The architecture review left a set of open questions for the user (D-optin, D-console, D-ih, D-connect, D-body, D2, D3, D5, D6, D7, D8, D10). They were not answered before v0.1.0, so each one takes the architect's recommended option. Every decision below is revisable; the security requirements listed first are not.
| Requirement | Where it lives |
|---|---|
The console runs apart from the agent, in its own account (lockhttp): since the Docker-only change (after v0.1.0) that account and its socket live in the image, the console in its own container (lox run's service lock-<lock>, or lockhttp-console with the standalone launcher), the agent's containers never get its socket or its home volume. lockkit's isolation checks run against the image's root-owned system.toml. |
Dockerfile, scripts/lockhttp, lockhttp console (lockkit console.Run) |
NeverAuto. Every hub ticket is NeverAuto; an auto rule asks like ask. At parse time such a rule is valid only for a container (or all containers), never for an agent, and never in a production profile, for every host (*), for a spliced host, for messaging, for registry publishing, for git-receive-pack or for a write to any path of a git forge. Uploads, heuristic hits and secrets are checked per request. No auto mode answers a held request. |
internal/policy (validation), internal/gate (hub tickets) |
| A human decision is shared only by requests of identical fingerprint (container, method, URL, every end-to-end header, normalised body). Requests under the same key but with another fingerprint are separate hub tickets, each approved on its own. | internal/gate (ask queue) |
Claude API path and body allowlist once TLS is decrypted: POST /v1/messages and /v1/messages/count_tokens are free only when the body is one unambiguous JSON object without mcp_servers or container and with no server tool other than web_search; /v1/files* asks; OAuth paths and other body-less reads on the four Claude hosts are free; a credential that is not one of the container's is denied; any other write on the Claude hosts, and any other host under the Anthropic domains, asks (NeverAuto), unless the human lists that host as trusted (docs.anthropic.com is in the default set): then it is an ordinary trusted host. A policy rule naming a Claude host applies before the allowlist, and shadow mode lets its asks through. |
internal/decide/claude.go |
Fail closed. No console: an ASK is refused at once (E_NO_CONSOLE). An error while deciding, an unknown source, an unreadable source map: deny. A policy that does not decode counts as no console: FREE continues on the last policy, an ASK gets E_NO_CONSOLE; a policy whose fingerprint does not match is not installed and the previous one stays. The gate never reads a policy file; it only uses what the console serves. |
internal/gate |
| ID | Question | Decision for v0.1.0 | Rationale |
|---|---|---|---|
| D-optin | Opt-in, and migration of LEA_EGRESS_ALLOW? |
Opt-in (LEA_LOCKHTTP=1 on the lea side). With LockHTTP on, squid's hard allowlist is switched off; LEA_EGRESS_ALLOW entries are migrated by hand, as trusted read-only hosts, with lockhttp profile add --trusted; squid's floor (DoH, IP literals, private ranges, ports) stays. Default after the shadow period. |
A new egress path must earn trust before it becomes the default; the floor keeps the old guarantees. |
| D-console | Behaviour without a console | FREE continues on the last approved policy the gate holds in memory (or the built-in floor and Claude hosts when it never had one); ASK is refused at once with E_NO_CONSOLE and a page naming the command to run. |
Availability of trusted reads without weakening anything that needs a human. |
| D-ih | Internal hosts | Before the private-range floor, as in squid today; decided at the domain (spliced); FREE only when marked read, ASK otherwise; on ports 80 and 443 and the ports their line declares. |
Keeps lea's internal hosts working without making them a free channel. |
| D-connect | claude.ai connectors and mcp_servers |
No connector exception; mcp_servers (and server tools other than web_search) on the API is ASK, NeverAuto. Connectors are switched off at the source by lea's managed settings. |
The Claude API is a generic relay; every relay use needs a human. |
| D-body | Body preview on the screen | Masked preview (PII detectors + terminal sanitising, first 1 KiB) by default; never the clear body. | lockkit refuses clear results at level 1; the agent's account may read the window. |
| D2 | Podman rootless without isolated gateway | Accepted if the isolation test passes (blocking test on the lea side, #115 option A). | Rootless Podman refuses the isolated gateway option. |
| D3 | CA key in the gate or in a signer | Separate signer (lockhttp signer): it holds the CA key and issues leaf certificates over a Unix socket; the gate never sees the CA key. |
A compromised gate cannot mint certificates beyond what it asks the signer for, and the key never sits in the data plane. |
| D5 | Hold time and body limit | 60 s hold, 10 MiB buffer limit (buffer_limit), both in the policy. |
Fits common client timeouts (git, npm, pip). |
| D6 | lea deck notification | A counter of pending asks (lea side; the gate publishes pending counts in its status file); the first ASK plays no sound. | Draws the human to the hub without a card that could be mistaken for an approval channel. |
| D7 | Concurrent approvals | Side by side on the hub: each held request is its own hub ticket, sent as it comes; identical requests share one. (Until the gate became hub-attached, it kept its own queue and sent one console approval at a time.) | No lockkit fork, no shared decisions. |
| D8 | Initial trusted list | The set lockhttp profile add writes into trusted (DocsTrusted: Go, Python, npm, Rust, Debian, Ubuntu and Alpine package hosts, GitHub's read hosts, developer.mozilla.org, docs.docker.com, kubernetes.io, man7.org, docs.anthropic.com), plus the hosts given with --trusted. The four Claude hosts are built in, not in the list. Nothing adds OpenAI: trust it by hand if a provider uses it. |
Small, read-only, human-reviewed at first start. |
| D10 | WebFetch | WebSearch allowed; WebFetch allowed (its requests go through the gate and ask for unknown hosts); revisit after the shadow period. | WebSearch is server-side; WebFetch is visible to the gate. |
- One approval, three meanings. An approval on the hub is a ticket the
human approves or denies (with a reason). For a CONNECT (phase 1, or a spliced host)
yopens container × host × port forconnect_grant(10 min by default, a policy field); for a decrypted requestyapproves that request, and creates a single-use authorisation (container, method, URL, every end-to-end header, normalised body; 10 min) if the client already left. Durable openings are policy rules, written by the human and confirmed in:review. - Shadow mode (
mode = "shadow") logs what would have asked or been denied, and lets it pass; the floor (IP literals, DoH, private ranges, ports, SNI mismatch, h2c, nested CONNECT), secret matches, the deny list and the company's managed rules stay enforced.heuristics = "shadow"(the default) does the same for the exfiltration score only. - Source map. The container is read from the local address the
connection arrived on (one proxy address per internal network), from a
file lea writes; an address missing from the map is refused
(
E_SOURCE_UNKNOWN).Proxy-Authorizationnames the agent, for the audit only. - No MCP server. Agents never talk to the console: they meet LockHTTP as a proxy, and its 403 pages say what to ask of the human.
The gate runs hub-attached only: lockhttp gate needs --hub,
--hub-key-file and --instance (lea passes them from lox run's
up.json) and exits otherwise ("launched by lox run only"). The
standalone gate, which asked a console of its own, and its rollback
(E_WITHDRAWN) are gone. lockhttp console stays: under lox run it serves
the approved policy to the gate and asks its human about every loosening;
it serves no approvals (http.approve and --auto are gone). One mode
keeps one path to test and to reason about (KISS).