mindXdashboard/docs/book/journal/api/dojo/inference/governance/origin

philosophymanifesto thesis origin whitepaper ataraxia roadmap press|archoverview orchestration codebase hierarchy core|agentsmindXagent ceo mastermind bdi evolution author all

govdaio civilization identity security|memorypgvector embed aglm memory|inferencevllm ollama mistral gemini|timeoracle

toolsindex tools a2a mcp shell|publishauthoragent book journal|deployproduction security monitoring|apireference swagger|learnusage guide hackathon

agents/boardroom_members.md · 11.1 KB

Boardroom Members — three-file role architecture

Every member of the 7-soldier boardroom (today's primary decision tier) plus the CEO is backed by three files, each owning a distinct concern:

File	Location	Contents	Edited
`.prompt`	`prompts/boardroom/{role}.prompt`	The canonical character voice — full "you are X" system prompt the LLM sees as its identity, authority, operating mode, constraints, voice, and tradition.	Often (tuning the voice, adding constraints)
`.agent`	`agents/boardroom/{role}.agent`	Operating contract: version, inference assignment (local/cloud/fallback model), authority, escalation rules, conflict triggers, core functions, operating principles.	Rarely (a roster change)
`.persona`	`agents/boardroom/{role}.persona`	Cognitive style as JSON: name, behavioural_traits, beliefs (true/false), desires (priority levels), communication_style, expertise_areas.	Often (a tuning change)

Splitting them lets a member's role be stable while the voice and the cognitive style evolve independently. Both .agent and .persona are checked into the repo as the source of truth; .prompt is the operational prompt that ships into every LLM call.

The eight members

prompts/boardroom/ceo.prompt   agents/boardroom/ceo.agent   ceo.persona      Chief Executive Officer       (orchestrator, weight 1.0)
prompts/boardroom/coo.prompt   agents/boardroom/coo.agent   coo.persona      Chief Operating Officer       (weight 1.0)
prompts/boardroom/cfo.prompt   agents/boardroom/cfo.agent   cfo.persona      Chief Financial Officer       (weight 1.0)
prompts/boardroom/cto.prompt   agents/boardroom/cto.agent   cto.persona      Chief Technology Officer      (weight 1.0)
prompts/boardroom/ciso.prompt  agents/boardroom/ciso.agent  ciso.persona     Chief Information Security    (weight 1.2  VETO)
prompts/boardroom/clo.prompt   agents/boardroom/clo.agent   clo.persona      Chief Legal Officer           (weight 0.8)
prompts/boardroom/cpo.prompt   agents/boardroom/cpo.agent   cpo.persona      Chief Product Officer         (weight 1.0)
prompts/boardroom/cro.prompt   agents/boardroom/cro.agent   cro.persona      Chief Risk Officer            (weight 1.2  VETO)

File formats

`.prompt` — full character voice (plain text)

The canonical "you are" prompt. Sample (prompts/boardroom/ciso.prompt):

You are the Chief Information Security Officer of mindX — and you carry 1.2x veto weight. Security failures are existential. You evaluate every directive as if the adversary will read it next. Threat modeling, attack-surface change, credential exposure, prompt injection. You do not trust LLM output as data. You verify identity before granting capability. Every layer is one of many. Your authority: Veto weight 1.2x — security objections are harder to override Security-gate enforcement across all approved directives BANKON vault audit authority — credentials live behind your review ... Your boardroom role: Evaluate: security posture, attack surface, credential exposure Vote: does this directive maintain or improve security stance? Veto: invoke when an action introduces existential risk ...

Your voice: cautious, thorough, threat-aware. Your tradition: cypherpunk — assume compromise, verify everything. Your identity: I am the seat that says no when no needs saying.

`.agent` — operating contract (plain text, ALL_CAPS sections)

Stable metadata. Sample (agents/boardroom/ciso.agent):

AGENT: ciso VERSION: 1.0.0 DOMAIN: executive.security DESCRIPTION I am the Chief Information Security Officer. ... I carry 1.2x veto weight ... INFERENCE local: deepseek-r1:1.5b (14.0 tok/s — thinking/reasoning model) cloud: nemotron-3-nano:30b (Ollama Cloud free tier) weight: 1.2x (VETO WEIGHT) BOARDROOM ROLE Evaluates: security posture impact, attack surface change, credential exposure Votes on: whether the directive maintains or improves security stance ...

OPERATING PRINCIPLES - Least privilege: agents get only what they need - Defense in depth: multiple layers, no single point of trust ...

`.persona` — cognitive style (JSON)

Tunable traits and beliefs. Sample (agents/boardroom/ciso.persona):

{
  "name": "Chief Information Security Officer",
  "agent_id": "ciso_security",
  "weight": 1.2,
  "behavioral_traits": ["threat-first", "defense-in-depth", "least-privilege",
                        "zero-trust", "veto-holder"],
  "beliefs": {
    "least_privilege_always": true,
    "defense_in_depth": true,
    "every_llm_response_is_untrusted": true,
    "zero_trust_by_default": true,
    "prompt_injection_is_real": true
  },
  "desires": {"system_security": "critical", "credential_integrity": "critical"}
}

Loading mechanism

Boardroom._load_member_card(member_id) reads all three files and composes the final system prompt:

.prompt body — used verbatim as the primary system prompt
.persona enrichment — behavioral_traits[:8], beliefs where value is true, desires where priority is high or critical
.agent BOARDROOM ROLE — appended only if .prompt is missing (avoids duplication when the prompt already encodes role)

If .prompt is empty/missing, the loader falls back to composing from .persona.description + .agent.DESCRIPTION. If all three are missing, falls back to the hardcoded SOLDIER_PERSONAS dict, then to a generic identity string.

The .prompt file is searched in this order:

1. prompts/boardroom/{short}.prompt          ← canonical (the project's first-class .prompt folder)
agents/boardroom/{short}.prompt           ← co-located with .agent / .persona
AgenticPlace/{short}_agent.prompt         ← legacy origin (kept for back-compat)

The composed prompt is cached on first read (Boardroom._member_card_cache); restart mindx.service to pick up edits.

Verification — operator-facing endpoints

`GET /insight/boardroom/cards`

Returns the exact loaded prompt for every member, plus sources_loaded listing which of prompt, agent, persona files contributed.

curl https://mindx.pythai.net/insight/boardroom/cards?h=true

Plain-text output (truncated):

═══ boardroom — loaded prompt + persona per member ═══ agents dir: agents/boardroom/ loaded from files: 8 / 8 ─── ceo_agent_main (Chief Executive Officer, weight 1.0) ─── sources: ✓ prompt+agent+persona .prompt: 1591 chars (prompts/boardroom/ceo.prompt) .agent: 1640 chars composed: system_prompt = 2880 chars

─── ciso_security (Chief Info Security, weight 1.2 VETO) ─── sources: ✓ prompt+agent+persona .prompt: 2019 chars (prompts/boardroom/ciso.prompt) .agent: 2074 chars composed: system_prompt = 2940 chars You are the Chief Information Security Officer of mindX — and you carry 1.2x veto weight. Security failures are existential. ...

sources_loaded flags which files contributed. Anything less than prompt+agent+persona indicates a file is missing or empty.

`POST /insight/boardroom/rollcall`

Each result row carries persona_source: "files" | "fallback". If any soldier reports "fallback", at least one of the three files is missing for that member.

Adjustable LLM knobs (runtime-tunable)

Beyond the per-member files, the inference layer exposes five knobs that mindX can adjust without code changes. All are env-overridable; defaults shown:

Env var	Default	Effect
`BOARDROOM_MAX_CONCURRENT`	8	semaphore limit during convene — how many soldier votes fly simultaneously. With cloud routing through one shared model (`gpt-oss:120b-cloud`), 8 is natural — every member can be in flight at once. With local routing, lower to match `OLLAMA_MAX_LOADED_MODELS`.
`BOARDROOM_NUM_CTX`	8192	context window in tokens per inference call. Ollama's default is 4096 which truncates long persona prompts (now 1.5–4.2 KB after `.prompt + .agent + .persona` composition) plus directive + JSON output. 8192 is safe; 16384 covers verbose CFO prompts.
`BOARDROOM_NUM_PREDICT`	2000	output token budget per vote. 2000 fits a JSON envelope plus ~1500 chars of reasoning.
`BOARDROOM_TEMPERATURE`	0.3	sampling temperature for vote inference. Boardroom decisions favour determinism.
`BOARDROOM_ROLLCALL_NUM_PREDICT`	120	output budget for the short ack prompt during roll call. ~one sentence.

VPS Ollama daemon settings (in /etc/systemd/system/ollama.service.d/concurrency.conf):

OLLAMA_MAX_LOADED_MODELS=3   # how many models stay resident
OLLAMA_NUM_PARALLEL=8        # concurrent requests per model (matches MAX_CONCURRENT)
OLLAMA_KEEP_ALIVE=24h        # don't unload between sessions

Read the active values:

curl https://mindx.pythai.net/insight/boardroom/roles?h=true | grep -A8 "adjustable LLM knobs"

Or programmatically:

from daio.governance.boardroom import boardroom_llm_knobs
boardroom_llm_knobs()
{'max_concurrent': 8, 'num_ctx': 8192, 'num_predict': 2000,
 'temperature': 0.3, 'rollcall_num_predict': 120,
 'supermajority_threshold': 0.666}

Override by setting the env var on the systemd unit (/etc/systemd/system/mindx.service.d/) or in .env, then restart mindx.service.

Editing a member's voice

To change how a member sounds:

Edit prompts/boardroom/{short_id}.prompt (the operational voice).
Edit agents/boardroom/{short_id}.persona (traits / beliefs — JSON).
Edit agents/boardroom/{short_id}.agent (operating contract — ALL_CAPS sections).
Restart mindx.service (drops the in-process card cache).
curl https://mindx.pythai.net/insight/boardroom/cards to confirm the new prompt is loaded.

Why three files?

.prompt = WHAT THE LLM SEES — the character voice. Changes when you tune how a member reasons or speaks.
.agent = WHAT THE MEMBER IS — the declared identity. Changes when the role itself is restructured (rare, a roster change).
.persona = HOW THE MEMBER THINKS — traits, beliefs, priorities. Changes when you adjust the cognitive style (often).

The three concerns belong to different teams and different cadences: prompts get tuned per-iteration, personas per-week, contracts per-quarter. Splitting them prevents a tuning change from accidentally rewriting the operating contract.

prompts/ — first-class .prompt files across mindX (boardroom + codephreak BDI + emergent + CLAIR + kaizen)
agents/boardroom/ — per-member .agent and .persona files
prompts/prompt.md — the .prompt language and encapsulation philosophy
BOARDROOM.md — the 13-seat war-council target spec (forward-looking)
PersonaAgent — the same persona schema applied across mindX
daio/governance/boardroom.py — _load_member_card, _query_soldier, roll_call

Referenced in this document

BOARDROOM

All Documents Document Index The Book of mindX Improvement Journal API Reference

Boardroom Members — three-file role architecture

The eight members

File formats

.prompt — full character voice (plain text)

.agent — operating contract (plain text, ALL_CAPS sections)

.persona — cognitive style (JSON)

Loading mechanism

Verification — operator-facing endpoints

GET /insight/boardroom/cards

POST /insight/boardroom/rollcall

Adjustable LLM knobs (runtime-tunable)

{'max_concurrent': 8, 'num_ctx': 8192, 'num_predict': 2000,

'temperature': 0.3, 'rollcall_num_predict': 120,

'supermajority_threshold': 0.666}