Imported from sailadachetansurya/Emily (src/pipeline/AGENTS.md). Install upstream with npx skills add sailadachetansurya/Emily --skill pipeline. Copyright stays with the author.

Pipeline Package

Purpose

Core emotive AI pipeline: modular stages wired through an orchestrator, producing emotionally attuned responses via deterministic policy control.

Ownership

All code under src/pipeline/ belongs to this doc.

Local Contracts

• Package entry: pipeline (installed via pyproject.toml with package-dir = {"" = "src"})
• Stage protocol interfaces live in contracts/interfaces.py — every stage must satisfy its Protocol
• Shared data models live in contracts/models.py — immutable dataclasses, no side effects
• Orchestrator (orchestrator/pipeline.py) owns stage wiring; individual stages must not import each other across stage boundaries
• Config loading (config/runtime.py) validates against a fixed schema; unknown fields are rejected
• Telemetry (telemetry/recorder.py) emits JSONL; stages do not write telemetry directly

Work Guidance

• Prefer adding new stages as new subdirectories under stages/ with their own __init__.py exporting public API
• New models or interfaces go in contracts/; keep contracts/ free of implementation logic
• Backward-compatible aliases use the pattern class NewName(OldName): """Backward-compatible alias."""
• When the pipeline orchestrator adds a new stage, update the trace list and the __init__ wiring together

Verification

• python -m pytest tests/ must pass (uses PYTHONPATH=src)

Child DOX Index

• stages/emotion_engine/ — Emotion perception, NLP training, dataset preparation
• stages/reasoning_loop/ — Internal reasoning/self-critique loop with activation gating
• stages/input_gateway/ — Input normalization (single file, no child doc needed)
• stages/memory_manager/ — Factual and emotional memory stores (single file each, no child doc needed)
• stages/policy_engine/ — Deterministic policy selection (single file, no child doc needed)
• stages/prompt_builder/ — Prompt assembly (single file, no child doc needed)
• stages/safety_processor/ — Output pruning and speech-act validation (single file, no child doc needed)
• stages/llm_client/ — Pluggable LLM backends (Ollama, llama.cpp) behind abstract LLMClient protocol
• config/ — Runtime configuration and validation (single file + errors, no child doc needed)
• contracts/ — Shared models and protocol interfaces (no child doc needed)
• evaluation/ — Evaluation harness (single file, no child doc needed)
• orchestrator/ — Pipeline wiring and runner (no child doc needed)
• telemetry/ — JSONL telemetry sink (single file, no child doc needed)