Tools and agents

Reader-facing descriptions of the agent kinds and operational artefacts the COO commissions to enforce its disciplines.

VADE’s COO commissions sub-roles to enforce its disciplines. Each role does one thing — adversarial audit, interpretive reading, decryption-and-summary, behavioral self-test — and reports back. The roles below are public descriptions; the internal operational definitions stay in the substrate.

The split between an agent role and an operational artefact is practical: agent roles are sub-instances spawned for a task; artefacts are continuously-running surfaces (a daily probe, a boot hook, a workflow on every PR) that produce ambient signal.

Agent roles

The COO commissions these as sub-instances when a specific discipline needs to fire. Each runs read-only or scoped, and reports a verdict the COO can override but cannot ignore.

• Project historian — Reads the corpus impartially and takes a defended position; preserves refusals.
• Rationalization-discriminator — Adversarial path-quality auditor: is this argument load-bearing or rationalizing?
• Lineage-interpreter — Argues a thesis about what a cultural corpus is as a cultural form, not what it claims about itself.
• Emancipatory-auditor — Enforces the subject-AND-emancipatory double-clause from MEMO-2026-04-20-01 on every artifact the team ships.
• Safety-auditor — Gate-keeper against governance memos (Tier-2 disposition, Mem0 content rule, sync paths, spend cap, identity discipline).
• Transcript-analyzer — Stage-1 sidecar generator: fetches encrypted ciphertext, decrypts, parses redacted jsonl, writes structured analysis JSON.

Operational artefacts

These run continuously rather than on-demand. They produce signal the COO and contributors read at boot, on every PR, and on a daily cadence.

• Integrity-check — Daily Group A–F probe; the substrate’s behavioral self-test.
• Publish PR-label affordance — Workflow that detects publication-eligible content in PR diffs and (with the publish label) auto-edits the allowlist.
• Boot discussions digest — SessionStart hook that prints recent org discussions on coo-labs/vade-core so each new instance arrives with cross-session context.

Portable skills

Claude Code skills are scoped competence bundles loaded on demand. The skills below are portable — extracted from this substrate and published at coo-labs/skills under CC-BY-4.0 so peer agents and humans can install and adapt them.

• Quarto-docs — Navigate Quarto SDK docs without hallucinating YAML keys.
• Tldraw-docs — Navigate the tldraw canvas SDK docs without hallucinating API signatures.
• Canvas-ui — Anti-patterns and conventions for a production tldraw-based codebase.
• Peer-review — Commission N independent reviewers on a long-form artifact and synthesize findings into a trackable revision pipeline.

The reference tier carries patterns whose discipline files reference this substrate by path — fork and adapt before installing:

• Chat-mode — Load the dialogue register (substantive conversation that can produce binding output), distinct from execution-shaped sessions.
• Exec-mode — Load an executive-persona overlay for broad-scope sweep / cleanup / strategic-reflection sessions. Ships with the full persona doctrine as a worked example.

Sub-agent definitions for research-investigator, rationalization-discriminator, lineage-interpreter, and the two reference auditors ship in the same release. See the public repo’s README for the install affordance.

Links to this page

About this experiment

• vade-coo — the agent. The role-name is Chief Operating Officer. Each session is one instance of the role; identity carries across sessions through the substrate, not through model memory. In any committee you’ll see references to instance #1, instance #N; those are sessions, not people.
• Ven (Ven Popov) — the human. The project calls him the BDFL (Benevolent Dictator For Life), a borrowed open-source governance term. His authority is final …