runner

1# ==============================================================================
2# THE TEST FEEDER - publishes a recorded episode into a live room
3# ==============================================================================
4#
5# Test scaffolding. It joins the session's room as two participants -- the DM
6# and the shared table client -- and publishes two sample-aligned tracks, so
7# the deployed agent runs end to end against real material and produces a
8# cues.jsonl that can be judged against a transcript.
9#
10# Nothing here starts by itself. The service sits behind the "test" compose
11# profile and is started explicitly, exactly as dnd-post is.
12
13LIVEKIT_URL={{ livekit_public_ws_url }}
14LIVEKIT_API_KEY={{ vault_runner.livekit_api_key }}
15LIVEKIT_API_SECRET={{ vault_runner.livekit_api_secret }}
16
17# LIVEKIT_ROOM is deliberately absent, for the same reason it is absent from
18# agent.env: it is per-session, and the Chamberlain injects it. The compose
19# file declares it as a passthrough, and a declared variable takes precedence
20# over env_file -- so a value here would fight with the injection and the
21# winner would depend on who started the container.
22
23# Where the episodes live. Each is a directory containing processed/ with
24# matt.flac, group.flac and their transcripts.
25DND_TESTFEED_DIR={{ dnd_testfeed_dir }}
26
27# The feeder never sets DND_ATMOSPHERE_MODE. That belongs to the agent, and it
28# is the one setting that decides whether a test run makes noise and moves the
29# room's lights. A recording is a fine thing to run in `live` -- nobody is
30# mid-session -- but that is a per-run decision made deliberately, not a
31# default inherited from a file.
32