narratio/docs/config.md

Configuration Reference

Purpose

Narratio resolves three YAML documents:

• pipeline.yml: pipeline/runtime settings
• campaign.yml: campaign identity and stable input defaults
• session.yml: session identity, metadata, and audio source selection

Discovery and Selection

pipeline.yml

When --config is omitted, search order is:

1. /usr/local/etc/narratio/pipeline.yml
2. /etc/narratio/pipeline.yml

campaign.yml

Selection rules:

• if --campaign-file is set, use that path;
• else if --campaign <id> is set, use {pipeline.campaigns.root}/{id}/campaign.yml;
• else use {pipeline.campaigns.root}/{pipeline.campaigns.default_campaign_id}/campaign.yml.

session.yml

When --session is omitted, local search order is:

1. /usr/local/etc/narratio/session.yml
2. /etc/narratio/session.yml

If local session discovery fails and a session_id is known, Narratio attempts remote session loading from:

• {root_prefix}/campaigns/{campaign}/sessions/{session_id}/session.yml

using configured object storage.

The downloaded remote session file is command-scoped: Narratio removes it after the command finishes and records only the remote object provenance alongside the durable copied session input.

Identity segments

Campaign IDs (campaign_id and default_campaign_id), session IDs, previous session IDs, and Narratio run IDs are opaque portable segments. They must use only ASCII letters, digits, ., _, and -; empty values, ./.., path separators, drive forms, whitespace, control characters, and non-ASCII text are rejected. Narratio does not trim or rewrite these values. Existing manifests or remote state with an unsafe legacy identity must be migrated before use.

Validation and Merge Rules

• YAML decode is strict (KnownFields(true)) and accepts exactly one document: unknown fields or trailing documents fail load.
• A pipeline file may explicitly import additive YAML fragments through the root-only composition.imports list. Imported files contribute fields to one logical pipeline document; they do not override fields supplied by the root or another import.
• A root pipeline may declare named profiles. Exactly one profile is selected by an option-aware caller or by composition.default_profile; a caller's explicit selection takes precedence. Declaring profiles without either form of selection is an error.
• Configured timeout and retry-delay durations must be positive. An omitted artifact timeout continues to inherit its configured Scriptorium timeout.
• Session files must be concrete; unresolved {{ ... }} placeholders fail load.
• Pipeline defaults are applied before validation.
• Campaign and session identities must agree.
• Required stable files (speakers_file, autocorrect_file, glossary_file, party_file) and the optional spell_catalog_file resolve from session overrides when provided, otherwise from campaign defaults. An empty or omitted session spell-catalog value inherits the campaign value.
• party_file is classified when pipeline and campaign configuration are combined. A versioned canonical party is campaign-owned, derives the players input internally, and forbids both a separate players_file and a session party_file override. An unversioned party remains a bounded legacy input and requires players_file; its normal campaign/session overrides continue to apply.
• Exactly one audio mode must be configured in session input:
  • local (audio_dir or audio_files), or
  • S3 (audio_s3.prefix).

Pipeline composition

Large pipeline configurations may be split into explicitly named fragments and may declare one overlay per selectable profile:

composition:
  imports:
    - config/storage.yml
    - config/integrations.yaml
  default_profile: production
  profiles:
    production:
      overlay: profiles/production.yml
    testing:
      overlay: profiles/testing.yml

campaigns:
  root: /usr/local/share/narratio/campaigns

Imports are resolved relative to the directory containing the root pipeline file and are loaded in declaration order. Narratio does not scan directories or infer fragments. Each import must be a confined regular .yml or .yaml file: absolute paths, traversal, symlinks, directories, duplicate files, and an import of the root pipeline itself are rejected. Only the root pipeline may contain composition; nested composition is rejected.

Composition is additive. A map may be extended by multiple files when every leaf is distinct, but a scalar, list, or map/list/scalar kind cannot be claimed more than once, even when the repeated values are identical. Conflict errors name the full field path and every source that claimed it. The assembled YAML is then decoded against the normal strict pipeline schema and defaults are applied once.

Profile names are case-sensitive, non-empty, trimmed, and cannot contain control characters. If profiles is present, it must contain at least one entry and every entry must contain only an overlay path. An explicit profile selection overrides default_profile; unknown and explicitly empty selections fail. Narratio never selects the first profile implicitly and does not read a profile selection from the environment.

Every declared overlay is resolved relative to the root pipeline directory and must satisfy the same confined regular-YAML-file rules as an import. Narratio parses every declared overlay even when it is not selected, then applies only the selected one. Maps merge recursively, overlay scalars replace base scalars, and overlay lists replace base lists completely. Explicit false, zero, empty lists, and empty maps remain meaningful. YAML null cannot delete a value, and kind changes are rejected. Profiles cannot inherit from or stack with other profiles, and overlays cannot import files or declare profiles.

After composition, Narratio strictly decodes the result, applies centralized defaults once, resolves ordinary paths, and computes a deterministic effective configuration digest. The digest represents the normalized, secret-free runtime pipeline mapping; it excludes composition declarations, source provenance, profile identity, and raw environment secret values. Equivalent effective mappings therefore have the same digest regardless of how fields are split among the root and imports.

An imported field has the same meaning it would have in a monolithic root pipeline. In particular, ordinary relative pipeline paths continue to resolve from the root pipeline directory, not from the importing fragment's directory.

Minimal Working Configuration

pipeline.yml

campaigns:
  root: /usr/local/share/narratio/campaigns
  default_campaign_id: sample-campaign
whisperx:
  transcribe_url: https://transcription.example.com/transcribe

campaign.yml

campaign_id: sample-campaign
inputs:
  speakers_file: ./speakers.yml
  autocorrect_file: ./autocorrect.yml
  glossary_file: ./glossary.yml
  players_file: ./players.yml
  party_file: ./party.yml

session.yml (local audio)

session_id: 2026-05-03
inputs:
  audio_dir: ./audio

Secrets Handling

• Do not place raw secrets in YAML.
• Use env var names in config (for example pipeline.audita.llm_api_key_env).
• Optionally load credential files from pipeline.secrets.env_dir. Each valid environment-variable filename supplies one value; trailing CR/LF is removed.
• An existing process environment value takes precedence over a credential file.
• Credential directories and files must not be symlinks and must be regular, bounded files (at most 8 KiB per value). On POSIX, provision the directory with no group/other access (normally 0700) and files with no group/other access (normally 0600).
• Commands that need storage/auth load filesystem secrets before constructing adapters.

Publish Configuration Summary

Publish rules live under pipeline.publish.

publish:
  enabled: true
  upload_run: true
  outputs:
    - source: narratio.transcript.final_trimmed
      dest: transcripts/final.trimmed.json
      required: true
    - source: narratio.transcript.final_markdown
      dest: transcripts/final.md
      required: true
    - source: narratio.transcript.final_trimmed_markdown
      dest: transcripts/final.trimmed.md
      required: true
    - source: narratio.artifact.session_recap
      dest: artifacts/session_recap.md
      required: true
  locks:
    - source: narratio.artifact.session_recap
      reason: manual post-publish edits

Rules:

• outputs[].source is required.
• outputs[].dest may be omitted when derivable from source.
• extraction sources require an explicit outputs[].dest and publish only when a rule names that source; the Notarius index and complete bundle are not publish sources.
• outputs[].required defaults to true.
• static locks (pipeline.publish.locks) merge with remote locks ({session_prefix}/locks.yml), with static locks taking precedence on duplicates.

Full Schema

Pipeline

Field	Type	Required	Default / Rule
composition.imports[]	list of strings	No	explicit additive pipeline fragments relative to the root pipeline directory; .yml or .yaml regular files only
composition.default_profile	string	Conditional	selected when profiles exist and no caller explicitly selects one; must name a declared profile
composition.profiles.<name>.overlay	string	Conditional	required for every declared profile; one confined .yml or .yaml overlay relative to the root pipeline directory
pipeline.workspace.root	string	No	/var/lib/narratio
pipeline.workspace.cleanup_after_publish	bool	No	false
pipeline.campaigns.root	string	No	/usr/local/share/narratio/campaigns
pipeline.campaigns.default_campaign_id	string	No	empty
pipeline.secrets.env_dir	string	No	empty
pipeline.storage.backend	string	No	local; supported values are local and s3 (case-insensitive)
pipeline.storage.s3.bucket	string	Conditional	required when backend is s3 and S3 session-audio or publish upload is enabled
pipeline.storage.s3.root_prefix	string	No	dnd
pipeline.storage.s3.region	string	No	empty
pipeline.storage.s3.endpoint	string	No	empty
pipeline.storage.s3.force_path_style	bool	No	false
pipeline.storage.s3.access_key_id_env	string	No	OBJECT_STORAGE_KEY_ID
pipeline.storage.s3.secret_access_key_env	string	No	OBJECT_STORAGE_KEY
pipeline.spool.root	string	No	/var/spool/narratio
pipeline.spool.delete_audio_after_publish	bool	No	false
pipeline.cache.root	string	No	/var/cache/narratio
pipeline.cache.s3_audio	bool	No	true
pipeline.publish.enabled	bool	No	true
pipeline.publish.upload_run	bool	No	true
pipeline.publish.outputs[]	list	No	defaults to final trimmed JSON plus final and final-trimmed Markdown outputs
pipeline.publish.outputs[].source	string	Yes (per rule)	must reference built-in or configured artifact source
pipeline.publish.outputs[].dest	string	Conditional	derived if omitted and source supports derivation
pipeline.publish.outputs[].required	bool	No	true
pipeline.publish.locks[]	list	No	empty
pipeline.publish.locks[].source	string	Yes (per lock)	must reference supported publish source
pipeline.publish.locks[].reason	string	No	empty
pipeline.whisperx.transcribe_url	string	Yes	absolute http or https URL
pipeline.whisperx.language	string	No	en
pipeline.whisperx.timeout	duration	No	30m
pipeline.whisperx.retries	int	No	3
pipeline.whisperx.retry_delay	duration	No	2s
pipeline.whisperx.concurrency	int	No	2
pipeline.seriatim.binary	string	No	seriatim
pipeline.seriatim.timeout	duration	No	10m
pipeline.seriatim.output_schema	string	No	seriatim-intermediate
pipeline.seriatim.coalesce_gap	float	No	3.0
pipeline.seriatim.report	bool	No	true
pipeline.seriatim.env.overlap_word_run_gap	float	No	unset
pipeline.seriatim.env.overlap_word_run_reorder_window	float	No	unset
pipeline.seriatim.env.backchannel_max_duration	float	No	unset
pipeline.seriatim.env.filler_max_duration	float	No	unset
pipeline.audita.binary	string	No	audita
pipeline.audita.timeout	duration	No	3h
pipeline.audita.llm_api_key_env	string	No	empty
pipeline.audita.modules[]	list[string]	No	empty
pipeline.audita.base_url	string	No	empty
pipeline.audita.model	string	No	empty
pipeline.audita.total_llm_concurrency	int	No	unset
pipeline.audita.proposal_llm_concurrency	int	No	unset
pipeline.audita.validation_model	string	No	empty
pipeline.audita.validation_llm_concurrency	int	No	unset
pipeline.audita.transcript_description	string	No	empty
pipeline.audita.config_path	string	No	empty
pipeline.audita.output_schema	string	No	empty
pipeline.audita.work_dir_retention	string	No	empty
pipeline.audita.report	bool	No	true
pipeline.normalize.output_path	string	No	transcripts/final.json
pipeline.normalize.output_schema	string	No	seriatim-intermediate
pipeline.normalize.report	bool	No	true
pipeline.trim.enabled	bool	No	true
pipeline.trim.output_path	string	No	transcripts/final.trimmed.json
pipeline.trim.bounds.prompt_id	string	No	dnd.session_bounds
pipeline.trim.bounds.profile_id	string	No	empty
pipeline.trim.bounds.transcript_input_name	string	No	transcript
pipeline.trim.bounds.output_path	string	No	artifacts/session_bounds.json
pipeline.trim.bounds.timeout	duration	No	10m
pipeline.trim.bounds.render_debug	bool	No	false
pipeline.trim.bounds.render_output_path	string	Conditional	required when render_debug is true
pipeline.trim.seriatim.report	bool	No	false
pipeline.notarius.enabled	bool	No	false
pipeline.notarius.binary	string	No	notarius
pipeline.notarius.config_path	string	Conditional	required when enabled; relative paths resolve from the pipeline file directory
pipeline.notarius.pipeline_id	string	Conditional	required when enabled
pipeline.notarius.timeout	duration	No	3h; must be positive
pipeline.notarius.working_directory	string	No	directory containing resolved config_path; relative paths resolve from the pipeline file directory
pipeline.notarius.references	map[string]string	No	empty; maps normalized Notarius selectors to supported prepared Narratio source IDs; maximum 256 entries
pipeline.notarius.outputs	map	Conditional	at least one entry when enabled
pipeline.render.enabled	bool	No	true
pipeline.render.format	string	No	markdown (only supported value)
pipeline.render.title	string	No	empty (falls back to session.title when set)
pipeline.render.include_timestamps	bool	No	true
pipeline.render.include_segment_ids	bool	No	true
pipeline.render.include_metadata	bool	No	false
pipeline.scriptorium.binary	string	No	scriptorium
pipeline.scriptorium.config_path	string	No	empty
pipeline.scriptorium.timeout	duration	No	10m
pipeline.scriptorium.render_debug	bool	No	false
pipeline.scriptorium.artifacts	map	No	empty
pipeline.scriptorium.artifact_families	map	No	empty; expands one ordinary artifact per canonical party character
pipeline.notification.mode	string	No	noop; the only supported notification mode until a provider is implemented

Notarius Reference Bindings

pipeline.notarius.references maps a Notarius CLI selector to a prepared Narratio source, not to a filesystem path:

notarius:
  references:
    glossary: narratio.input.glossary
    party: narratio.input.party
    players: narratio.input.players
    spell_catalog: narratio.input.spell_catalog

Supported sources are narratio.input.party, narratio.input.players, narratio.input.glossary, and narratio.input.spell_catalog. Each map entry is required by its presence: omit a binding when the selected Notarius pipeline does not need it. A spell-catalog binding additionally requires an effective campaign or session spell_catalog_file.

Selectors accept Notarius's slot, chunk.slot, lane.slot, lane.extract.slot, lane.merge.slot, and lane.normalize.slot forms. Narratio trims whitespace around selectors and their dot-separated components, rejects empty components and =, rejects duplicate normalized selectors, and limits the map to 256 entries. It validates only selector structure and the prepared source vocabulary; Notarius owns target-slot declarations and media compatibility.

Before extraction, Narratio resolves every binding from the current prepared session manifest and streams it into a verified invocation-local snapshot whose absolute path is passed to Notarius. Missing, unsafe, empty, changed-during-copy, or checksum-inconsistent prepared evidence fails with guidance to force prepare. Bindings are sorted by normalized selector and are part of extraction fingerprint and resume identity. See the Notarius integration contract for the subprocess boundary and the complete example for a copyable configuration.

Notarius Output Entries

For each pipeline.notarius.outputs.<name>:

Field	Type	Required	Rule
lane_id	string	Yes	unique Notarius lane ID
media_type	string	Yes	exact accepted descriptor media type
schema_id	string	Yes	exact accepted descriptor schema ID
schema_version	string	Yes	exact accepted descriptor schema version
module_key	string	No	exact accepted module key when set

Output names must match ^[a-z][a-z0-9_]*$ and become selectable sources named narratio.extraction.<name>. Lane IDs must be unique. Every declared output is required from a successful Notarius result; a missing, rejected, duplicate, or contract-incompatible lane fails extraction. See the complete maintained example for the current ten-lane D&D mapping and the Notarius contract for compatibility ownership.

Scriptorium Artifact Entries

For each pipeline.scriptorium.artifacts.<name>:

Field	Type	Required	Rule
enabled	bool	No	false if omitted
depends_on[]	list[string]	No	must reference configured artifact keys; no self-reference; configured graph must be acyclic
render_debug	bool	No	per-artifact override
prompt_id	string	Conditional	required when artifact is enabled
profile_id	string	No	empty
output_path	string	Conditional	required when enabled; also required when referenced by publish/output/input rules
timeout	duration	No	artifact override
inputs	map	No	input key names must be non-empty
vars	map	No	values must be string or bool; session_id is reserved and overwritten by Narratio

Narratio adds session_id=narratio-session-<session_id> to every Scriptorium request for sticky upstream LLM routing. If an artifact config sets vars.session_id, Narratio replaces that value before invoking Scriptorium. Use a different variable name if a prompt needs the raw Narratio session ID as content.

Without --artifacts, analyze executes enabled configured artifacts. With an explicit --artifacts list, the exact named configured artifacts are the one-invocation targets even if their enabled values are false. Analyze closes those targets over depends_on: a current prerequisite is reused, while a stale, missing, failed, or legacy prerequisite is rebuilt before its dependent. Unrelated artifacts are not executed. Named targets and any prerequisite that may require rebuilding must therefore have valid executable fields. This override affects analyze planning only; publish uses the list only to filter configured narratio.artifact.<name> output rules.

For each artifact input pipeline.scriptorium.artifacts.<name>.inputs.<input_name>:

Field	Type	Required	Rule
source	string	Yes	built-in runtime source, prepared input source, narratio.extraction.<name>, narratio.artifact.<name>, or narratio.previous_session.artifact.<name>
required	bool	No	optional input requirement

artifact and path are obsolete and rejected by strict configuration loading. Use the canonical source identifier to select the input; Narratio does not provide adapter-specific input passthrough fields.

Scriptorium Artifact Families

pipeline.scriptorium.artifact_families declares a shared artifact template for every canonical campaign character. Configuration resolution expands each family into ordinary pipeline.scriptorium.artifacts entries before analyze planning or Scriptorium invocation. A legacy party cannot be used for a family.

For each pipeline.scriptorium.artifact_families.<name>:

Field	Type	Required	Rule
enabled, prompt_id, profile_id, timeout, render_debug, depends_on, inputs, vars	ordinary artifact fields	No	copied to each generated artifact under the corresponding ordinary rules
for_each	string	Yes	exactly party.characters
output_path_pattern	string	Yes	safe path beneath artifacts/ with exactly one {character_id} token and no other brace syntax
member_vars	map	No	maps an ordinary Scriptorium variable name to a supported canonical character selector
member_dependencies	list	No	unique family keys; each generated member depends on the corresponding generated member of each listed family
publish	map	No	typed family publish metadata (enabled, required, dest_pattern); it does not yet create publish outputs

Generated keys are <family>_<character_id> and generated output paths must not collide with explicit artifacts or another generated artifact. Families expand even when disabled; normal analyze selection still omits disabled artifacts unless they are explicitly selected by their concrete key.

Supported member_vars selectors are character_id, player.name, character.name, character.class_summary, and character.alias_summary. Their resolved values are strings. A member variable may not reuse a static vars name; session_id remains owned and overwritten by Narratio as for any other Scriptorium artifact.

Within a family only, an input source may use narratio.member_artifact.<family>. The referenced family must be named in that family's member_dependencies; resolution rewrites the source to the corresponding ordinary narratio.artifact.<family>_<character_id> source. This syntax is rejected in explicit artifacts and never reaches runtime stages or Scriptorium.

Notifications

Narratio currently supports only notification.mode: noop, which is also the default when the section is omitted. The notify stage performs no delivery in this mode. Backend, recipient, timeout, and other provider settings are rejected by strict configuration loading until Narratio has a provider integration.

Campaign

Field	Type	Required	Notes
campaign_id	string	Yes	canonical opaque campaign identity
session_template_file	string	No	used by session init when set
inputs.speakers_file	string	Yes	stable input default
inputs.autocorrect_file	string	Yes	stable input default
inputs.glossary_file	string	Yes	stable input default
inputs.players_file	string	Conditional	required only with an unversioned legacy party_file; forbidden for a canonical party
inputs.party_file	string	Yes	stable campaign party source; relative paths resolve from campaign.yml
inputs.spell_catalog_file	string	No	optional spell-catalog overlay default; required when a Notarius reference selects narratio.input.spell_catalog

Session

Field	Type	Required in session file	Notes
session_id	string	Yes	opaque identity; must match CLI session target when provided
previous_session_id	string	No	opaque identity; must not equal session_id
campaign	string	No	opaque identity; filled from campaign_id during resolve if omitted
date	string	No	metadata
title	string	No	metadata
inputs.speakers_file	string	No	overrides campaign stable input
inputs.autocorrect_file	string	No	overrides campaign stable input
inputs.glossary_file	string	No	overrides campaign stable input
inputs.players_file	string	No	legacy-party override; forbidden for a canonical party
inputs.party_file	string	No	legacy-party override; forbidden for a canonical campaign party
inputs.spell_catalog_file	string	No	overrides the optional campaign spell catalog; empty or omitted inherits the campaign value
inputs.audio_dir	string	Conditional	local audio mode
inputs.audio_files[]	list[string]	Conditional	local audio mode
inputs.audio_s3.prefix	string	Conditional	S3 audio mode

Audio rules:

• configure local mode (audio_dir or audio_files) or S3 mode (audio_s3.prefix), not both.
• audio_s3 requires pipeline.storage.backend: s3 and a configured S3 bucket.

Storage backend selection

local is the default and disables remote object-store operations. Configure s3 explicitly before supplying storage.s3; a populated S3 block does not select a backend on its own. Unknown backend names and an S3 block paired with local are rejected during configuration validation.

Previous-session expectation

previous_session_id is optional in a session file. When a command supplies --previous-session-id, however, the session file must contain the same value; an omitted or different value is rejected before the command performs work.

Maintained Examples

See the maintained examples index for complete pipeline, campaign, session, template, and input fixtures. Keep complete copyable files there rather than duplicating them in this reference.