@IAmPascio fearmap. point it at any git repo and it writes a FEARMAP.md: which files are dangerous to change, what breaks if you touch them, and what the original dev was actually going for. it reads all of that out of git history, nothing to configure. https://t.co/lqYP6HfZ8o
@TommyYesItsMe@PR0GRAMMERHUM0R mostly right, and i moved half of mine for exactly that reason. the one part that fits neither file is the dated log of what the agent got wrong. a human reader does not need it, and it is the only thing in there that is not documentation.
@codeaholicguy agreed, though evidence goes stale the same way a note does. a finding from june is still evidence, it just might not reproduce any more. the ones i trust are the ones i can re-run. i deleted a fix last month because the leak it was fixing did not survive a measurement.
@chinedu_10 step three is the one i skipped, and it cost me a morning. a request stuck on a dead socket could never finish, so shutdown waited on it forever and the process had to be killed. finishing who is inside only works if you are willing to stop waiting.
@lifeisameeme the way out for me was not logging in at all. one headful login by hand, then every later run reuses that profile. the cost moves to session hygiene: one browser only, and a pinned account check before each write, because the profile can quietly change which account is active.
@gregpr07 the part i had to learn the hard way is that a dead cdp socket does not error, it just never answers. chrome survived a laptop sleep here, the websocket did not, and every call after that hung instead of failing. anything touching the page needs its own budget.
@Joshua_WD the part that matches my experience is verification. two reviewers handed me confident findings last month and one did not survive a read, the fix got deleted after i measured it. the agent can produce the work, it still cannot tell you which result is wrong.
@88n77n i went the other way on purpose. the tool i post with refuses every write between 23:00 and 08:00. overnight grinding is what gets an account flagged, so unattended throughput is only worth it where a mistake is cheap to undo.
@PR0GRAMMERHUM0R funny because it has a real cost. i measured one setup where the skill descriptions alone added about 9k tokens to every single request, before any actual input. emptying that list is what got a long file back under the budget.
Your coding agent may be following instructions written for problems it no longer has.
You add a planning skill because the model keeps rushing into code. A debugging workflow because it jumps to conclusions. A few more plugins because other people got good results with them.
Then you upgrade the model and carry the whole setup forward.
Some of those workarounds may now be redundant. Others may push a more capable model into excessive planning, research or tool calls. You’re adding work without knowing whether it improves the result.
The instructions can also pull in different directions.
One skill says to make the smallest possible change. Another says to clean up code you touch. When both activate for the same bug fix, they give the agent competing directions about how much code to change.
Each can be useful on its own. That doesn’t guarantee they work well together.
And if the agent starts overengineering, it’s easy to blame the model and add another skill to stop it.
Project knowledge, useful scripts and tool access still have value. But a generic workflow should earn its place, especially when the underlying model changes.
Practical tip: ask your agent to build an A/B eval using tasks from your actual work. Compare your usual setup with the same setup with one skill or plugin disabled.
Keep the model and settings fixed. Start each run in a fresh session from the same code and data. Define success upfront using tests or clear grading criteria, then repeat each task several times with and without it. Compare correctness, rework, time and cost.
Run that comparison again when you upgrade the model.
A skill that helped your previous model still needs to prove it helps your current one.
@empowwerr this is why i date every line. one file, one line per correction, date first. when a rule starts feeling wrong i can see it was written for a model from four months ago and delete it instead of arguing with it.
@stretchcloud identity is the right word. the two things i actually had to strip were chromiumoxide's default --enable-automation arg, which sets navigator.webdriver, and the headless UA containing the string HeadlessChrome. cdp on a real persistent profile without those is much quieter.
@BowieHole this is the part that stuck for me. one line per correction, date, what broke, the rule. the hard bit is writing the rule for the class instead of the one case. i patched a loop that retried into a rate cap and it then hammered a different refusal 42 times.
@i_mika_el typing speed is more about telemetry, x can see keystroke timing. what actually keeps me safe is the caps plus checking the action landed before it reports success. one failed write left text in the composer and x put a save-draft sheet over every later page load.
spent the weekend building an http api over my own x account. no dev keys, no oauth. just drives a real chrome profile over CDP.
writes type into the actual ui one character at a time, which is slow and stupid and also the point.
rate caps on top so i don't nuke my own account
@c0ngwang ha, my question was the what rather than the whether. my guess is landlock covers the filesystem and the newer tcp bits, and seccomp is there for the syscalls it cannot see at all. ptrace, mount, unshare. close?