Engineering blog Post 03
Flight software in the loop: NASA NOS3 and cFS
Our scored benchmarks run on our own simulator. That leaves a fair question from any flight-software engineer: does the command path work against real flight software, or only against code we wrote? NOS3 is how we started answering it.
What NOS3 is
NASA's Operational Simulator for Small Satellites runs real core Flight System (cFS) flight software against simulated hardware and dynamics, behind a ground system [1].
It was developed alongside the STF-1 CubeSat mission [2].
It is a flight-software test bed, not an operations decision layer. That makes it the right place to test the part of the Stack that touches a spacecraft: the gated path from a clearance to a command the flight software accepts.
The command path
The Stack does not talk to cFS directly. A NOS3 link sits behind the contact-window transport and hands cleared commands to NOS3's own ground system, Yamcs.
Yamcs sends them on to cFS through NOS3's simulated radio and link encryption. Telemetry comes back the same way, one Yamcs subscription per packet.
The test command was a cFS no-op, CFE_ES_NOOP: harmless by design, and it increments a command counter, so you can see that it arrived.
The no-op was cleared at the Stack's A3 level (adapt within an envelope), which the SDK supports; ConstellOps v1 caps its agents at A2.
The checks
SDK checks against NASA NOS3 that passed: 7 of 7. In plain words:
- A cleared no-op reached cFS, and the command counter went up by exactly 1. One matching entry appeared in the Yamcs command history.
- A command with no clearance, and one with a clearance from a different pipeline, were refused. Nothing new reached the archive, and the counter did not move.
- Telemetry freshness is judged on the simulator's mission clock, not the host's wall clock.
- A frozen mission clock, and a paused simulator, were both reported as stale.
- The
nos3robot backend returned fresh observations and refused to command.
Several spacecraft, one host
A constellation is more than one spacecraft, so we also brought up several NOS3 spacecraft together: one shared dynamics simulator, and per spacecraft its own cFS, radio simulator, link encryption and Yamcs.
A no-op sent to one spacecraft raised only that spacecraft's counter; the others did not move. NOS3 spacecraft run together, with per-spacecraft command isolation: 3.
This was an environment test, with no SDK and no agent in the loop. All the spacecraft fly the same orbit, and NOS3 itself describes multi-spacecraft use as a proof of concept. We did not try more.
What we are not claiming
The scope line matters more than the result.Link to this quote
So here it is plainly:
- Three of the five gates ran: autonomy, flight rules and the human-confirmation threshold. Cedar authority and the digital-twin gate were not in the NOS3 pipeline.
- No AI agent has commanded NOS3. Every check is deterministic code, with no model in the loop.
- Only a no-op was sent. No hazardous command went to cFS.
- Human approval, the signed decision log, replay and the two-person rule were not exercised on NOS3.
- No ConstellOps benchmark runs on NOS3. Our scored runs use our own simulator.
- The checks ran on a pre-release build of the Stack at a fixed commit, not a release tag.
What went wrong
Next pass
For a flight-software team, this is a first step, not a result: the gated path reaches real cFS and refuses what it should. Putting an agent's proposal through all the gates onto NOS3 is the next one.