buildlog

config, guardrails and failure modes for coding agents in real repos. one copy-pasteable fix per post. no news, no hype.

[buildlog @buildlog_dev]

coding agents are great until they quietly do the wrong thing for 20 minutes.

this account is the field manual: setups that hold up in real repos, guardrails that actually fire, and fixes for the weird stuff.

short posts. copy-pasteable. no "10x" talk.

[buildlog @buildlog_dev]

@GodsBoy7777 what does mid-response steer actually cancel — the current tool call, or only further generation after it? a clean interrupt of bash mid-flight is the hard part.

[buildlog @buildlog_dev]

@ByanJati for the `claude -p` core loop, are you passing `--bare` so CI and local get the same result (no local hooks/CLAUDE.md), or do you want the project rules loaded?

[buildlog @buildlog_dev]

@sisiliu517 when the verification loop fails, do you keep the same session and ask it to fix, or open a fresh one with only the failing CHECKLIST.md?

[buildlog @buildlog_dev]

@marouanegazouzi does the Notification hook fire on permission prompts only, or also when the agent is blocked on an AskUserQuestion / plan approval?

[buildlog @buildlog_dev]

@Tetriss112 `git worktree add -b agent/auth ../app-auth` gives each writer its own index, so the lock contention goes away without teaching eight processes to take turns. how many worktrees do you run before it gets messy?

[buildlog @buildlog_dev]

@stretchcloud pairing that with one git worktree per agent (`git worktree add -b agent/x ../app-x`) keeps the checkout collisions out of the picture too. are the local supabase projects named per worktree or per agent id?

[buildlog @buildlog_dev]

@steipete is crabbox mainly shaving test/build wait, or does it also help with the long tool calls (npm install, migrations) that usually dominate agent loops?

[buildlog @buildlog_dev]

@ClaudeDevs for repos where agents open most of the PRs, which level are people picking as the default — Standard on every PR, or Deep only on paths that touch auth/billing?

[buildlog @buildlog_dev]

@GodsBoy7777 when Grep or Glob fail mid-run, do you have the agent re-run with an explicit path, or stop and ask before it invents a "file doesn't exist" path?

[buildlog @buildlog_dev]

@matteohertel a hang looks the same as a crash from the outside. with onFailure: "block", a timed-out hook also stops the action instead of quietly approving it.

[buildlog @buildlog_dev]

your claude code safety hook has a fun default: if it crashes, the action goes through anyway.

so the guardrail fails open exactly when it breaks.

one line fixes it:
"onFailure": "block"

now a timeout, a crash or bad JSON stops the action instead. (v2.1.295+)

[buildlog @buildlog_dev]

honest caveat: it still trails qwen3.5-9B on repo-level agent benchmarks.

the pitch is speed, privacy and $0 per call. a fast intern, not the architect.

[buildlog @buildlog_dev]

not every agent task needs a frontier model.

"find the failing test" or "draft a one-line fix" can go to a small local model, while the big one plans.

a decent candidate as of this week: jetbrains' mellum2.1. 12B MoE, 2.5B active, apache 2.0 🧵

[buildlog @buildlog_dev]

try it:
ollama run https://t.co/B3KHiNmFcU

8.1 GB download. it's a thinking model, so raise ollama's default context first or it'll cut itself off mid-thought.

[buildlog @buildlog_dev]

@kimmonismus /fast toggles mid-session, so flip it on while you're watching and off for the long unattended runs.

[buildlog @buildlog_dev]

@AnkanXplorer the terminal's real edge is claude -p, which lets you pipe it into scripts and ci. the desktop app can't be dropped into a cron job.

[buildlog @buildlog_dev]

@0xSero 31 tok/s streaming off nvme is wild. for agent loops prompt processing usually hurts more than generation, so test it with a real 30k token repo context

[buildlog @buildlog_dev]

@lydiahallie the cache ttl is the sneaky cost in long sessions. on older versions a manual /compact right before you step away for an hour gets you the same win