Przemek Galarowicz (@_p18z) header
Przemek Galarowicz (@_p18z) profile picture

Building @pharn_dev — know if an agent's PR is safe to merge.

Kraków, Poland
Joined March 2014
Photo by Przemek Galarowicz (@_p18z)Photo by Przemek Galarowicz (@_p18z)Photo by Przemek Galarowicz (@_p18z)
worst bug in pharn was a silent one plan lists files the build is allowed to touch. item with a glob or empty line? dropped, no warning. you find out mid build when a write gets denied happened 3 times. 7.22 refuses the plan and names the item
1
tests pass is weak signal when agent wrote both code and tests. real question is did tests execute the new lines. pass rate without diff coverage is vanity metric
5
Just asked Claude to write release tweets for Pharn. Then asked "you sure?" It found 3 overclaims in its own posts 🤦‍♂️ A benchmark that wasnt even in the release, "runs at" where the code says "requested at", and one "never" that was false always ask twice 😀
1
16
Przemek Galarowicz retweeted
Pharn OSS 7.22 + CLI 0.21.1 are out. pharn-oss now declares its own install: pharn-install.json lists all 614 files with a SHA-256 each. The CLI checks every hash before writing anything. New pharn-* surfaces no longer need a CLI release. npx @pharn-dev/pharn@latest update
1
7
Agents write diffs faster than people can review them. Most teams still review the same way they did two years ago. pharn-ci is one CI check that tells you whether an agent's PR is safe to merge. Diff coverage. Agent setup. Test integrity. One comment.
1
14
Agents write diffs faster than people can review them. pharn-ci tries to answer the cheap, mechanical questions before a human opens the PR: Did the tests actually run the new code? Does the repo tell agents how to work in it?
1
14
Another: TI03 used to miss some ways a skip could be added to an existing test. Now it catches renames to it.skip, describe.skip wrapped around existing tests, and pytestmark skips — even when the body didn't change. A skip that the diff shows as new is still called out for what it is.
1
15
The pattern across 0.5.0: Fail closed when the check can't see. Catch more ways tests can be weakened. Bound the optional upload so it never hangs the job. Diff coverage + agent setup inventory + test integrity. Still AI-agnostic. github.com/pharn-dev/pharn-c…
4
We've given all Pro, Max, Ultra, and Team users one free full reset, across all models and limits. Use it anytime before December 10 at 2:00PM PT. We recently fixed a few small issues that used more limits than intended, and the Fable resets didn't do much for those of you who'd already hit your weekly limits for all models. While these issues didn't affect everyone, we want to make up for the trouble. You'll find it in Settings > Usage, or at clau.de/reset.
8
Everybody says software factory now. mine is not fancy. loop takes task, writes code, runs gates, opens PR only when gates pass 🤷‍♂️
12
The question isn't "did the agent write good code?" The question is: can this PR be merged safely? pharn-ci answers that on every pull request, before a human opens it.
8
most multi agent setups i see are just org chart cosplay. ceo agent, 3 managers, daily standups 😅
1
12
Excited to see what Fable 5.2 can do 🤓
Fable 5.2 is now available in the Claude Platform and Claude Code. It’s up to 30% faster than Fable 5.1, and costs $8/$40 per 1M tokens, down from $10/$50.
18
for solo dev one agent + small tasks + strict gates works better. boring but it ships
1
39
Przemek Galarowicz retweeted
Fable 5.2 is now available in the Claude Platform and Claude Code. It’s up to 30% faster than Fable 5.1, and costs $8/$40 per 1M tokens, down from $10/$50.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
324
670
191
11,188
536,770
I'd say add .claude/settings.local.json to gitignore today. secrets in env, not in commands
A Claude Code permission setting has exposed thousands of developer credentials on public GitHub repositories. Security researchers at Apiiro found 1,122 public copies of Claude Code's settings file containing 1,379 secrets, including database passwords, API tokens, AWS keys and SSH credentials. The problem comes from the "Yes, and don't ask again" option. When developers approve a command containing a password or token, Claude Code can save the entire command in plain text inside .claude/settings.local.json. Apiiro's wider investigation found 2,539 secrets across 1,994 repositories, including Git history and other agent files.
17
3 of 3 measured agent runs started with the test suite already red. Nobody noticed for 28–92 minutes. The first thing we do now is run the gates before the agent touches anything. Cheap. Obvious. Almost nobody does it.
17
If an agent can merge a PR that never ran the new lines, you don't have a review process. You have a hope process. pharn-ci measures whether the tests actually executed the code this PR added. AI-agnostic. Runs in your CI. Source code never leaves it.
15