lazaret/docs/specification.md

# Lazaret Specification

This document is the normative product specification for Lazaret.

### Document authority

| Document | Authoritative for |
| :--- | :--- |
| This specification | Product north star, trust model, stable vocabulary and identifiers, below-abstraction architecture, and product requirements |
| [TLA+ application protocol model](protocol/Lazaret.tla) and [TLC config](protocol/Lazaret.cfg) | Protocol lifecycle, communication semantics, publication, registration, escalation, recovery, and checked invariants |
| [Prepared-lower TLA+ model](protocol/PreparedLower.tla) and [TLC config](protocol/PreparedLower.cfg) | Prepared-artifact publication, generation leases, mounts, reclamation, stage recovery, and checked invariants |
| [Design schemas](schema.md) | Logical records, protocol data shapes, network-policy documents, readiness evidence, and their relationships |

This specification does **not** restate modeled actions or invariants. The TLA+
models are authoritative above their abstraction boundaries. This specification
owns physical refinements below those boundaries. A new above-abstraction rule
belongs in the applicable model before a product refinement relies on it.

## 1. Problem Statement

AI coding agents need broad local capabilities: they run arbitrary tools, modify a
project, use Git, install dependencies, and access selected network services. Code
produced or invoked by an agent is not trusted to run directly on the developer's
host. A conventional container reduces risk, but poor terminal behavior, transient
tool state, restricted Git access, and manual setup can make the safe path too
inconvenient for routine work.

Lazaret's north star is:

> Make isolated agent execution the easiest way to work, while ensuring that every
> transition from the sandbox to host authority is explicit, attributable, and
> controlled by the developer.

Lazaret therefore provides a rootless gVisor sandbox with a bidirectionally shared
workspace, persistent sandbox-local tooling, controlled network egress, and a
human-in-the-loop host escalation path.

### 1.1 Trust Model

The following components are trusted:

- The developer and host operating system.
- The Lazaret CLI and the domain's Session Coordinator (Host Coordinator, Runtime
  Supervisor, and attached runsc client).
- Host configuration selected by the developer.
- Domain network-policy installation and connectivity helpers under Runtime
  Supervisor ownership, and the policy Lazaret installs for the Domain.

The following components are untrusted:

- The agent and every process it launches.
- Code, tools, and package installation scripts executed inside the Agent Sandbox.
- Data received from the network by sandboxed processes.

All Client Sessions in the same Domain intentionally share an Agent Sandbox and
are one sandbox trust domain. Session identity distinguishes ownership and routes
terminal interactions; it is not an isolation boundary between processes in that
sandbox.

The Agent Workspace is intentionally writable. Lazaret protects the host outside
the explicitly mounted workspace and volumes. Tracked and ordinary untracked
workspace paths remain writable, while Git-ignored paths are isolated into
Domain-scoped sandbox state by default. The developer may explicitly disable that
isolation in trusted host configuration.

The physical refinement of model configuration confinement canonicalizes the
configuration file and every bind-mount source, compares their resolved ancestry,
and omits configuration bytes from the sandbox mount specification.

## 2. Glossary

These terms and identifiers are stable across the specification, protocol model,
implementation, logs, and user interfaces.

| Term | Definition |
| :--- | :--- |
| **Agent Profile** | A developer-selected host configuration describing an agent image and its granted storage, network, environment, and escalation capabilities. |
| **Agent Workspace** | The canonical host project directory mounted read-write at `/workspace` in the Agent Sandbox. |
| **Agent Sandbox** | The rootless gVisor container in which an agent and its child processes execute. One sandbox is shared by all Client Sessions in a Domain. gVisor sentry, gofer, and guest processes are the sandbox, not the Session Coordinator. |
| **Agent Sandbox ID** | A runtime-unique identifier assigned to one Agent Sandbox instance and bound to its Domain protocol identity. |
| **Client CLI ID** | A runtime-unique identifier assigned to one host Lazaret CLI process. It is authenticated as a protocol identity and bound to one Client Session. |
| **Client Session** | One active host Lazaret CLI process and the interactive or non-interactive command it attaches to the Agent Sandbox. |
| **Session ID** | A cryptographically random identifier unique to one Client Session. |
| **Session Anchor** | An ephemeral host-only signing key whose public key binds developer decisions to a registered Session ID and owning Client CLI. |
| **Session Request Capability** | An ephemeral Session Coordinator-signed binding of one live Client Session to the exact Agent Sandbox generation and its protocol credentials. HC physically holds the signing key. It is a routing capability within the shared sandbox trust domain, not a process-isolation boundary. |
| **Session Coordinator** | The logical Domain role defined by the protocol model. The selected physical refinement maps it to Host Coordinator, Runtime Supervisor, and attached runsc processes. |
| **Session Coordinator ID** | A runtime-unique identifier for the logical Session Coordinator as a protocol peer, authenticated on the Host Coordinator face. Generation scoping and stale-credential refuse follow the protocol model. |
| **Host Coordinator** | The constitutive Session Coordinator process that runs in the host user and mount namespaces. It is the protocol face: Domain lock, durable state store, Session PTY masters, escalation mailbox and routing, publication and session admission, and lifetime authority over the Runtime Supervisor. |
| **Runtime Supervisor** | The constitutive Session Coordinator process that runs in the Domain generation's user, mount, and network namespaces under a Host Coordinator lifetime lease. It owns supervised filesystem and network resources for the generation, arms the parent-death edge to attached runsc, and launches the attached runsc client. |
| **Attached runsc** | The constitutive Session Coordinator process that is the non-detached `runsc` client child of the Runtime Supervisor. It supplies the causal coupling edge to the complete Agent Sandbox. |
| **Domain** | The shared runtime and persistent-state scope for one `(Agent Profile, canonical Agent Workspace)` pair. |
| **Domain ID** | A deterministic identifier derived from the Agent Profile name and canonical Agent Workspace path. It namespaces Domain resources but is not an authentication credential. |
| **Domain generation** | One attempt to prepare, publish, operate, and tear down or recover a Domain's transient runtime resources under a single Session Coordinator instance and durable generation identity. |
| **Domain Root** | The host-only directory containing persistent state and metadata for one Domain. |
| **Network Sidecar** | The logical Domain network-policy role for one Domain generation: derived-policy application, enforcement network-namespace preparation, connectivity-helper supervision, and network readiness attestation. Under the selected architecture this role is realized on the Runtime Supervisor-owned install path (with a supervised connectivity helper). It is not a constitutive Session Coordinator process and need not be a separate OS process. The Agent Sandbox never holds network-administration capability. |
| **Network Sidecar ID** | Protocol component identifier for the Domain generation's Network Sidecar role instance (not a Runtime Supervisor or connectivity-helper process id, and not a filesystem path). Generation scoping follows the protocol model. |
| **Connectivity helper** | The supervised process that provides user-space connectivity for the Domain enforcement network namespace. Production uses `slirp4netns`. Ownership and lease coupling follow the protocol model; process identity is product recovery material, not the Network Sidecar ID. |
| **Derived network policy** | The generation-bound policy document Lazaret derives from trusted host configuration (default and profile allow/deny prefixes, DNS reachability, address-family disposition). Only derived policy is installed; the Agent Sandbox never supplies policy. |
| **Network readiness fingerprint** | Product encoding of protocol network readiness: a cryptographic digest over Domain ID, generation, configuration fingerprint, Network Sidecar ID, derived-policy digest, and enforcement network-namespace identity material. |
| **Prepared lower** | A Lazaret-owned immutable root filesystem tree materialised from the selected base image. Content is identified by a prepared-artifact identity. |
| **Prepared-artifact ID** | A Lazaret-assigned identity for one immutable prepared-lower tree (content-addressed or otherwise unique among prepared lowers under the Storage Root). |
| **Prepared-lower lease** | Product binding of a Domain generation to a prepared-artifact ID for State Anchor use (not a container-engine mount refcount). Lease lifecycle follows the protocol model. |
| **Durable intent** | Protocol durable intent that Lazaret intends to create or has created a named external resource for a Domain generation (see protocol model). |
| **Publication** | Protocol publication of a Domain generation for Client Session registration and service (see protocol model). |
| **Git Sentry** | The protective mount layout that preserves bidirectional workspace and Git access while preventing writes to the host's default Git hooks and repository configuration. |
| **Ignored Workspace State** | Domain-scoped sandbox-local storage presented at Git-ignored Agent Workspace paths instead of exposing the corresponding host content. |
| **State Anchor** | The Domain-scoped persistent writable filesystem layer merged over selected paths from the sandbox base image prepared lower. |
| **Trust Anchor** | The ephemeral Domain-generation trust package containing separate TLS certificate-authority and protocol-credential issuer keys. |
| **Protocol Credential** | A Trust Anchor-issued binding of a protocol-signing public key to an exact Domain generation, component role, and runtime component ID. |
| **Protocol Identity** | An authenticated binding of a component role, Agent Profile, Domain ID, and runtime component ID. Identifiers, paths, filenames, and process IDs name resources but do not authenticate protocol peers by themselves. |
| **Escalation Client** | The Lazaret-provided executable through which a sandboxed process requests a host action. |
| **Escalation Mailbox** | Host-private, Domain-scoped regular-file message paths mounted into the Agent Sandbox for signed escalation requests and correlated results. |
| **Escalation Request ID** | A cryptographically random identifier used to correlate one host escalation request and response. |
| **Escalation Result** | The terminal response to an accepted Host Escalation, correlated by Escalation Request ID and carrying its outcome plus command exit status and output when execution occurred. |
| **Host Escalation** | A request by a sandboxed process for Lazaret to execute a developer-approved or auto-approved action with host authority. |
| **Configuration Fingerprint** | A cryptographic digest of the validated host configuration that created or joined a running Domain. |
| **Storage Root** | The trusted host directory beneath which Lazaret stores Domain Roots and profile-scoped state. |
| **Storage Catalog** | The versioned redb store for prepared artifacts, preparation operations, and cross-Domain generation leases. Domain coordinators serialize short catalog transactions with the Storage Catalog lock; no process keeps the catalog open as global runtime authority. |

## 3. Architecture

### 3.1 System Boundaries

```mermaid
flowchart TB
    subgraph Host[Trusted host]
        CLI[Client CLI]
        Config[Host configuration]
        HostWorkspace[Agent Workspace]
        HC[Host Coordinator]
        RS[Runtime Supervisor]
        Runsc[Attached runsc]
        State[State Anchor / FUSE]
        Net[Network policy + connectivity helper]
        Mailbox[Escalation Mailbox]
        HostExec[Approved host process]
    end

    subgraph Sandbox[Untrusted Agent Sandbox - rootless gVisor]
        Agent[Agent and child processes]
        Workspace[/workspace mount]
        Escalate[Escalation Client]
    end

    Config --> CLI
    HostWorkspace <--> Workspace
    CLI <--> HC
    HC --> RS
    RS --> Runsc
    Runsc --> Sandbox
    RS --> State
    RS --> Net
    HC <--> Mailbox
    Agent <--> Workspace
    Agent --> Escalate
    Escalate <--> Mailbox
    HC --> HostExec
    Agent --> Net
    Net --> Network[Permitted destinations]
```

The physical Session Coordinator refinement uses an HC→RS parent/child edge and
lifetime pipe. The RS creates the generation namespaces, arms the parent-death
edge, and launches a non-detached runsc client. Client CLI protocol descriptors
terminate in the HC process.

The Agent Sandbox uses gVisor's application-kernel isolation under rootless
operation. The product launches runsc directly through the HC→RS→runsc process
chain; an external container engine does not own the Domain's Agent Sandbox.
Rootless operation limits the host authority of runtime helpers, while gVisor
interposes a separate userspace kernel between untrusted processes and the host
kernel. Network policy installation and connectivity helpers are trusted and are
not part of the untrusted agent environment.

Lazaret is daemonless: it does not require a persistent global Unix service or
cross-Domain runtime authority that survives independently of Domain generations.
Runtime paths and durable rows carry Domain ID, generation, configuration
fingerprint, role identity, and physical ownership fields. Bootstrap recovery is
initiated by a host Lazaret CLI invocation, not by a resident scheduler.

### 3.2 Domain and Session Lifecycle Refinement

The Host Coordinator implements Domain ownership with a non-blocking host
`flock`, Domain-scoped runtime paths, an exclusive Domain Store handle, and
generation identity records. Client CLIs reach the Host Coordinator through its
host-private protocol face rather than a global control plane.

Protocol lifecycle—startup, registration, cancellation, recovery, teardown,
publication, escalation, generation-scoped identities, durable intent, network
readiness, prepared-lower lease, and connectivity-helper ownership—is defined by
the [TLA+ application protocol model](protocol/Lazaret.tla). Product
implementation refines that model with physical resources (prepared-lower trees,
mounts, `slirp4netns`, PTYs, runsc clients, durable store encoding, and exact
ownership records). Physical recovery enumerates and inspects those resources
independently of process liveness.

#### Host process identity and causal acquisition refinement

Lazaret supports host-process authority only on an exactly source/build-qualified
Linux x86-64 kernel at or above 6.9. Each supported upstream or vendor build
preserves the boot-monotonic 64-bit pidfs inode allocation used by the qualified
build and passes its qualification suite. A release string, version floor,
partial backport, or successful runtime smoke cannot substitute for exact build
qualification. Linux 32-bit, other architectures, kernels below 6.9, and
unqualified vendor builds refuse before a generation or external runtime
resource is created.

On every boot Lazaret runs the finite `PidfsHostSupportProbe` before the first
generation and caches success only under the current boot ID. The probe has the
repository-registered operation sequence and bounds. It checks pidfd operations
and pidfs magic, same-live-process concurrent and last-close/reopen identity,
distinct simultaneous processes, stale-handle non-action, and
exit/reap/drop-all/later-process non-reuse. Any missing operation, failed
observation, cleanup uncertainty, timeout, or unexpected equality refuses
support for that boot. This bounded smoke checks the running path; it does not
prove global inode non-reuse or qualify the kernel source/build.

`PidfsProcessIdentityV1` separates three kinds of data:

- `PidfsProcessAuthorityV1 = (boot_id, pidfs_st_dev, pidfs_st_ino)` is the only
  positive durable host-process authority. The boot ID is stored as the 16 UUID
  bytes in canonical UUID field order; both pidfs values are unsigned 64-bit
  types distinct from general filesystem device and inode identities.
- `host_pid` is an unsigned 32-bit lookup locator interpreted only in the host
  PID namespace where Host Coordinator and bootstrap recovery run.
- `start_time_ticks` is unsigned 64-bit `/proc/<pid>/stat` negative evidence.
  Equality of PID and start time adds no authority; inequality or an unreadable
  required observation refuses.

A retained pidfd is the primary live authority. Polling, waiting, signaling,
and termination act through that exact retained descriptor. Production process
control has no numeric-PID signal or reopen-and-signal fallback.

`CausalChildAcquisitionV1` applies to every Lazaret-owned child that can become
durable process ownership. Each Host Coordinator and Runtime Supervisor has one
exclusive in-process broker for all spawn, `SIGCHLD`, and reap operations. The
process has default non-auto-reaping `SIGCHLD`; no same-parent spawn or reap and
no dependency process API bypasses the broker while acquisition is active. The
broker performs this sequence:

1. Persist the owning participant or terminal intent without positive process
   ownership.
2. Hold process-wide spawn/reap exclusion and create an inherited fail-closed
   launch barrier.
3. Spawn a direct repository-owned wrapper that blocks before target work;
   barrier EOF or malformed release exits without executing the target.
4. Capture the host-visible child PID, call `pidfd_open` exactly once, and call
   `waitid(P_PIDFD, WEXITED | WNOHANG | WNOWAIT)` on that descriptor. `ECHILD`
   or any other error refuses; `WNOWAIT` never reaps the child.
5. `fstat` the same pidfd, read the current boot ID and diagnostic start time,
   and form `PidfsProcessIdentityV1`.
6. Persist `Staged` ownership. An RS-launched child is sent as the exact
   `StagedProcessOwnership` tuple to HC, and RS waits for an authenticated
   `ProcessOwnershipCommitAck` that echoes the same Domain, generation,
   generation-scoped coordinator ID, ownership slot, and complete process
   identity. Lost, mismatched, unknown-version, unauthenticated, or duplicate
   after release acknowledgment is never success.
7. Release the wrapper barrier only after the exact staged ownership commit is
   acknowledged. The receiver consumes the acknowledgment once while that
   exact barrier is pending.
8. The wrapper arms its parent-death action, rechecks the expected parent pidfs
   authority to close the arming race, and execs the foreground,
   non-daemonizing target in place.
9. Promote ownership to `Committed` only after target-specific readiness.
   Staged ownership cannot satisfy readiness, publication, dependent-resource
   creation, or adoption. It remains positive authority for exact same-handle
   cleanup and `PidfsRecoveryMatch`, including bootstrap recovery.

Every durable host-process slot is exactly `Unassigned`,
`Staged(PidfsProcessIdentityV1)`, or `Committed(HostProcessIdentity)`. Only a
current pidfs identity may be staged. A committed legacy-unverified identity is
retained only for migration and refusal. Current staged and committed identity
both authorize exact matched cleanup; only committed current ownership may
satisfy an active role, readiness, publication, or adoption.

A role-bound Participant is first committed as bare `Intent` with no ownership
snapshot. Once physical acquisition begins, its role snapshot is present and
never returns to absent. Every non-process physical field is optional while
acquisition is partial; every process field uses the three-phase slot above.
Within Intent, optional fields move only absent-to-present. A newly acquired
child slot moves only `Unassigned` to `Staged` to the exact same current
identity in `Committed`. An already committed current owner or supervising
process may enter a newly created dependent-role snapshot directly, but that
recorded identity then remains immutable. Legacy data cannot use this
existing-process rule. Recorded evidence cannot be replaced or cleared.
The direct existing-process edges are Network RS, Helper RS, Sandbox HC, and
Session Terminal HC. State Anchor daemon, Git-routing daemon, Helper process,
Sandbox RS, attached runsc, and Session Terminal exec client are acquired child
edges and require staging.
No bare Intent moves directly to an active state: a non-empty acquisition
snapshot is committed first. PreparedLower additionally records
`lease_state=Intent` before the catalog insertion changes both participant and
lease phase to `LeaseHeld`.

The active Participant pairs are `prepared_lower=LeaseHeld`,
`state_anchor=Prepared`, `git_routing=Prepared`, and
`network|helper|sandbox|publication=Committed`. An active snapshot contains all
role-required physical evidence and current committed pidfs ownership in every
required process slot. `SessionTerminal=Running` likewise requires its
container, PTY, guest identity, and current committed HC and exec-client slots.
Staged, unassigned, or legacy process ownership cannot satisfy either boundary.
Network readiness evidence must also name the enclosing generation and match
its policy, fingerprint, namespace, and committed RS identity internally.

Any partial or active snapshot moves to `ReleaseIntent` while preserving every
recorded identity. A staged current identity remains exact cleanup and recovery
authority in ReleaseIntent. `ReleaseIntent` becomes `Clean` only after exact or
independent absence proof; that transaction clears every process slot to
Unassigned and every optional physical field to absent. Clean retains the
logical generation key, role, and state as an explicit empty role snapshot. A
bare Intent may become that explicit Clean snapshot only after the sealed
runtime-absence check. No ordinary lifecycle transition may clear evidence or
produce Clean.

An exited, unreaped child remains pinned while acquisition completes. The
descriptor-based child check rejects a foreign replacement, while the exclusive
same-parent broker prevents a replacement child from passing the same-parent
control case. Implementations preserve these as safety conditions rather than
timing assumptions.

`PidfsRecoveryMatch` is the only positive recovery reacquisition path:

1. Read and validate the current boot ID.
2. On boot mismatch, do not look up the stored PID. Reconcile only resources
   whose independent exact ownership survives reboot.
3. Verify that the stored locator is a host-PID-namespace PID; nested or
   ambiguous mappings refuse before lookup.
4. Call `pidfd_open` exactly once for that locator. `ESRCH` or any other error
   cannot authorize process action.
5. `fstat` the resulting pidfd and compare its device and inode exactly with the
   durable authority. Either mismatch refuses and closes it without action.
6. Apply start-time and expected namespace observations only as negative checks.
7. Use that same matched pidfd for every action. An action-time `ESRCH` or
   observation failure refuses without numeric fallback.

Domain Store v1 PID/start-only values dispatch to
`LegacyUnverifiedProcessIdentityV1`. They cannot be observed, converted, or
infallibly migrated into positive authority. They refuse live process action
until reboot, independently proven absence, or explicit operator resolution
allows the ordinary recovery state machine to reconcile the surrounding record.
Boot mismatch fences all old pidfs authority before PID lookup.

The contract accepts two residual assumptions: Linux boot-ID collision is
operationally negligible but not mathematically impossible, and the qualified
boot-monotonic 64-bit pidfs inode counter does not wrap during one boot. Codecs
and comparisons support `u64::MAX` and refuse every unequal authority tuple;
they do not claim to recognize a hypothetical equal tuple after wrap.

#### Protocol cryptography refinement

Before starting the Agent Sandbox, the Host Coordinator creates one ephemeral
Trust Anchor package containing separate Ed25519 TLS-CA and protocol-issuer keys.
Its signed Trust Manifest binds both roots to the generation identity defined in
the schema. The Session Coordinator TLS and protocol credentials use distinct
leaf keys physically held by HC; the Agent Sandbox uses another request-signing
key. The publication identity pins the CA, manifest, and coordinator TLS leaf in
the validated host-private Domain path, while the exact sandbox credential mount
pins the same manifest. Protocol credential IDs are SHA-256 digests of their
signed claims. `SessionCoordinator` remains the logical protocol role; HC is its
physical key holder, not another principal.

The host-private protocol socket uses TLS 1.3 with the generation CA, fixed ALPN
`lazaret-control/1`, no early data, and no session resumption. A connection with
no client certificate may perform only a credential enrollment of at most 8,192
bytes from the trusted developer UID beneath that runtime directory. HC verifies
the signed CSR and issues a one-hour `clientAuth` leaf; all other operations
require a new mTLS connection whose leaf matches the issued Client CLI and
generation. The HC leaf is `serverAuth`-only for
`coordinator.lazaret.invalid`. Peer PID and process start time are not protocol
identity evidence.

The Client CLI keeps its Session Anchor private key in host memory. HC delivers
the signed Session Request Capability with guest execution and mounts the Agent
Sandbox signing key, credential, Trust Manifest, and Session Coordinator
verification credential read-only at the fixed credential path. No TLS,
protocol-issuer, coordinator-signing, or Session Anchor private key enters the
sandbox. HC unlinks the sandbox-key source after the running sandbox holds the
mount and retains only its public key.

All Trust Anchor and session cryptographic material is ephemeral. Normal teardown
withdraws publication before dropping keys and credential artifacts. HC loss
destroys the in-memory package; bootstrap never restores it into another
generation.

### 3.3 Session Terminals

The Host Coordinator owns one host PTY master per interactive Client Session. The
PTY slave is attached to non-detached `runsc exec` standard streams with a guest
TTY. The root attached `runsc run` client remains terminal-neutral and is not the
Session PTY attachment path. Production must not depend on `runsc exec
-console-socket`.

Session terminal ownership is the Host Coordinator; a Client CLI never inherits
the sandbox Session PTY. For each Client Session the coordinator maintains the
session identity needed for routing and cleanup, including Session ID, PTY master,
`runsc exec` client process identity, and guest process or process-group identity.

Interactive terminal control requires:

- Each interactive command starts as the leader of a generation-bound guest
  process group. The Host Coordinator records both the leader PID and process
  start time and the process group ID relative to the exact runsc container
  before admitting terminal events.
- Window resize applies `TIOCSWINSZ` to the host PTY master and then delivers
  guest `SIGWINCH` to that process group through runsc. Rows and columns are
  nonzero unsigned 16-bit values; pixel dimensions may be zero when unknown.
- Client `Interrupt`, `Quit`, and `Suspend` events translate respectively to
  `SIGINT`, `SIGQUIT`, and `SIGTSTP` for that process group. Raw control bytes do
  not substitute for these events. Host CLI suspension is local: the CLI
  restores its terminal before stopping itself and sends the current window size
  after resuming; it does not synthesize a guest signal.
- Per-session exit status comes from the non-detached `runsc exec` client. An
  interactive PTY produces one merged output byte stream, so the protocol does
  not label its bytes as separate stdout and stderr streams.

Disconnected Client Sessions close rather than reattach to an unbounded scrollback
owned by the coordinator.

The session connection carries the terminal events and messages defined by the
normative design schema. One Host Coordinator event loop serializes input,
resize, signal, close, prompt-routing, PTY output, and exec-client completion for
each Session ID. It uses bounded queues and PTY backpressure; terminal data for
one session cannot consume another session's control capacity. Input, resize,
and signal events are accepted only while the modeled terminal route is the
sandbox route. An invalid Session ID, inactive route, stale generation, or
terminal in release refuses the event.

PTY output is forwarded only while the modeled terminal route is the sandbox
route. During `Prompt` or `HostProcess` routing the Host Coordinator stops reading
the PTY master and relies on bounded kernel PTY buffering and backpressure; it
does not accumulate an unbounded user-space queue. Forwarding resumes only after
the model restores the sandbox route.

Normal guest completion begins the modeled session close and physical release,
so no new terminal or escalation event is admitted. The Host Coordinator
first cancels any active prompt or exact host process and serializes that
cancellation before terminal output. It then collects and reaps the exec client,
drains and forwards PTY bytes in order until EOF or a bounded drain deadline,
and emits at most one `SessionExit` after the last forwarded byte on the closing
connection. This close handshake does not restore session routing. If the
connection is already lost, the outcome is not retained for reattachment.

Physical Session Terminal acquisition is ordered as follows:

1. The Host Coordinator records `SessionTerminal=Intent` in the Domain Store.
2. It allocates the PTY and starts the exec wrapper through
   `CausalChildAcquisitionV1` with the expected committed HC pidfs authority.
   The wrapper arms its HC parent-death edge, rechecks that authority, and aborts
   before invoking runsc if it differs. It may create the guest process only
   after staged exec-client ownership is durably committed and the barrier is
   released.
3. It establishes the guest process group and obtains the guest leader PID,
   start time, and group ID.
4. One Domain Store transaction promotes the exact Host Coordinator and exec
   client ownership to committed, records the PTY, container, guest leader, and
   guest process-group identities, and sets `SessionTerminal=Running`.
5. Only then may the Host Coordinator accept terminal events for the session.

Failure before `Running` never admits the terminal. A live Host Coordinator uses
its in-memory exact handles to terminate any partial exec and guest group and
records release intent. If HC itself is lost, the armed exec edge stops the host
client and the generation lease-loss path stops the guest sandbox; bootstrap
recovery cleans the retained Intent only after runtime absence.

Session close, Client connection loss, PTY loss, or exec-client loss starts the
same physical release:

1. Apply the modeled session close, remove terminal and escalation routing, and
   record `SessionTerminal=ReleaseIntent` while retaining all exact identities.
2. If the exact generation is still serviceable and the guest leader identity
   and group still match, send `SIGHUP`, wait for a bounded grace period, then
   send `SIGTERM`, wait again, and send `SIGKILL` if the group remains. A leader
   mismatch refuses the numeric process-group signal. If the leader is absent,
   an empty recorded group needs no signal; a nonempty group is left in release
   intent for Domain teardown.
3. Terminate and reap the exact exec client when it is still a Host Coordinator
   child and close the exact PTY master. Record `SessionTerminal=Clean` only
   after the guest group, exec client, and PTY are absent; otherwise retain
   release intent.

The process-group boundary is the per-session cleanup contract, not a security
boundary. A sandbox process may daemonize or move into another process group;
the final Client Session leaving the Domain triggers Domain teardown, which is
the containment boundary for such processes.

Hard Host Coordinator loss closes its PTY descriptors and fails modeled open
sessions; the exec-client lifetime edge prevents an unrecorded or recorded exec
client from surviving that loss. The surviving Runtime Supervisor follows the
generation lease-loss teardown: it stops the complete Agent Sandbox, unmounts
the exact State Anchor while retaining the prepared-lower lease, and exits.
Bootstrap recovery waits until the Runtime Supervisor and attached runsc are
absent, never recreates a PTY or reattaches a Client Session, releases the
retained prepared lower under the durable generation refinement, verifies that
the exact generation container and guest processes are gone, terminates any
still-live exact exec-client host process, and then records each retained Session
Terminal `Clean`. An `Intent` without physical identities becomes `Clean` only
after that runtime-absence check. A process, container, PTY, or generation
mismatch fails closed and retains the record; it is never adopted or signaled as
the recorded session.

### 3.4 Filesystem and Git Model

The Agent Workspace is mounted read-write. Tracked paths and untracked paths that
are not ignored propagate in both directions. The Git index, `HEAD`, references,
and objects are shared with the host in real time. Lazaret does not create a
separate worktree or branch and does not promise that the host working tree
remains unchanged.

By default, paths matched by the repository's effective Git ignore rules are
redirected to Ignored Workspace State. Existing host content at those paths is not
exposed to the Agent Sandbox, and sandbox writes remain Domain-scoped and persist
across restarts. This prevents agent-modified dependency trees, virtual
environments, generated executables, caches, and other ignored artifacts from
later being executed by host tools. Trusted host configuration may disable this
behavior when fully bidirectional ignored paths are required.

Ignored-path isolation is defense in depth and does not replace the mandatory Git
Sentry protections below.

Lazaret applies narrow protective mounts over host integration files that must not
be modified by the agent:

- The host repository's `.git/config` is exposed read-only.
- The host repository's `.git/hooks` is hidden behind sandbox-local transient
  storage.
- Supported host editor configuration directories may be hidden behind
  sandbox-local transient storage.

These protections do not make the workspace immutable. Their purpose is to stop
the sandbox from writing the protected host Git configuration and default hook
directory while preserving normal Git operations and bidirectional source edits.

### 3.5 Persistent Sandbox State

The State Anchor preserves developer tooling and selected sandbox system paths for
one Domain. It combines a Lazaret-owned prepared lower from the selected base
image with a Domain-scoped writable upper directory.

#### Prepared lower

Preparation resolves the selected base image, applies layers with a supported
OCI-aware preparer, and materialises an immutable Lazaret-owned root filesystem
tree under the Storage Root. The tree carries a prepared-artifact ID and a source
image identity. Live container-engine image mounts are not a production State
Anchor lower. Archive export alone is not a complete preparation boundary;
container tooling may participate only as transport into Lazaret-owned
preparation. Ad hoc unsupported tar extractors are not a production preparer.

Prepared-artifact identity is independent of its storage pathname. The Domain
Store and Storage Catalog encode the model's generation-to-artifact lease with
separate Domain and generation fields; an engine mount refcount is not its
physical representation. Preparation uses descriptor-relative staging,
publication, quarantine, and deletion beneath validated Storage Root handles.

The prepared rootfs preserves OCI modes and is exposed only as a read-only State
Anchor lower. Lazaret enforces immutability through artifact ownership and
lifecycle rather than by rewriting image permission bits.

#### Durable generation document refinement

Durable Domain state is an embedded store opened exclusively by the Host
Coordinator process that holds the Domain's exclusive lifecycle lock for write.
No other process opens the store for read or write: Client CLIs, the Runtime
Supervisor, recovery observers, and status tools obtain state only through Host
Coordinator IPC or by becoming the lock-holding Host Coordinator. Domain flock
serializes which process may be that exclusive opener. Store mutations use
atomic transactions; an interrupted write leaves the last successfully committed
document, never a half-applied generation decision.

Protocol durable decision vocabulary, multi-role intent, and mismatch refuse are
defined by the protocol model. The product store encodes that generation document
in typed tables under a schema/db version (unknown newer versions fail closed
until an explicit Host Coordinator migration), including generation identity,
minted runtime component IDs, and participant ownership material (for example
prepared-artifact ID, staged or committed pidfs process ownership, mount
identity, and network readiness fingerprint).
Participant roles are `prepared_lower`, `state_anchor`, `git_routing`, `network`,
`helper`, `sandbox`, and `publication`. The `helper` participant independently
records durable intent and exact committed connectivity-helper pidfs ownership
before the `network` participant records readiness. Sessions and the generation mailbox
are separate generation records. Deterministic paths are lookup hints for
physical recovery, not ownership proofs.

The normative logical records and relationships are defined in the
[design schemas](schema.md). Their implementation representation must preserve
the distinct Domain Store and Storage Catalog boundaries. Opaque generation
documents are not a product storage format.

Each catalog operation acquires host-wide `storage-catalog.lock`, opens the
catalog, performs one redb transaction, closes it, and releases the lock. The
lock covers lease changes, prepare publication, and transition to `Reclaiming`.
This permits concurrent Domain coordinators without a persistent global daemon
or concurrent redb opens.

The physical lease-acquisition refinement is:

1. Commit `prepared_lower = Intent` in the Domain Store with the intended
   prepared-artifact identity.
2. Under `storage-catalog.lock`, insert the exact `(prepared-artifact ID, Domain
   ID, generation)` lease in one catalog transaction.
3. Commit `prepared_lower = LeaseHeld` in the Domain Store.
4. Commit `state_anchor = Intent`, create the exact mount, and then commit
   `state_anchor = Prepared` with its mount identity.

The physical unmount and lease-release refinement is:

1. Commit `state_anchor = ReleaseIntent` while retaining its exact mount and
   prepared-artifact identity.
2. Verify the recorded mount identity and unmount it. A mismatched mount refuses
   release and remains untouched.
3. After independently observing that the exact mount is absent, atomically
   commit `state_anchor = Clean` and `prepared_lower = ReleaseIntent` in the
   Domain Store.
4. Under `storage-catalog.lock`, recheck the matching Domain release intent and
   absence of the exact mount, then remove only the exact generation lease in one
   catalog transaction.
5. Commit `prepared_lower = Clean` in the Domain Store.

Host Coordinator hard loss uses the model's split lease-loss refinement. The
surviving exact Runtime Supervisor stops the sandbox and sidecar, verifies and
unmounts its exact State Anchor, and exits without opening the Domain Store or
Storage Catalog. The original durable anchor intent and exact Prepared ownership
authorize rollback of the resource the supervisor created; they do not authorize
catalog lease removal. The catalog lease remains held after unmount.

Bootstrap recovery begins only after the Runtime Supervisor and attached runsc
are absent. If the exact mount remains, the bootstrap Host Coordinator records
`state_anchor = ReleaseIntent` and completes normal exact unmount. If the
surviving Runtime Supervisor already unmounted it, bootstrap independently
verifies exact mount absence. It then atomically commits `state_anchor = Clean`
and `prepared_lower = ReleaseIntent`, removes the exact catalog lease under the
catalog lock, and commits `prepared_lower = Clean`. Thus lease-loss unmount and
lease release are distinct: only a lock-holding Host Coordinator releases the
prepared-lower lease.

Recovery resumes only from matching durable identity. `state_anchor =
ReleaseIntent` retains the catalog lease while recovery verifies and completes
the unmount. `prepared_lower = ReleaseIntent` permits removal of a still-present
exact catalog lease only after the exact mount is absent; an already-absent lease
permits the Domain participant to become `Clean`. A catalog lease with missing,
unreadable, non-release, or mismatched Domain evidence is retained and reported.
A Domain `LeaseHeld` record whose exact catalog lease is absent is inconsistent
and fails closed; recovery neither recreates the lease nor mounts the State
Anchor. A failed generation with a Prepared anchor, an absent exact mount, and a
still-present exact catalog lease is the conservative HC-loss state above; only
bootstrap recovery may advance it to release intent. These rules make every
interrupted release conservative and prevent a mounted State Anchor from
becoming unleased.

Store version acceptance occurs before physical resource inspection. The
Domain Store v1-to-v2 migration preserves every PID/start-only value as
legacy-unverified and never observes it into pidfs authority. Migration steps
create and validate destination rows in one redb transaction and update the
version last. Future versions follow the same fail-closed ordering.

Files not represented in the upper layer follow the selected base image. Files,
directories, and whiteouts recorded in the upper layer remain authoritative when
the lower image changes. This permits installed tools to survive restarts while
allowing unmodified base-image content to follow the selected image.

Persistent state is sandbox state, not host configuration. It must not contain or
override Lazaret's host control plane.

### 3.6 Network Isolation

The Network Sidecar role is realized on the Runtime Supervisor-owned install
path. The Runtime Supervisor holds network-administration capability while it
installs and attests policy and while attached runsc imports the Domain
enforcement network namespace. It irreversibly drops that capability after the
initial runsc handoff and before publication. Stock attached runsc retains
network-administration capability in the subordinate Domain user namespace for
its generation lifetime; it is a trusted host runtime process and has no
authority over initial-user-namespace host network resources. The Agent Sandbox
OCI process does not receive the capability and may join only that prepared
enforcement namespace, not the host network namespace. Production connectivity
uses `slirp4netns`. Installing policy by driving nftables through runsc into the
guest is not a production path. Protocol helper ownership, network readiness,
and publication gates follow the protocol model.

#### Derived policy and readiness encoding

The policy, attestation, readiness-identity projection, and fingerprint inputs
are defined by the normative [design schemas](schema.md).

The Host Coordinator passes canonical policy bytes through a write-sealed
inherited memfd. The Runtime Supervisor verifies the seals and digest, installs
the policy, and returns observed route, helper, process, and namespace evidence
over an inherited Host Coordinator-created private `SOCK_SEQPACKET` channel.
Before helper launch, the Host Coordinator commits the independent `helper`
participant as `Intent`. It acquires the helper through
`CausalChildAcquisitionV1`; after the exact helper and supervising Runtime
Supervisor pidfs ownership is committed, it may accept network readiness
evidence and advance the `network` participant.
The Host Coordinator independently reads the Runtime Supervisor's namespace
identity and persists the attestation, fingerprint, and generation-bound
deny/permit/DNS probe result. Persisted readiness files are evidence after this
exchange, not its transport.

#### Default policy and deny representation

IPv4 destinations are permitted by default except for the RFC 1918 private
address ranges (`10.0.0.0/8`, `172.16.0.0/12`, `192.168.0.0/16`), which are denied
by default. Explicit, more-specific allow rules may grant access within those
ranges. Lazaret also installs the reachability needed for the Domain's selected
DNS resolvers. Other permitted and denied destination prefixes come from trusted
host configuration.

Default and configured deny prefixes are enforced with **loopback-unicast deny
routes** in the Domain enforcement network namespace (routes that direct matching
destinations to a loopback unicast address so packets are not forwarded). Linux
`blackhole` routes are not the production deny representation for this path.
Equal-prefix allow/deny conflicts must be rejected or resolved in favor of denial.

Normal Linux longest-prefix routing determines the most specific matching route.
IPv4 and IPv6 must either receive equivalent policy or an address family must be
disabled.

The OCI configuration joins only the Domain enforcement network namespace, and
`slirp4netns` attached to that namespace is the sole external egress mechanism.
Dynamic policy changes are outside the current physical refinement.

### 3.7 Host Escalation

The Escalation Client exchanges request and result records through the
Escalation Mailbox: regular files written by atomic rename beneath host-private,
Domain-scoped paths. Request and result paths may use different mount modes. The
mailbox is not an authentication boundary. Its paths include Domain and
generation identity and are opened through Host Coordinator-owned directory
handles. Host-only CLI/coordinator control may use host-private Unix sockets that
are never mounted into the Agent Sandbox.

Mailbox files carry the signed protocol records defined by the normative design
schema; TLS is not layered over the regular-file transport. HC opens each ready
entry relative to a confined directory handle, rejects links and non-regular
files, bounds the complete file before allocation, and reads it once. It verifies
the signature over those owned bytes before parsing. Pathnames, file ownership,
mailbox identity files, PIDs, and start times are not authentication evidence.

Requests, cancellations, acknowledgements, developer decisions, chunks, and
terminal results use the signer relationships in the schema. Postcard payloads
are signed with `lazaret/v1/` plus the lower-kebab schema record name and their
exact encoded length and bytes. A complete record is at most 65,536 bytes; a
result chunk contains at most 49,152 output bytes.

The Escalation Client acknowledges the highest contiguous result sequence with an
Agent Sandbox signature. HC retains at most eight unacknowledged chunks per
request and pauses the exact host process's output pipes at that bound. Session
loss, cancellation, or acknowledgement timeout terminates and reaps the process
before removing its result files.

The physical host-process launcher invokes an executable with an argument vector
and contains no shell-expansion path.

Escalation authentication, correlation, approval, execution, cancellation, result
delivery, and terminal routing are defined by the
[TLA+ application protocol model](protocol/Lazaret.tla).

### 3.8 State Inspection and Reset

The host CLI provides read-only discovery, inspection, structural verification,
and explicit reset operations for Domain State Anchors. These operations resolve
a Domain through trusted host metadata and never accept an arbitrary host
filesystem path as an upper-layer target.

Inspection reports metadata and overlay changes without printing file contents by
default. Structural verification requires an inactive Domain and exclusive
lifecycle ownership so its result is not invalidated by concurrent writes. A
reset may target one configured persistent container path or all State Anchors in
the Domain. It discards the selected upper-layer state so the next launch follows
the current prepared lower; it does not modify the prepared lower artifact, Agent
Workspace, Ignored Workspace State, or unselected State Anchors.

Reset is permitted only after Lazaret holds the Domain's exclusive lifecycle lock
and has established that no Client Session, Session Coordinator constitutive
process, sandbox, network helper, host process, or overlay mount is using the
selected state. Confirmation flags may suppress the human prompt but may not
bypass these safety conditions. Reset must be crash-consistent and must never
follow a symlink outside the Domain Root.

## 4. Requirements

These requirements cover physical mechanisms and encodings below the TLA+
abstraction boundaries. Modeled safety and liveness requirements are intentionally
not duplicated here.

### 4.1 Security Requirements

| ID | Requirement |
| :--- | :--- |
| **SR-1** | Agent processes shall execute in a rootless gVisor sandbox and shall not receive host root authority. |
| **SR-3** | Host paths outside the Agent Workspace, developer-configured volumes, and Lazaret-owned fixed control mounts shall not be accessible to the Agent Sandbox. |
| **SR-4** | The sandbox shall be unable to modify the host repository's `.git/config` or default `.git/hooks` directory through the workspace mount. |
| **SR-5** | The Runtime Supervisor shall irreversibly drop network-administration capability after initial attached-runsc network import and before publication. Trusted attached runsc may retain that capability only in the subordinate Domain user namespace for its generation lifetime. The Agent Sandbox process specification shall omit the capability and set `noNewPrivileges`. |
| **SR-6** | Network policy shall cover every enabled address family and shall deny equal-specificity policy conflicts. Default and configured deny prefixes shall be enforced with loopback-unicast deny routes in the Domain enforcement network namespace; Linux blackhole routes shall not be the production deny representation. |
| **SR-10** | Domain sockets, mailboxes, and coordination files shall reside in a host-private directory owned by the developer. |
| **SR-11** | Runtime and persistent path construction shall begin from validated Domain roots, include the Domain ID in resource identities, and verify stored Domain identity before reuse. |
| **SR-12** | The runsc OCI specification shall create private PID, IPC, UTS, mount, and user namespaces and set its network namespace path to the independently verified Runtime Supervisor namespace descriptor. It shall contain no host namespace path entry. |
| **SR-13** | Git-ignored Agent Workspace paths shall use Ignored Workspace State by default, preventing the sandbox from reading or modifying corresponding host content; only trusted host configuration may disable this behavior. |
| **SR-14** | State Anchor maintenance shall remain confined to the resolved Domain Root. Destructive maintenance shall require exclusive lifecycle ownership and an inactive, unmounted target. |
| **SR-15** | Lazaret shall not use a live container-engine image mount as the production State Anchor lower; the lower shall be a Lazaret-owned prepared lower. Archive export alone shall not be treated as a complete preparation boundary. |

### 4.2 Functional Requirements

| ID | Requirement |
| :--- | :--- |
| **FR-1** | Lazaret shall derive a stable Domain ID from the selected Agent Profile and canonical Agent Workspace. |
| **FR-3** | Tracked and non-ignored workspace edits and supported Git operations shall propagate bidirectionally between host and sandbox. |
| **FR-4** | Selected sandbox paths and Ignored Workspace State shall persist across Domain restarts without copying Lazaret host configuration into the sandbox. |
| **FR-7** | Product recovery shall inspect external runtime resources independently of owner-process liveness. Host-process action shall require the retained pidfd or `PidfsRecoveryMatch` against staged or committed current `PidfsProcessIdentityV1`; legacy PID/start-only ownership, unequal pidfs authority, boot or namespace ambiguity, and unreadable negative evidence shall refuse without numeric-PID signaling. Staged ownership authorizes only exact cleanup/recovery, never service or adoption. Recovery shall also compare typed namespace, mount, artifact, path, and inode evidence before physical cleanup. |
| **FR-8** | Lazaret shall support interactive terminal sessions without degrading supported host terminal capabilities. |
| **FR-9** | Trusted host configuration shall allow the developer to disable Git-ignored path isolation for an Agent Profile. |
| **FR-11** | The host CLI shall list, inspect, structurally verify, and reset all or selected State Anchors without requiring direct manipulation of Lazaret's storage layout. |
| **FR-14** | The Host Coordinator shall own one host PTY master per interactive Client Session; the PTY slave shall attach through non-detached `runsc exec` standard streams. The root attached `runsc run` client shall remain terminal-neutral. Production Session PTYs shall not depend on `runsc exec -console-socket`. |
| **FR-15** | For each interactive Client Session, Lazaret shall preserve the Session Terminal record and protocol in the normative design schema; serialize bounded terminal events and merged output by Session ID; map resize and the `Interrupt`, `Quit`, and `Suspend` events to the specified exact process-group operations; collect exit status from the non-detached `runsc exec` client; and follow the specified release and hard-loss recovery ordering without reattachment or mismatched cleanup. |
| **FR-16** | Domain network connectivity shall use `slirp4netns` as the production connectivity helper. The Domain Store shall record independent `helper` intent before launch and exact committed helper and supervising Runtime Supervisor pidfs ownership before network readiness. |
| **FR-17** | The Domain Store shall be opened only by the Host Coordinator process holding the Domain `flock`; mutations shall use whole redb transactions. Storage Catalog opens shall be bounded by `storage-catalog.lock` and one transaction. Cross-store lease release shall follow the durable release-intent, exact-unmount, catalog-delete, Domain-clean ordering defined by the durable generation refinement. |
| **FR-18** | Product network-readiness evidence shall conform to the normative design schema and include the readiness fingerprint plus the generation-bound deny/permit/DNS probe result. |
| **FR-19** | The Host Coordinator shall derive generation-bound network policy from trusted host configuration; the Runtime Supervisor shall install only that derived policy and attest readiness by observed namespace state. Durable network participant state shall record the network readiness fingerprint. |
| **FR-20** | The physical network probe shall execute in the Agent Sandbox network view or an equivalent generation-bound process using the same enforcement namespace and record configured deny, permitted reachability, and selected DNS resolver observations. |
| **FR-21** | The Domain Store and Storage Catalog shall preserve the records, ownership variants, identities, and relationships in the normative design schema. |
| **FR-22** | Preparation shall use a supported OCI-aware preparer that materialises a Lazaret-owned immutable rootfs with source identity and prepared-artifact ID. |

### 4.3 Developer-Experience Requirements

| ID | Requirement |
| :--- | :--- |
| **DX-1** | A developer shall be able to launch an Agent Profile in the current workspace with one command. |
| **DX-2** | Tools installed in configured persistent sandbox paths and dependencies stored in Ignored Workspace State shall remain available on later launches of the same Domain. |
| **DX-3** | Normal source editing, `git status`, staging, committing, and reference updates shall work from inside the Agent Sandbox. |
| **DX-4** | Escalation prompts shall identify the requesting profile, session, command, and justification without mixing terminal input or output between sessions. |
| **DX-5** | Failures shall report actionable diagnostics and shall not leave the developer's terminal in a corrupted mode. |
| **DX-6** | State inspection shall provide human-readable and machine-readable output without exposing persisted file contents by default, and reset shall show its resolved Domain and exact scope before confirmation. |

### 4.4 Operational Requirements

| ID | Requirement |
| :--- | :--- |
| **OR-4** | Security-sensitive behavior shall be covered by integration tests on supported runsc, filesystem, and network configurations. |
| **OR-5** | An interrupted State Anchor reset shall recover to either the prior generation or a new empty generation; it shall not leave a partially reset active upper layer. |
| **OR-6** | Lazaret shall not require a persistent global daemon or multi-Domain resident process. Shared Storage Catalog operations shall use bounded lock/open/transaction/close critical sections. |
| **OR-7** | Product GC shall provide policy-controlled reclamation for prepared artifacts after all protocol leases for the prepared-artifact ID are released. Shared-artifact reclaim shall wait until no lease remains for that artifact. In-progress prepare trees and foreign trees shall not be deleted; exact-owned abandoned prepare trees may be reclaimed only through fail-closed recovery backed by durable prepare intent. |