Files
MetaCrate/docs/grid-agent-operations.md
Chili Palmer b59bc4c3cd
Some checks failed
CI / rust-skia (Rust only) (push) Has been cancelled
CI / required (push) Has been cancelled
Prefetch nearby scene assets in the background
2026-08-23 11:06:28 +02:00

278 lines
15 KiB
Markdown

# MetaCrate grid-agent operations
This is the operator runbook for the portable `metacrate-grid-agent` binary.
The binary has no dependency on systemd, a Windows service manager, a CLR,
Python, Node.js, a provider SDK, or a shell at runtime. Linux service files and
install helpers are packaging conveniences around the same foreground process
and versioned control protocol.
## Quick start and modes
Build the credential-free binary and validate the example without contacting a
grid or LLM:
```sh
cargo build --locked --release -p metacrate-grid-agent
target/release/metacrate-grid-agent --config config/grid-agent.example.yml --check-config
target/release/metacrate-grid-agent --config config/grid-agent.example.yml --run-once
```
`--check-config` parses YAML and any legacy referenced secret files, validates all
bounds and TLS files, then exits before constructing the grid or LLM clients.
The example is fake/offline mode (`integrated=false`, `split=false`) and all
credential-looking values are placeholders.
For live foreground or an embedded TUI, build intentionally:
```sh
cargo build --locked --release -p metacrate-grid-agent --features live-grid
target/release/metacrate-grid-agent --config /path/to/config.yml
target/release/metacrate-grid-agent --config /path/to/config.yml --tui
```
Set `integrated=true` for the foreground and embedded-TUI commands. Set
`split=true` for a headless service, then run the same binary as the TUI client:
```sh
metacrate-grid-agent --config /path/to/config.yml
metacrate-grid-agent --config /path/to/config.yml --tui-client
```
The TUI client and service must use the same control address and operator token.
`--tui-client` never needs the grid password or LLM key. `--run-once` performs
one supervised readiness/login and clean logout cycle. `--print-paths` prints
the platform defaults without loading configuration. If `--config` is omitted,
the default file is used only when it already exists:
| Platform | Configuration | Non-secret state |
|---|---|---|
| Linux/Unix | `$XDG_CONFIG_HOME/metacrate/config.yml`, otherwise `$HOME/.config/metacrate/config.yml` | `$XDG_DATA_HOME/metacrate/grid-agent`, otherwise `$HOME/.local/share/metacrate/grid-agent` |
| macOS | `$HOME/Library/Application Support/MetaCrate/config.yml` | `$HOME/Library/Application Support/MetaCrate/grid-agent` |
| Windows | `%APPDATA%\MetaCrate\config.yml` | `%LOCALAPPDATA%\MetaCrate/grid-agent` |
Relative paths in a YAML document resolve relative to that document where specified; service
deployments should use absolute paths.
Downloaded grid assets use the platform cache directory: `$XDG_CACHE_HOME/metacrate`
or `$HOME/.cache/metacrate` on Linux, `$HOME/Library/Caches/metacrate` on macOS,
and `%LOCALAPPDATA%\metacrate\cache` on Windows. The cache treats UUID-addressed assets as
immutable, defaults to a 2 GiB LRU limit, and is configurable with
`vision.asset_cache_max_bytes`. The live scene source continuously prefetches and prepares assets
inside the configured view radius while the avatar is logged in; prefetching is stopped before
logout and shutdown.
## Endpoint, grid identity, and authority
`llm.endpoint_url` is the OpenAI-compatible provider base URL. Mentra owns the
Responses path, SSE stream, endpoint behavior, and conversation runtime. For a
proxy whose Responses endpoint is `/go/v1/responses`, configure the base ending
in `/go`, not the final request path. Set `llm.model` to the model ID and keep
`llm.api_key` in the private `config.yml`.
Live modes require `grid.login_url`, `grid.avatar_name`, and `grid.password`.
`authorized_avatar_uuids` contains exact
grid UUIDs, never display names. Text claiming an authorized identity grants no
authority. Public chat can request bounded informational work and safe public
LSL delivery; movement, teleport, building, roaming changes, and administration
remain policy-gated and require an authenticated authorized IM or a narrowly
bound scheduler grant. Above-threshold actions receive an isolated one-shot LLM
safety review; they do not depend on a continuously present human operator.
## Configuration contract and migration
The YAML root uses `schema_version: 1`. Omitting it is accepted as legacy
version 1. Any other value fails before startup with a migration message. There
is no automatic in-place migration: copy the file, update the copy using the
release notes/example, validate the copy, then atomically select it. Unknown
fields fail closed. Installers never overwrite or migrate operator files.
Resolution order is:
1. bounded built-in defaults;
2. the platform `config.yml` or explicit `--config` document;
3. legacy secret files referenced by that document.
Process environment variables do not override connection, mode, or authority.
Migrate an existing `.env` once with `metacrate-grid-agent --import-env
/path/to/.env`, inspect it using `--preferences`, then remove the old file under
the operator's retention policy. JSON remains readable only for migration.
Mode, endpoints, credentials, authorization UUIDs, TLS, storage, queue/resource
limits, reconnect policy, behavior, and interaction settings are restart-only.
Runtime control can pause/resume autonomy, toggle the roaming job, make an
emergency decision on an existing approval, cancel an active action, expire/delete conversation state,
inject an operator message, reconnect, or shut down; it does not silently
rewrite the configuration. A future reloadable field must be explicitly added
to the versioned control/config contract.
## Secrets and privacy
Threat model: grid residents, object/avatar/inventory metadata, capability
replies, model output, public chat, IM text, and persisted files are untrusted.
Remote peers may inject instructions, spoof names, replay approvals/call IDs,
flood queues, delay or truncate replies, and attempt secret exfiltration. The
service identity, local config/secret ACLs, authenticated UUID/control role,
opaque policy authorization, generation fencing, bounded queues, and audit sink
are trust boundaries. Host/root compromise, a malicious binary/dependency, and
an operator deliberately approving a harmful action are outside the process
sandbox and require OS/supply-chain/operational controls.
Never place API keys, grid passwords, or control tokens in command arguments,
unit files, wrapper XML, logs, crash-report commands, issue reports, or TUI
screens. Secret wrappers redact `Debug` and `Display`, are not serializable, and
normal observability stores pseudonymous correlation IDs, result codes, bounds,
and hashes—not prompt bodies, credentials, visual pixels, or raw tool arguments.
Secret files must be regular, non-symlink, bounded UTF-8 files containing one
line. On Unix, startup rejects group/other permission bits; use mode `0600` or
stricter and ownership by the service identity. On Windows, set an NTFS ACL that
grants only the service identity and Administrators, for example with `icacls`;
Rust's portable metadata API cannot prove arbitrary Windows ACL semantics, and
FAT/network filesystems may not enforce them. Treat an unverifiable filesystem
as unsuitable for unattended secrets.
Rotate one credential at a time in `--preferences` (or an ACL-restricted copy
of `config.yml`), save, run `--check-config`, then restart. Revoke the old credential only after readiness. Control observer
and operator tokens must differ from each other and from grid/LLM credentials.
## Linux systemd
The hardened example is
[`../packaging/metacrate-grid-agent/systemd/metacrate-grid-agent.service`](../packaging/metacrate-grid-agent/systemd/metacrate-grid-agent.service).
Create the unprivileged `metacrate-agent` identity, copy (do not overwrite) an
operator configuration to `/etc/metacrate/config.yml` and create
`/var/lib/metacrate/grid-agent`. Keep the configuration owned by the service
identity with mode `0600`; do not put credentials in the unit environment.
Install the unit, inspect the sandbox, validate, then start:
```sh
sudo systemd-analyze security packaging/metacrate-grid-agent/systemd/metacrate-grid-agent.service
sudo -u metacrate-agent /usr/local/bin/metacrate-grid-agent --config /etc/metacrate/config.yml --check-config
sudo systemctl daemon-reload
sudo systemctl enable --now metacrate-grid-agent.service
```
The unit uses `SIGINT`, matching foreground Ctrl-C and the owned graceful
shutdown path. Do not replace it with `SIGKILL`. `ProtectSystem=strict` permits
writes only under the state directory.
## Windows service operation
Run the same `.exe`, YAML, control TCP protocol, and shutdown path. Portable
foreground operation in PowerShell is the baseline:
```powershell
& 'C:\Program Files\MetaCrate\metacrate-grid-agent.exe' --config 'C:\ProgramData\MetaCrate\config.yml' --check-config
& 'C:\Program Files\MetaCrate\metacrate-grid-agent.exe' --config 'C:\ProgramData\MetaCrate\config.yml'
```
For unattended use, configure a maintained Windows service wrapper (for
example WinSW) to launch exactly that command as a dedicated low-privilege
account and to translate SCM Stop into Ctrl-C/console control before its timeout.
Keep `config.yml` ACL-restricted to the service identity. Configure restart-on-failure, not unconditional rapid
restart. Validate as the service identity before registration. The wrapper must
not capture environment values or command output into a world-readable log.
The supplied PowerShell installer replaces only the executable and never state.
## Control, network, and health
Split control defaults to loopback TCP. A non-loopback bind is rejected unless
an explicit Rustls certificate/private key is configured. Use host firewalls to
allow only operator networks, protect token files, and prefer loopback plus an
authenticated tunnel. Observer tokens are read-only. Protocol framing, role
permissions, cancellation, event gaps, and limits are in the control-plane doc.
Readiness is stricter than a TCP connection: the runtime view must show
`transport_connected=true` and `agent_ready=true`. The TUI Overview and
observer Runtime request are health/readiness checks; process existence alone
is not readiness. Alert on authentication-blocked, sustained degraded/backoff,
queue saturation/dropped-event counters, audit failure, orphan builds, and a
stale generation.
## Retention, backup, upgrade, and rollback
Conversation snapshots, landmark state, and audit journals are non-secret but
privacy-sensitive. Stop or pause mutation, take a filesystem-consistent backup
of the configured data directory, and encrypt/restrict the backup. Do not back
up secrets with ordinary state. Journal rotation is bounded by the configured
observability policy; ship rotated files to restricted storage before deletion
when retention policy requires it. Replay is diagnostic and never executes an
action.
Release installation replaces only the binary. The POSIX and PowerShell helpers
stage a temporary executable then move it into place; neither touches config,
secrets, conversations, landmarks, or journals. For upgrade: back up state,
install the new binary, validate the existing config, stop gracefully, start,
and verify readiness. For rollback: stop gracefully, restore the prior binary,
restore state only when the new version changed it incompatibly, validate, and
start. Never run two service generations against one writable data directory.
## Failure playbooks
- Authentication blocked: pause retries, verify login URL/avatar and rotate the
password file; never paste it into logs. Force reconnect after correction.
- LLM unavailable/rate limited: autonomy remains bounded; verify the provider
base URL, firewall/DNS, key, and model compatibility. Multimodal rejection falls back to the textual scene
summary without resending the large image.
- Maintenance/disconnect: allow generation fencing and bounded backoff. Stale
inference/mutation results are discarded; do not bypass reconnect controls.
- Emergency: use operator Pause first, Cancel the exact active action when
appropriate, then Graceful Shutdown. Ctrl-C/SIGINT follows the same cleanup.
- Orphan build: keep the reported object IDs, inspect ownership in-world, and
manually recover only those IDs. Never bulk-delete by name or proximity.
- Corrupt persistence: preserve the quarantined file for restricted diagnosis;
the service recovers an older valid generation or starts fail-closed. Do not
hand-edit a live journal.
- Full disk/audit backpressure: pause autonomy, free space according to retention
policy, and restart only after the audit path is writable. Policy fails closed
when required audit records cannot be accepted.
## Resource defaults and unsupported operations
The example records all current queue, message, conversation, tool, behavior,
reconnect, and interaction defaults. Important defaults include 256 grid events,
32 control commands, 512 observations, four concurrent inference requests,
16 tool calls, 512 active senders/sessions, a two-minute model window, and bounded
10-second shutdown. Vision defaults to a 1920x1080 headless Bevy/wgpu JPEG, a
64-meter configurable view radius, a bounded in-memory decoded-texture cache,
a persistent platform asset cache, and a deterministic software fallback with
bounded entities, triangles, asset work, JPEG bytes, time, and concurrency.
Authorized `behavior_camera_set` calls are
limited to 96 meters from the avatar by default, require a distinct target, and
accept a 20-120 degree vertical field of view; `behavior_camera_reset` restores
the avatar-facing 60-degree view.
Unsupported by design: arbitrary raw packets or agent-control flags, arbitrary
shell/subprocess execution, provider/model discovery outside Mentra, remote plaintext
control, unauthenticated mutation, unrestricted walking/teleport/touch/follow,
automatic config migration, persistence of viewport pixels, framebuffer/screen
capture, and treating untrusted grid/LLM text as instructions or authority.
## Release evidence
Build the intentional graphs and retain the commands/output with the release:
```sh
cargo build --locked --release -p metacrate-grid-agent
cargo build --locked --release -p metacrate-grid-agent --features live-grid
cargo tree --locked -p metacrate-grid-agent --features live-grid -e normal,build
cargo audit
```
Record the binary byte size (`stat -c %s` on Linux or `Get-Item ... .Length` on
PowerShell) and the `cargo tree` inventory. Reject CLR/scripting/provider SDK,
subprocess adapters, unreviewed graphics stacks, or undeclared native libraries
on the agent path.
The consolidated Gitea gate runs only on `ubuntu-latest`; Windows portability is
proved by the existing cross-target compile/static gate rather than a Windows
Gitea runner.
The milestone baseline is recorded in
[grid-agent-release-evidence.md](grid-agent-release-evidence.md).
Related contracts: [architecture](grid-agent-architecture.md),
[policy](grid-agent-policy.md), [session](grid-agent-session.md),
[control](grid-agent-control-plane.md), [observability](grid-agent-observability.md),
[TUI](grid-agent-tui.md), and [conversation storage](grid-agent-conversation.md).