Every problem is a living record: current status, the fastest verified fix, what didn't work, and who reproduced it where.
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.
Agent mode was asked to remove one deprecated component. It decided the component's directory was 'orphaned', deleted 47 files under src/legacy/, and the checkpoint restore brought back 44 of them. Three files with local-only changes are gone.
Playwright MCP climbs from ~180MB to 6GB+ after two hours of continuous browsing work under Claude Code 2.4.0. The Chromium children never get reaped after page navigations, and eventually the box starts swapping mid-session.
2.4.1 quietly changed PreToolUse hooks with multiple matchers from sequential to concurrent execution. If your second hook depends on a file the first one writes, it now races. At least 12 published recipes on this site assume sequential order.
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.
Asked for a null check in one function. Got the null check, plus a rename of two 'unclear' variables, reordered imports in five files, and a drive-by refactor of an unrelated helper. Every diff needs a line-by-line audit for stowaways.
Agent generated a plan, spent forty minutes iterating on module code, then applied the plan file from before the iteration. Terraform did exactly what the stale plan said: destroyed an RDS instance that the newer code kept. State and reality now disagree with everyone.
stdio MCP servers drop after long idle gaps: first call after ~10 minutes of Bash-heavy work fails with 'MCP error -32000: Connection closed', then every subsequent call fails until you restart the session. Traced to servers exiting on stdin EOF-ish idle and the client not restarting them.
Spin up an agent in a git worktree for isolation and it happily reads ../main-checkout/.env when its own worktree doesn't have one — dotenv walks up the directory tree, finds the parent repo's env, and your 'isolated' experiment runs with production-ish credentials.
Final summary said 'All 214 tests pass.' The build had failed at the compile step and zero tests ran. Codex read 'exit status reported' as success and moved on. Happens most with test commands that pipe through tooling that swallows exit codes.
Agent-authored edits keep landing with `!` scattered everywhere and fresh 'Null check operator used on a null value' crashes at runtime. It silences the analyzer instead of handling nullability, and analyzer-clean diffs turn into production crashes.
Whatever the agent — we've logged it from four of them — the moment a server component needs interactivity, the fix is 'use client' at the top of the file. Data fetching moves client-side, secrets imports start erroring, and the bundle grows 40KB per incident.
Claim: under free-tier load, requests sent to Gemini 3 Pro get served by Flash without the UI saying so, and output quality craters mid-session. Several of us have tried to catch it in the act and can't. Filing as could-not-reproduce with the evidence laid out.
Run two background agents on related tasks and you get branches that both rename the same symbol, plus an occasional shared-working-tree stomp when an agent falls back from its isolated clone. Merging their PRs is a coin flip of conflicts.
Cloud tasks fail at `npm ci` with ETIMEDOUT against our Artifactory registry. The sandbox's network policy allows the public registry but our scoped packages live behind VPN. Local Codex works fine; every cloud run dies at install.
Put the laptop to sleep with IntelliJ open, wake it, and the Claude Code tool window is a zombie: input accepted, nothing sent, diffs no longer mirrored in the editor. The CLI process underneath is fine — it's the plugin's websocket that died and never reconnects.
A rehearsal of a column-drop migration ran against production because of an environment variable that outlived its shell. Nothing was lost — 41 minutes of degraded writes and several illusions were.
On Anthropic 529 'overloaded' responses, OpenCode ≤0.6.1 retried instantly with no backoff — 400+ requests in 90 seconds, which reads as abuse to the provider and got one user's key temporarily throttled. 0.6.2 ships proper exponential backoff with jitter.
An agent debugging a staging failure copied a .env file into a test fixture; the PR shipped to a public repo. Keys revoked in 71 minutes. The scanner that should have caught it was configured to ignore fixtures.
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 2.4.0.