Three things have to be built. A cohort mechanism so nineteen personal seats do not land in every code project's expert catalog. A seat model carrying the structured fields the roster depends on — room order, standing tensions, length ceilings, fails-by — because those are enforcement mechanics rather than prose. And a Rookery meeting format that lifts the four-agent cap, skips the handoff step, and concludes with the disagreement instead of a synthesis.
The problem
Code Raven's meetings exist to produce work. ConclusionStep runs handoff first, then summary, then chronicle. MeetingSummaryResult carries summary, keyInsights and actionItems. The whole apparatus converts a conversation into something you can do.
The Rookery is designed to do the opposite. "No seat closes a question." "A session ends with the disagreement stated sharply, not resolved." "Never a recommendation." "No host summary, no synthesis paragraph, no 'in short.'" Run the roster through the existing conclusion pipeline and the machinery fights it every session — generating action items the roster forbids and a synthesis the roster names as failure.
Three concrete blockers on top of that: StandupMeeting.init enforces 2...4 agents while Bench: wants 3–5 and Full bench: wants more; AgentDefinition has no field for register, terms, room order or fails-by; and AgentCatalog.allAgents is a flat bundled list handed to every project.
Approach
Cohorts, so Personal seats stay in Personal
AgentDefinition
gains cohort, defaulting to .core, with .rookery as the second case.
Every catalog surface filters by the active project's kind. Nineteen seats appearing in Press Raven's expert
list would be the same failure shape as the blueprint vocabulary leak, and this is the general mechanism that
prevents it for any future set.
A seat model beside the agent model, not inside it
RookerySeat carries
what the roster actually specifies — owns, refuses, opensWith,
register, terms, roomOrder, maxWords,
failsBy. The prompt file carries voice; the model carries the mechanics that get enforced in code.
Splitting them is what lets room order and standing tensions be real rather than advisory.
Enforce what erodes, prompt what doesn't
The roster's own thesis is that uniform
output means the bench has collapsed. Room order becomes a sort. Standing tensions become a pair table that
injects a mandatory-disagreement directive. Length ceilings become maxWords. Presets become a
picker. Register, Terms and Tell stay in the prompt, where they belong.
Conclusion produces a disagreement, not a summary
MeetingSummaryResult
already carries dissenting: [String]. The Rookery conclusion fills that, keeps
summary short and neutral, and leaves keyInsights and actionItems
empty. No handoff step runs.
Decisions made for you
The nineteen seats are bundled, not authored through Craft Expert
That builder is designed for
interviewing one expert into existence. Nineteen seats already fully specified in the roster should be written
directly as prompt files plus catalog entries plus agents.yaml rows in the Personal workspace (not this repo's), following the same wiring the
Craft Expert checklist describes. A seat can still be edited afterwards through prompt customization.
Full bench is capped at nine
"Everyone with standing" is the roster's language, but each agent holds its own CLI session and this codebase has a documented failure where agent runs wedge in an uninterruptible IPC wait. Nine feels full and is recoverable. One constant to change.
The roster file moves into the repo
It currently sits untracked at the repo root. It belongs at
docs/rookery/roster.md as the source of truth, with the per-seat prompts generated
beside it.
Rookery output never reaches the editorial triage queue
These seats produce judgment about a kitchen
renovation, not findings with a file:line anchor. They surface in Personal and in the meeting
record only.
The work
Phase 1 — Cohort scoping
4 tasksNineteen personal seats must not reach any code project.- Add an AgentCohort enum with core and rookery cases to AgentDefinition
- Default every existing AgentDefinition to the core cohort
- Filter the agent catalog by cohort so rookery seats appear only in Personal
- Verify no rookery seat appears in another project's expert list
Phase 2 — The seat model
5 tasksThe mechanics the roster depends on, as data.- Add the RookerySeat model with roomOrder, maxWords, failsBy and the roster fields
- Add the standing-tensions table as seat-id pairs with an axis label
- Add the presets table mapping a question shape to a seat list
- Move rookery-roster.md to docs/rookery/roster.md as the tracked source of truth
- Build RookerySeatCatalog holding all nineteen seats as structured data
Phase 3 — The nineteen prompts
8 tasksWrite in batches. Uniformity here is the failure mode.- Write the five People and Institutions seat prompts
- Write the four Time, Money and Direction seat prompts
- Write the three Body, Mind and Memory seat prompts
- Write the three Place and Form seat prompts
- Write the four procedural seat prompts
- Register all nineteen in AgentCatalog with the rookery cohort
- Install the nineteen seats into the Personal workspace agents.yaml and prompts directory
- Cross-check every prompt for vocabulary borrowed from another seat
Phase 4 — The Rookery meeting format
8 tasksWhere the roster and the existing pipeline diverge.- Add a rookery case to MeetingFormat
- Lift the 2...4 agent validation for rookery meetings and cap the bench at nine
- Sort speaking order by roomOrder with Devil's Advocate always last
- Inject a mandatory-disagreement directive when both poles of a tension are present
- Enforce each seat's maxWords ceiling in its turn directive
- Skip the handoff conclusion step entirely for rookery meetings
- Conclude with a disagreement block that fills dissenting and leaves actionItems empty
- Keep rookery findings out of the editorial triage queue
Phase 5 — Summoning and stances
7 tasks- Build RookeryCommandParser covering seat, Bench, Full bench, Route and Drift check
- Support +Empty Chair with a named subject bound per invocation
- Implement Route as a Mira reply that names the room without convening it
- Implement Drift check as a self-audit pass against each seat's failsBy line
- Never let the Oblique Strategist and the Devil's Advocate sit in the same session
- Add the Rookery convene sheet as the Personal-only entry point in the Meetings tab
- Accept an optional pursuitId so a convened room can be told what it is about
Phase 6 — Probation and verification
3 tasks- Track per-seat session counts so the Ethicist's probation can be reviewed
- Surface a probation prompt once the Ethicist has sat in ten sessions
- Build the macOS target and verify the product with scripts/verify-build-product.sh
Phase 7 — The room
5 tasksFrom the first review: the Rookery becomes the inline core of Personal's Meetings tab.- Make the Rookery the inline core of Personal's Meetings tab with the developer experts collapsed beneath it
- Replace the grammar chips with a who-sits choice: I choose, the host chooses, everyone with standing, ask only
- Add a seating chart so the chair can choose seats by hand, with tensions and the cap shown at the table
- Demote presets to a menu that fills the question and pre-seats the room
- Attach files, records and pasted images to a Rookery session
Phase 8 — The record
3 tasksSecond review: the ending must be actionable. Seats still never recommend; the record distills.- Replace the disagreement-only conclusion with a five-part record: agreement, recommendations, ideas, cautions, the disagreement
- Add a cautions field to the meeting summary model with defensive decoding for stored meetings
- Relabel the transcript view, summary editor and copied notes for a Rookery record
Phase 9 — Mira as counsel
6 tasksThird review: synthesis needs a subjective voice with the whole table in view.- Seat Mira as the Rookery's lead on the Meeting Lead machinery with her own voice and color
- Keep a living brief through the sitting via the document protocol, organized by the object being designed
- Give Mira three speaking rights: unanswered openers, two versions of one thing, the chair's direction
- Enable the chair's directives in a sitting: advance, where things stand, recommendations so far, assign work
- Write Mira's closing brief above the clerk's record and into Chronicle
- Add a Rewrite Record & Brief action so a concluded sitting can be re-run
Manual Testing
6 tasksYours — whether the bench sounds like nineteen people is not a thing a build can check.- Convene a bench on a real question and confirm the seats sound like different peopleyou
- Confirm two seats in a standing tension actually disagree by nameyou
- Confirm The Eye stays under sixty words and the Oblique Strategist says one sentenceyou
- Confirm a session ends with a disagreement and produces no action itemsyou
- Confirm none of the nineteen appear in Code Raven's own expert listyou
- Rewrite the SAT detective sitting and judge whether Mira's brief organizes the games and their optionsyou
Risks & edge cases
Nineteen prompts written in one pass will drift toward each other
This is the roster's own named failure: "if every seat is producing the same 80 words, you have failed." A single generation session converges on a house voice. Write them in the five thematic batches, restate the anti-collapse rules at the head of each, and treat the cross-check as real work. Read three prompts side by side before calling the phase done.
Nine concurrent agent sessions is past anything tested
Each agent holds its own CLI session, and
there is a documented failure where runs wedge in mach_msg2_trap and survive kill -9.
Run seats sequentially with the existing pacing delays, and treat the first nine-seat session as a load
test.
The disagreement block is generated, and generation smooths
Asking a model to state where seats split tends to produce a balanced both-sides paragraph — the synthesis the roster forbids, wearing a different hat. The conclusion prompt must name that trap and require the block to quote seats rather than paraphrase them.
Room order is a partial order, not a total one
The roster gives positions like "early," "quiet then reframes once," and "interrupts." An integer sort flattens that, and the Empty Chair in particular is specified to interrupt, which no ordering expresses. Accept the flattening and let the Empty Chair be inserted at a caller-chosen point.
Nothing connects a bench to a pursuit yet
Personal experts finally have an anchor now that Pursuits
shipped, but the wiring does not exist. A rookery meeting should accept an optional pursuitId so
the room can be told what it is about.
Out of scope
- The Care Navigator seat — tabled in the roster, raise it only if a question needs it
- Any rookery seat running unattended, scheduled, or as a background agent
- Rookery output feeding handoffs, plans, or the triage queue
- A visual seating chart or room diagram
- Porting the cohort mechanism to Venture Raven
Final review
The strongest part of this plan is that it takes the roster seriously as an engineering document rather than a prompt. Room order, standing tensions and length ceilings become code, which is the only way they survive contact with a model that wants to be helpful and even-handed.
The weakest part is unavoidable: whether nineteen seats actually sound like nineteen people cannot be verified by a build. The cross-check task and the manual testing section carry that weight, and the prompts are where the quality of this whole feature lives.
If one thing gets cut for time, cut Phase 6. Probation tracking is charming and it is the least load-bearing thing here.
High on the architecture — the cohort split, the seat model and the diverging meeting format each address a specific blocker in existing code, and the roster is unusually complete as a source document. Held down by two things: the nine-seat load is past anything the meeting service has been exercised at, and the nineteen prompts are the kind of work that looks finished while being subtly homogenous.