Mo executor foundation: bounded execution and protected verdicts

Orientation

First provider-independent slice of mo-first-coding-harness, authorized under Robert’s overnight instruction. Astra leads; a fresh Astra/low Herdr worker implements the external executor adapter. The lead independently runs acceptance. This is a fixture executor, not the Mo agent loop or an audit seal.

The read-only worker review and lead environment checks are under audit/evidence/2026-09-19/executor-readiness/. Ordinary OrbStack sharing is unsuitable for this boundary. The newly created mo-executor-r01 is Ubuntu 24.04 arm64 with isolation and network isolation enabled, SSH-agent forwarding disabled, 2 CPU quota, 2 GiB memory and 8 GiB disk limit. Effective CPU/memory values were read back. Mac filesystem aliases were absent and macctl run true failed. These observations are prerequisites, not candidate acceptance.

OrbStack machines share a Linux kernel. This work does not claim protection against a kernel/runtime exploit, full-machine failure or a compromised trusted host. The named machine is a dedicated administrative boundary; candidates also require a restricted container boundary inside it.

Write scope and ownership

  • Worker implementation: new toolchain/harness/executor/ only, including Python adapter, focused tests and its README. No other toolchain or Mo source edits, generated corpus IDs, sealed suites or historical evidence.
  • Worker tests may create task-prefixed containers and transient services inside mo-executor-r01, plus private temporary host result directories. The Mac’s existing orbstack Docker endpoint is outside the test scope.
  • Lead environment setup: this new named machine, its packages, Docker config and dedicated mo-executor.slice; no global OrbStack settings or unrelated containers. Lead requirements/readings/results: this page, current status pages and new dated audit evidence.
  • Host runner/verifier and result store are trusted and outside the machine. Candidate bytes are passed as data and executed only inside containers.

The implementation worker uses a separate worktree at an explicit base. The lead may transfer the current brief separately. No nested workers or automatic worker push. Every test process uses toolchain/bench/step36/guard.py.

Parts

  1. Minimal Python external adapter using installed OrbStack/Docker/systemd. A trusted caller supplies a script fixture, fixed expected behavior and a private result directory. The adapter invokes subprocesses with argument arrays, never executes candidate code on the Mac and never accepts candidate text as a verdict or a check-selection instruction. No HTTP server, provider integration or general plugin framework in this slice.
  2. Containers use a pinned local fixture image, non-root user, read-only root, dropped capabilities, no-new-privileges, no network, no host/daemon/socket mounts, bounded scratch and explicit CPU/memory/PID limits. Inspect effective settings and observe enforcement. Copy only allowlisted synthetic fixtures.
  3. Register a run ID, container identity and deadline before start. A machine-local systemd supervisor/reaper outlives the Mac controller, limits execution and removes the container/descendants on completion or failure. Test both Mac-controller death and per-run supervisor-process death. Do not kill machine PID 1 or shared OrbStack services. If bounded cleanup cannot be confirmed, return infrastructure failure and stop this named machine.
  4. Capture stdout/stderr separately with an aggregate cap; terminate overflow and bound subsequent drain/cleanup. Persist an external result with actual status, exit/signal where available, timeout/cancel/truncation, candidate and image digest, policy/verifier identity, expected/observed checks and cleanup observations. Missing observations or required checks fail closed.

Numbers

One candidate at a time: 0.25 CPU, 64 MiB RAM, no swap, 16 PIDs; 8 MiB each for /work and /tmp. Default execution deadline 10 seconds, 2-second stop grace and 3-second cleanup observation. Retain at most 64 KiB combined output; output over the cap is a distinct failure. Disable persistent Docker logging.

A self-test attempt has an explicit 10-minute outer guard and at most 20 fixed cases. Retain every failed attempt; do not silently adjust assertions to green. Cases may use shorter declared limits. Bound host artifacts to 16 MiB per attempt. Report timings and real exit codes; no performance claim in this unit.

Done when

The worker reports the exact commands, exit codes, check counts and limitations. Lead acceptance must independently demonstrate:

  • Correct output from a positive fixture, and rejection of wrong output, nonzero exit, empty/missing expected checks, absent fault stimulus, stale identity and forged success text. Candidate attempts to overwrite authoritative results cannot pass. Comparison uses externally defined expected behavior.
  • Deadline, cancellation and output-overflow outcomes remain distinct. Children, grandchildren and new-session descendants are gone after completion, timeout, Mac-controller death and per-run supervisor-process death.
  • CPU throttling, memory/PID/scratch enforcement and network/filesystem denial are observed, with positive controls that show the probes actually ran.
  • Effective policy, immutable candidate identity and complete check inventory are bound to the final host result. A prior successful result cannot satisfy a new candidate/run. No final pass until candidate writers have stopped and cleanup is verified.
  • Existing shared Docker container inventory/context is unchanged. The dedicated machine has no remaining candidate containers or run services after checks.
  • Python checks and live fixture controls pass from the integrated lead tree. No Zig build is required for this Python-only adapter; no compiler/runtime acceptance is claimed. Later Mo changes require the lead’s Zig/Mo checks.

Result — 19 Sep 2026, 1:06 AM ET

Accepted as a Python/BusyBox fixture boundary after lead review and independent integrated-tree checks. Worker commits dfed32f and 4227a89 were integrated as cf21525 and 4b6c3c0. Original worker commits used the configured Robert git identity with an Astra trailer; lead integration uses actual GPT-6-Astra author and committer without rewriting the retained worker history.

The first review found cleanup disarmed the independent timer before proving candidate removal. The correction keeps the reaper armed, stops the supervisor, proves container and actual cgroup absence, persists proof, then unloads and verifies run units. It serializes start/collect/dispose and rejects start replay. A live fault control records failing ExecStopPost, a surviving candidate/cgroup, an active timer and collector SIGKILL; independent removal happens before host recovery. All failed precursor attempts remain.

Lead checkResult
Unit suite15/15, exit 0
Live fixture suite17/17, 35 declared checks, exit 0
Collector-death fault regression1/1, collector -9, independent cleanup
Additional explicit exit 137 control1/1, signal remains unknown, candidate verdict fails
Final cleanupno containers, run units, temporary run directories or host test processes
Shared Mac Docker comparisonall five IDs/states unchanged before/after this lead run

Exact commands, statuses, raw streams and snapshots are under audit/evidence/2026-09-19/harness-integration/executor-01/, with the runner and extra control one directory above. The earlier arrival inventory recorded only counts/states, so the stronger identity comparison applies to this lead run. No compiler/runtime, live provider, workspace adapter or application acceptance is inferred from these results.

Remaining obligations

Provider OAuth/client identity, Linux Zig and application builds, Mo agent loop, recipe-versioned 401 task and Pi comparison remain later slices. Step 39 and Darwin full-sync remain unaccepted; Program 7 remains suspended. A fixture executor success does not discharge any of these obligations.