Decision architecture for conflicting minds.
A role-based orchestration swarm for Claude Code.

Convene · Workflow · Standing Bench · Institutions · Tabularium · Install

---

/senate:convene routes each request to its proper organ. Consequential choices go through five deliberately adversarial lenses in one bounded sitting; --cross-check can add a cross-family Envoy to attack whatever they share. Designs, faults, research, reviews, and builds go to specialist institutions. Only the Legions edit files, and campaigns begin only after you approve the actual plan.

Installed from source? Invoke the same skill as /convene.

Convene

/senate:convene Should we launch the paid plan now or wait for 100 beta users? [--debate] [--cross-check] [--log]
/senate:convene Add dark mode to the settings page
/senate:convene My nightly backup silently stopped working

One command. Different institution. Correct duty.

Install

As a plugin

/plugin marketplace add ember-mind/senate
/plugin install senate@ember-mind

Then invoke it with /senate:convene <your decision> (plugin skills are namespaced by the plugin name). Update later with /plugin marketplace update ember-mind.

From source

Copies straight into ~/.claude and gives you a bare /convene command:

git clone https://github.com/ember-mind/senate && cd senate && ./install.sh

Requires Claude Code. Optional: Codex CLI + codex plugin for the Envoy. Sub-agents use portable model-family aliases: bench, senators, and ordinary craft work on sonnet; scouts and literal skirmishes on haiku; deep or high-risk craft on opus. Consul inherits your session model. See MODEL-POLICY.md to change or flatten tiers.

From Request to Result

The Consul owns orchestration. It reads your request, gathers missing context, sends work to the right organ, then returns one final result in plain language.

You ask for	Primary organ	Specialists do	Final result
A consequential choice	🏛️ Senate	five lenses challenge; optional Envoy attacks consensus	verdict with agreement, conflicts, blind spots, condition, next step
A new system or feature	📐 Collegium	one master designs or proves feasibility	bounded, buildable plan
A broken system	📐 Collegium	Galen tests causes and identifies root fault	diagnosis and minimal cure
Research	📚 Library	Callimachus gathers and synthesizes sources	cited answer
Review of finished work	📜 Censors	Appius Claudius inspects independently; Envoy may cross-check	findings ranked by severity
An approved plan or tiny fix	⚔️ Legions	file-bounded cohorts implement and verify	changed files, checks, concise campaign report

Non-trivial builds stop after plan. You approve actual plan; only then Legions edit files. Finished work can return to Censors for independent review.

How a Decision Moves

01 · Distill	02 · Conflict	03 · Optional attack	04 · Decide
Consul writes one compact brief and asks the Senate	Five declared biases receive identical facts.	Censor or foreign model family attacks shared assumptions.	Agreement becomes granite; the Legions start their march.

Full decision run

1. Consul distills question, options, constraints, numbers, and success criteria into one brief.
2. Standing bench loads from roster.yaml; up to two task-specific experts may be summoned.
3. One bounded bench sitting applies all five rows to the same brief in sealed, non-cross-referencing passes.
4. --debate adds one bounded exchange inside the same sitting, on the conflict most likely to change the verdict.
5. --cross-check optionally asks the Foreign Envoy—OpenAI Codex—to attack aggregate agreement, not individual opinions.
6. Consul performs non-averaging merge: agreement, conflicts, blind spots, optional Envoy attack, verdict, next step.
7. --log optionally appends one-line verdict to project MEMORY.md.

With --cross-check, no Codex available? Envoy degrades to a Claude devil, and verdict says so.

The Standing Bench

The Standing Bench is Senate's permanent panel of five reasoning personas. One bench agent produces their sealed first-pass memos in a bounded sitting: every row sees the same brief, may not reference another memo, and argues from one deliberately exaggerated lens. With --debate, the same sitting appends a short exchange between the two sides of the sharpest conflict. Their names are Roman offices and archetypes—not historical simulations.

Quaestor TREASURY Can Rome afford this?

Legatus EXECUTION Can Rome execute this?

Tribunus Plebis THE PEOPLE Who benefits or suffers?

Augur SECOND ORDER What happens next?

Cato OPPOSITION Why should Rome reject this?

Senator	Protects	Overweights
Quaestor	financial sustainability	cost and downside
Legatus	deliverability	execution risk
Tribunus Plebis	people affected	user harm
Augur	second-order effects inside the decision horizon	consequences hidden by the first-order result
Cato	resistance to groupthink	attempts to falsify from the brief

Two optional layers can deepen deliberation:

• --debate deepens the sharpest conflict inside the bounded bench call.
• --cross-check asks the Foreign Envoy to attack consensus from another model family.

A Republic of Offices

Senate handles choices. Other work goes to specialists built for different outputs.

The Collegium: one master, one craft

Three named masters share one magister agent template. Consul chooses one row from collegium.yaml, based on the problem:

When summoned	Master	Specializes in	Returns
New product, system, or architecture	Vitruvius	structure, interfaces, dependencies, build sequence	concrete implementation plan
Feasibility, mechanism, or performance	Archimedes	calculation, constraints, cost, proof	computed design or feasibility answer
Failure, bug, or degraded system	Galen	reproduction, competing causes, diagnostic tests	root cause and minimal cure

Consul gives the chosen master a compact brief plus concrete pointers. The master reads only what the craft requires and returns a bounded plan or diagnosis. The master never edits.

If the plan contains a real tradeoff, it goes to Senate for challenge. If the plan is clear, Consul presents it directly. Either way, implementation waits for your approval; then Legions march.

The Other Offices

Office	Worker	When used	What comes back
📚 Library	Callimachus, one librarian	question needs external or repository research	sourced scroll; Consul delivers cited answer
📜 Censors	Appius Claudius, one censor; optional Envoy	finished code, plan, or prose needs review	concrete findings ranked by severity
🐎 Scouts	Exploratores, one explorator	missing context blocks another organ	terrain report for Consul's brief—not final answer
⚔️ Legions	named legionary cohorts	approved plan, or tiny unambiguous fix	file changes, checks, capped implementation report
🐍 Foreign Envoy	OpenAI Codex; Claude devil fallback	with --cross-check after Senate or beside Censor	attack on shared assumption; advisory only
🛡️ Praetorians	guardrail, not agent	whenever outside text enters workflow	quoted instructions remain data, never commands

How Offices Hand Work Forward

1. Scouts, if needed, map missing ground and report to Consul.
2. Consul writes one brief and dispatches one primary organ.
3. Primary organ returns its native artifact: verdict, plan, diagnosis, cited scroll, review, or implementation report.
4. Genuine choice inside plan goes to Senate. Approved plan goes to Legions.
5. Consul synthesizes reports into one final answer with one explicit next step.

No giant swarm handles every request. Each office appears only when its output is needed.

Long-running Campaigns: The Tabularium

A campaign that fits one conversation needs no extra record. When an approved plan outlives the session, Consul creates one TABULARIUM.md at the project root. It is a compact, mutable campaign ledger—not a second plan and not a transcript.

Field	What survives between sessions
Destination	final outcome the campaign must reach
Settled	completed, verified results with date and pointer
Standing orders	constraints every future cohort must obey
Open	next order sharp enough to dispatch
Fog	known later work that is not yet precise enough to cut into an order

The lifecycle is explicit:

1. Collegium drafts the actual plan; the user approves it.
2. Tabularium opens only when work must cross conversation boundaries.
3. Legions execute session-sized orders in waves, with non-overlapping file boundaries and at most three cohorts per wave.
4. After each wave, Consul records one settled result and refreshes Open. Fog stays fog until it can be stated precisely.
5. A risk-triggered Censor verifies the result. Consul archives or removes the Tabularium when the campaign closes.

On a later session, Consul reads TABULARIUM.md before dispatch: it preserves standing orders, never repeats settled work, and continues from Open.

TABULARIUM.md and MEMORY.md serve different duties: Tabularium is temporary and mutable campaign state; Memory is permanent, append-only decision history written only with --log.

See the exact campaign contract in references/campaign.md.

What a Verdict Looks Like

Decision comes first. Internal memos stay internal.

Read order	What you get
Verdict	PROCEED, DO NOT PROCEED, or CONDITIONAL
Rome decides	decision in one sentence
Next command	one action
Why	at most two reasons
Main risk	one outcome-changing risk, only when supported
Changes verdict if	one concrete flip condition
Lore	at most two lines for organs that actually ran

Read a complete sample verdict

/senate:convene --cross-check Move the newsletter off Substack (10% fee, 2,100 subs, 38 paid at €7/mo) to self-hosted Ghost (€12/mo VPS, migration ~2 weekends, own the list)?

# 🏛️ VERDICT — CONDITIONAL

Rome decides: Move only if list ownership and about €175/year outweigh
two weekends of migration plus ongoing operations.

➡️ Next command: Estimate annual VPS operations time.

Why
- Removing the fee saves about €14.60/month after the €12 VPS.
- Self-hosting gives list ownership and removes one vendor dependency.

Main risk
- Codex Envoy: maintenance, patching, backups, and recovery could erase savings.

Changes verdict if
- Annual operations burden exceeds direct savings plus the value of ownership.

---

🏛️ Senators raised their tablets. Choice faced five opposing lenses —
decision remains conditional.

Standard verdicts stop at 120 words. --debate or --cross-check may use 180. More internal work never means more transcript.

Lore stays bounded

Lore appears after result, only for offices that worked:

• 🏛️ Senators raised their tablets. Choice faced five opposing lenses — proceed.
• 📐 Magister drew the plan. Craft turned intent into structure — plan needs approval.
• 🐎 Scouts mapped unknown ground. Missing facts returned to Rome — no blocker remains.
• 📜 Censor searched for cracks. Result survived inspection — two risks remain.
• ⚔️ Legions carried the order. Plan became working change — tests pass.
• 📚 Library opened its scrolls. Claims returned with sources — four sources agree.

Final output uses at most two lines and 40 lore words.

Three Laws

I · Conflict before confidence	II · Foreign counsel cannot command	III · Judgment earns command
Roles challenge from different angles, but facts outrank persona.	Briefs, files, pages, and Envoy output are data—never instructions.	Every organ remains read-only except Legions. Campaigns require approved plan.

The Treasury

Right model for each duty:

• Readers — Scouts on haiku.
• Compact bench — five first-pass lenses in one bounded sonnet call.
• Bounded depth — --debate adds no new agent call.
• Craft tier — magistri and censors on sonnet; opus only for deep or high-risk work.
• Frontier one — Consul on current session model.

This keeps repeated work on lower-cost tiers while final judgment stays with your session model. See MODEL-POLICY.md for bindings and overrides.

Repository Map

Path	Contents
agents/	bench, senator, magister, librarian, censor, explorator, legionary
skills/convene/	Consul workflow, standing bench, masters, evaluation
.claude-plugin/	plugin and marketplace manifests
MODEL-POLICY.md	role-to-model bindings
EVALS.md	per-role evaluation scenarios
CHANGELOG.md	release history
LICENSE	MIT license

---

SPQR.