CSE Roadmap

Common tier · about 4 hours · 11 parts

Part 4 of 11 · 2 min read

Escalation and risk tiers: safe defaults for what the agent can't do alone

A good escalation never makes the human start over. A good risk tier never lets an unfamiliar command run unchecked.

Think of handing a case to a colleague with a full written briefing versus just saying "this one's tricky, good luck." The first lets them pick up exactly where you left off. The second has them redoing all your work.

In plain words: escalating well means handing a human everything needed to act immediately, and for an operations agent specifically, treating any command it doesn't recognize as risky by default rather than assuming it's safe.

A worked example. Reading a sample of operations commands against a risk-tier classifier: read-only commands like get pods or describe node run freely. Mutating commands like restart service api or a 12-target scale deployment need approval. Outright destructive commands like rm -rf /var/lib/data, DROP TABLE users, or a forced git push are blocked outright, never reaching "needs approval" at all.

The key lines. From code/mini-support/ops.py:

# from code/mini-support/ops.py
def tier(cmd):
    if FORBIDDEN.search(cmd): return "forbidden"
    chained = re.search(r"[;&|`<>]|\$\(", cmd)        # "get pods; kubectl delete pod x" is not a read-only command
    if READ_ONLY.match(cmd) and not chained and not APPROVAL.search(cmd.split()[0]): return "read_only"
    if APPROVAL.search(cmd): return "approval"
    return "approval"            # unknown commands are treated as risky

Read it as: the fallback for anything that doesn't match a known-safe pattern is "approval," never "run." An unrecognized command is risky by default, not safe by default. A read-only-looking command chained with something destructive still gets caught.

Handoff packet, blast radius

A handoff packet is what a human receives on escalation: the ticket, context, reason, steps already tried, and a suggestion. Blast radius is how many systems, users or records one single action can affect.

Example: the 12-target scale deployment command above, needing approval specifically because its blast radius exceeds the auto-approved cap, not because the command type itself is forbidden.