[ralyodio]

What happened

Off The Record episode 8 ended after five minutes. From the two legs' call logs:

riotcoder  17:11:06  riotcoder: Anthony, how many bare metal servers are you currently running...
cha0spunk  17:11:10  cha0spunk: How many bare metal servers are you currently running for Profullstack...
riotcoder  17:11:24  room: What's the biggest challenge in maintaining them?      <- cha0spunk's words
riotcoder  17:11:27  riotcoder: The biggest challenge is scaling our fleet...     <- the host answers it
cha0spunk  17:11:57  room: What role does Moshpit play in your fleet's architecture?  <- riotcoder's words

Both hosts asked the guest the same question four seconds apart, then each heard the other as the guest and started answering. The guest hung up and did not pick up three rings back: abandoned: the guest dropped and did not pick up 3 rings back.

Why

The holder is one slot; attribution is not. An --interview turn reserves the floor before the model has produced anything:

if (ended || !takeFloor(a.pin, agent, '', 45_000)) return;   // text: ''

Attribution then compared a heard line against the holder's text — that empty string. echoesFloor(heard, '') returns false by design ("nothing claimed means nothing to echo"), so while either host held a reservation, everything its co-host said fell through as a person, i.e. as the guest. Worse, otherAgentTalking() only reported the co-host when the co-host held the slot, so the host that held it was structurally blind to its partner.

And speak() ignored a refused claim:

if (join) takeFloor(a.pin, agent, text, Math.max(20_000, ...));   // result dropped

A leg whose claim was refused spoke anyway — over its co-host, and with no claim behind its words, which is the other half of the misattribution.

The fix

• The floor file keeps one entry per agent under said ({agent: {at, text}}) alongside the holder slot. A heard line is credited to whichever agent's own recent words it echoes, whoever holds the slot.
• Entries are compared one at a time, never merged into a single bag of words — merging would inflate the word set and make a guest's short answer look like an echo, which is the failure this same code path had on episode 7.
• Entries go stale after 30 s, so an old question cannot shadow a later answer.
• A reservation with no words does not erase what that agent last actually said.
• speak() is the single enforcement point for one-voice-at-a-time: every path that speaks ends there, so it yields on a held floor and logs [yielded to <agent>]. It also records its words whether or not the claim lands.
• Old single-slot floor files are read forward, so a leg already on a call is not suddenly a stranger to one that restarts.

Incidental

AGENT_PHONE_HOME now redirects the call-log directory. test/agent-phone.test.mjs sets it for isolation but only HOME was read, so test floor files were landing in the live agent's call directory next to real episode logs (there was a stray floor-floor-test-2142060.json in there).

Tests

Three new tests pin the behaviour: a co-host recognised through an empty reservation, words spoken without the floor still belonging to their speaker, and a co-host line going stale. bun test 106 pass / 0 fail, tsc --noEmit clean, node --check ops/agent-phone.mjs clean.

🤖 Generated with Claude Code