Skip to main content

Execution environments

An execution environment keeps a session's file operations and shell commands in one workspace. Choose the default profile for workspace tasks or no-fs for sessions that need no local file access.

Availability

The default profile is available in mecated, mecak8s, mecatui's embedded server, and engine embeddings. A session can instead select the no-fs profile for research, coordination, or remote deployments that must not expose a local filesystem.

Default workspace

Create a default session with a workspace root. Read, ListDir, Write, Edit, Copy, Move, Remove, Grep, Glob, and Shell all use that root. The workspace requires a read before overwriting an existing file. Edits use exact, unique matches, and writes fail if the file changed after the read. Removal is non-recursive, copy accepts only regular files, and copy and move refuse an existing destination.

A new run or process may require another Read before Edit or an existing-file Write because the version record belongs to the live environment. This fail-safe check does not mean file data was lost.

The no-filesystem profile

Create a no-filesystem session by setting profile: "no-fs":

curl -s -X POST http://127.0.0.1:8081/v1/sessions \
-d '{"profile":"no-fs"}'

The profile requires an empty workspace. Any other profile value is rejected; there is no silent fallback. A no-FS catalog removes Read, ListDir, Write, Edit, Copy, Move, Remove, Grep, Glob, Shell, ShellStatus, Parallel, and SkillDraft. It retains web tools, memory, MCP tools, skills, Subagent, and Team; children use the same file-less tool set and cannot create a shell or fork a workspace.

The profile is fixed at session creation. The model cannot switch it during a run. See Core tools for the complete catalog.

Child environments

Different delegation modes use different environment strategies:

  • Read-only Subagent and team members get a child environment. When the workspace is trusted, the harness can use a worktree so child changes are isolated. An untrusted workspace withholds the read-only child shell because creating a worktree requires operating on the repository's .git.
  • Parallel branches use hardened force-copy environments. The initial copy does not invoke Git; each branch then works in its own copied namespace.
  • Mutating members and direct-write Subagents use the parent environment and modify the real workspace. They are serialized against sibling tool calls; there is no merge-back step.

All agent-facing shells use a secret-scrubbed environment. Provider keys, MECATL_* credentials, cloud credentials, and other secret-shaped variables are removed before a command runs.

Persistence and reattachment

The environment identity is persisted with a session snapshot. Mecatl reconstructs local and no-FS environments. A non-local identity can be reattached only when the deployment supplies an EnvironmentResolver; a missing resolver, mismatched identity, or nil workspace returns an error instead of using a local workspace.

Native Kubernetes lifecycle and retention

The optional Kubernetes execution provider stores environment ownership in a namespaced ExecutionEnvironment. Provider replicas coordinate through Kubernetes resource-version compare-and-swap; a replica restart does not clear another replica's operation. Operation lease expiry fences the environment and retains the unresolved operation identity for administrator recovery.

Workspace PVCs are retained by default. Removing the last session reference, deleting a session, uninstalling the chart, or deleting the provider does not delete a workspace PVC. Executor replacement and retirement are private, mutually authenticated administrator operations. They require exact environment, execution-epoch, Pod-UID, and PVC-UID values. The provider waits for a kubelet terminal Pod phase and terminated state for every container before removing its Pod finalizer. A missing Pod or unreachable kubelet is not termination proof and leaves the environment fenced.

Retirement stops the executor but retains the PVC. PVC deletion requires the separate DeleteRetiredEnvironment administrator RPC with the exact retained PVC UID and no references or claims. Deleting the custom resource outside this provider workflow is unsupported. Do not remove the retention or executor finalizers manually; follow the operator's external-fencing runbook when the provider cannot independently observe terminal compute.

Persisted prototype resources use an explicit administrator migration. Normal operations reject schema versions other than 2. MigrateEnvironment accepts only recognized versions 0 and 1, verifies the exact live Pod and PVC identities, and requires a healthy, idle environment. Unknown or malformed state is retained unchanged rather than reset.

Production security material

The production chart requires one projected Secret containing the TLS identity, client CA bundle, and Ed25519 grant keys, plus provider.securityManifest. The chart does not generate keys or certificates. The manifest is strict JSON with this shape:

{
"version": 1,
"generation": 42,
"issuer": "https://execution.example.com",
"audience": "mecatl-execution",
"activeKeyID": "k1",
"grantTTL": "1m",
"clockSkew": "5s",
"keys": [
{
"id": "k1",
"version": 1,
"file": "grant-k1.pem",
"publicKeySHA256": "0000000000000000000000000000000000000000000000000000000000000000",
"activateAt": "2027-01-01T00:00:00Z",
"verifyUntil": "2027-01-02T00:00:00Z",
"state": "active"
}
],
"tls": {
"certificateFile": "tls.crt",
"privateKeyFile": "tls.key",
"clientCAFile": "clients.pem"
},
"clients": [
{
"uri": "spiffe://cluster.example.com/ns/mecatl/sa/mecak8s",
"mayAttestOwner": true,
"administrator": false
},
{
"uri": "spiffe://cluster.example.com/ns/mecatl/sa/execution-admin",
"mayAttestOwner": false,
"administrator": true
}
]
}

The all-zero fingerprint and 2027 dates are non-secret example values. Replace them with the public-key fingerprint and a reviewed active window before use.

Use only basename file names. The projected Secret keys in this example are grant-k1.pem, tls.crt, tls.key, and clients.pem; an external secret manager owns their bytes. Kubernetes projected-volume ..data symlinks are supported, but paths escaping the mounted directory are rejected. Increase generation for every authority change. Key IDs and (id, version) fingerprints cannot be reused; the provider persists a bounded high-water ledger in its authority ConfigMap. Keep retired keys as verify-only until all grants expire, then mark them revoked. Invalid, incomplete, rolled-back, or newly expired material makes readiness fail and denies new RPC authorization until corrected. The provider re-verifies the peer certificate and URI policy against the current client CA on every RPC, including RPCs on an existing HTTP/2 connection. Profile-resource or controller-cache startup failures terminate the provider with that bounded reason class before it accepts traffic. A failed /ready response reports security-authority-or-expiry. Inspect provider logs and the named RuntimeClass, StorageClass, and authority ConfigMap metadata; the endpoint never returns key or certificate contents.

RevokeEnvironment is an administrator-only, exact-reference CAS. Supply the current positive grant generation; success increments it without changing the execution epoch. Old grants then cannot authorize another operation or renewal. An operation already accepted may still finish and clean up, so revocation is not termination proof and a new run needs a fresh claim after the old claim is released or fenced.

The gRPC port remains the only execution data plane. /live and /ready are separate, unauthenticated operational HTTP endpoints intended only for Kubernetes probes; do not expose that port outside the cluster. Readiness requires current security material, the authoritative generation, synchronized controller caches, and Kubernetes access. Liveness reports process health only.

Production values require two or more provider replicas, explicit provider-client selectors, API-server and DNS CIDRs, and finite ResourceQuota/LimitRange values. Executor pods are default-deny for ingress and egress. Add only profile-specific CIDR and port egress under networkPolicy.workloadProfiles; there is no implicit DNS, metadata-service, API-server, provider, or peer access. Each profile requires an explicit RuntimeClass supplied by the cluster, but do not treat a RuntimeClass name alone as a hostile-workload isolation guarantee. CPU, memory, ephemeral storage, and /tmp are explicitly bounded per profile. Each profile also requires maxEnvironments; the namespace quota is the final global bound.

For an existing single-key installation, first create the authority ConfigMap and a generation-1 manifest that names the existing TLS and grant files, then upgrade the chart. Confirm /ready before rotating. Do not reuse the old key ID with new key bytes. Schema-2 environment migration is independent of keyring migration: complete both before enabling new sessions.

Live qualification (experimental)

The repository keeps real-provider qualification separate from the default mock suite. It is explicit opt-in, uses an already-owned retained Kind cluster, and is not part of task test:

MECATL_EXECUTION_CREDENTIAL_FILE=/absolute/path/to/provider-key \
MECATL_EXECUTION_QUAL_STATE=/absolute/path/to/owned-state \
task e2e:k8s:execution:live

The credential file must be a private regular file (no group or other access). A trusted helper loads it only at runtime and creates a run-scoped Kubernetes Secret through the API; it never renders the value into a manifest or reads the Secret back. Only the mecak8s harness container receives the OpenRouter key. The execution provider, controller, OIDC fixture, and executor workloads do not. The task first runs the deterministic mock qualification, then runs one bounded real-model coding smoke against https://openrouter.ai/api/v1. It verifies positive token usage, required file and shell tool calls, file contents, and a successful go test independently through the typed gRPC execution service. Finally it restores the mock deployment and removes only its run-scoped Secret; the owned cluster and execution workspaces remain for inspection.

Limitations

  • No-FS sessions cannot use local file tools, shell commands, workspace forks, or parallel branches.
  • Read-only child shells depend on project trust; this is separate from the parent's ability to run its own shell.
  • Direct-write children can leave partial edits if cancelled or interrupted; the parent workspace and Git are the rollback boundary.
  • Remote environment reattachment requires an explicit resolver.

Next steps