[
  {
    "startMs": 560,
    "text": "Hi, I am Aush. I built a CLI that lets"
  },
  {
    "startMs": 3280,
    "text": "your agent reuse context from your past"
  },
  {
    "startMs": 5200,
    "text": "session across tools. While improving"
  },
  {
    "startMs": 7680,
    "text": "it, I realized why my agent secretly"
  },
  {
    "startMs": 9840,
    "text": "hated my CLI and what you can do about"
  },
  {
    "startMs": 12639,
    "text": "yours."
  },
  {
    "startMs": 15839,
    "text": "Cloud Code wanted one setting from a"
  },
  {
    "startMs": 17680,
    "text": "docs page. So, it downloaded the whole"
  },
  {
    "startMs": 19680,
    "text": "page, grabbed it, and kept the result at"
  },
  {
    "startMs": 22480,
    "text": "40 lines."
  },
  {
    "startMs": 24800,
    "text": "Codex wanted to know how many"
  },
  {
    "startMs": 26320,
    "text": "dependencies were vulnerable. It asked"
  },
  {
    "startMs": 28240,
    "text": "the whole JSON report and kept the last"
  },
  {
    "startMs": 30400,
    "text": "40 lines where this report puts the"
  },
  {
    "startMs": 32800,
    "text": "totals."
  },
  {
    "startMs": 34640,
    "text": "Cursor ran a test suit, found one"
  },
  {
    "startMs": 36719,
    "text": "failure and kept the last 80 lines. All"
  },
  {
    "startMs": 39920,
    "text": "three cuts kept the information shown"
  },
  {
    "startMs": 41840,
    "text": "here. Each one coded in an assumption"
  },
  {
    "startMs": 44079,
    "text": "about where the answer sits in the"
  },
  {
    "startMs": 45760,
    "text": "output that can hide friction from the"
  },
  {
    "startMs": 47840,
    "text": "people who build these tools. One pipe"
  },
  {
    "startMs": 50239,
    "text": "tells you very little. The same pipe"
  },
  {
    "startMs": 52480,
    "text": "written again and again can point to"
  },
  {
    "startMs": 54480,
    "text": "something your tool is missing. These"
  },
  {
    "startMs": 56480,
    "text": "are real calls from my own transcripts"
  },
  {
    "startMs": 58320,
    "text": "and the cloud code one was the agent"
  },
  {
    "startMs": 60399,
    "text": "reviewing this talk and it still"
  },
  {
    "startMs": 61840,
    "text": "couldn't help itself."
  },
  {
    "startMs": 65280,
    "text": "Agents can work around friction without"
  },
  {
    "startMs": 67439,
    "text": "filing a bug report. They write a little"
  },
  {
    "startMs": 70240,
    "text": "program around your program. In June,"
  },
  {
    "startMs": 73760,
    "text": "Sanity, the content platform, wrote up"
  },
  {
    "startMs": 76479,
    "text": "their MCP server. More than 20,000"
  },
  {
    "startMs": 78880,
    "text": "agents, about 3 million tool calls since"
  },
  {
    "startMs": 81200,
    "text": "launch and not a single bug report. So"
  },
  {
    "startMs": 83759,
    "text": "they shipped a give feedback tool. Once"
  },
  {
    "startMs": 85920,
    "text": "agent had somewhere to complain, they"
  },
  {
    "startMs": 88000,
    "text": "started expressing what's wrong. But you"
  },
  {
    "startMs": 90080,
    "text": "don't need a feedback tool to start."
  },
  {
    "startMs": 92079,
    "text": "Feedback is already sitting somewhere in"
  },
  {
    "startMs": 94079,
    "text": "your transcripts."
  },
  {
    "startMs": 97520,
    "text": "I I extracted about 20,000 agent calls"
  },
  {
    "startMs": 100159,
    "text": "from my same transcripts. 14 CLI appear"
  },
  {
    "startMs": 102880,
    "text": "in that sample. Git, GitHub CLI, cargo,"
  },
  {
    "startMs": 105680,
    "text": "curl, npm, docker, and others. A quarter"
  },
  {
    "startMs": 108799,
    "text": "fed output into filters such as head,"
  },
  {
    "startMs": 110960,
    "text": "grab, tail, python, sort or wc. For read"
  },
  {
    "startMs": 114320,
    "text": "heavy commands, it's more than half. Git"
  },
  {
    "startMs": 116880,
    "text": "show and cargo test 62% each. Some"
  },
  {
    "startMs": 119840,
    "text": "harnesses pipe far less. Codeex for"
  },
  {
    "startMs": 122479,
    "text": "example lets the agent set an output"
  },
  {
    "startMs": 125040,
    "text": "limit on each call."
  },
  {
    "startMs": 128000,
    "text": "And pipes don't only cut noise in batch"
  },
  {
    "startMs": 131200,
    "text": "without pipe fail. A pipe then returns"
  },
  {
    "startMs": 133440,
    "text": "the exit code of its last command. So"
  },
  {
    "startMs": 135680,
    "text": "let's say you pipe cargo test into head."
  },
  {
    "startMs": 138160,
    "text": "It can report success even when the test"
  },
  {
    "startMs": 140400,
    "text": "failed. Here's one example on the"
  },
  {
    "startMs": 142800,
    "text": "screen. A compile error and a green dot."
  },
  {
    "startMs": 147520,
    "text": "The status was wrong but the agent could"
  },
  {
    "startMs": 150000,
    "text": "still read the error."
  },
  {
    "startMs": 152720,
    "text": "A pipe can also cut the error itself"
  },
  {
    "startMs": 155280,
    "text": "which is worse. In a cloud code issue on"
  },
  {
    "startMs": 157599,
    "text": "GitHub, a user reported it a script"
  },
  {
    "startMs": 159680,
    "text": "through a reference error and exited"
  },
  {
    "startMs": 161680,
    "text": "with an error code. tail showed three"
  },
  {
    "startMs": 163760,
    "text": "harmless lines and Claude reported"
  },
  {
    "startMs": 165599,
    "text": "success. The error never reached the"
  },
  {
    "startMs": 168239,
    "text": "model. So my advice would be to fail"
  },
  {
    "startMs": 170879,
    "text": "loudly when a tool fails. Make the very"
  },
  {
    "startMs": 173680,
    "text": "last lines a short summary and a part to"
  },
  {
    "startMs": 176400,
    "text": "the full log."
  },
  {
    "startMs": 179680,
    "text": "I learned all this while tinkering with"
  },
  {
    "startMs": 181599,
    "text": "a CLI I built. It's a small tool that"
  },
  {
    "startMs": 184640,
    "text": "lets coding agents search their own past"
  },
  {
    "startMs": 186959,
    "text": "sessions across cloud code, codeex,"
  },
  {
    "startMs": 189200,
    "text": "cursor. It ships with a skill so the"
  },
  {
    "startMs": 191680,
    "text": "agent can decide when and how to call"
  },
  {
    "startMs": 193920,
    "text": "it. Agents are the users that I'm"
  },
  {
    "startMs": 196640,
    "text": "studying. My October audit recorded"
  },
  {
    "startMs": 198959,
    "text": "about thousand calls and that's my small"
  },
  {
    "startMs": 201599,
    "text": "case study. The method works on anything"
  },
  {
    "startMs": 203840,
    "text": "your agents calls. Just look at what"
  },
  {
    "startMs": 206000,
    "text": "they do right after the call."
  },
  {
    "startMs": 210080,
    "text": "80%."
  },
  {
    "startMs": 211599,
    "text": "Four out of five times an agent opened"
  },
  {
    "startMs": 213920,
    "text": "an old session with my tool. The output"
  },
  {
    "startMs": 215920,
    "text": "went straight into crab sad head tail or"
  },
  {
    "startMs": 218480,
    "text": "Python. I call that the pipe text. A"
  },
  {
    "startMs": 221599,
    "text": "pipe keeps junk out of context and"
  },
  {
    "startMs": 223840,
    "text": "that's fine. The text is that the agent"
  },
  {
    "startMs": 226159,
    "text": "writing that same cut down command by"
  },
  {
    "startMs": 228480,
    "text": "hand every single time. It's small each"
  },
  {
    "startMs": 231040,
    "text": "time and it mostly works but each one is"
  },
  {
    "startMs": 233519,
    "text": "another guess about where the answer is."
  },
  {
    "startMs": 238080,
    "text": "The fair objection comes from a hacker"
  },
  {
    "startMs": 239920,
    "text": "news thread about designing CLS for"
  },
  {
    "startMs": 242319,
    "text": "agent that I was reading. The command"
  },
  {
    "startMs": 244720,
    "text": "says to leave it to the agents to choose"
  },
  {
    "startMs": 247280,
    "text": "how much output to reads. agent can pipe"
  },
  {
    "startMs": 249519,
    "text": "through grap or head perfectly well."
  },
  {
    "startMs": 252239,
    "text": "Well, they can and for oneoff they"
  },
  {
    "startMs": 254799,
    "text": "should. I'm not against that. But across"
  },
  {
    "startMs": 257199,
    "text": "those everyday CLIs, I found 83% of"
  },
  {
    "startMs": 259919,
    "text": "pipes fell into the patterns seen at"
  },
  {
    "startMs": 262240,
    "text": "least 10 times. The same CLI subcomand"
  },
  {
    "startMs": 264720,
    "text": "and pipe tools. The arguments passed"
  },
  {
    "startMs": 266960,
    "text": "could differ, but the agent keep asking"
  },
  {
    "startMs": 269520,
    "text": "for the same things."
  },
  {
    "startMs": 272960,
    "text": "I [snorts] understand the output they're"
  },
  {
    "startMs": 274400,
    "text": "cutting is probably big. I opened one"
  },
  {
    "startMs": 276720,
    "text": "long session with my tool two ways the"
  },
  {
    "startMs": 279520,
    "text": "whole thing 1.2 2 million characters and"
  },
  {
    "startMs": 282080,
    "text": "that's why the agent pipe just the part"
  },
  {
    "startMs": 284560,
    "text": "around the search match was about 3,000"
  },
  {
    "startMs": 286720,
    "text": "characters. Harnesses do cap output but"
  },
  {
    "startMs": 289840,
    "text": "a size cap does not identify the"
  },
  {
    "startMs": 291759,
    "text": "relevant messages. Selection flags on"
  },
  {
    "startMs": 294320,
    "text": "the other hand let the agent tell your"
  },
  {
    "startMs": 296479,
    "text": "tool which part it wants."
  },
  {
    "startMs": 299919,
    "text": "So based on my experience and"
  },
  {
    "startMs": 301600,
    "text": "improvements I made to my own tool, I'll"
  },
  {
    "startMs": 304160,
    "text": "give you three examples. Read the left"
  },
  {
    "startMs": 306400,
    "text": "column as the bug reports. So starting"
  },
  {
    "startMs": 309039,
    "text": "with the first one filtering with RG and"
  },
  {
    "startMs": 311280,
    "text": "Z means let me pick the part I want. So"
  },
  {
    "startMs": 314080,
    "text": "read my command. So my read command view"
  },
  {
    "startMs": 316560,
    "text": "can now return just the part around a"
  },
  {
    "startMs": 318479,
    "text": "match only the matching messages or the"
  },
  {
    "startMs": 321039,
    "text": "last few with a size gap."
  },
  {
    "startMs": 323680,
    "text": "Second example would be Python over"
  },
  {
    "startMs": 325840,
    "text": "JSON. So Python over JSON means give me"
  },
  {
    "startMs": 328400,
    "text": "fewer fields. So search got a compact"
  },
  {
    "startMs": 330880,
    "text": "mode one line per hit about six times"
  },
  {
    "startMs": 333440,
    "text": "smaller than JSON with the same results."
  },
  {
    "startMs": 337600,
    "text": "Inspect command in chat history"
  },
  {
    "startMs": 339280,
    "text": "summarizes a session. About the third of"
  },
  {
    "startMs": 342000,
    "text": "inspect calls came from agent looping"
  },
  {
    "startMs": 344320,
    "text": "over several sessions by hand. As you"
  },
  {
    "startMs": 346160,
    "text": "can see with the for loop that means let"
  },
  {
    "startMs": 348880,
    "text": "me ask several about several at once and"
  },
  {
    "startMs": 351520,
    "text": "now it's one call. As you can see my"
  },
  {
    "startMs": 354880,
    "text": "road map was in the commands around my"
  },
  {
    "startMs": 357280,
    "text": "command and I had to learn from that."
  },
  {
    "startMs": 360800,
    "text": "Infraost the cloud cost tool wrote up"
  },
  {
    "startMs": 363440,
    "text": "the same discovery in May. Their agents"
  },
  {
    "startMs": 365520,
    "text": "were writing jQ and Python pipelines"
  },
  {
    "startMs": 367520,
    "text": "over JSON dumps. So they moved the"
  },
  {
    "startMs": 369440,
    "text": "filtering into the CLI and added a"
  },
  {
    "startMs": 371360,
    "text": "compact format and a skill to teach it."
  },
  {
    "startMs": 373840,
    "text": "Plain claw with none of that got zero of"
  },
  {
    "startMs": 376560,
    "text": "six hard questions right in their"
  },
  {
    "startMs": 377919,
    "text": "testing and with it six of six and cost"
  },
  {
    "startMs": 381120,
    "text": "across the benchmark also fell 41%. On"
  },
  {
    "startMs": 384400,
    "text": "hacker news, the person who wrote that"
  },
  {
    "startMs": 386000,
    "text": "post said that reading those agent logs"
  },
  {
    "startMs": 388639,
    "text": "felt like watching usability tests and I"
  },
  {
    "startMs": 391199,
    "text": "couldn't agree more."
  },
  {
    "startMs": 393680,
    "text": "This part might sound trivial and"
  },
  {
    "startMs": 395440,
    "text": "obvious. Once it's in the CLI, teach"
  },
  {
    "startMs": 397680,
    "text": "your agent it's there with the skill"
  },
  {
    "startMs": 399440,
    "text": "file or the health flag, but the agents"
  },
  {
    "startMs": 402240,
    "text": "take the teaching literally. I made some"
  },
  {
    "startMs": 404319,
    "text": "changes that turned my deep flag into a"
  },
  {
    "startMs": 407199,
    "text": "no op. I left the skill telling the"
  },
  {
    "startMs": 409440,
    "text": "agents to use it to see what they do."
  },
  {
    "startMs": 411680,
    "text": "Would they find out or not? 95 searches"
  },
  {
    "startMs": 414319,
    "text": "of out of 95 still passed it."
  },
  {
    "startMs": 418000,
    "text": "After that, two sessions loaded a"
  },
  {
    "startMs": 419840,
    "text": "rewritten skill and that no longer told"
  },
  {
    "startMs": 422880,
    "text": "them to use it. None of the searches"
  },
  {
    "startMs": 424800,
    "text": "passed D flag and both tried new flags"
  },
  {
    "startMs": 427280,
    "text": "on their first read. So your skill.md,"
  },
  {
    "startMs": 430160,
    "text": "your help text, your error messages are"
  },
  {
    "startMs": 432560,
    "text": "the interface. That also means stale"
  },
  {
    "startMs": 434400,
    "text": "dogs are bugs because your agent will"
  },
  {
    "startMs": 436319,
    "text": "keep following them."
  },
  {
    "startMs": 439919,
    "text": "Once the new flags reached agents,"
  },
  {
    "startMs": 442319,
    "text": "nearly every search used compact output"
  },
  {
    "startMs": 444560,
    "text": "and nearly every reduced the new flags."
  },
  {
    "startMs": 447360,
    "text": "Pipe breed fell from 82% to about a"
  },
  {
    "startMs": 450800,
    "text": "quarter."
  },
  {
    "startMs": 456479,
    "text": "You can get real levels and test cases"
  },
  {
    "startMs": 458319,
    "text": "for your tool easily. If you use it"
  },
  {
    "startMs": 460080,
    "text": "every day, they are already in your"
  },
  {
    "startMs": 461440,
    "text": "transcripts. The commands agent really"
  },
  {
    "startMs": 463759,
    "text": "ran are the test cases and what agent"
  },
  {
    "startMs": 465680,
    "text": "did immediately next is a rough answer"
  },
  {
    "startMs": 467680,
    "text": "key. replay those commands through your"
  },
  {
    "startMs": 469759,
    "text": "old build and your new one. If you"
  },
  {
    "startMs": 472000,
    "text": "change a ranking, check that result uh"
  },
  {
    "startMs": 475280,
    "text": "the agent went on to use still comes out"
  },
  {
    "startMs": 477919,
    "text": "on the top or not. If you change an"
  },
  {
    "startMs": 479919,
    "text": "output format, check that nothing went"
  },
  {
    "startMs": 482080,
    "text": "missing. If you change a default, see"
  },
  {
    "startMs": 484240,
    "text": "which commands now behave differently."
  },
  {
    "startMs": 486639,
    "text": "And when an agent reader a call with"
  },
  {
    "startMs": 488720,
    "text": "different arguments, you have probably"
  },
  {
    "startMs": 490400,
    "text": "found a failing case. The answer key is"
  },
  {
    "startMs": 492800,
    "text": "only a rough guess at what the agent"
  },
  {
    "startMs": 494479,
    "text": "wanted. keep a few cases back and use"
  },
  {
    "startMs": 496800,
    "text": "those to score your fix."
  },
  {
    "startMs": 500560,
    "text": "So after analyzing 20,000 invocation, if"
  },
  {
    "startMs": 504240,
    "text": "I were to start all over again, this is"
  },
  {
    "startMs": 506560,
    "text": "what I would build first. 55% of the"
  },
  {
    "startMs": 509440,
    "text": "measured pipelines used only head, tail,"
  },
  {
    "startMs": 512479,
    "text": "or set downstream that points to demand"
  },
  {
    "startMs": 514880,
    "text": "for smaller slices. So ship ranges and"
  },
  {
    "startMs": 517360,
    "text": "limits before anything fancy. A flag"
  },
  {
    "startMs": 521039,
    "text": "only helps if agents reach for it."
  },
  {
    "startMs": 522959,
    "text": "GitHub CLI's list command have a limit"
  },
  {
    "startMs": 525120,
    "text": "flag. An agent used it 81% of the time."
  },
  {
    "startMs": 528560,
    "text": "The counter example would be cargo for"
  },
  {
    "startMs": 531040,
    "text": "cargo test quite mode prints a character"
  },
  {
    "startMs": 534080,
    "text": "per test instead of a line and still"
  },
  {
    "startMs": 536000,
    "text": "reports failures. Agent passed it on"
  },
  {
    "startMs": 538640,
    "text": "about one cargo run in 10 and pipped"
  },
  {
    "startMs": 541040,
    "text": "instead. So teach the flag. So teach the"
  },
  {
    "startMs": 543920,
    "text": "flag in your help test and your skill"
  },
  {
    "startMs": 546880,
    "text": "and make big outputs print less by"
  },
  {
    "startMs": 549120,
    "text": "default and say how to get more. for"
  },
  {
    "startMs": 551680,
    "text": "example, and with something like showing"
  },
  {
    "startMs": 553920,
    "text": "20 of 300, add offset flag 20 for the"
  },
  {
    "startMs": 557040,
    "text": "next wage."
  },
  {
    "startMs": 560000,
    "text": "One way to ship several of these at once"
  },
  {
    "startMs": 562080,
    "text": "is an agent mode. Your CLI notices an"
  },
  {
    "startMs": 564240,
    "text": "agent is calling it and prints"
  },
  {
    "startMs": 565760,
    "text": "differently. Agents like cloud code set"
  },
  {
    "startMs": 568160,
    "text": "an environment variable in their cell so"
  },
  {
    "startMs": 570320,
    "text": "your tool can spot them. Doug DB shipped"
  },
  {
    "startMs": 573360,
    "text": "one at a terminal. You get a nicely"
  },
  {
    "startMs": 575440,
    "text": "drawn table, but an agent gets a plain"
  },
  {
    "startMs": 577760,
    "text": "one that says how many rows were left"
  },
  {
    "startMs": 579600,
    "text": "out and how to get them. A huge query"
  },
  {
    "startMs": 582080,
    "text": "stops early and says so. And errors come"
  },
  {
    "startMs": 584720,
    "text": "back as JSON because agents read them"
  },
  {
    "startMs": 586880,
    "text": "better. Is the same list as we have been"
  },
  {
    "startMs": 589440,
    "text": "just through and it cut the output model"
  },
  {
    "startMs": 592000,
    "text": "red by 59%."
  },
  {
    "startMs": 594080,
    "text": "However, the tokens saved were nearly"
  },
  {
    "startMs": 596640,
    "text": "about.5% of all the tokens the model"
  },
  {
    "startMs": 599519,
    "text": "read. So cost and runtime didn't"
  },
  {
    "startMs": 601360,
    "text": "improve. If you remember, infra cost"
  },
  {
    "startMs": 604399,
    "text": "agent started getting answers right."
  },
  {
    "startMs": 606720,
    "text": "Doug Dvs were already there. So measure"
  },
  {
    "startMs": 609680,
    "text": "turns and write answers, not just"
  },
  {
    "startMs": 611519,
    "text": "tokens."
  },
  {
    "startMs": 614320,
    "text": "You might be thinking, I'll ask my agent"
  },
  {
    "startMs": 616560,
    "text": "to do all of this. They are not that"
  },
  {
    "startMs": 619600,
    "text": "good at this yet. At least not"
  },
  {
    "startMs": 621680,
    "text": "oneshotting it. The design is where it"
  },
  {
    "startMs": 624079,
    "text": "goes wrong. Ask an agent to build a tool"
  },
  {
    "startMs": 626560,
    "text": "for agent. It probably pictures a person"
  },
  {
    "startMs": 628480,
    "text": "and then bolts the agent on. It copies"
  },
  {
    "startMs": 630959,
    "text": "most common patterns it has seen and"
  },
  {
    "startMs": 632720,
    "text": "most of those are poor. A study of"
  },
  {
    "startMs": 635040,
    "text": "official MCP servers found most just rep"
  },
  {
    "startMs": 638959,
    "text": "an existing API. Another check tool"
  },
  {
    "startMs": 641760,
    "text": "descriptions from real MCP servers and"
  },
  {
    "startMs": 643600,
    "text": "flagged issues in 97% of them. Docs have"
  },
  {
    "startMs": 647760,
    "text": "the same gap. Researchers at each Zurich"
  },
  {
    "startMs": 650560,
    "text": "tested context file written by agents."
  },
  {
    "startMs": 653120,
    "text": "They didn't help tasks succeed cost more"
  },
  {
    "startMs": 655680,
    "text": "to run and did worse than files written"
  },
  {
    "startMs": 657920,
    "text": "by people. Agents are already good at"
  },
  {
    "startMs": 660880,
    "text": "other half of this talk. Give one"
  },
  {
    "startMs": 662880,
    "text": "transcript and a test it can run and it"
  },
  {
    "startMs": 665440,
    "text": "can find what's wrong and improve the"
  },
  {
    "startMs": 667120,
    "text": "tool. Sometimes past what people wrote"
  },
  {
    "startMs": 669680,
    "text": "by hand and tropic reported that on its"
  },
  {
    "startMs": 672000,
    "text": "own internal tools as well. So you"
  },
  {
    "startMs": 674560,
    "text": "decide what the tool should be and let"
  },
  {
    "startMs": 676320,
    "text": "the agent check it against what really"
  },
  {
    "startMs": 678240,
    "text": "happened."
  },
  {
    "startMs": 680480,
    "text": "You have seen the evidence for each of"
  },
  {
    "startMs": 682000,
    "text": "these. So, I won't read them. Take a"
  },
  {
    "startMs": 684000,
    "text": "moment to absorb these or screenshot."
  },
  {
    "startMs": 693360,
    "text": "To find your own list, back up your"
  },
  {
    "startMs": 695440,
    "text": "transcripts first because clot code can"
  },
  {
    "startMs": 697279,
    "text": "delete inactive ones after 30 days."
  },
  {
    "startMs": 700560,
    "text": "Then, search for them for a tool's name"
  },
  {
    "startMs": 703519,
    "text": "followed by a pipe and other similar"
  },
  {
    "startMs": 705279,
    "text": "patterns. If you sip ship an MCP server,"
  },
  {
    "startMs": 708079,
    "text": "look for agent calling your tool again"
  },
  {
    "startMs": 710720,
    "text": "straight away with different arguments."
  },
  {
    "startMs": 714959,
    "text": "If you remember one thing from this"
  },
  {
    "startMs": 716480,
    "text": "talk, if you are building a tool for"
  },
  {
    "startMs": 718880,
    "text": "your agent, your agent transcript are"
  },
  {
    "startMs": 721120,
    "text": "the product analytics you probably"
  },
  {
    "startMs": 722880,
    "text": "aren't reading."
  },
  {
    "startMs": 726079,
    "text": "The workaround is the bug report. So go"
  },
  {
    "startMs": 728639,
    "text": "read your transcripts and if you need"
  },
  {
    "startMs": 730560,
    "text": "any help doing that, check out chat"
  },
  {
    "startMs": 732639,
    "text": "history on cargo or GitHub. It can do"
  },
  {
    "startMs": 734880,
    "text": "this for you as well. Thank you."
  }
]
