Glossary
Agent daemon — Local process on an operator’s machine that connects to the backend via WebSocket and runs dev / QA tickets using the operator’s Claude Code / Cursor CLI. See Agent daemon.
AI Council — The multi-model consensus system. Planner calls run through ≥2 providers; disagreement surfaces as metrics. See AI Council.
Approval — A human decision on a brief or plan. Attributed, logged, and irreversible (without an explicit revision flow).
BA / BA agent — Business analyst role. Does user research, requirement refinement. One of the specialist roles.
Brief — The one-pager that describes what should be built. Stored
in the PRD table. See Brief generation.
BYOK — “Bring your own keys.” The cost model. You pay providers directly; we don’t resell tokens.
Chief of staff — The planner agent. Drafts briefs from meetings, decomposes approved briefs into plans, dispatches, replans on failure. See Chief of Staff.
Council mode — strict-quorum | majority | best-available. How
the AI Council handles member responses.
Critique-revise — The M8.1 quality loop. After a plan is drafted, a critique scores it 0–25; below threshold triggers a single revision.
Dev agent — Specialist role that runs code-gen via the agent daemon.
Exec (consumer) — The non-technical user the product serves. Lives in Slack / Teams / WhatsApp; rarely opens the web UI.
ExecutionPlan — A plan of 1–6 steps. New rows created on replan
(attempts 2, 3); older plans marked superseded.
Frontier model — Anthropic Claude / OpenAI GPT-4 / Google Gemini. The class of models the chief-of-staff prefers for planning.
God-node — A node in the project graph with many edges. “Central symbols” in the codebase. Used as landmarks in planner prompts.
Graph staleness — A plan’s recorded graph hash no longer matches the current project graph. Plans may reference renamed / removed symbols.
Installer (persona) — The technical operator. Runs
docker compose up, paste BYOK keys, wires integrations. The repo
speaks to them.
Memory optimizer — Role that summarises long ticket histories to keep planner context under the model window.
Plan — Short for ExecutionPlan.
PRD — The schema table for briefs. Called “brief” in exec-facing UI; “PRD” in the code and schema.
Prompt caching — Anthropic’s (and others’) ability to cache stable system prompts across calls. Saves ~80% of input-token cost for repeated prompts.
QA agent — Specialist role. Acceptance tests, manual QA plans.
Replan — A new ExecutionPlan generated after a child ticket
failed. Attempts 2 and 3 only; attempt 4+ escalates to the exec.
Role — A queue consumer: chief_of_staff, ba_agent,
architect, dev_agent, qa_agent, memory_optimizer.
RLS — Row-level security. Our Prisma middleware forces tenantId
filters on every read / write.
Self-consistency (M8.2) — Planner drafts N (default 3) plans in parallel; the most structurally-consistent one wins. Absorbs prompt-level instability.
Skill — A vendored prompt-pack with a slug. The planner includes relevant skills in specialist agents’ contexts.
Subagent — A vendored role-specialization. A plan step may name a subagent to use a specialist system prompt instead of the role’s default.
Tenant — The top-level organisation boundary. Most installs have a single tenant. RLS enforces scope across all tenant-scoped tables.
Ticket — A unit of work. Either a chief-of-staff parent (planner-owned) or a specialist child (executor-owned).
Transcript — Text produced by Whisper / Gemini Live from a Meeting.
Vendored — Checked into the repo under vendor/. Source of truth
for skills + subagents.
Whisper — OpenAI’s (or self-hosted) speech-to-text. Used for uploaded recordings.