Investigate and safely reclaim disk space on this machine, especially on NixOS systems with heavy Nix, Rust/Haskell, Docker, and Podman usage. Use when disk is low, builds fail with no-space errors, /nix/store appears unexpectedly large, or the user asks for easy cleanup wins without deleting important data.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Investigate and safely reclaim disk space on this machine, especially on NixOS systems with heavy Nix, Rust/Haskell, Docker, and Podman usage. Use when disk is low, builds fail with no-space errors, /nix/store appears unexpectedly large, or the user asks for easy cleanup wins without deleting important data.
Disk Space Cleanup
Reclaim disk space with a safety-first workflow: investigate first, run obvious low-risk cleanup wins, then do targeted analysis for larger opportunities.
Bundled helpers:
scripts/rust_target_dirs.py: inventory and guarded deletion for explicit Rust target/ directories
references/rust-target-roots.txt: machine-specific roots for Rust artifact scans
references/ignore-paths.md: machine-specific excludes for du/ncdu
Execution Default
Start with non-destructive investigation and quick sizing.
Prioritize easy wins first (nix-collect-garbage, container prune, Cargo artifacts).
Propose destructive actions with expected impact before running them.
Run destructive actions only after confirmation, unless the user explicitly requests immediate execution of obvious wins.
Capture new reusable findings by updating this skill before finishing.
Workflow
Establish current pressure and biggest filesystems
Run easy cleanup wins
Inventory Rust build artifacts and clean the right kind of target
Investigate remaining heavy directories with ncdu/du
Investigate /nix/store roots when large toolchains still persist
Summarize reclaimed space and next candidate actions
Record new machine-specific ignore paths, Rust roots, or cleanup patterns in this skill
Step 1: Baseline
Run a quick baseline before deleting anything:
df -h /
df -h /home
df -h /nix
Optionally add a quick home-level size snapshot:
du -xh --max-depth=1 "$HOME" 2>/dev/null | sort -h
Step 2: Easy Wins
Use these first when the user wants fast, low-effort reclaiming:
sudo -n nix-collect-garbage -d
sudo -n docker system prune -a
sudo -n podman system prune -a
Notes:
Add --volumes only when the user approves deleting unused volumes.
Re-check free space after each command to show impact.
Prefer sudo -n first so cleanup runs fail fast instead of hanging on password prompts.
On macOS, a successful Xcode archive can still fail during -exportArchive with a generic Copy failed when the APFS data volume is full. Inspect the generated .xcdistributionlogs for No space left on device, and preserve several GiB of extra staging headroom for XcodeDistPipeline beyond the archive itself.
If root is still tight after these, run app cache cleaners before proposing raw rm -rf:
Do not start with a blind find ~ -name target or with hard-coded roots that may miss worktrees. Inventory explicit target/ directories first using the bundled helper and the machine-specific root list in references/rust-target-roots.txt.
Inventory the biggest candidates:
python /srv/dotfiles/dotfiles/agents/skills/disk-space-cleanup/scripts/rust_target_dirs.py list --min-size 500M --limit 30
Focus on stale targets only:
python /srv/dotfiles/dotfiles/agents/skills/disk-space-cleanup/scripts/rust_target_dirs.py list --min-size 1G --older-than 14 --output tsv
Use cargo-sweep when the repo is still active and you want age/toolchain-aware cleanup inside a workspace:
nix run nixpkgs#cargo-sweep -- sweep -d -r -t 30 <workspace-root>
nix run nixpkgs#cargo-sweep -- sweep -r -t 30 <workspace-root>
nix run nixpkgs#cargo-sweep -- sweep -d -r -i <workspace-root>
nix run nixpkgs#cargo-sweep -- sweep -r -i <workspace-root>
Use direct target/ deletion when inventory shows a discrete stale directory, especially for inactive repos or project-local worktrees. The helper only deletes explicit paths named target that are beneath configured roots and a Cargo project:
Run rust_target_dirs.py list to see the largest target/ directories across ~/Projects, ~/org, /srv/dotfiles, and other configured roots.
For active repos, prefer cargo-sweep from the workspace root.
For inactive repos, abandoned branches, and .worktrees/*/target, prefer guarded direct deletion of the explicit target/ directory.
Re-run the list command after each deletion round to show reclaimed space.
Machine-specific note:
Project-local .worktrees/*/target directories are common cleanup wins on this machine and are easy to miss with the old hard-coded workflow.
cargo-sweep is installed through the NixOS code.nix package set, but stale manually-installed binaries under ~/.cargo/bin can shadow /run/current-system/sw/bin/cargo-sweep. If cargo sweep fails with a missing loader or No such file or directory, run type -a cargo-sweep and remove the stale ~/.cargo/bin/cargo-sweep entry.
cargo-sweep sweep -i/--installed can fail when rustup toolchain list contains stale toolchains whose rustc no longer exists. On this machine, 1.68.2-x86_64-unknown-linux-gnu caused failed to determine fingerprint ... 'rustc': No such file or directory.
/home/imalison/Projects/codex/codex-rs/target can be dominated by current-looking target/debug/incremental data that cargo-sweep sweep -a and --maxsize report as not removable. If it is stale and space pressure is high, use the guarded rust_target_dirs.py delete ... --yes workflow for that explicit target directory.
/home/imalison/Projects/hypr-workspace-history/target is a small non-Cargo false positive; the guarded delete workflow correctly rejects it because there is no Cargo project above the directory.
nixos/cargo-sweep.nix defines a user timer for every user, cargo-sweep-rust-targets.timer, that every 6 hours runs cargo-sweep sweep -r --hidden --time 2 across $HOME/Projects, $HOME/org, /srv/dotfiles, and $HOME/.cargo/build, then deletes centralized cargo build dirs (~/.cargo/build/<xx>/<hash>) untouched for 2 days. CARGO_BUILD_BUILD_DIR (set system-wide in nix-shared/system/essential.nix) redirects cargo intermediate artifacts to ~/.cargo/build/, so per-project target/ dirs only hold final artifacts.
cargo-sweep does not discover centralized $HOME/.cargo/build/<xx>/<hash> directories as Cargo workspaces; the timer's wrapper handles those separately. During acute pressure, a one-time 12-hour cutoff over those hash directories, skipping any directory with open files, reclaimed 19.57 GiB on 2026-07-20. Keep the persistent timer at its two-day safety window unless the user explicitly requests a configuration change.
Step 4: Investigation with ncdu and du
Avoid mounted or remote filesystems when profiling space. Load ignore patterns from references/ignore-paths.md.
Prefer the local safe_ncdu wrapper over raw ncdu invocations. It writes compressed exports under the predictable, shared /srv/disk-space-assessments/ location (users-group readable, not a per-user ~/.cache dir), records the exclude list beside each export, excludes mounted descendants of the scan root, and keeps a latest-<root>.json.zst symlink (plus .excludes/.meta sidecars) pointing at the most recent run for each root so later analysis and later sessions — and other accounts in the users group — never have to guess a path or rescan:
safe_ncdu /home/imalison
sudo -n env HOME=/home/imalison safe_ncdu /
safe_ncdu /nix/store
safe_ncdu top /srv/disk-space-assessments/latest-root.json.zst 30 /home/imalison
safe_ncdu open /srv/disk-space-assessments/latest-root.json.zst
Only fall back to raw, non-persistent ncdu -x "$HOME" / sudo ncdu -x / for a quick interactive look when safe_ncdu itself is broken — never use it as the recorded evidence for an assessment. If ncdu is missing entirely, use nix run nixpkgs#ncdu -- -x "$HOME" the same way, as a one-off, not as a substitute for a safe_ncdu snapshot.
For quick, non-blocking triage on very large trees, prefer bounded probes:
timeout 30s du -xh --max-depth=1 "$HOME/.cache" 2>/dev/null | sort -h
timeout 30s du -xh --max-depth=1 "$HOME/.local/share" 2>/dev/null | sort -h
Machine-specific heavy hitters seen in practice:
2026-07-24 ryzen-shine online partition reclaim: GPT partition numbers
were not in physical order. /boot (nvme0n1p3) and root
(nvme0n1p4) physically preceded the obsolete Microsoft-reserved
(p1) and /shared NTFS (p2) partitions. After preserving an
sfdisk text dump and raw GPT sectors, unmounting /shared, and
validating the exact replacement table with sfdisk --no-act, writing
the table required sfdisk --no-reread because root was mounted.
partx --delete --nr 1:2 /dev/nvme0n1 removed the deleted, unmounted
partition devices from the running kernel, and
partx --update --nr 4 /dev/nvme0n1 exposed the enlarged root boundary.
resize2fs /dev/nvme0n1p4 then grew ext4 online. Always revalidate the
disk ID, partition starts, PARTUUIDs, mount state, and exact physical
order before reusing this pattern.
2026-07-22 zero-free-space failure mode: with / reporting zero bytes
available, apply_patch truncated an untracked shell function to zero bytes
before reporting its write failure. Stop all source-file writes as soon as
the filesystem reaches 100%, reclaim headroom first, and re-check the exact
byte size of any file involved in the failed write before continuing.
2026-07-10 railbird-sf incident: K3s reported DiskPressure even with tens of GiB free because its container imagefs shares /, kubelet image GC used the default 85% high-water mark, and the K3s config overrode only nodefs eviction thresholds. crictl imagefsinfo showed only ~677M of images, so image GC could not reclaim its requested ~149G and repeatedly evicted application pods. Set matching imagefs.available values alongside nodefs.available in eviction-hard, eviction-soft, and eviction-soft-grace-period; verify via K3s eviction-manager logs rather than trusting df alone.
~/.cache/uv can exceed 20G and is reclaimable with uv cache clean.
If uv cache clean reports that the cache is currently in use, do not add --force; leave it for a later idle cleanup so an active environment or install is not disrupted.
Step 5: /nix/store Deep Dive
When /nix/store is still large after GC, inspect root causes instead of deleting random paths.
Many .direnv/flake-profile-* symlinks under ~/Projects and worktrees keep nix-shell-env/ghc-shell-* roots alive.
On the macOS host, .direnv symlinks on mounted external volumes still register roots for the internal /nix/store. A 2026-07-13 audit found seven direnv profiles across the active home and /Volumes/Extreme SSD retaining 17.3 GiB of store paths unique to direnv roots; clearing the external direnvs reclaims internal Nix space, not meaningful space on the external volume.
2026-07-13 macOS direnv cleanup: preserving only the main ~/railbird-mobile/.direnv while removing six other profiles plus scrobble-scrubber result roots made 6,262 paths dead; nix-collect-garbage -d reported 8.3 GiB actually freed versus a 12.9 GiB preflight narSize estimate.
Old taffybar constellation repos under ~/Projects can pin large Haskell closures through .direnv and result symlinks. Deleting gtk-sni-tray, status-notifier-item, dbus-menu, dbus-hslogger, and gtk-strut and then rerunning nix-collect-garbage -d reclaimed about 11G of store data in one validated run.
find_store_path_gc_roots is especially useful for proving GHC retention: many large ghc-9.10.3-with-packages paths are unique per project, while the base ghc-9.10.3 and docs paths are shared.
NixOS system generations and a repo-root nixos/result symlink can pin multiple Android Studio and Android SDK versions. Check /nix/var/nix/profiles/system-*-link, /run/current-system, /run/booted-system, and /srv/dotfiles/nixos/result before assuming Android paths are pinned by project shells.
~/Projects/railbird-mobile/.direnv/flake-profile-* can pin large Android SDK system images. Removing stale direnv profiles there is a more targeted first step than deleting Android store paths directly.
2026-05-27 Railbird GHC audit: the Railbird backend flake did not explicitly reference Haskell, but its dev shell had derivation-time GHC edges through inputs.secrets.devShells.${system}.default -> agenix -> shellcheck -> ShellCheck -> ghc and through shell-packages.nix's . Railbird Mobile had similar non-app-code GHC edges through / and . The and shells did not show GHC edges in their derivation graphs, only Rust/Cargo build tooling from packages such as .
safe_ncdu /nix/store
safe_ncdu top /srv/disk-space-assessments/latest-nix_store.json.zst 30
nix-store --gc --print-dead plus the Nix SQLite database is a fast way to estimate immediate GC wins before deleting anything:
If counts are high and the projects are inactive, propose targeted .direnv cleanup for user confirmation.
2026-08-14 jimi-hendnix (multi-user railbird box) — single largest validated win in this skill's history: 251.8 GiB freed by one nix-collect-garbage -d, moving / from 90% used (94G free) to 44% (499G free). Three cleanup classes fed that single GC, and none of them freed meaningful space until the GC ran:
Dead nixtheplanet macOS VM, ~167 GiB. Three plain symlinks (/home/imalison/macos-ventura-base-image.qcow2, /home/kat/macos-ventura-base-image.qcow2, /var/lib/private/nixtheplanet-macos-ventura/macos-ventura-base-image.qcow2) each pinned a ~40 GiB mac_hdd_ng.qcow2. Because this host sets keep-derivations = true and keep-outputs = true, they also transitively pinned four InstallAssistant.iso/.pkg build inputs (45 GiB) and BaseSystem.img (3 GiB). Those build inputs have zero referrers — nix-store --query --referrers-closure returns only the path itself — so only nix-store --query --roots <path> reveals the retention. Use --query --roots, not referrers, whenever a huge store path looks unrooted but is not in --print-dead. services.macos-ventura.enable was already false and the 19 GiB runtime disk was last touched 2024-11-01.
13 stale .direnv dirs across six users' homes, ~56 GiB collectively direnv-only, all 600–800 days old. On disk they were only 76–248 KiB each — the size is entirely what they pin. The four per-user railbird checkouts each had a ~14.8 GiB closure but ~3 MiB marginal-unique, so they only paid off removed together.
/var/lib/private/gitea-runner, 32 GiB, removed outright along with the service (see below).
2026-08-14 nix-store --gc --print-dead returned zero dead paths on the first run because a concurrent nix process held ~15k temp GC roots (nix-store --gc --print-roots was dominated by {temp:NNNN} entries). After that process died and its temproots file went stale, the same command found 345 dead paths. If --print-dead says zero on a store you have reason to believe is dirty, check the roots listing for domination and re-run.
Removing a Service Entirely (not just its caches)
Validated 2026-08-14 on jimi-hendnix for gitea-runner. When the user wants a service gone rather than trimmed, disable it in Nix and rebuild before deleting state, so nothing recreates the directory:
Confirm no job is mid-flight: systemctl status <unit> should show the bare daemon with no child job processes, and pgrep -a -u <service-user> should list only the daemon.
Flip the flag in the host config (nixos/machines/<host>.nix) to false rather than deleting the line — this host already expresses services.macos-ventura.enable = false and myModules.railbird-k3s.enable = false the same way.
just switch from /srv/dotfiles/nixos. A successful run prints removing user 'gitea-runner'.
Only then rm -rf the state tree.
Note the server-side consequence: the runner's registration (.runner token) is destroyed, so the runner will show as offline in the Gitea instance and must be re-registered if ever reinstated.
just switch gotchas hit during that run
safe_switch refuses to start when a tmux session already exists on the nixos-switch socket, and instead silently tails the previous run's log — so it looks like a successful switch that did not include your change. Check tmux -L nixos-switch list-panes -t switch -F '#{pane_pid} #{pane_dead}'; pane_dead=1 means the old run finished and the session is just lingering. tmux -L nixos-switch kill-session -t switch, then re-run. Always verify with readlink /run/current-system plus a systemctl is-enabled check rather than trusting the tail output.
/boot on this host is deliberately kept ro by another workflow (the journal shows a mount -o remount,rw /boot immediately followed by remount,ro). switch-to-configuration then fails with OSError: [Errno 30] Read-only file system: '/boot/loader/entries/...' and Failed to install bootloaderafter the config has already built. Before assuming corruption, check dmesg for fat-fs/device errors against the actual /boot device — in this run the I/O errors were all on sdd/sdc3 (removable rescue media), and boot-time systemd-fsck reported /dev/nvme1n1p1 clean. Remount rw, re-run the switch, then restore ro to leave the machine as found.
Safety Rules
Do not delete user files directly unless explicitly requested.
Prefer cleanup tools that understand ownership/metadata (nix, docker, podman, cargo-sweep) over rm -rf.
For Rust build artifacts, deleting an explicit directory literally named target is acceptable when it is discovered by the bundled helper; Cargo will rebuild it.
Present a concise “proposed actions” list before high-impact deletes.
If uncertain whether data is needed, stop at investigation and ask.
Learning Loop (Required)
Treat this skill as a living playbook.
After each disk cleanup task:
Add newly discovered mountpoints or directories to ignore in references/ignore-paths.md.
Add newly discovered Rust repo roots in references/rust-target-roots.txt.
Add validated command patterns or caveats discovered during the run to this SKILL.md.
Keep instructions practical and machine-specific; remove stale guidance.
Later on 2026-07-20, repeating that acute-pressure workflow after verifying no active Cargo/Rust processes or open files reclaimed another 65.94 GiB from 32 centralized build shards and moved / from 98% used (18 GiB free) to 91% used (84 GiB free). Check freshness recursively within each hash directory, recheck immediately before deletion, and leave the persistent two-day timer unchanged.
On 2026-07-20, / reached 100% with only 1.5 GiB free while all centralized build shards had recursive activity within 12 hours. After verifying no active Cargo/Rust processes or open files, a one-time six-hour cutoff removed the oldest 20 shards (last touched about 9.9 hours earlier) and reclaimed 29.24 GB of filesystem space, leaving 30 GB free. Reserve this shorter cutoff for acute pressure, validate the exact <xx>/<hash> paths, and leave the persistent two-day timer unchanged.
On 2026-07-21 this host had no uv executable on PATH; after confirming no open references or writes in the last hour, directly removing the idle ~/.cache/uv directory reclaimed 8.2G.
2026-07-23 Railbird Mobile Yarn Berry/Nix rebuild: regenerating a large fetchYarnBerryDeps offline cache and then materializing node_modules exhausted the filesystem with only 299 MiB free. Removing 803 MiB of task-created conversion scratch and running unprivileged nix store gc reclaimed 15.4 GiB; the final node_modules derivation then succeeded with about 18 GiB free. Preserve roughly 18 GiB of headroom for this dependency build.
~/.cache/yarn and ~/.cache/cabal can together exceed 4G on this machine; after checking for open files and recent writes, clearing these dependency caches is a safe reclaiming step (dependencies will be redownloaded or rebuilt).
~/.cache/pypoetry can exceed 7G across artifacts, repository cache, and virtualenvs; inspect first, then use Poetry cache commands or targeted virtualenv removal.
~/.cache/google-chrome can exceed 8G across multiple Chrome profiles; close Chrome before clearing profile cache directories.
~/.cache/spotify can exceed 10G; treat as optional app-cache cleanup.
~/.gradle can exceed 8G, mostly under caches/; prefer Gradle-aware cleanup and expect dependency redownloads.
~/.local/share/picom/debug.log can grow past 15G when verbose picom debugging is enabled or crashes leave a stale log behind; if picom is not running, deleting or truncating the log is a high-yield low-risk win.
~/.local/share/Trash can exceed several GB; empty only with user approval.
On the macOS Codex host, ~/.cache/keepbook-cargo-target can grow to several GiB as a Rust target cache and is safe to remove when not actively building keepbook.
On the macOS Codex host, nix run github:ccusage/ccusage/v20.0.17 attempted a source build of 213 Rust derivations (about 1.5 GiB unpacked) and exhausted the nearly full APFS container. For ccusage-fleet remotes without npx, prefer nix shell nixpkgs#nodejs --command npx --yes ccusage@<version> ...; the npm package uses a prebuilt native binary. After aborting the Rust build, unprivileged nix store gc removed 670 dead paths (estimated at 4.5 GiB) and increased APFS free space from 207 MiB to 6.2 GiB.
On the macOS Codex host, large low-risk cache wins have included ~/Library/Caches/Homebrew via brew cleanup --prune=all -s, Yarn/npm/CocoaPods caches, ~/Library/Caches/com.spotify.client, ~/Library/Caches/ms-playwright, and stale app updater caches such as ~/Library/Caches/com.anthropic.claudefordesktop.ShipIt.
2026-07-20 macOS cache cleanup: yarn cache clean reclaimed the 3.1 GiB ~/Library/Caches/Yarn/v6 cache. Removing stale Sparkle payloads under ~/Library/Caches/com.openai.codex/org.sparkle-project.Sparkle/Installation (1.9 GiB) and ~/Library/Caches/com.anthropic.claudefordesktop.ShipIt/update.MKJITXP (751 MiB) increased available Data-volume space from 4.1 GiB to 11 GiB.
2026-07-20 macOS Railbird iOS build: a clean signed simulator build used about 5 GiB of task-specific DerivedData and failed at GenerateDSYMFile when less than 1 GiB remained. Removing only that task's DerivedData before rebuilding, then running sudo -n nix-collect-garbage -d, provided enough headroom; preserve at least 6 GiB free for a clean build.
2026-07-21 macOS Railbird TestFlight archive: clearing the regenerable ~/Library/Caches/Google cache reclaimed about 1.1 GiB while Xcode was archiving. Chrome may retain a live cache directory, so partial-removal "Directory not empty" messages are expected; remeasure reclaimed space rather than treating them as a cleanup failure.
2026-07-21 macOS Railbird worktree release: symlinking node_modules to the primary checkout caused Xcode duplicate React-Fabric copy outputs. A worktree-local hard-link copy (cp -al <primary>/node_modules node_modules) kept CocoaPods paths local while adding minimal data volume.
On the macOS Codex host, allowing direnv in a fresh railbird-mobile/.worktrees/* worktree can materialize the full Android SDK dev shell even for workflow-only edits. When free space is already low, use the existing root worktree's tools or direct formatters first; a failed shell realization followed by sudo -n nix-collect-garbage -d reclaimed about 6.3 GiB on 2026-07-12.
2026-07-13 macOS Railbird worktree cleanup: a fresh Android dev-shell realization repeatedly filled the disk while unpacking NDK r23b. After the failed build exited, 463 paths were dead; sudo -n nix-collect-garbage -d reclaimed 12.0 GiB. Avoid retrying that worktree shell and reuse the main worktree's dependencies/toolchain when the task only needs formatting, linting, or TypeScript checks.
2026-07-09 macOS runner cleanup: sudo -n nix-collect-garbage -d removed 11,455 unreferenced store paths and reclaimed 24.4 GiB, increasing shared APFS container free space from about 3 GiB to 30 GiB. The preflight estimate from the Nix DB reported only 15.6 GiB of dead-path narSize, so actual reclaimed filesystem space can be materially larger.
2026-07-09 Xcode runner update: mas outdated correctly detected Xcode 16.3 -> 26.6, but mas upgrade 497799835 stalled when no active App Store purchase account was present (needsAuthentication because appleID is nil in unified logs). Verify App Store sign-in before relying on unattended mas upgrades.
On the macOS Codex host, /private/var/folders/*/*/X/*.code_sign_clone can accumulate stale app framework copies, especially Chrome and Codex clones. Remove only stale clone directories after checking age; avoid blind deletion of the whole /private/var/folders tree.
On the macOS Codex host, ~/Library/Developer/CoreSimulator/Caches/dyld can exceed several GiB. Clean it only after confirming simulator shutdown is acceptable, because CoreSimulator services may be running.
On the macOS Codex host, xcrun simctl runtime delete <UUID> can remove an installed simulator runtime while its source MobileAsset remains under /System/Library/AssetsV2/com_apple_MobileAsset_iOSSimulatorRuntime. In a 2026-07-13 iOS 18.4 cleanup, the runtime and its mounted Cryptex image disappeared but the 8.3 GiB source asset remained; do not raw-delete system-managed AssetsV2 content, and measure actual APFS free-space change rather than assuming the runtime's displayed size is fully reclaimed.
/var/lib/private/gitea-runner can exceed 50G and is not visible to an unprivileged ncdu / scan; use sudo -n env HOME=/home/imalison safe_ncdu / when /var looks undercounted.
Validated cleanup pattern: stop gitea-runner-nix.service, remove cache/work directories under /var/lib/private/gitea-runner (.cache, .gradle, action-cache-dir, workspace, stale nested gitea-runner, and nested nix/.cache/nix/.local), recreate action-cache-dir, workspace, and .cache owned by gitea-runner:gitea-runner, then restart the service.
Preserve registration/config-like files such as /var/lib/private/gitea-runner/nix/.runner, /var/lib/private/gitea-runner/nix/.labels, /var/lib/private/gitea-runner/.docker/config.json, and SSH/Kube material.
On the macOS runner, state is split between /private/var/lib/gitea-runner and /Users/gitea-runner; the daemon label is org.nixos.gitea-runner-nix. Stop it with sudo -n launchctl bootout system /Library/LaunchDaemons/org.nixos.gitea-runner-nix.plist, clear only regenerable action clones/caches/temp data, recreate required directories as gitea-runner:staff, then restore it with launchctl bootstrap and verify it is running.
The iOS workflow may mount /Volumes/Extreme SSD/railbird-mobile-ci/build.sparseimage inside an action worktree. A failed cleanup can delete the sparse-image pathname while leaving its disk image attached. If hdiutil info shows that missing path under mounting user: gitea-runner, unmount the exact stale worktree mount, identify the parent disk device from the same hdiutil info image block, and detach that device with sudo -n hdiutil detach /dev/diskN. Never infer or detach a disk from its number alone.
Preserve .runner, configuration, SSH material, keychains, and provisioning profiles. Validated regenerable targets include action-cache-dir, runner .cache and tmp contents, Cargo registry cache, CocoaPods repos, and runner-owned Xcode caches/DerivedData when no job is active.
/Volumes/Extreme SSD is ExFAT with noowners, so it is not a safe direct home or worktree filesystem for Git, Nix, CocoaPods, or Xcode. Use a capped APFS sparse bundle stored on the SSD, keep registration and credential state internal, and make runner startup fail closed rather than falling back to the internal disk when the image is unavailable.
2026-07-22 external SSD audit: stale Cargo target directories and JavaScript node_modules under /Volumes/Extreme SSD/Projects occupied about 60 GiB. These are the first candidates for creating runner-image headroom, but delete them only after approval and after checking for active builds.
2026-07-22 validated migration: deleting seven approved target/node_modules trees reclaimed 59.1 GiB and moved the external SSD from 46 GiB to 105 GiB free. A 64 GiB APFS sparse bundle at /Volumes/Extreme SSD/railbird-runner/runner.sparsebundle mounted successfully at /private/var/lib/gitea-runner-external; its initial physical size was 37 MiB. The live runner used the external work directory plus external Cargo, Rustup, Gradle, CocoaPods, Yarn, npm, Go, pip, uv, XDG, runtime, and temp caches, while registration and credential state stayed internal.
2026-07-22 internal cleanup: clearing inactive user build/package caches, deleting the superseded nix-darwin generation, and running nix-collect-garbage moved the internal APFS pool from about 3.4 GiB to 17 GiB free; the final Nix GC alone reported 8.6 GiB reclaimed. The active configuration schedules daily GC and optimisation at 04:15 and reserves free space so future builds trigger cleanup before the disk reaches zero.
2026-07-22 T3 Code rebuild: compressing the pnpm dependency store exhausted the APFS pool even after a failed build had recovered 5.3 GiB free. sudo -n nix-collect-garbage -d, brew cleanup --prune=all -s, and npm cache clean --force reclaimed another 4.3 GiB. A subsequent rebuild that needed both the active and replacement T3 closures was unblocked by pnpm store prune, which removed 133,104 files / 1,795 packages and reclaimed 5.32 GiB. Preserve roughly 10 GiB of build headroom for this derivation.
2026-08-02 runner-image cleanup: with no active job or open files, stale nix-shell.* scratch under /private/var/lib/gitea-runner-external/cache/tmp occupied 8.8 GiB. Boot out org.nixos.gitea-runner-nix, terminate any orphaned mount-gitea-runner-storage or job log stream processes that still reference the mount, clear only the validated temp contents, detach the exact device identified by the runner.sparsebundle block in hdiutil info, and run hdiutil compact before bootstrapping the runner again. Compaction reclaimed 18.3 GB of physical SSD space; verify both the mount and runner PID afterward.
2026-08-02 external Paseo cleanup: /Volumes/Extreme SSD/Projects/paseo/node_modules consumed 70 GiB on ExFAT because its 1 MiB allocation units massively amplified a dependency tree of small files. After confirming the running Paseo app and daemon came from installed Nix paths and that no process or open file referenced the checkout, deleting that exact regenerable directory reclaimed 70 GiB. Long ExFAT deletions may outlive an exec yield while continuing under the session manager; locate and monitor the original rm PID to completion instead of starting a competing deletion.
2026-08-05 railbird-sf remote cleanup (/ at 100%, 109 MiB free → 86%, 124 GiB free): the three big reclaimable pools were podman image storage, agent worktrees, and Nix-store roots those worktrees pinned. sudo -n podman system prune -a -f reclaimed 31.5 GB (three 10.5 GB registry-tagged railbird-full images plus stale org-agenda-api tags — all re-pullable from us-central1-docker.pkg.dev; running kokoro/org-agenda-api containers kept their images). The validated worktree sweep over ~imalison/Projects/{paseo,subtr-actor,rocket-sense,t3code}/.worktrees removed ~60 clean worktrees (~85 GiB apparent; paseo rebase-pr-* worktrees are ~3 GiB each from node_modules) and skipped 13 with real uncommitted work. Rerunning nix-collect-garbage -d AFTER the worktree sweep freed 10.4 GiB vs 0.4 GiB before it — removed worktrees' .direnv profiles were the roots, so always GC again after a worktree purge. Left intact: 144 GiB of live k3s PVCs under /var/lib/rancher/k3s/storage (rocket-sense postgres + storage — application data, never deletable), gitea-runner .gradle (33 GiB, an Android CI job was actively running — Gradle/Kotlin daemons visible in ps), and /tmp/paseo-* agent caches. journalctl --vacuum-size=500M freed 0B (all 4 GiB was active journals). New failure pattern: git worktree remove --force can FAIL and leave an orphan when the checkout contains root-owned files (container-run __pycache__); the fix is sudo rm -rf of the verified-clean leftover, then git worktree prune.
On mac-demarco-mini, sudo -n requires a password for imalison and root SSH is denied, so privileged cleanup (root nix-collect-garbage, other users' caches, simulator runtime deletion) needs an interactive session; unprivileged nix store gc freed only 250 MiB. Remaining large admin-only candidates seen 2026-08-05: AssetsV2 Apple Intelligence models ~21 GiB (system-managed, do not raw-delete), iOS simulator runtimes 16 GiB (18.4 + 26.5), kat's .yarn cache 6.3 GiB.
On macOS, safe_ncdu scan currently assumes Linux findmnt and cannot make a fresh snapshot. Use ncdu -0 -x -c -o <snapshot>.json.zst /System/Volumes/Data, then use safe_ncdu top with zstdcat available to query it.
~/Projects/*/target directories can dominate home usage. Recent example candidates included stale target/ directories under scrobble-scrubber, http-client-vcr, http-client, subtr-actor, http-types, subtr-actor-py, sdk, and async-h1.
2026-05-26 cleanup: deleting explicit Cargo-backed targets under ~/Projects/{keepbook,subtr-actor,rlru,rocket-sense,boxcars,rumno} plus stale subtr-actor/.worktrees/*/target reclaimed about 65G by helper sizing and moved / from 100% used to 89% used. A final all-depth scan left no ~/Projects Rust target/ directories over 500M.
2026-05-26 cleanup: when cargo test is actively running in ~/Projects/subtr-actor, leave subtr-actor/target alone and delete only inactive Cargo-backed targets. Deleting keepbook, rlru, rocket-sense, rumno, and stale subtr-actor/.worktrees/*/target reclaimed about 24.5G by helper sizing.
2026-05-26 cleanup: ~/Projects/nixpkgs/.worktrees/*/result symlinks pinned several GiB of Nix closures, and clean registered nixpkgs worktrees were about 460M each. Removing stale result symlinks, running GC, and removing clean worktrees while preserving dirty ones moved / from 100% used to about 90% used.
2026-05-27 cleanup: under ~/Projects, hypr-workspace-history/target can be a Rust-style build cache even though the guarded helper rejects it because no Cargo.toml is present; inspect and remove that explicit cache manually if present. Preserve ~/Projects/Hyprland/src/layout/target, which is source code, not a build artifact.
2026-07-18 agent-worktree purge (railbird/.agents/worktrees, railbird-mobile/.worktrees+.claude/worktrees, subtr-actor/.worktrees, rocket-sense/.claude/worktrees): validated safety workflow for bulk git worktree remove across many agent-generated worktrees:
Check for an active session first, before checking git state. Scan /proc/*/cwd for every candidate worktree path; a live Claude/Codex agent process with cwd inside one (e.g. railbird-mobile/.claude/worktrees/pr-1840-ci-comments-7df0a9 in this run) means skip it outright — removing the worktree out from under a running agent is a correctness hazard regardless of git state.
git worktree remove (even --force) never deletes the branch or its commits, only the linked checkout — so a clean worktree on an unpushed named branch is still safe to remove; the commits remain reachable in the main repo's refs afterward. Only uncommitted working-tree changes are actually at risk of being lost.
For a clean worktree in detached HEAD, check git branch --contains <commit> -a before removing — if the commit is already an ancestor of master or another branch (pushed or local), it's safe; if not, it's only reflog-recoverable and riskier.
A single-line M <submodule-path> diff (e.g. M gql, M vendor/subtr-actor) is routine submodule-pointer drift from a stale checkout, not real work — treat it as safe to discard, unlike a real source-file diff.
Some worktree directories can be orphaned: still present on disk but no longer in git worktree list --porcelain (their .git gitdir file points at a .git/worktrees/<id> entry git has already forgotten, so git status fails with fatal: not a git repository). These can't be cleaned with git worktree remove; rm -rf the directory directly, then git worktree prune.
Result: 4/5 railbird/.agents/worktrees removed (28G), 6/7 railbird-mobile worktrees removed (skipped the active session and a /tmp worktree with unexplained untracked credential-like files: access_tokens.db, credentials.db, legacy_credentials/), 2/3 subtr-actor/.worktrees removed, 13/15 rocket-sense/.claude/worktrees removed (2 were already-orphaned). Net effect on this run: / went from 94% used (54G free) to 82% used (155G free), combined with rancher/gradle/tmp/node_modules cleanup below.
2026-07-18 cleanup: /var/lib/rancher (4.7G) was orphaned state from a fully uninstalled rancher-desktop/containerd/k3s (all systemctl is-enabled returned not-found) — safe to sudo rm -rf outright, no service to stop first.
2026-07-18 cleanup: root /tmp (not $HOME/tmp) had accumulated 9,450 nix-shell.* ephemeral TMPDIR dirs (4.6G, almost all empty) plus ~500 named agent/session scratch dirs (codex-*, t3code-*, *-audit, *-validation-*, installer .dmgs) totaling ~12G. Before bulk-deleting /tmp, cross-check git worktree list output for entries rooted in /tmp (a railbird-mobile-rebase worktree lived at /tmp/railbird-mobile-rebase in this run) and check /proc/*/cwd / lsof +D /tmp for live processes (found an active rumno service writing rumno.pid/.out/.err, and 730 small live com.google.Chrome.* sandbox dirs from running Chrome/Electron apps) — exclude those from the sweep. Some extracted scratch dirs (from .asar unpacking, tarball extraction) have read-only directory bits (dr-xr-xr-x, mtime epoch-0) that make plain rm -rf fail with Permission denied; chmod -R u+rwX <dir> before retrying rm -rf fixes it.
2026-07-20 acute unblock: when manual shard deletion is unavailable, systemctl --user start cargo-sweep-rust-targets.service is a sanctioned first move — one run swept project target/ dirs and freed 43G (/ 100% -> 95%) in under a minute, without touching the 2-day ~/.cargo/build window.
2026-07-19 cleanup: inactive /tmp/rmk-*, /tmp/glove80-*, and /tmp/codex-asar.* build/extraction directories accumulated about 85G in one day. Before deleting these explicit top-level patterns, scan /proc/*/{cwd,root,exe,fd/*} for references and check the candidates for nested .git entries; with neither present, removing them reduced /tmp from 63G to 4.2G and restored 85G of filesystem headroom.
2026-07-22 T3 Code review-worktree unblock: many detached .worktrees/pr-* checkouts each materialized about 2.7G of node_modules while repeated PR review cycles were running. After checking /proc/*/cwd and selecting only completed, inactive worktrees, git -C <worktree> clean -fdx -- node_modules on five explicit targets restored about 1.9G of physical space. Preserve source and foreign/active worktrees, and prefer relinking the repository's existing dependencies over creating another independent install.
2026-07-24 T3 Code stack cleanup: deleting all 174 node_modules directories under /home/imalison/src/t3code (including workspace-local link farms) removed about 30.7G of apparent data but reclaimed only 1.7G of filesystem space because the large pnpm trees shared content/hard links. Enumerate exact paths first, verify /proc/*/{cwd,maps,fd/*} has no node_modules users, and use git -C <worktree> clean -fdx -- <relative-node_modules-path> for each validated path. The Nix package/system rebuild remained reproducible without repository-local dependencies.
2026-07-29 abandoned-git-tempfile pattern (highest-yield find of that run): ~/Projects/nixpkgs/.git/objects was 40G against only 6.36 GiB of real packs. git count-objects -vH reported garbage: 3288 / size-garbage: 33.05 GiB — abandoned .git/objects/pack/tmp_pack_* files left behind by interrupted fetches/repacks over months. Always run git count-objects -vH on any repo whose .git dwarfs its worktree; git names the waste itself. Check ps for a live git fetch/gc and the newest tmp_pack_* mtime before clearing, since an in-flight fetch legitimately owns one. Some tmp_pack_* files are mode -r--------, so rm needs -f or a prior chmod. A scan of every other repo under ~/Projects found this pattern essentially unique to nixpkgs (t3code had a single 0.1G stray), so scan broadly but expect one offender. Validated removal: find .git/objects/pack -maxdepth 1 -name 'tmp_pack_*' -delete reclaimed 33.8 GiB in seconds and left the repo intact (garbage: 0, all 9,581,748 objects still in-pack, git rev-parse HEAD and git status fine). This is far cheaper than git gc, which would repack 9.5M objects to remove data git already classifies as garbage — prefer the targeted delete.
2026-07-29 counterexample worth remembering: a large .git is not automatically waste. ~/Projects/railbird/.git is 6.7G but reports garbage: 0 with 6.42 GiB of legitimate packs across only 54,745 objects — i.e. large binary blobs committed to history, nothing to reclaim. Run git count-objects -vH to tell the two cases apart instead of assuming; the in-pack vs size-pack ratio is the tell (tiny object count + huge pack = real blobs, not garbage).
2026-07-29 /tmp caveat that overrides the "ephemeral scratch" assumption: descriptively-named /tmp dirs can be full git checkouts holding the only copy of real work. t3-integ, t3-integ2, and t3-4506-compat each contained a nested .git with staged and untracked source changes and no remote at all (git remote get-url origin empty), so deletion would have been unrecoverable. Always run find <dir> -maxdepth 3 -name .git on each /tmp candidate and then git status --porcelain plus git remote get-url origin on every hit — the registered-worktree scan (git worktree list --porcelain | grep ^worktree /tmp) does NOT catch these, because standalone clones are not worktrees of anything.
2026-07-29 validated safe /tmp subset on this machine: metro-cache, node-compile-cache, a stray tmp.*/app-release.apk, and 4078 nix-shell.* dirs (no .git in any) together reclaimed 1.5 GiB. The nix-shell.* dirs are individually tiny but a handful reach ~100M each.
2026-07-29 hard-link accounting reminder: safe_ncdu top reported railbird-mobile at 49.3 GiB and t3code at 38.4 GiB where du -x said 27G and 15G. The gap is pnpm/node_modules hard links. Trust du -x (or measured df deltas) for reclaimable space and treat ncdu's apparent sizes as an upper bound only.
2026-07-29 agent-worktree accumulation recheck: .worktrees dominated every large JS/Python repo — railbird-mobile 22G/5, t3code 14G/41, org-agenda-api 5.8G/1, railbird 2.3G/26. Also on this machine: railbird/.venv 7.1G, railbird/.git 6.7G, influence-sensitivity/.lake 6.9G + .elan 2.6G (Lean toolchain/build, regenerable). Preserve subtr-actor-mechanics/ml (9.3G) and data (1.2G) — those are ML datasets/checkpoints, not build output.
2026-07-29 worktrees live OUTSIDE the repo more often than expected, and du on the repo will not find them. Always cross-check git worktree list --porcelain against on-disk .worktrees/, because the registry is the only thing that reveals external checkouts. Found this run: 8 registered t3code worktrees under ~/.t3/worktrees/t3code/ (5.4G, the T3 Code desktop app's own state dir, 11 dirs on disk so 3 orphans), and two registered railbird worktrees at sibling paths ~/Projects/railbird-agents-architecture and ~/Projects/railbird-alert-tuning. Assign these to the agent that already owns the parent repo — never a second writer, or the two race on .git/worktrees metadata.
2026-07-29 railbird/.claude is a symlink to .agents, so .claude/worktrees/ and .agents/worktrees/ are the SAME directory (same inode). Check with ls -ld before reporting them as two separate cleanup targets or double-counting their size.
2026-07-29 detached-HEAD reachability: git branch --contains -a is NOT sufficient — it misses tags. Use git for-each-ref --contains <commit> as well. In t3code this distinguished thread-picker-build (reachable only via refs/tags/t3code-thread-picker/*, so safe to remove) from six worktrees whose commits no ref contained at all (reflog-only — correctly skipped).
2026-07-29 orphan verification without git: when an orphan's gitdir pointer targets a parent repo that no longer exists at all (here ~/dotfiles/dotfiles/emacs.d/straight/repos/org-agenda-api/.git, a deleted straight.el package clone), git status cannot help. The validated technique is to content-hash every file in the orphan and check blob presence in a surviving object store: 54/54, 34/35, and 29/29 matched, with the only misses being generated artifacts (.elc bytecode whose .el source was present, and a gitignored expo-env.d.ts). For a nested submodule whose blobs correctly do NOT live in the superproject, trace .gitmodules to the canonical repo and verify the pinned gitlink commit exists and is reachable there. This is much stronger evidence than an mtime heuristic.
2026-07-29 git worktree remove can fail with fatal: working trees containing submodules cannot be moved or removed even on a completely clean worktree. --force is correct there — it is overriding a submodule-mechanics objection, not discarding uncommitted work. Do not read that particular --force as risky, but do confirm git status --porcelain was empty first, and verify branches still resolve afterward.
2026-07-29 measured yields, one agent per repo (df deltas are cross-contaminated when agents run concurrently — always have each report an attributable du -sk taken immediately BEFORE deleting): t3code 26 of 33 worktrees removed (7 skipped: 1 real diff, 6 reflog-only detached), railbird-mobile 5 of 5, railbird 22 of 24 (2 skipped for genuine uncommitted source work), org-agenda-api 4 of 4 (5.73 GiB attributable, of which 100% was mova/node_modules 3.9G + mova/android 1.9G). Expect roughly 1 in 8 worktrees to hold real unpushed work — the skip rate is not zero, so never bulk-remove without the per-worktree status check.
2026-07-29 ~/.t3 (T3 Code desktop app) is 9.0G: worktrees/ 5.4G and userdata/logs/ 2.8G, of which logs/provider/ is 2.6G across 332 per-session <uuid>.log files. Only ~45 files (377M) were older than 7 days. The app rotates desktop.trace.ndjson.N / server.trace.ndjson.N / server-child.log.N at 10M each. Check ps for a running T3 Code node process and scan /proc/*/fd/* before touching any of it; leave userdata/secrets, attachments, and chats alone entirely.
rdma-core -> pandoc-cli -> ghc
inputs.secrets
agenix
nixGLIntel -> shellcheck
railbird/gql
railbird-mobile/src/gql
just
For a repeatable /nix/storencdu snapshot without driving the TUI, use safe_ncdu rather than a one-off /tmp export, so the snapshot lands at the predictable /srv/disk-space-assessments/latest-nix_store.json.zst path instead of a location later sessions won't know to check:
