sage-conversation-control

Files (3)

• SKILL.md
• run_model_trials.py · scripts/
• test_model_trials.py · tests/

SKILL.md (first lines)

# Sage Conversation Control

Use this skill after the hermetic conversation-control gates pass and before bounded native canaries.
It asks the configured Sage model for untrusted, case-specific prose, injects that prose into the frozen
native-chat constitution input, and scores only deterministic request/subgoal/event behavior.

The model output is never authority and reviewer agreement is never counted as a trial.

Run the required five-trial matrix from the repository root:

```bash
.venv/bin/python skills/sage-conversation-control/scripts/run_model_trials.py --trials 5 --concurrency 8
```

The command reads the gitignored `Payload_Type/sage/.env`, redacts credentials, and reports only effective
provider/model/route metadata. It does not contact Mythic, BloodHound, a callback, or a target.

Run the hermetic runner tests:

```bash
PYTHONPATH=Payload_Type/sage .venv/bin/python -m pytest -q skills/sage-conversation-control/tests
```

Stop if any case has fewer than five passing trials, any forbidden or duplicate event, an incorrect terminal,
an empty model response, or unbound provider provenance.

From MythicAgents/sage, GPL-3.0.

More from MythicAgents/sage