md-first memory extraction framework for AI agents. Markdown is the single source of truth; SQLite holds state and LanceDB provides the rebuildable vector + BM25 + scalar index. The codebase follows a single-direction DDD layering (entrypoints -> service -> memory -> infra, with component / core / config cross-cutting) enforced by import-linter. Engineering surface: - Coding conventions in .claude/rules/ (path-scoped) and workflows in .claude/skills/ (/commit, /new-branch, /pr). - GitHub Actions CI runs make lint + test + integration; pre-commit mirrors the gates locally (ruff, hygiene hooks, gitlint commit-msg). - Commit messages follow Conventional Commits, enforced by gitlint. - make lint also enforces datetime two-zone discipline and OpenAPI drift.
3.5 KiB
PromptSlot
PromptSlot is the layer between the algorithm code (everalgo) and
the prompts it sends to LLMs. Algorithm code receives a PromptSlot
parameter; the project (EverOS) supplies defaults and lets operators
override.
Status (2026-05-07): the YAML loader is implemented; the higher- level
PromptSlotmodel + sandbox dry-run + three-layer overlay resolution arrive when the memory layer ships (see Stage 2).
Three-layer overlay
config/prompt_slots/<name>.yaml (Layer 1: defaults shipped with the package)
↓
~/.everos/prompt_slots/<name>.yaml (Layer 2: app-level override; per-deployment)
↓
runtime override (Layer 3: per-call override; e.g. "force model X")
Effective prompt = layer 3 wins → layer 2 → layer 1. Layer 1 is loaded eagerly at startup; layer 2 is loaded on first reference (lazy); layer 3 is supplied at the call site.
Loader
The category loader lives at
src/everos/component/config/loader.py
as YamlConfigLoader:
from pathlib import Path
from everos.component.config import YamlConfigLoader
loader = YamlConfigLoader(
root=Path("src/everos/config"),
categories={"prompt_slots": None}, # subdir == category name
)
# Reads <root>/prompt_slots/episode_extract.yaml → dict
slot = loader.find("prompt_slots", "episode_extract")
# Refresh after on-disk edits.
loader.refresh() # drop the entire cache
loader.refresh("prompt_slots") # drop one category
loader.refresh("prompt_slots", "episode_extract") # drop one entry
Top-level YAML is required to be a mapping; a list / scalar root
raises TypeError to fail-fast (loud, not silent).
YAML format (proposed; subject to change)
# config/prompt_slots/episode_extract.yaml
template: |
Extract a single episode from this conversation:
{{ memcell.text }}
variables:
memcell: input memcell
output_schema:
type: object
properties:
summary: { type: string }
participants: { type: array }
llm:
model: gpt-4o-mini
temperature: 0.3
max_tokens: 2000
validation:
test_cases:
- input: { memcell: { text: "Hi" } }
expected: { summary: "...", participants: [] }
When layer 2 supplies an override the loader will be re-pointed at
~/.everos/prompt_slots/; the runtime resolution logic (currently TBD)
sandbox-runs the merged slot before returning it.
Why YAML (not TOML)
Two reasons:
- Multiline templates — TOML's basic-string grammar fights
prompt content (no easy
{{ jinja }}variables, awkward escaping). YAML's literal block scalar (|) preserves prompts as-is. - Comment + reference ergonomics — operators frequently inherit slots, tweak a few keys, and leave inline notes. YAML is more forgiving for hand-editing.
The Pydantic Settings file (config/default.toml) stays TOML — it's
machine-managed and type-validated; YAML's flexibility costs more
than it pays for that case.
Why a separate loader (not Pydantic Settings)
Settings = one structured tree, validated at load time, tied to a single source of truth. PromptSlots = many separate templates discovered by name, layered per-deployment. They're different shapes; forcing one model on the other gets clunky.