Improve agent resiliency
Improve workspace agent resiliency by tuning child process CPU and out-of-memory priority.
On this page
Coder's agent can automatically lower the scheduling priority and raise the out-of-memory (OOM) kill score of user processes so the agent itself stays alive under resource pressure.
Prerequisites
- Linux — The feature is ignored on other operating systems.
Environment variables
Configure the feature with environment variables in the environment that launches the agent binary.
These must be set on the workspace container or host, not in the coder_agent resource's env block.
The agent reads them from its own process environment at startup.
| Variable | Required | Default | Description |
|---|---|---|---|
CODER_PROC_PRIO_MGMT | Yes | None | Set to enable the feature. The agent checks whether the variable is present, not its value, so even an empty string enables it. Use 1 by convention. To disable, unset the variable entirely. |
CODER_PROC_OOM_SCORE | No | Computed from agent's score | Explicit oom_score_adj value for child processes. Range: -1000 to 1000. |
CODER_PROC_NICE_SCORE | No | Agent nice + 5 (capped at 19) | Explicit nice value for child processes. Range: -20 to 19 (higher = lower priority). A value below the agent's own nice value is not applied. Refer to Nice value is not applied. |
OOM score defaults
If you do not set CODER_PROC_OOM_SCORE, the agent computes a value based on its own oom_score_adj:
Agent's oom_score_adj | Child score | Rationale |
|---|---|---|
| Negative (< 0) | 0 | Children are treated as normal processes. |
| >= 998 | 1000 | Children get the maximum score (killed first). |
| Any other value | 998 | Children get a near-maximum score. |
The goal is for the kernel's OOM killer to target child processes before the agent, keeping remote connectivity alive even when a workspace runs out of memory.
Nice score defaults
If you do not set CODER_PROC_NICE_SCORE, the agent sets children to its own nice value plus 5, capped at 19. This gives the agent more CPU scheduling priority than user workloads.
By default, Linux does not let the agent give child processes a nice value below its own.
Adding CAP_SYS_NICE to the container does not change this, because Coder drops the child process's Linux capabilities before it sets the nice value.
Example
The following Kubernetes template snippet enables process priority management on the workspace container:
resource "kubernetes_deployment" "workspace" {
# ... other configuration
spec {
template {
spec {
container {
name = "dev"
image = "codercom/example-base:ubuntu"
env {
name = "CODER_AGENT_TOKEN"
value = coder_agent.main.token
}
env {
name = "CODER_PROC_PRIO_MGMT"
value = "1"
}
env {
name = "CODER_PROC_OOM_SCORE"
value = "10"
}
env {
name = "CODER_PROC_NICE_SCORE"
value = "1"
}
}
}
}
}
}CODER_PROC_OOM_SCORE=10gives child processes a slightly elevated OOM score while keeping them well below the maximum.CODER_PROC_NICE_SCORE=1gives children a mildly lower CPU priority than the agent.
Troubleshooting
OOM score adjustment fails
If you see failed to adjust oom score in stderr but the process still starts, the agent likely lacks permission to write to /proc/self/oom_score_adj.
Make sure the process is dumpable (the agent handles this automatically, but it can fail if the container runtime restricts prctl calls).
Nice value is not applied
If you see coder-agent: failed to adjust niceness before a command's output, the most likely cause is a CODER_PROC_NICE_SCORE that is lower than the agent's current nice value.
Coder writes the message to the stderr of the command, so it appears with that command's output, for example in the terminal session that ran it.
The command still runs with the agent's nice value.
Set CODER_PROC_NICE_SCORE to a value at or above the agent's nice value, or remove it to use the default.
For details, refer to Nice score defaults.
Environment variables leak to nested Coder agents
The agent strips the CODER_PROC_PRIO_MGMT, CODER_PROC_OOM_SCORE, and CODER_PROC_NICE_SCORE variables from child environments automatically.
This prevents interference in "Coder on Coder" development scenarios where a workspace runs another Coder agent.
Verify the feature is enabled
The agent logs whether process priority management is active at startup. Look for these lines in the agent log:
"process priority management enabled"
"process priority management not enabled (linux-only)"The log entry includes the CODER_PROC_PRIO_MGMT value and the operating system.
Check the agent log file at <log-dir>/coder-agent.log or stderr output.
Feature has no effect on macOS or Windows
Process priority management is Linux-only.
Setting CODER_PROC_PRIO_MGMT on other operating systems is safe but has no effect.