The cut-over dropped the per-bottle token flow, so an authed egress route on
the shared gateway failed with 'env var EGRESS_TOKEN_0 is unset' — the gateway
reads the token from its env, but a shared gateway has no per-bottle env.
Now the bottle's egress auth tokens travel to the gateway over /resolve and
the addon injects from them, mirroring what the per-bottle sidecar's env did:
- launch resolves the token values from the host env and hands them to the
orchestrator, which holds them IN MEMORY (keyed by bottle_id, never written
to the registry DB) and serves them on /resolve;
- PolicyResolver.resolve_policy_and_bottle_id + resolve_client_context now
return the token map alongside policy + bottle_id (one round-trip);
- the egress addon overlays the process env with the bottle's tokens per
request and uses that env for auth injection AND DLP — the agent never sees
the credential.
Secrets stay off disk (validated: /resolve returns the token, the registry DB
does not contain it). SecretProvider (#355) is the future hardening.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Rewire DockerBottleBackend.launch to the consolidated model, replacing the
per-bottle sidecar bundle. VALIDATED END-TO-END on real docker:
test_sandbox_escape passes all 5 attacks (egress DLP + git-gate gitleaks)
through the shared gateway.
launch now:
- mints git-gate dynamic keys (if any), then launch_consolidated() to
register the bottle + provision its git-gate state into the gateway and
get the agent's attach context (pinned source IP, gateway address);
- installs the SHARED gateway CA (gateway.ca_cert_pem) into the agent;
- renders the agent-only consolidated compose on the gateway network at the
pinned IP, proxied through the gateway;
- points the agent's git-gate insteadOf (http://<gw>:9420) and supervise
MCP (http://<gw>:9100) at the gateway (DockerBottlePlan.agent_git_gate_url
/ agent_supervise_url);
- teardown = compose down + teardown_consolidated (dereg + deprovision).
Dropped the per-bottle networks, per-bottle egress CA, and bundle service.
Fixes surfaced by the real e2e (couldn't be caught by unit mocks):
- provision_git_gate installs the bottle-agnostic pre-receive/access hooks
into the gateway (were cp'd per-bundle before);
- source-IP allocation reads the *actual* container IPs on the network
(gateway + orchestrator + agents), not just gateway + registry — the
orchestrator container sits on the network and was colliding.
Unit tests for the old bundle launch rewritten against the new collaborators.
NOTE for reviewers: the gateway/bundle image (bot-bottle-sidecars) must be
rebuilt when its flat sources change — `ensure_built` only builds-if-missing,
so a stale image silently runs the OLD single-tenant daemons. A content-hash
/ --rebuild path is a follow-up.
pyright 0 errors; pylint 9.83/10; unit suite green (1760 tests; the 13
test_sidecar_init /bin/sleep errors are pre-existing NixOS-local noise);
test_sandbox_escape green on real docker.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Real-on-host testing surfaced two issues the unit-mocked slices couldn't:
1. The host (NixOS) firewall DROPS container->host traffic, so a host-process
orchestrator is unreachable from the gateway container. Fix: run the
orchestrator AS a container on the shared gateway network (PRD 0070's
"virtualize the orchestrator") — the gateway reaches it by container name
over docker DNS (container<->container, no firewall), and the host CLI
reaches it via a published loopback port.
2. The control plane crashed the connection on a dispatch error (e.g. a
broker failure) instead of returning 500.
Changes:
- lifecycle: OrchestratorProcess (host process) -> OrchestratorService
(containers). Runs the control plane in the bundle image with the repo
bind-mounted (orchestrator is stdlib-only), register-only stub broker so it
needs NO docker socket (the backend launches agents; the host manages both
containers). Registry DB persists via a host-root mount. ensure_running is
an idempotent singleton over both containers.
- gateway: BOT_BOTTLE_ORCHESTRATOR_URL is now the orchestrator's *by-name*
URL on the shared network (dropped the host.docker.internal hack).
- control_plane: _serve wraps dispatch — a failure returns 500, never crashes
the connection.
- OrchestratorProcess default broker -> stub (register-only) for docker.
Validated live end-to-end: both containers up, gateway->orchestrator by name
OK, register -> resolve-by-source-IP returns the bottle's policy from inside
the gateway.
pyright 0 errors; pylint 9.83/10; unit suite green (1760 tests; the 13
test_sidecar_init /bin/sleep errors are pre-existing NixOS-local noise).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
The per-bottle model rendered a compose project with the agent + a sidecar
bundle on two per-bottle networks. Consolidated has no sidecars — one shared
gateway serves everyone — so this renders JUST the agent, attached to the
external shared gateway network at the orchestrator-allocated pinned IP, and
pointed at the gateway's address for egress.
- consolidated_agent_compose(plan, *, gateway_ip, source_ip, network):
agent service only; networks {<gateway>: {ipv4_address: source_ip}} on the
external gateway network; HTTPS_PROXY -> http://<gateway_ip>:9099; NO_PROXY
includes the gateway address so git-http + supervise (also on the gateway)
bypass the proxy and are reached directly. No depends_on/sidecars.
- Pure (takes the launch-time LaunchContext values), so it's unit-testable
without docker.
pyright 0 errors; pylint 9.83/10; unit suite green (the 13 test_sidecar_init
/bin/sleep errors are pre-existing NixOS-local noise).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Composes the orchestrator primitives into the register/teardown sequence
that replaces the per-bottle bundle:
launch_consolidated(egress_plan, git_gate_plan): ensure orchestrator +
gateway up -> read the gateway network's subnet + the gateway's own
address -> allocate the bottle a pinned source IP (skip the gateway +
live bottles) -> register (egress policy blob + slug metadata) ->
provision its git-gate repos/creds into the gateway. Returns a
LaunchContext (bottle_id, identity_token, source_ip, network, gateway_ip,
orchestrator_url) — everything the agent container needs to attach.
Rolls the registration back if provisioning fails, so no orphan.
teardown_consolidated(bottle_id): deregister + deprovision (idempotent).
The agent `docker run` itself stays the backend's job (provider
provisioning); this owns the orchestrator-facing wiring so the sequence is
unit-testable — collaborators mocked, next_free_ip + registration_inputs run
for real (IP allocation + policy blob asserted end to end).
pyright 0 errors; pylint 9.83/10; unit suite green (1755 tests; the 13
test_sidecar_init /bin/sleep errors are pre-existing NixOS-local noise).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Places a bottle's git-gate credentials + bare repos into the live shared
gateway container when it registers — the execution side of the 13c(i)
provisioning render.
- provision_git_gate(gateway, bottle_id, plan): mkdir the per-bottle creds
dir, docker cp each upstream's identity key (+ known_hosts when present)
into /git-gate/creds/<bottle_id>/, then docker exec the namespaced,
init-only script from git_gate_render_provision. No-op with no upstreams.
- deprovision_git_gate(gateway, bottle_id): rm the bottle's /git/<id> +
creds on teardown (idempotent).
- bottle_id is validated (shell/path-safe alphabet) BEFORE any docker cp/rm,
so a traversal id can never reach a path arg — the creds/repo dirs land in
docker cp/rm arguments. Registry ids are token_hex; defense in depth.
pyright 0 errors; pylint 9.83/10; unit suite green (the 13 test_sidecar_init
/bin/sleep errors are pre-existing NixOS-local noise).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Deterministic per-bottle address assignment on the consolidated docker
gateway's shared network — the address the gateway uses as its attribution
key. Pure ipaddress logic; the docker-specific inputs (subnet CIDR, the
gateway container's own address) are gathered by the caller and passed as
`taken`, keeping it testable without docker.
- next_free_ip(cidr, taken): lowest host address not reserved (network +
broadcast excluded by hosts(); docker's router .1 excluded here) and not
already assigned (gateway container + live bottles from the registry).
NoFreeAddressError when the subnet is exhausted.
pyright 0 errors; pylint 9.83/10; unit suite green (the 13 test_sidecar_init
/bin/sleep errors are pre-existing NixOS-local noise).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Create bot_bottle/paths.py as the canonical home for the app-root path
helpers (bot_bottle_root, host_db_path, HOST_DB_FILENAME) — foundational,
not supervise- or db-specific. `bot_bottle_root()` now honours a
BOT_BOTTLE_ROOT env override.
Repoint every consumer (supervise, supervise_types, db_store, queue_store,
audit_store, store_manager, bottle_state, cli/supervise, docker/cleanup,
orchestrator/registry) at paths; remove the definitions (and supervise's
duplicate host_db_path) and the now-dead `import sys`. Add paths.py to the
sidecar bundle (Dockerfile.sidecars) for the flat-import copies.
Tests: replace ~12 files' monkeypatching of supervise.bot_bottle_root (and
the flat/pkg/supervise_types triple-patch dance) with a single
`use_bottle_root()` helper that sets BOT_BOTTLE_ROOT — every module and
flat/package copy reads the same env var, so one override covers them all.
Net -97 lines. Behaviour-preserving: full unit suite unchanged (only the
pre-existing /bin/sleep sidecar-init errors remain).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
The pool params (size, IP base, iface prefix, nft table) were triplicated
— hardcoded in netpool.py, scripts/firecracker-netpool.sh, and
nix/firecracker-netpool.nix — plus the IP math (3x) and the nft ruleset
(2x). Nothing enforced agreement; changing the base (ce3fad9, off CGNAT)
forced a coordinated three-file edit, and a missed one would silently
provision a range the launcher doesn't expect.
Collapse to one source of truth:
* netpool.defaults.env — a plain KEY=VALUE file (bash-sourceable,
systemd EnvironmentFile-compatible, Python- and Nix-parseable) holding
the four defaults. A BOT_BOTTLE_FC_* env var still overrides any key.
* netpool.py reads it for the Python defaults (missing file = hard
error, not confusing empty defaults).
* the shell script falls back to it (no literal `:-8` / `10.243.0.0`),
and its `up` is now non-destructive/idempotent (only creates a
missing TAP), so re-running never cuts a live VM.
* the Nix module readFile-parses it for its option defaults and
delegates bring-up to the SAME shell script (dropping its duplicate
IP math, nft ruleset, and TAP loop) — passing every value as
Environment= so the store-detached script never needs the file.
Net: defaults 3x -> 1x, nft ruleset 2x -> 1x, TAP loop 2x -> 1x. The one
remaining IP-math dup (Python launch-addressing vs bash bring-up) is
justified — different runtimes. Tests now guard the invariant (Python
reads the shared file; the script/module hold no literals) instead of
pinning duplicated strings.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
The control SSH logs in as root, so the agent (dropped to node via
runuser) inherited cwd=/root. That was harmless while the rootless
rootfs left /root node-owned, but the dropbear fix (chown /root →
root:root 0700) made /root unreadable to node — so the agent's cwd was
inaccessible, breaking Node's process.cwd(), Claude Code's shell-snapshot
machinery, and `/doctor` (which reported /root/.claude EACCES and failed
every Bash command).
Run the agent from its workdir (default /home/node) via `env --chdir`.
Not a `sh -c 'cd … && exec "$@"'` wrapper: ssh space-joins everything
after the host into one string for the guest shell, so the quoted script
+ $@ get re-split and mangled (it exec'd the $0 placeholder → exit 127).
`env --chdir=DIR …` is all simple words, so it survives the join.
Verified end-to-end: a real headless claude bottle now runs its Bash tool
and exits 0 (was EACCES/127 before).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Make the network pool a persistent, distro-uniform resource instead of a
non-persistent per-distro shell command. systemd is the common
denominator across Debian/Ubuntu/Fedora/RHEL/Arch/… (and NixOS), so
`backend setup` on any systemd host now installs one bot-bottle-owned
oneshot unit (bot-bottle-firecracker-netpool.service): params pinned via
Environment= (so it doesn't depend on $SUDO_USER at boot), ExecStart/Stop
delegating to the bundled bring-up script (single source of logic).
- render_systemd_unit() in netpool.py (derived from the same constants
as the shell script + nix module).
- backend setup: install + enable the unit directly when run as root,
else print a self-contained copy-paste block. NixOS keeps its
declarative module (which produces the same-named unit); non-systemd
hosts fall back to the raw imperative script.
- backend teardown: symmetric — disable + remove the unit.
- backend status: report the unit's active/inactive state so you can
tell a persistent install from an imperative one.
Net: one install path (`backend setup`), persistent, identical across
distros; per-distro variance shrinks to installing nft + iproute2.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
First real end-to-end launch (there was no firecracker integration test)
surfaced three bugs in the control/provision path, all now verified fixed
by tests/integration/test_firecracker_launch:
1. Guest SSH rejected the correct key. The rootless rootfs build
(`docker export | tar` as a non-root user) can't preserve uid 0, so
every path — including /root — is owned by the build uid (node in
guest). dropbear refuses root's authorized_keys when /root isn't
root-owned, so auth fell back to password → denied. bb-init now
`chown -R 0:0 /root`.
2. The SSH client (newer OpenSSH) didn't reliably present the -i key
against the operator's ~/.ssh config; add `IdentitiesOnly=yes` to
ssh_base_argv so only the per-bottle key is offered.
3. cp_in double-wrapped the remote command as `sh -c <remote>`, but ssh
space-joins everything after the host into one string for the guest
shell, collapsing `sh -c mkdir -p X && …` to `mkdir` with no operand
("missing operand"). Pass the command as a single arg and let the
guest login shell run it (stdin still carries the tar).
Also stop discarding dropbear's stderr (`-E` now reaches the host-side
console.log) so future guest-auth issues are debuggable.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
`status()` hard-failed when it couldn't confirm the nft table, but listing
nftables usually needs root — so an unprivileged `backend status` reported
"not ready" even with the pool fully up, making it useless as a launch
gate (and skipping the firecracker integration test on a set-up host).
Base readiness on what the launch preflight actually hard-requires: the
TAP pool present (unprivileged, authoritative) + no range overlap. Report
the nft table state (present / unverified / not-confirmable-unprivileged)
but don't let it flip readiness — the post-boot isolation probe is the
authoritative isolation check, same as the preflight's deferral.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
The module used `networking.nftables.tables.*` (which forces
`networking.nftables.enable = true`, flipping the host firewall backend)
and `systemd.network.enable` (handing interfaces to systemd-networkd) —
disruptive on an iptables + Docker daily driver.
Rewrite it to a single systemd oneshot that brings the pool up: idempotent
`ip tuntap` for the TAPs + `nft -f` for the independent `inet bot_bottle_fc`
table (its own hooks at priority -10), with ExecStop teardown. Same as the
imperative script, but declarative. It touches neither the firewall backend
nor networkd, so it coexists with iptables/Docker/ufw/firewalld. Verified
through the NixOS module system (service present, firewall untouched).
Drop the now-redundant `render_nixos_module()` paste generator (the module
is a real importable file) and point `backend setup` at importing it (flake
output or the file path), noting it's non-invasive.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
`backend setup --backend=firecracker` only covered the network pool. Lead
with the other prerequisites: the firecracker binary (with a release
install pointer when it's not on PATH, plus a NixOS note), a /dev/kvm
check, and a reminder about the cached guest kernel + static dropbear and
mke2fs. The network pool is now step 2.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Add an abstract teardown() classmethod to BottleBackend — the inverse of
setup(), surfaced as `./cli.py backend teardown [--backend=NAME]`
(uninstall). Symmetric with setup: it prints the privileged commands /
declarative config change to remove the host prerequisites.
- firecracker: NixOS-aware — disable the flake module (or drop the
import) and rebuild, or `firecracker-netpool.sh down` imperatively.
- docker / macos-container: nothing to undo (no privileged host state);
print a short note.
Not called by the launch path or the test suite. Extends test_cli_backend
for the new dispatch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Add abstract `setup()`/`status()` classmethods to BottleBackend (same
per-host, no-instance shape as is_available) so host provisioning is
part of the backend contract, not a per-backend command. Replace the
`./cli.py firecracker {setup,status}` command with a generic
`./cli.py backend {setup,status} [--backend=NAME]` that resolves a
backend (flag / $BOT_BOTTLE_BACKEND / host default) and dispatches —
swapping backends is just a different --backend.
Implementations:
- firecracker: moved out of cli/ into backend/firecracker/setup.py
(network pool module/script + range-overlap check), unchanged output.
- docker: new backend/docker/setup.py — reports docker on PATH, daemon
reachability, and gVisor runsc; setup notes no privileged pool is
needed. Minimal placeholder; richer version tracked in #345.
- macos-container: new setup.py — container CLI + system-service checks.
Also retarget the launch-preflight / isolation-probe pointers to the new
command. New test_cli_backend covers dispatch + docker setup/status.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
The default TAP-pool base was 100.64.0.0/10 (RFC-6598 CGNAT) — chosen to
dodge RFC-1918, but that's exactly the range Tailscale assigns node
addresses from, so on a Tailscale host it's the worst pick. Move the
default to 10.243.0.0/16, an obscure RFC-1918 block that steers clear of
docker/libvirt/k8s/LAN and Tailscale.
No default is collision-proof, so add netpool.overlapping_routes(): it
parses `ip -json route show table all` and flags any route intersecting
the pool range (excluding our own bbfc* TAPs and the default route). The
launch preflight warns on overlap; `backend status` reports it.
Distribute the NixOS host setup as a flake module instead of a
copy-pasted blob: nix/firecracker-netpool.nix computes the taps / nft
table from typed options (poolSize, ipBase, ifacePrefix, owner) with a
/31-alignment assertion, and flake.nix exposes it as
nixosModules.firecracker-netpool. Defaults mirror the backend constants;
writeEnvFile emits the matching BOT_BOTTLE_FC_* so the host pool and the
launcher can't drift.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Delete the smolmachines backend (the whole bot_bottle/backend/smolmachines
package and its tests). It had fatal Linux issues (TSI networking under
sustained use, exec-channel contention, no SIGWINCH) and is superseded by
the Firecracker backend (issue #342).
Backend selection now:
- default is macos-container on macOS, firecracker on KVM-capable Linux
hosts, and docker as the last resort (was smolmachines).
- firecracker is selected on a KVM host even when the `firecracker`
binary isn't installed, so start routes through its preflight and
prints an install pointer (same UX as require_container), instead of
silently falling back. Split is_host_capable() (Linux + KVM) out of
is_available() (adds the binary check) to drive this.
Retarget the cross-backend tests (parity, print-parity, prepare,
workspace, freezer, selection) from smolmachines to firecracker rather
than dropping the coverage. Remove docker.util.image_id/save, which only
smolmachines used. Update README/AGENTS/example bottles and stale
comments; historical docs/prds are left as a point-in-time record.
BREAKING: BOT_BOTTLE_BACKEND=smolmachines now errors as unknown.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WBMWTEtQdJ4W5UrWuLHCck
Adds a Firecracker-based backend for Linux, providing mature KVM-based
microVM isolation to replace smolmachines/libkrun (issue #342, closes
the dead-end tracked in #332).
Architecture:
- Guest control over SSH (dropbear injected into the rootfs) on a
point-to-point TAP link. `ssh -t` forwards SIGWINCH natively, so no
resize bridge is needed.
- Networking: a one-time, root-provisioned pool of user-owned TAP
devices (no shared bridge → no docker0/virbr0/cni0 collisions) plus a
dedicated `table inet bot_bottle_fc` nftables table (independent of
Docker/ufw/firewalld rules). `./cli.py firecracker setup` prints the
host-appropriate config (NixOS module or sudo script).
- Rootfs: `docker export` → ext4 via `mke2fs -d` (rootless, no mount),
cached by image digest; per-bottle SSH pubkey + IP passed via the
kernel cmdline.
- Sidecar: reuses the Docker bundle, published on the slot's host TAP IP.
- Fail-closed isolation: TAP pool verified at preflight; the egress
boundary is proven empirically post-boot (before the agent runs) by a
canary probe — the VM must fail to reach the host directly, or launch
is refused.
Linux hosts with Firecracker + KVM now default to this backend;
macOS stays on macos-container.
Not yet validated end-to-end on live hardware (requires the one-time
network pool). Unit tests + pyright pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G8p32HJgPoS1hLPWubbftM
When a git-gate repo entry has no host_key configured, the prepare step
now either errors (headless) or fetches the key via ssh-keyscan, shows
it to the operator for confirmation, and optionally persists it to the
bottle config .md file on disk.
Closes#333
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
On Linux the guest kernel's LOCAL routing table routes all
127.0.0.0/8 to the guest's own loopback interface (priority 0,
checked before any main-table route), so TSI never sees
connections to the per-bottle loopback alias — the fix_guest_
loopback_routing approach confirmed this at the kernel level.
Use the per-bottle docker bridge gateway (192.168.N.1) instead.
It is not a loopback address, so the guest routes it via eth0
and TSI intercepts it normally. The TSI allowlist remains a
/32 that is distinct from the container IP (192.168.N.2), so
direct bypass to egress:9099 is still blocked by TSI.
Changes:
- Add _proxy_host() helper: returns bundle_gateway on Linux,
loopback alias on macOS
- Thread proxy_host through _start_bundle, _discover_urls,
_launch_vm, and _bundle_launch_spec (publish_host_ip)
- Remove _fix_guest_loopback_routing (no longer needed)
- Relax proxy-URL assertion in the integration test to accept
any http://IP:port (with a comment explaining the difference)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Port the smolmachines backend so BOT_BOTTLE_BACKEND=smolmachines
works on Linux (KVM), not just macOS:
- Preflight gates /dev/kvm presence + accessibility on Linux with
actionable remediation (kvm module, kvm group).
- smolvm state-DB path is platform-derived (XDG on Linux).
- force_allowlist runs on both platforms and is fail-closed: it
verifies the persisted TSI allowlist and dies rather than booting
a VM whose egress confinement it can't confirm. Previously it
no-oped on Linux, failing OPEN.
- allocate() does per-bottle 127.0.0.<N> scoping on Linux too (no
ifconfig needed — all of 127/8 is already loopback); only
ensure_pool's lo0 aliasing stays macOS-only.
- README documents Linux + NixOS host setup.
Linux/KVM integration (the sandbox-escape acceptance gate) is
pending verification on a NixOS host; unit tests cover the new
platform branches.
Issue: #283
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NkwFXLFff9PYPy4wgVBJp9
Apple's `container run --mount type=bind` only accepts directory
sources — a file source fails with "path '<file>' is not a
directory". Move the sqlite db into its own ~/.bot-bottle/db/
subdirectory so it can be bind-mounted the same way the CA/routes
mounts already are, without exposing the rest of ~/.bot-bottle.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- `bottle:` in agent frontmatter is now optional; agents without it
are portable and require bottles to be selected at launch.
- Adds `filter_multiselect` to `tui.py`: multi-select picker with
ordered selection list, Space/Enter to toggle, Ctrl-D to confirm.
- `ManifestIndex` gains `all_bottle_names` and `load_for_agent` accepts
`bottle_names: tuple[str, ...]` to merge bottles in order at runtime.
- `merge_bottles_runtime` in `manifest_extends.py` applies the same
field-merge rules as `extends:` to pre-resolved bottle objects.
- `BottleSpec` gains `bottle_names`; `_validate` and `write_launch_metadata`
thread it through so `resume` replays the same bottle configuration.
- `cmd_start` shows the bottle multiselect after agent selection,
pre-populated from the agent's `bottle:` field when present.
- Existing agents with `bottle:` declared continue to work unchanged.
`container build` resolves -f relative to the current working directory,
not the build context, so builds failed from any cwd other than the repo
root. Anchor a relative Dockerfile to the context before passing it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
EgressPlan gains a `canary: str` field (default "") populated in Egress.prepare()
using secrets.token_urlsafe(32). Each launched bottle:
- sidecar receives EGRESS_TOKEN_CANARY=<value> (literal env entry, scanned by
existing known-secrets detector without any detector code changes)
- agent receives BOT_BOTTLE_CANARY=<value> (visible fake secret that signals
exfiltration with zero false positives if it appears in outbound traffic)
Docker compose and macos-container backends updated; smolmachines shares docker
compose and so picks this up automatically. Unit tests cover canary uniqueness,
detection via scan_known_secrets, and EgressPlan backward-compat default.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
On resume from a committed snapshot, smolvm's pack process remaps all
file uids to the host uid (501 on macOS). Files in /tmp that were
created during the session (e.g. /tmp/claude-1000 owned by node=uid
1000) get remapped to 501. Claude Code then refuses to use the temp
directory because it's owned by a different uid.
Two-part fix:
- Exclude ./tmp and ./var/tmp from the tar in _exec_tar_to_file.
Both directories are ephemeral; a resumed VM should start with clean
temp directories identical to a fresh VM.
- Add mkdir -p /tmp /var/tmp to _init_vm before chown/chmod, so the
directories are created if the committed snapshot omitted them.
Replace the Popen/stdout=PIPE approach with a write-then-copy
strategy that avoids binary-stdout piping through the smolvm exec
channel entirely:
1. Probe connectivity with `machine_exec(machine, ["true"])` first.
If this fails while an interactive session is running, the error
now says "concurrent exec not available" instead of the opaque
"<no stderr>".
2. Run `tar --create --gzip --file=/var/tmp/.bot-bottle-commit.tar.gz`
inside the VM via machine_exec (same mechanism used during
provisioning). tar writes to a file in the VM, not stdout, so
smolvm never has to transmit binary data over the exec channel.
3. Copy the compressed archive to the host with machine_cp.
4. Dockerfile switches to ADD rootfs.tar.gz / — Docker decompresses
gzip tarballs automatically.
smolvm machine exec requires stdout to be a pipe, not a regular
file descriptor. Passing stdout=file caused smolvm to return
non-zero with no stderr (the error was silently swallowed or went
to the regular-file fd instead of reaching us).
Switch _snapshot_running_vm to a new _exec_tar_to_file helper that
uses Popen with stdout=PIPE and streams the tar to disk via
shutil.copyfileobj. A background thread drains stderr concurrently
to prevent deadlock when the stderr pipe buffer fills while we are
writing stdout data.
The terminal-decoration wrapper script is invoked with sh -lc, which
sources login-shell init files (/etc/profile, ~/.profile) rather than
interactive-shell files (~/.zshrc). smolvm is typically installed via
homebrew whose PATH setup lands in ~/.zprofile or ~/.zshrc — not picked
up by sh -l — so pty_resize.py's Popen(["smolvm", ...]) raises
FileNotFoundError, pty_resize exits non-zero, and the trailing reset-
printf makes sh exit 0. The caller sees "session ended (exit 0)"
immediately with no agent output.
Use sh -c instead. The calling process (./cli.py) inherits the user's
interactive shell PATH where smolvm is present, confirmed by the
provision steps (machine_exec) succeeding before exec_agent is reached.
smolvm pack create --from-vm requires the VM to be stopped, and stopping
a smolmachines VM terminates any running interactive session.
Instead, mirror the macos-container approach: exec into the running VM as
root and stream the root filesystem via tar (smolvm machine exec -- tar),
build a Docker image from the archive, push to an ephemeral local registry,
and run smolvm pack create --image to produce the .smolmachine artifact.
The VM stays running throughout the commit.
Remove the stop-confirm prompt and machine_is_running check that were
added in the previous commit — neither is needed when we no longer stop.
smolvm pack create --from-vm requires the VM to be stopped. Add
machine_is_running() to smolvm.py (via machine ls --json state field),
and add the same confirm-stop flow to SmolmachinesFreezer that was
originally designed for macos-container: if running, prompt the user,
stop the VM, then pack. Already-stopped VMs are packed directly.
Apple Container removes containers when they stop, making the
stop-then-export flow impossible regardless of the --rm flag.
Replace `container export` (requires stopped container) with
`container exec --user root <name> tar --create ... --file=- --directory=/ .`
streamed to a temp file, then build the committed image from that archive
as before. The bottle stays running after commit, which is better UX.
Drop the stop-confirm prompt from MacosContainerFreezer since we no longer
need to stop the container at all.
container stop was removing the container immediately (due to --rm)
before container export could run. The force_remove_container teardown
callback on the ExitStack already handles cleanup on normal exit, so
--rm was redundant. Without it, the stopped container stays available
for container export to snapshot.
Freezer._freeze only ever used bottle.name, which is always
f"bot-bottle-{agent.slug}". Remove the Bottle parameter from
commit() and _freeze(), derive the container name from agent.slug
directly in each subclass, and delete the _NamedBottle stub that
existed solely to paper over this.
Adds a Freezer ABC (backend/freeze.py) that encapsulates the
stop-commit-mark-preserved flow for all backends, following the same
pattern as BottleBackend. Each backend gets its own Freezer subclass:
DockerFreezer — docker commit
MacosContainerFreezer — container export + image rebuild; prompts
to stop if the container is running
SmolmachinesFreezer — smolvm pack create --from-vm
The base class owns write_committed_image, mark_preserved, and the
resume hint. Subclasses implement _freeze() and optionally override
_export_hint() for migration instructions.
Freezer.commit(agent, bottle) is the primary entry point for use
within a live launch context. Freezer.commit_slug(slug) is a
convenience wrapper for cmd_commit, which no longer branches on
backend names itself.
get_freezer(backend_name) is the factory, analogous to
get_bottle_backend(). CommitCancelled is raised by MacosContainerFreezer
when the user declines the stop prompt; cmd_commit catches it and
returns 0.
`container export` requires the container to be stopped first. When a
running bottle is detected, prompt the user to confirm, stop the
container, then commit. Adds `container_is_running` and
`stop_container` helpers to the macos-container util.
Addresses #240 (comment)
didericis
didericis-claude
didericis-codex