The room rotates because the code rotates it: the next speaker is chosen by who has been quiet longest, with no test of whether they have anything to say. Every expert is then told to include a concrete suggestion, so the one with nothing to add produces a small amendment instead of passing. Late in a long meeting almost every turn is an amendment to a sentence the room already agreed on. The short meeting in this sample does not have the problem at all, which is the clearest sign that length, not the experts, is what breaks it.
What the four meetings look like
| Meeting | Expert msgs | Avg length | Amendment | New direction | Asks a question | Disagrees |
|---|---|---|---|---|---|---|
| Sep 6 · Beckett / SAT game | 89 | 880 | 65% | 5% | 1% | 2% |
| Jul 27 · Game Concepts | 39 | 2,132 | 71% | 20% | 12% | 10% |
| Jul 26 · Game Mechanics | 281 | 885 | 38% | 6% | 4% | 2% |
| Aug 8 · NPC signal (healthy) | 19 | 873 | 26% | 26% | 5% | 0% |
Read the last two columns together. In the nineteen-message meeting, new directions and amendments appear at the same rate. In the eighty-nine-message meeting, amendments outnumber new directions thirteen to one, and the ratio gets worse as the meeting runs: twenty-one amendments in the first half, thirty-seven in the second.
The three mechanisms
The next speaker is chosen by silence, not by relevance
selectNextAgent weights each candidate as
1.0 + turnsSinceSpeaking × 0.3 and picks one at random from that. There is
no term for whether the expert's lane has anything to do with what was just said. The
weighting actively pushes toward equal shares, and it gets them: on 6 September the
four experts spoke 24, 22, 21 and 22 times; on 26 July, 70, 82, 64 and 65 times. This
is the "rotating turns for the sake of rotation" you described, and it is literally
what the function computes.
Every turn is required to produce a suggestion
All three expert prompt builders carry the line "Prefer concrete suggestions over philosophy. Include at least one concrete suggestion when possible." Combined with the first mechanism, an expert who is called on with nothing new to add is still obliged to suggest something. What comes out is a micro-amendment. In the late stage of the 6 September meeting, nine consecutive messages each appended a clause to the same sentence of the plan.
Nine turns, one sentence Game Feel proposes the slip → Learning Science adds a spacing line → Product Philosopher bans the word "Tonight" as a date in disguise → Game Feel withdraws the header → Learning Science resolves the revisit line late → Feature Strategist folds it into the migration → Feature Strategist adds a station id → Game Feel adds a drawer verb → Product Philosopher bans remarking on elapsed time.
Every one of those is individually reasonable and individually small. Together they are the room ratcheting, and no turn is permitted to say "this is settled, I have nothing."
The length rule contradicts the content rule, and loses
The same prompts say "Keep responses concise (1-3 sentences). Vary length naturally." A substantive concrete suggestion with its reasoning does not fit in three short sentences, so the model drops the length rule and keeps the content rule. The measured result is an average of 880, 885 and 873 characters across three different meetings with three different sets of experts. Five to six sentences, every time, from everyone. Uniform length across a whole bench is the signature of the bench collapsing into a single voice, which is the failure mode the Rookery roster names explicitly and prevents with per-seat word ceilings enforced in code.
Nobody may pass, and nothing is ever closed
A dev meeting has no equivalent of the Rookery's silence token, so an expert cannot decline a turn. There is also no record of what is settled, so an agreed point stays open to amendment forever. Concession language appears twice in eighty-nine messages and five times in two hundred eighty-one. Explicit disagreement runs at 2%. Positions accumulate; they never resolve.
The short meeting is fine
The 8 August meeting on NPC signalling ran nineteen expert messages and shows a one-to-one balance of new directions to amendments, with no ritual cross-referencing at all. Nothing is wrong with these experts or their charters. What degrades is the behaviour after roughly forty turns, once the topic has narrowed and the loop keeps calling on people anyway.
The brake was broken
Mira's turn had been failing silently since her eleventh
Her directive prompts embedded the entire meeting transcript, and her CLI session is resumed from turn to turn, so each request re-sent everything into a session that already held every previous copy. By her eleventh turn in the 6 September meeting she had been handed roughly 773,000 characters of overlapping transcript, and the next request carried another 186,000 on top. It failed, produced nothing, and the loop moved on to the next expert. That is exactly what you saw: a typing indicator, then an expert, then nothing from her.
She now receives only what was said since she last spoke, because her session already holds the rest. Across the same eleven turns that is 67,000 characters instead of 773,000, and the twelfth turn that failed would carry 16,000. A turn that still fails now retries on a fresh bounded session, says so in the error banner, and no longer costs her the next scheduled check-in.
What to change, most leverage first
These are prompt and loop changes, not model changes. The first three are small and reversible. None of them cap what an expert may explore: they change what happens when an expert has nothing to explore, which is where the noise comes from.
Let an expert pass
Give the dev bench the Rookery's silence token. An expert with nothing new replies with it and the turn is dropped rather than appended. This is already built and working in the Rookery loop, so it is a small port, and it directly removes the obligation that manufactures micro-amendments.
Replace "always suggest" with a turn contract
Each turn is one of four things, named: a proposal, a challenge, a question, or a pass. Never "an amendment because it is my turn." A challenge and a question are both productive and neither requires inventing a new constraint, which is what the current instruction forces.
Differentiate length per expert instead of capping everyone
Give each core expert its own ceiling the way each Rookery seat has one, and say explicitly that matching the other experts' length is itself the failure. The current uniform "1-3 sentences" is both ignored and, when it is followed, wrong for the experts who should run long.
Weight the next speaker by relevance, not only by silence
Add a term for whether the last two messages touched the expert's lane. Staleness stays in the formula so nobody disappears, but it stops being the only thing that decides.
Keep a settled list, and make reopening cost something
Mira already maintains the living document. Add a settled section, and tell the experts not to re-amend anything in it unless they can name what new evidence changed it. The nine-turn slip sequence would have ended at turn three.
Phase the meeting explicitly
The current prompts treat message 1 and message 105 identically. Have Mira open by declaring a diverge phase, where wide and speculative proposals are the point and the concision rules are relaxed, then declare the converge phase, where the turn contract and the settled list apply. Right now the room is asked to be concrete and concise from the first message, which is why outside-the-box proposals cluster in the first half and vanish in the second.
What I am unsure about
- The amendment and new-direction figures come from language patterns, not from judgment about each message. They are directionally solid and the gap between meetings is far too large to be measurement noise, but treat the exact percentages as approximate.
- I can prove Mira's turn failed and that her prompt was growing quadratically. I cannot prove from disk which error the CLI returned, because nothing logs a failed meeting turn durably. That is worth fixing separately.
- Whether recommendations 1 and 2 alone are enough is genuinely unknown until you run two or three meetings with them. My guess is they cut the amendment rate substantially and leave the exploration untouched, but that is a guess.