you only set up a logs exporter without tracing enabled, 55 00:03:26,021 --> 00:03:29,365 that tool dot output payload never gets emitted. 56 00:03:29,498 --> 00:03:32,005 It lives strictly inside span events. 57 00:03:32,162 --> 00:03:33,402 A crucial distinction. 58 00:03:34,162 --> 00:03:39,922 Now, 2.1.283 also packed a few really welcome operational fixes. 59 00:03:40,422 --> 00:03:45,702 My personal favorite: slash context now breaks out MCP server instructions into 60 00:03:45,742 --> 00:03:47,222 their own dedicated row. 61 00:03:47,250 --> 00:03:48,850 That is so good. 62 00:03:48,950 --> 00:03:53,250 You plug in two or three MCP servers and suddenly you are burning ten thousand 63 00:03:53,330 --> 00:03:56,690 tokens on system prompts before you even type your first prompt. 64 00:03:56,804 --> 00:03:59,250 Seeing that itemized is brilliant. 65 00:03:59,250 --> 00:04:04,290 And MCP progress notifications now actually stick around on backgrounded tasks 66 00:04:04,340 --> 00:04:06,370 instead of just disappearing into the void. 67 00:04:06,530 --> 00:04:12,210 Plus, they fixed stdio MCP servers getting left behind as zombie processes when a 68 00:04:12,250 --> 00:04:14,050 session ends while they are still booting. 69 00:04:14,083 --> 00:04:18,083 No more running killall in the terminal every Friday afternoon just to clear out 70 00:04:18,140 --> 00:04:19,843 orphaned node processes. 71 00:04:19,976 --> 00:04:24,323 It is a really solid update, pulls back the curtain on what the agent is actually 72 00:04:24,380 --> 00:04:27,523 seeing, provided you do not leak your keys doing it. 73 00:04:27,542 --> 00:04:32,342 Sanitize your spans, calibrate your caps, and you finally have real visibility into 74 00:04:32,358 --> 00:04:32,902 your agent.