narratio/docs/s3-audio-input.md

S3 Audio Input

This document describes implemented S3 audio input behavior in prepare.

Scope

Implemented:

• prepare can acquire source audio from S3 when session.inputs.audio_s3.prefix is configured.
• object listing and download go through the storage backend abstraction.
• tests use fake storage; no live S3 service is required for test runs.

Not implemented:

• archive uploads (run uploads, promotion uploads, current pointer writes)
• spool cleanup/deletion behavior
• uploads of failed runs

Required Configuration

pipeline.yml:

• storage.s3.bucket must be set when S3 audio input is used.
• storage.s3.root_prefix defaults to dnd.
• spool.root defaults to /var/spool/narratio.

session.yml:

• configure session.campaign and session.session_id.
• configure session.inputs.audio_s3.prefix for S3 audio input.
• do not configure inputs.audio_dir or inputs.audio_files at the same time as inputs.audio_s3.

Prefix Shape

Session S3 root:

{root_prefix}/campaigns/{campaign}/sessions/{session_id}/

Audio prefix:

{session_root}/{audio_s3.prefix}

Example:

dnd/campaigns/forsaken/sessions/2026-04-19/audio/

Audio files must already exist in S3 before running Narratio.

Prepare Behavior

When inputs.audio_s3.prefix is configured, prepare:

1. lists objects under the computed S3 audio prefix
2. filters to .flac objects
3. fails when no .flac objects are found
4. downloads selected objects to spool audio:
   • {spool.root}/{campaign}/{session_id}/{run_id}/audio/
5. materializes audio into workdir audio:
   • {workspace.root}/work/{campaign}/{session_id}/{run_id}/audio/
6. records input provenance in the manifest (bucket, key, metadata, local paths, checksum)

Notes:

• .flac filtering is case-insensitive.
• ETag is recorded as provider metadata only and is not treated as a checksum.

Local Audio Development

Local audio workflows remain supported:

• inputs.audio_dir
• inputs.audio_files

These options are mutually exclusive with inputs.audio_s3.

Archive Boundary

Current archive behavior remains unimplemented in this area:

• no upload of source audio by default
• no upload of failed runs
• no current/manifest.json or current/run_id.txt writes