One turn, one invocation
Classic AWS Lambda and Azure Functions on Container Apps are episodic runtimes: each turn is one invocation of a function you deployed, and nothing runs between turns. They differ in how the turn is carried and how long it may take.
sequenceDiagram
participant D as Grove daemon
participant R as Episodic driver
participant P as aws or az
participant F as Function
participant W as Worker in the function
D->>R: run_turn(spec, turn)
alt Lambda
R->>P: aws lambda invoke, synchronous
P->>F: one invocation
else Azure Functions
R->>P: az account get-access-token
P-->>R: Entra ID token
R->>F: POST the turn, with the token
end
F->>W: run the worker
W->>W: restore from the branch, run one turn, push
W-->>R: session id, head commit, outcome
R-->>D: turn result
Classic Lambda
One synchronous invocation is one turn, and the function's own deadline is the hard edge.
- A function runs at most 15 minutes, so a turn must fit inside one invocation.
- Lambda sends no SIGTERM, so the worker must push before its own deadline. Whatever it pushed is what survives.
- The response is capped at 6 MB, which is why the result stays small and the branch carries the work.
- A function error is a failure even though the call succeeded. Lambda reports it beside a zero exit status, and Grove reads the two separately so it is never mistaken for a result.
- A throttled invocation is a refusal. Grove surfaces it and does not retry. The function's
/tmpmust be large enough to hold the restored worktree.
Azure Functions on Container Apps
One HTTPS request is one turn, sent with an Entra ID token minted per turn.
- You deploy the app once from the emitted
kind: functionapptemplate, withaz containerapp create --yaml. - Grove mints a token per turn with
az account get-access-tokenfor the applicationtoken_resourcenames, so nothing long-lived is held. - The app has no long-running guarantee, so a turn must fit one request. Platform maintenance may end a replica with SIGTERM and then SIGKILL 30 seconds later.
- No replica for the request is a refusal, surfaced on the first answer.
What both share
- There is nothing to freeze. An idle workspace costs only its branch, and the next turn is the thaw.
- A turn cut short is a failure, never a made-up result.
- There is no mid-turn steer. The only input is the next turn.
- Every later turn continues the session the previous result named.
Configuration
.grove/config.json
{
"runtimes": {
"aws-turns": {
"kind": "lambda_episodic",
"region": "us-east-1",
"function": "grove-turns",
"image": "000000000000.dkr.ecr.us-east-1.amazonaws.com/grove-runtime:latest"
},
"azure-turns": {
"kind": "aca_functions_episodic",
"subscription": "00000000-0000-0000-0000-000000000000",
"resource_group": "grove-rg",
"environment": "grove-env",
"app": "grove-turns",
"url": "https://grove-turns.example.azurecontainerapps.io/api/turn",
"image": "acme.azurecr.io/grove/runtime:latest",
"token_resource": "api://grove-turns"
}
}
}
Every field is in the Configuration Reference.