# Decepticon operator examples

These examples separate host commands, Decepticon's interactive slash commands,
and MCP tool calls. Confirm target scope before active testing.

## Inspect a running CLI engagement from Claude Code or Codex

The operator started `decepticon start`, selected an engagement, and began a
run in the interactive CLI. From the coding agent:

```text
decepticon_list_engagements(limit=20)
  -> [{thread_id: "th_1", engagement_name: "example-assessment", ...}]
decepticon_engagement_state(thread_id="th_1")
decepticon_transcript(thread_id="th_1", after_index=0, limit=40)
  -> {next_index: 8, messages: [...], run_status: "running"}
decepticon_transcript(thread_id="th_1", after_index=8, limit=40)
```

Report only new, observed activity. A `decepticon_watch` call can sample a
short live window; it is not a durable event subscription.

## Steer after the current run

```text
decepticon_send_message(
  thread_id="th_1",
  message="Within the approved scope, focus on /api/v2 before other endpoints."
)
```

This queues a later turn. Check the subsequent transcript before claiming that
the agent changed its current activity. `decepticon_cancel_engagement` stops an
active run when the operator requests that.

## Start a local Blue Cell

Inside the Decepticon interactive CLI, the operator can enter:

```text
/blue up 127.0.0.1:3000
/blue verify
/blue events 20
/blue incidents 20
/blue analyze
```

Tell the operator to send test traffic to `http://127.0.0.1:18080`. These are
interactive CLI commands, not MCP tools or shell commands. The coding agent
may inspect the target's local logs and the selected workspace when it has
host access, but it must not claim that MCP started the sensor.

## Read results without hiding a gap

`decepticon_engagement_findings(engagement_name="example-assessment")` reads
`graph.json` only. If it returns `available=false`, inspect the selected host
workspace's `findings/FIND-*.md` and `report/` files. The installed MCP bridge
cannot currently read those files from its container. Report the actual
finding IDs, evidence, verification status, and affected targets from files
you opened; do not turn `available=false` into a no-findings claim.

## A separately mapped LangGraph server

For a deployment that explicitly maps the MCP engagement name to its own
workspace, the agent can call `decepticon_start_engagement` with the user's
approved targets and a full in-scope and out-of-scope instruction. Save its
`thread_id`, then use the transcript cursor, status, and cancel tools as above.
Confirm that mapping before using this flow on an installed Docker stack.