feat(box): support explicit host backend (#2498)

skills/skills/langbot-testing/cases/sandbox-skill-authoring-e2e.yaml

@@ -13,6 +13,7 @@ tags:
   - tools
   - e2b
   - nsjail
+  - host
 skills:
   - langbot-env-setup
   - langbot-testing
@@ -23,7 +24,7 @@ env:
   - LANGBOT_LOCAL_AGENT_PIPELINE_NAME
 preconditions:
   - "LANGBOT_LOCAL_AGENT_PIPELINE_URL or LANGBOT_LOCAL_AGENT_PIPELINE_NAME points to the local-agent pipeline under test."
-  - "LangBot is started with the sandbox backend intended for this run, such as e2b or nsjail."
+  - "LangBot is started with the Box backend intended for this run, such as e2b, nsjail, or explicit host development mode."
   - "The selected model route supports tool/function calling strongly enough to invoke sandbox tools."
 steps:
   - "Start LangBot with the target sandbox backend and confirm the Box status UI or LANGBOT_BACKEND_URL /api/v1/box/status reports the expected backend."
@@ -33,7 +34,7 @@ steps:
 checks:
   - "UI: Debug Chat final assistant response contains E2E_OK:<skill-name>."
   - "Logs: The model called exec, register_skill, activate, then exec again from the activated skill path."
-  - "Logs: The selected backend name is the expected one, such as e2b or nsjail."
+  - "Logs: The selected backend name is the expected one, such as e2b, nsjail, or host."
   - "Skill store: The registered package and activated writeback match references/sandbox-skill-authoring.md."
   - "Box status: recent_error_count is 0 after the run."
 evidence_required:

skills/skills/langbot-testing/references/sandbox-skill-authoring.md

@@ -4,7 +4,7 @@
 
 Verify that Local Agent can use sandbox tools to create, register, activate, and use a LangBot skill package through the same path a user would exercise in Debug Chat.
 
-This flow applies to Docker, nsjail, and E2B backends. API calls are useful diagnostics, but the primary pass/fail signal is the model-driven Debug Chat tool sequence.
+This flow applies to Docker, nsjail, E2B, and the explicit host development backend. Host runs commands directly as the Box Runtime user and must never be treated as sandbox-isolation coverage. API calls are useful diagnostics, but the primary pass/fail signal is the model-driven Debug Chat tool sequence.
 
 ## Preconditions
 
@@ -13,6 +13,7 @@ This flow applies to Docker, nsjail, and E2B backends. API calls are useful diag
    - `BOX_BACKEND=e2b` when validating E2B.
    - `BOX_BACKEND=nsjail` when validating nsjail.
    - `BOX_BACKEND=local` or `docker` when validating local container fallback.
+   - `BOX_BACKEND=host` only when validating explicit, trusted local direct execution.
 3. Confirm `/api/v1/box/status` reports `available: true` and the expected backend name.
 4. Confirm Debug Chat uses a model with function-calling ability.
 5. Confirm backend logs say native sandbox tools are available.
@@ -71,7 +72,7 @@ Backend logs should show:
 - `register_skill`
 - `activate`
 - a second `exec` whose workdir is `/workspace/.skills/<skill-name>`
-- `backend=e2b`, `backend=nsjail`, or the expected local backend
+- `backend=e2b`, `backend=nsjail`, `backend=host`, or the expected local backend
 
 After the run, verify the skill store through the UI or API:
 
@@ -125,6 +126,8 @@ For E2B raw HTTP diagnostics, include a valid template id such as `base`; a miss
 - Session metadata should keep LangBot logical paths such as `/workspace`; storing provider-internal paths can make later requests look incompatible.
 - nsjail versions differ. Some expose only `--disable_clone_new*` flags and use `--bindmount` instead of `--rw_bind`.
 - On WSL, cgroup v2 may exist but not be writable. The backend should warn and fall back to rlimits rather than fail the sandbox.
+- The host backend does not honor sandbox image, network, rootfs, process, or
+  resource isolation. Use a disposable workspace and low-privilege account.
 - If `ALL_PROXY` uses a SOCKS URL and `socksio` is not installed, some Python HTTP clients can fail during startup. Prefer consistent HTTP proxy variables unless SOCKS support is installed.
 
 ## Related Troubleshooting

skills/skills/langbot-testing/troubleshooting/sandbox-native-tools-unavailable.yaml

@@ -3,7 +3,7 @@ title: "Native sandbox tools are unavailable even though a backend is configured
 date: 2026-05-18
 symptoms:
   - "Backend logs show Native sandbox tools (exec/read/write/edit/glob/grep) are NOT available."
-  - "The Box runtime later reports that E2B, nsjail, or Docker is configured."
+  - "The Box runtime later reports that E2B, nsjail, Docker, or explicit host mode is configured."
   - "Debug Chat does not expose exec, register_skill, or activate as usable tools."
 patterns:
   - "Native sandbox tools ... are NOT available"
@@ -19,6 +19,7 @@ fix_steps:
   - "Ensure the Box runtime reselects a backend when get_backend_info is called and the cached backend is empty."
   - "For E2B, verify the key without printing it and confirm any required template setting."
   - "For nsjail, run nsjail --help and confirm the binary is on PATH for the LangBot process."
+  - "For trusted local development only, explicitly set box.backend=host; never use host as a production sandbox fallback."
 verification: "Run sandbox-skill-authoring-e2e. Logs should show Native sandbox tools are available and /api/v1/box/status should report available=true with the expected backend."
 related_cases:
   - sandbox-skill-authoring-e2e