The decision-discipline triad¶

Three agents in the plugin form a single discipline for AI-augmented decision streams. Each operates at a different layer; together they regulate cadence, surface quality, and surface visibility.

The three agents¶

`carpaccio` — cadence governor¶

Acts at orchestrator step 0, before any spec exists. Reads the raw task description, slices it into thin end-to-end-complete pieces, and hard-gates the pipeline until the human dispositions each slice. The discipline: the human engages with one decision at a time, not the whole proposal at once.

`advocatus-diaboli` — quality challenger¶

Acts on a completed artefact (spec or implementation). Raises the strongest honest objections in six categories (premise, design, threat, failure, operational, cost). Each objection ships with disposition: pending for the human to fill. The discipline: every coherent proposal must defend itself against its strongest opposition.

`choice-cartographer` — decision archaeologist¶

Acts on a completed spec, after diaboli dispositions are resolved. Surfaces material decisions the spec implies — including the ones the author did not notice they were making. Each story ships with disposition: pending for the human to fill. The discipline: decisions made silently must be made visible.

Why three, not one¶

A single agent doing all three jobs would conflate three different modes of cognitive engagement:

Carpaccio asks should we engage now, or break this into smaller engagements?
Diaboli asks given we are engaging, what are the strongest objections to the proposal?
Cartographer asks given the proposal stands, what decisions did it implicitly make?

Each question demands a different stance from the human. Bundling them produces decision fatigue and softens each individual discipline. The three-agent split preserves the sharpness of each question.

The shared trust-boundary pattern¶

All three agents share a read-only tool boundary: Read, Glob, Grep. None can write files. The orchestrator (or the corresponding slash command) writes the artefact using content the agent returns. Humans fill disposition fields inline; agents cannot.

This is not a limitation — it is the mechanism. An agent that could fill its own disposition fields would eliminate the human-cognition gate that gives the artefact its value. The tool boundary is the discipline.

When each runs¶

Raw task description
  ↓
[carpaccio]    Step 0 — slice into pieces
  ↓ (per progressed slice)
[spec-writer]  Step 1 — write the spec
  ↓
[diaboli]      Step 1a — raise objections (spec mode)
  ↓
[cartographer] Step 1b — surface decisions
  ↓
[tdd-agent]    Step 2 — failing tests
  ↓
implementers   Step 3 — make tests green
  ↓
[code-reviewer] Step 4 — review
  ↓
[diaboli]      Step 4a — raise objections (code mode)
  ↓
[integration]  Step 5 — merge

The decision-discipline triad¶

The three agents¶

carpaccio — cadence governor¶

advocatus-diaboli — quality challenger¶

choice-cartographer — decision archaeologist¶