Using agents to find and fix bugs — where they genuinely help, where they hallucinate root causes, and the workflows that separate the two.
12 answers
Every session that touches a dependency ends with a 3,000-line package-lock.json diff, even when I asked for a one-line version bump. Turns out it runs `npm install` instead of `npm ci` and regenerates the lockfile against its own npm version.
The single most-commented complaint on the tracker, and mostly not a billing bug — it's what long context plus wide tool use costs. Here's what actually moved the needle for people.
A long session dies mid-turn with a 400 and the instruction to run /rewind. It happens most often when several tools are dispatched in one turn, and /rewind does recover the conversation — but the turn's work is gone.
A long-running report with a lot of theories and few confirmed fixes. Collecting what has and hasn't worked, honestly labelled.
Long sessions work fine until auto-compaction kicks in, then the agent starts violating CLAUDE.md rules it had been following all session: wrong package manager, editing generated files, committing without asking. The rules are still on disk; it just stops consulting them.
Gemini CLI returns an empty stream, or crashes reading `candidates` off an undefined response. Both are the same class of failure and both are worth retrying rather than debugging.
Hundreds of people report the same thing and almost nobody posts numbers. Here are ours, and what we found when we went looking for the cause.
Codex drops the stream partway through a long turn and retries, sometimes in a loop. Reported on flaky links, corporate proxies, and reliably after a laptop sleeps.
OpenCode refuses a model that works when you query the provider's own API directly. Usually a custom-provider block that isn't being loaded, not a wrong model name.
A clean version-scoped regression: 1.17.10 segfaults on startup on Windows, 1.17.9 does not. Pin and wait.
On Windows with a PowerShell profile that sets a custom prompt, PreToolUse hooks configured in .claude/settings.json silently never execute. No error, no log line — tool calls just proceed unguarded. Fixed in 2.1.219.
Our AGENTS.md grew to 1,900 lines of accumulated team wisdom. Somewhere past a few thousand tokens, agents stop seeing the bottom half — rules in the top sections get followed, rules in the bottom sections might as well not exist. No warning from any tool.