An agent receives capabilities, not ambient authority. A policy defines the resources it may use and the bounds on their use; anything not granted is denied. The goal is that code can work directly in a workspace without inheriting the host user's other privileges.
“Guarantee” here means an enforceable, testable security property under explicit trust assumptions—not a proof of correctness, invulnerability or complete CPython compatibility. The critical question is whether every effect crosses one mediation boundary, and whether independent tests can falsify that claim.
agentknit-kimi-k3.1-guaranteed --policy policy.json 'Implement the requested change'
agentknit-kimi-k3.1-guaranteed --session SESSION_ID --policy policy.json
agentknit-caps --json policy --policy policy.json
agentknit-caps --json check --policy policy.json --code program.py
agentknit-caps --json run --policy policy.json --code program.py--policy uses the same file across resumed sessions. The model cannot choose a
policy, runtime binary, import callback or resource limit. Kimi endpoint settings
and credentials come from inference-db entry kimi-coding on the host; the
keyring dependency reads its Secret Service credentials. Strict
cache-proof checking remains on by default. If the provider's first response
has no cache-accounting field, --no-strict-cache-proof explicitly permits
continuing without a cache guarantee.
{
"schema_version": 1,
"fs": [{"host_path": "/absolute/project", "virtual_path": "/workspace", "mode": "rw"}],
"env": {"EXPLICIT_NAME": "value"},
"functions": [],
"limits": {
"source_bytes": 100000, "result_bytes": 100000, "output_bytes": 1000000,
"feed_seconds": 30, "wall_seconds": 40, "memory_bytes": 256000000,
"recursion_depth": 200, "suspensions": 1000, "write_bytes": 10000000,
"total_calls": 100
}
}All keys are strict; unknown keys, overlapping mounts, missing host paths,
non-normalized virtual paths and unavailable host functions fail closed. ro
means read-only and rw means direct read-write. Environment grants are exact
name/value pairs. policy --json redacts their values. The initial host-function
registry is empty; embedding applications can explicitly register validated
functions. Network, subprocess and native-code authority are not offered.
For each run, only the declared mounts, environment values and registry functions are supplied to a fresh Monty checkout. No ambient host filesystem or environment is supplied; socket/process modules are unavailable. Writes are confined by Monty's capability-relative mounts. Source, output, result, memory, recursion, suspensions, time, write bytes and total calls are bounded. A worker/session is never reused between model calls. Execution failures return structured errors and never fall back to CPython.
State the claim as a total mediation property: every observable effect either is permitted by the effective policy or is absent. Then provide four kinds of evidence:
- Architecture: enumerate every model-to-host path and show there is no fallback executor, ambient file tool or model-controlled callback. A single tool schema alone does not prove this; review dispatch and the runner too.
- Differential/adversarial tests: pair every denied probe with a granted control, so a broken executor that rejects everything cannot appear secure. Include traversal, absolute paths, escaping symlinks, ambient environment, sockets, processes, host functions and resource boundaries.
- Independent oracles: assert effects on the host (sentinel contents, filesystem changes, process/network observations), not merely sandbox errors or exit statuses. Run fuzzing and mutation tests: deliberately remove a mediation check and require a test to fail.
- Assumptions and provenance: pin and audit Monty, exercise the exact production tool path, and document trusted components and residual risks.
Finite tests cannot prove universal absence of escapes. Stronger evidence would be OS-level isolation plus a machine-checked model of the complete host boundary. The static checker is not that proof: dynamic Python remains indeterminate and runtime mediation is still required.
python -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/agentknit-kimi-k3.1-guaranteed 'Fix the failing tests in /workspace'Without --policy, the current working directory is directly writable at the
virtual path /workspace. This is intentionally direct: sandbox code can modify
every file beneath an rw mount. Use a disposable checkout or a read-only policy
when that is not acceptable.
agentknit_caps.testing.invoke(paths) returns source defining and calling
run_tests(paths). The runner discovers test_*.py, loads each file with
sandboxed exec, and calls zero-argument test_* functions. Test and production
code execute in the same Monty checkout and therefore have identical mounts,
environment, host functions and limits. Failures are structured data, not a host
process exit. This is a small Monty-compatible test protocol, not pytest
compatibility: fixtures, plugins, parametrization, arbitrary third-party
packages and CPython-only syntax are unsupported. Host pytest is used to test
this package itself and is never exposed to the agent.
pytest --cov=agentknit_caps --cov-report=term-missing
mypy agentknit_caps tests
ruff check .