What I gave back

community-scripts / ProxmoxVE

changedetection: migrate the Python install to a uv venv

∷ insight The install script relied on pip's --ignore-installed, which leaves duplicate dist-info metadata and a deferred crash on the next service restart. I reproduced it live on my own container (46 duplicated packages, certifi ×3) and migrated install + update onto the project's own setup_uv helper (Python 3.13 venv), with automatic migration of the existing pip-global layout. Fixes the root cause their earlier #13548 only patched over. (+33 −17)

community-scripts / ProxmoxVE

homelable: preserve the MCP server config across updates

∷ insight The Homelable update routine overwrote the MCP server configuration on every run, wiping local customizations. Reworked the update path to preserve the existing config instead of clobbering it.

community-scripts / ProxmoxVE

infisical: fix update abort caused by a credentials field mismatch

∷ insight Found on my own CT: the Infisical update aborted mid-run and left the service down. ct/infisical.sh grepped for 'Database Password:' but setup_postgresql_db writes 'Password:' — a one-line grep mismatch. Fixed the field; the update completes cleanly again.

RightNow-AI / openfang

fix(security): unify SSRF protection for WASM host calls

∷ insight The WASM sandbox's host_net_fetch() carried its own SSRF check that had drifted from the canonical check_ssrf() in web_fetch.rs — WASM agents could reach internal targets that builtin tools correctly block. Unified both call sites onto the single canonical implementation (+56 −98).

ublue-os / homebrew-experimental-tap

claude-code-linux cask — livecheck npm vs GCS

∷ insight The official Anthropic GCS /stable file lags up to 13 versions behind the npm registry. A livecheck pointing at npm catches releases the same day they ship.

requarks / wiki

`render IS NULL` in DB causes silent HTTP 500 with no recovery path

∷ insight Hit in production after a migration: pages with a NULL render column returned HTTP 500 instead of falling back to re-rendering from the source. Minimal repro + root cause pointer to server/models/pages.js#L952-L969 + suggested fix.

grafana / alloy

docs: systemd journal example for Promtail → Alloy migration

∷ insight The official migration guide only covered file-based scrape configs, skipping the most common Linux source — systemd journal. Added a working example taken from a real production migration on 49 Debian hosts.

wazuh / wazuh-documentation

warn that wazuh-agent conflicts with wazuh-manager on same host

∷ insight The wazuh-agent package silently uninstalls wazuh-manager via dpkg Conflicts/Replaces when both are installed on the same machine. No warning in the install doc. Cost us a 17-hour silent outage (see the blog post-mortem).

ublue-os / bluefin

∷ insight The default JXL wallpaper renders pure white on Fedora 44: libjxl ships as a codec, but there's no gdk-pixbuf loader module for it, so GNOME can't decode the .jxl the default gschema points at. Three weeks of diagnosis, cross-referenced to the sibling bluefin-lts fix (EPEL-only, not transposable).

requarks / wiki

∷ insight A pages.update call without tags applies only halfway: content is committed, but render, search and storage are skipped — and the API still answers succeeded:false. Root cause: associateTags isn't guarded the way createPage guards it. The angle the 2021 report missed is the non-atomicity, not the crash.

BerriAI / litellm

∷ insight MiniMax-M3's max_input_tokens was set to 512000 — but that's the billing-tier boundary, not the model's 1M window (1048576). Sourced from OpenRouter, official pricing and the sibling M2.x entries. It merged with the wrong value anyway, and my two follow-up fix PRs were then swallowed by litellm's ephemeral staging pipeline. A clean lesson in reading a repo's merge flow before contributing.

rtk-ai / rtk

∷ insight `rtk gh run view --job <id>` still failed after the run-id crash was patched: a two-layer bug where --job isn't in the passthrough allowlist, so it falls into the wrong formatter. Sourced onto the existing PR rather than opening a duplicate.

gadievron / raptor

∷ insight Under rootless podman/distrobox the sandbox half-engages and raptor reported '0 findings in 0 files' instead of failing loudly — a false clean. The maintainer declined my env-override approach on trust-model grounds but shipped his own fail-loud fix, which I validated on our setup. I closed the PR myself: the real win had already landed.