[magma-beef]

In the post step, run fstrim on the sticky disk before it is unmounted and committed, so freed blocks are discarded from the backing image.

• The agent sends a trim budget in GetStickyDiskResponse.fstrim_timeout_secs (field 8). main reads it (from the typed field, or from unknown field 8 when the bundled proto package predates it) and stores it in action state; 0/absent disables the trim.
• post evaluates the full commit decision (mode, early deny, if-missing, on-change verdict, previous-step failures) before unmount and trims only when the budget is positive and a commit will be requested. Order: sync → df sample → trim → drop_caches → umount → flushbufs.
• Best-effort: timeout <budget> sudo fstrim -v; a missing binary, unsupported discard, or timeout only logs and never fails the step.

Start here: src/post.ts (commit decision, trim ordering), then src/mount.ts (trimFilesystem, fstrimTimeoutSecsFrom). dist/ is regenerated.

Checks: npm run lint, npx tsc --noEmit -p ., npm test, npm run build.

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]

✅ Devin Review: No Issues Found

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

[blacksmith-staging]

Post step trims freed blocks before unmount

After, main saved the budget and post ran fstrim between the df sample and drop_caches/umount, shrinking the backing image from 1026 MiB to 2 MiB; before, no trim ran and the image stayed at 1026 MiB.

_Merge-base

_{This PR 0afafedb}

Also checked · 4 match

No trim when no commit will be requested
Budget 0 or absent disables the trim
Unsupported discard only logs
Host min extent passed as fstrim -m

_{Proof on 0afafedb (from b19c797c)}

View full Proof report

[piob-io]

Waiting 60s is too long. How short can it be while staying useful. 5s? Also... since this code gets rolled out by customers, I feel like we should make it so we can inject this delay from the fa agent side, so we can do a slow rollout, and tune this in the future.

[devin-ai-integration]

Agreed on both. Plan:

• fa: add fstrim_timeout_secs to GetStickyDiskResponse, sourced from a new live-config key (guest_fstrim_timeout_secs, built-in default 0 = guest does not trim). That gives a slow rollout + per-host/global tuning from the agent side without a new action release.
• this PR: drop the hard-coded 60 s; the post step only trims when the agent sent a positive timeout and uses that value. Until the fa change is rolled out the action is a no-op here. I'll start with 5 s as the value we set via live config; a timed-out trim is still useful since ext4 FITRIM keeps the extents it already discarded.

The action can't compile against the new field until the proto lands on buf.build, so I'll link the fa PR here and update this one once it merges.

[devin-ai-integration]

Re @piob-io's comment (#82 (comment)) — implemented as of c4a2f7b:

• No fixed timeout in the action any more. The VM agent sends the budget in GetStickyDiskResponse.fstrim_timeout_secs (FastActions/fa#6600: live-config key guest_fstrim_timeout_secs, default 0 = no trim), main saves it in action state, post trims only when it is positive and the commit will be requested, bounded by timeout <secs>.
• Works on this action version even before its proto package is bumped: when the typed field is absent the value is read from unknown field 8 (protobuf-es keeps unknown binary fields over gRPC), so a slow per-host rollout and later tuning need no new action release.
• On "how short": a trim of an already-trimmed disk is sub-second; the first trim of a large dirty disk is what costs time, and with a budget it just stops early and the rest is picked up next run. Proposed start is 5 s per host; the post step logs fstrim <mount>: trimmed|timed_out|failed in <ms> so we can see the distribution before raising it.

Same shape for the docker (useblacksmith/setup-docker-builder) and git-mirror (useblacksmith/checkout) actions, branches pending PRs; the web schema PR adds the key to AgentConfigSchema.