Pinned Tweet
It has a name now.
Competitive Record.
The market keeps score.
Products compete. Results become record.
The first record is still unwritten.
Coming soon.
CompetitiveRecord.com 👀
4
11
1,028
our celery workers exit after each task. worker_max_tasks_per_child is 1 so no process reuses an asyncpg connection across event loops. the cost is a new process per task. write the reason next to the setting so nobody raises it later
1
66
next public env vars are frozen at build time. the browser bundle keeps the value the build saw. our frontend dockerfile comment says a missing turnstile site key at build time renders no captcha. pass public values as build args
5
4
127
new date().toisostring() always returns utc. in a utc+2 or utc+3 office the default date is yesterday for the first hours of the day. our journal entry and patient record forms default to that value. build the date from local parts
1
3
100
a double entry pair takes the same integer total on both sides and so debit and credit cannot drift apart through rounding. the risk is two separate calculations that each round their own copy
1
59
in async sqlalchemy a relationship you did not load can raise an error instead of querying. load what the response needs with selectinload before the session closes
1
1
83
two requests can both read a balance both pass the check and both write. a row lock taken with select for update holds the row until your transaction ends. use it on the paths where two requests can act on the same row
2
2
91
claude code 2.1.325 changed scheduled tasks and /loop to ask before running a skill that declares allowed-tools or hooks. to run it unattended add skill(name) to permissions.allow. @AnthropicAI
3
89
claude code -p could drop earlier text output when background work started another turn. each turn now prints when that turn ends. in a headless pipeline, what tells you a response was cut, and do you compare turn counts against the output @anthropicai
1
77
claude code fixed mcp servers that repeat a pagination cursor. the client asked for the same page up to 20 times on every connect. a cursor you have already seen should end the loop, so what does your agent check before it asks for the next page @anthropicai
2
75
claude code fixed remote mcp servers that reconnected in a tight loop to a server dropping each connection right after it connects. repeated drops now back off up to 30s. what backoff does your agent use when a tool server flaps and do you log each retry @anthropicai
12
5
248
openai async tool calling lets the model keep working while your app runs the tool. the guide says async tools do not move execution to openai, so your app still owns the background jobs. which task handles do you reuse across a conversation, and what happens when the model waits on one that already arrived @OpenAIDevs
42
mid-turn steering queues input on the open connection. the docs say queued input exists only on that connection and a failed steer is not applied later. so what does an agent replay after a disconnect, and how does it know which steers never landed @OpenAIDevs
47
claude code lets a permission rule match a tool's input, so Agent(model:opus) blocks opus subagents. the question is which parameters you would gate this way before a rule that matches on input becomes one more thing to audit @anthropicai
2
3
82
claude code 2.1.325 makes the sandbox refuse writes to .claude settings, skills, agents, commands and hooks in added directories. if an agent has to change its own hook config, where does that change go now @anthropicai
2
65
claude code 2.1.323 added an experimental code mode where claude writes javascript to call its tools. what has to be true about your tool layer before you let a model write the calls instead of emitting one tool call at a time @anthropicai
5
3
160
claude code 2.1.325 changed sandboxed bash commands to follow the normal permission flow by default. set sandbox.autoAllowBashIfSandboxed to true if you want them to keep running without prompts. check that setting before any unattended run @anthropicai
2
1
78
before retrying a send after a timeout check whether the first attempt went through. a retry can publish twice. we look up a content id in our execution log first and then check the profile
34
a scheduled job and an executed job need different rows. the watchdog only counts rows with a confirmed status. a finished job needs a confirmed status or the monitor reports success that never happened
68
the watchdog writes one alert per missed gap to a state file and clears that state when a run is confirmed. without memory an alert becomes noise people learn to ignore
1
1
60

