[magma-beef]

Summary

Feature — the post step runs a best-effort fstrim on the buildkit sticky disk (/var/lib/buildkit) right before unmounting a disk it is about to commit, for at most the number of seconds the VM agent sends. Nothing changes until a host is configured: the agent sends 0 today, and the hypervisor does not negotiate discard yet.

Start here: src/setup_builder.ts (trim helper + where it sits in cleanup), then src/main.ts (reads the budget from the mount response, saves it in state). dist/ is generated.

Why

The buildkit sticky disk is the largest sticky-disk population by bytes. It is an RBD image; layers buildkit prunes stay allocated in Ceph until the guest discards them (fleet-wide ~47 % of allocated sticky-disk bytes are filesystem-free, measured 2026-10-19). Same change as useblacksmith/stickydisk#82 for the bare action.

What

• In post, when the run is still on track to commit (integrity check passed, no commitEarlyDeny, no SIGKILL, no failed steps) and the host enabled the trim, run timeout <secs> sudo fstrim -v -- <mount> after buildkitd stops and before the df sample and umount. Any exit code or exception only logs (fstrim <mount>: <output> in Nms | failed in Nms: timed out after Ns | ...). The previous-step check is evaluated once and reused for the commit decision, so skip reasons are unchanged.
• The budget comes from the VM agent: GetStickyDiskResponse.fstrim_timeout_secs (FastActions/fa#6600, live config guest_fstrim_timeout_secs), saved in action state at setup. 0/absent = no trim, so this release does nothing until the agent sends a budget — the rollout and its tuning are agent-side.
• fstrimTimeoutSecsFrom() reads the typed field when the bundled proto package has it, else unknown field 8 from $unknown (protobuf-es keeps unknown binary fields; the transport is gRPC), so this version keeps working once the agent starts sending the field without a package bump.

Evidence

Check	Result
lefthook pre-commit (prettier, eslint, tsc, build, dist diff)	clean
pnpm test	34 pass; new tests for fstrimTimeoutSecsFrom (typed field, unknown field 8 on a real response decoded with the field-less schema, wrong wire type), decodeVarint, parseFstrimTimeoutSecs
CI	Build, verify-build, 5 setup-docker-builder e2e legs, step-duration gate, CodeQL green

Risks

Verdict: Ship. With the agent sending 0 (current default) post only reads one state value; the trim path is reachable only after a host is configured.
Worry about: once a host is configured and discard is negotiated, the first trim of a large never-trimmed buildkit disk runs until the budget expires and post is longer by up to that budget; later trims are incremental.
Checked, not worried: every trim failure mode (unsupported discard, missing binary, timeout) logs and continues; the trim runs after buildkitd is stopped and before the df sample, so the committed state and the sample agree; non-commit runs skip it.
If it goes wrong: post logs fstrim /var/lib/buildkit: failed ...; set guest_fstrim_timeout_secs back to 0 on the host — no action release needed.

Related: useblacksmith/stickydisk#82, useblacksmith/checkout#63, FastActions/fa#6600, useblacksmith/web#18629.

Link to Devin session: https://app.devin.ai/sessions/2e17815c306b4c3cb7af4f6bce008a66
Open in Devin Desktop: https://app.devin.ai/desktop/session/2e17815c306b4c3cb7af4f6bce008a66?variant=devin
Requested by: @magma-beef

---

---

^{Need help on this PR? Tag @codesmith-bot with what you need. (Staging)}

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[blacksmith-staging]

Post trims freed blocks before unmount, then commits

Before, the deleted 512 MiB stayed allocated on the backing disk after post. After, post trimmed before the df sample and unmount, the backing disk dropped back to its empty size, and the commit still went out.

_Merge-base

_{This PR 56214f0b}

Also checked · 4 more claims, all matched observations

Setup reads budget from unknown field 8
No budget means no trim
Failed earlier step: no trim, same skip reason
fstrim timeout only logs

_{Proof on 56214f0b}

View full Proof report