Amadeus Replay

AI-DLC Session Replay

Purpose

Turn a workflow's audit trail and artefacts into a readable story: what was decided, in what order, and why. For async review, for stakeholders who weren't present, or as a post-session record. Not a raw log dump.

Classification

Read-only. This skill renders the narrative to the terminal and writes no file. It never advances the workflow stage pointer and never emits an audit event.

The counting rule

All counts and aggregates — number of stages, per-phase breakdown, duration, approved/failed/pending tallies, learnings captured — come from the tool, not from eyeballing files:

bun .codex/tools/amadeus-runtime.ts summary --json

The narrative prose (what happened, key decisions, reasoning) is yours to synthesise from audit.md and the artefacts. The skeleton numbers are the tool's. Never hand-count stages or artefacts when the tool already reports the figure.

Steps

Step 1: Read the aggregates

Run bun .codex/tools/amadeus-runtime.ts summary --json.

If it exits non-zero (no runtime-graph.json yet), print:

No session data yet — start a workflow with /amadeus before running
/amadeus-replay.

and STOP. Otherwise keep the parsed JSON; you'll cite its fields for every number in the report.

Step 2: Read the narrative sources

• amadeus-docs/audit.md — the full event trail (for ordering and the

reasoning behind each decision).

• amadeus-docs/amadeus-state.md — the active-stage cursor.
• The artefacts under amadeus-docs/// — for what each

stage produced.

These are your sources for prose. Do not derive counts from them when Step 1's JSON already carries the count.

Step 3: Render the replay

Print the narrative to the terminal in this shape (write no file):

# Session Replay
**Workflow**: {summary.workflow_id}
**Scope**: {summary.scope}
**Duration**: {summary.duration_minutes} min   (or "in progress")
**Stages**: {summary.stages.approved} approved / {summary.stages.total} total

## Executive Summary
{3-5 sentences: what was built or decided, key choices, constraints, outcome}

## Timeline
{For each phase in summary.by_phase, in workflow order:}

### {Phase} Phase  —  {by_phase[phase].approved}/{by_phase[phase].total} stages approved

#### {Stage Name}
**What happened**: {1-2 sentences from the audit trail}
**Key decisions**: {bullets, with reasoning drawn from audit.md}
**Artefacts produced**: {list with one-line descriptions}

{...repeat per stage that executed...}

## Decisions Register Summary
{Table: decision | alternatives considered | chosen option | rationale}

## Learnings Captured
From orchestrator: {summary.learnings.from_orchestrator}
From user additions: {summary.learnings.from_user_addition}
{Then narrate the notable ones from the stage memory.md diaries.}

## What's Next
{Outstanding open threads from the last audit entries / open questions}

Step 4: Offer adjustments

You may offer: "Want a different tone (more technical / more executive) or an added section?" — re-render to the terminal if asked. Do not loop on it; one offer is enough. If the user wants the replay saved, point them at /amadeus-outcomes-pack (the skill that writes a file) rather than writing one here.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: amadeus-dlc
• Source: amadeus-dlc/amadeus
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.