[eersnington]

Fixes #2083. Reported and diagnosed in #2083.

On a 1-CPU host, a guest HTTPS download larger than the VM's buffer limits made the sidecar unresponsive, and the client killed it after 30 s. Every VM on the sidecar was lost.

• ResourceLedger::reserve charged the process ledger first and undid the charge when the VM ledger was full. The undo called notify_one(), which stores a permit when nobody is waiting yet.
• reserve_when_available then consumed its own permit, retried, failed, and stored a new one. It never returned Pending, so on one runtime worker the task that frees VM capacity never ran. The waiter here is AsyncCompletionSender::send (crates/native-sidecar/src/state.rs).
• CapabilityRegistry::reserve had the same loop on a single ledger: it took the capability and ready-handle counts, failed on Sockets, and released them.

flowchart LR
    R["reserve / CapabilityRegistry::reserve / HTTP/2 admission"] --> A["ResourceLedger::reserve_all"]
    A --> L["lock process ledger, then VM ledger"]
    L --> C{"all requests fit<br/>in every scope?"}
    C -- no --> E["LimitError: no charge, no notify"]
    C -- yes --> OK["charge all, return Reservations"]

Loading

• New ResourceLedger::reserve_all(&[(ResourceClass, usize)]) checks every request in every scope under ordered locks, then charges. A rejected admission changes nothing, so a waiter can no longer wake itself.
• reserve calls reserve_all with one request. CapabilityRegistry::reserve and reserve_http2_resources admit their counts in one call.
• Locks are always taken root first. The helpers called while locked (maybe_warn, observe_usage) take no ledger lock.
• A near-limit warning no longer fires for a reservation that is then rejected.

17 MB guest fetch, Linux	main	This PR
1 CPU (taskset -c 0)	0/3, sidecar unresponsive	3/3
16 CPUs	3/3	3/3

[railway-app]

This PR was not deployed automatically as @eersnington does not have access to the Railway project.

In order to get automatic PR deploys, please add @eersnington to your workspace on Railway.

[github-actions]

📟 Pip-Boy · 339e226

📦 Packages

Skipped: this PR doesn't change any published package.

🧹 Rust dependencies · cargo deny + cargo shear

check	new in this PR	fixed	existing on base
advisories	✅ 0	0	28
licenses	✅ 0	0	0
sources	✅ 0	0	0
bans	✅ 0	0	0
unused	✅ 0	0	16

🦀 Binaries · publish Dockerfiles, --release

Base: main@3c60c9a, the merge base.

artifact	npm @latest	base	this PR	Δ vs base
agentos-native-sidecar linux-x64-gnu	143.8 MB	143.8 MB	143.8 MB	−1.6 KB (−0.0%)

⏱ Timing

job	wall time	slowest step
zizmor	10s	checkout · 4s
changes	37s	checkout · 32s
native · linux-x64-gnu	14m52s	Build with the publish Dockerfile · 13m47s
hygiene	1m01s	checkout · 25s
⏭ footprint	—	skipped
⏭ self-test	—	skipped
report	29s	Write the comment and run summary · 10s

🔭 Full CI trace for this commit

One OpenTelemetry trace of CI and Pip-Boy for 339e226: every job and step, with the sizes above attached. Run after CI finishes.

pnpm install --frozen-lockfile --filter 'pipboy...'
pnpm --filter pipboy pipboy trace --sha 339e2269e330d1acebb1dcea00e37184dec570fb

Add --otlp <url> to send it to any OTLP/HTTP collector.

Agent instruction: From a rivet-dev/agentos checkout, run the commands above, then read the trace file the last command prints. List the slowest jobs and steps and explain where CI time went.