Slot Manifests and Pipelines Reference

Bootstrap, input, and output slots are loaded from the concrete core's runtime.surface_root. With the default surface_root: agent, the directory contract is:

agent/bootstrap/<slot_id>/
  module.py
  slot.yaml
agent/input/<slot_id>/
  module.py
  slot.yaml
agent/output/<slot_id>/
  module.py
  slot.yaml

The slot id is the directory name. Slot metadata lives in slot.yaml; phase ordering lives in <surface_root>/pipelines.yaml.

agent/pipelines.yaml

The loader requires pipelines.yaml in runtime.surface_root:

schema_version: 1
bootstrap:
  serial:
    - session_context
input:
  serial:
    - base_input
  parallel: []
output:
  serial:
    - base_output
  parallel: []

schema_version must be 1. Supported phases are bootstrap, input, and output.

When adding a slot, edit the existing phase list. Do not replace the whole file unless you intend to rewrite all pipelines.

Pipeline Rules

Rule	Behavior
Unknown phase	Core load fails.
Unknown phase key	Core load fails.
Unknown slot id in a pipeline	Core load fails.
Duplicate slot id across phase directories	Core load fails.
Duplicate slot id inside one pipeline	Core load fails.
bootstrap.parallel	Core load fails.

Bootstrap supports only serial. Input and output support both serial and parallel.

Lane Semantics

Phase/lane	Semantics
bootstrap.serial	Runs once per session before the first turn.
input.serial	Runs before the provider call and can modify the current prompt.
input.parallel	Background input side effects; cannot modify the current prompt.
output.serial	Runs after the provider response and can write history or result data.
output.parallel	Background output side effects; cannot write session history or result data.

base_input and base_output are editable seed slots from the default core. They are not hidden host built-ins.

slot.yaml

Accepted fields are exactly:

entrypoint: module:process
description: "Adds current-turn context."
input_schema: {}
capabilities: []
timeout_seconds: null
failure_policy: soft
default_placement: pre_current_user
history_policy: persist

Field	Default	Meaning
entrypoint	module:process	Slot handler. Use module:function relative to the slot directory, or a core-root-relative Python file path plus function.
description	""	Description for inspection and docs.
input_schema	{}	Optional authored metadata.
capabilities	[]	Capabilities available to this slot through ctx.capability.require(...).
timeout_seconds	null	Loaded as metadata; the current slot runtime does not enforce it.
failure_policy	soft	soft logs and continues; hard raises the slot failure.
default_placement	pre_current_user	Default placement metadata for legacy context contribution shapes.
history_policy	persist	Default delivery history policy.

Unknown fields are rejected. Legacy aliases such as run and failure are not accepted.

Entrypoints

The common shape is:

entrypoint: module:process

def process(ctx):
    ...

Relative imports inside a slot directory are isolated per slot. Shared helper code for the default surface can live under agent/lib/.

Failure Policy

Use failure_policy: soft for optional behavior. A soft failure emits module failure events and the turn continues.

Use failure_policy: hard only when the turn cannot proceed without the slot, such as the seed base_input slot that writes the current user message.

History Policy

Valid values are:

• persist
• model_hidden
• transient

persist writes visible output into model-visible session history. model_hidden writes session history that is not included in later model context. transient delivers live output without writing assistant history.