tai/docs/ARCHITECTURE.md

Architecture

This document describes tai's current runtime architecture, module responsibilities, and data flow.

High-Level Flow

1. User runs tai run with issue text and target host settings.
2. CLI validates input and opens a shared SSH session.
3. Probe and collection run against a read-only command plan.
4. Collection output is converted into diagnostic chunks.
5. Optional RAG retrieval selects top-k chunks per question.
6. Optional runbook retrieval selects top-k runbook chunks from ChromaDB.
7. Prompt builder composes system + user message.
8. AI completion returns analysis.
9. Guardrails validate response quality signals.
10. Optional session logger writes JSONL events.

Module Layout

• src/tai/cli.py
  • Command definitions (run, runbooks sync/list/add)
  • Orchestration across SSH, collection, RAG, prompts, AI, and logging
• src/tai/input_parser.py
  • User input validation and request normalization
• src/tai/models.py
  • Core dataclasses (TroubleshootRequest)
• src/tai/ssh_client.py
  • SSH invocation
  • Read-only command policy validation
  • Probe and command execution helpers
• src/tai/plan.py
  • Issue keyword/service extraction
  • Command plan generation
  • Service/subsystem presence probes (unit files, binaries)
• src/tai/collectors.py
  • Executes command plans and builds CollectionReport
• src/tai/rag_retriever.py
  • Command-output chunking
  • Embedding wrapper structures
  • Similarity retrieval and scoring
• src/tai/runbook_store.py
  • Persistent ChromaDB runbook indexing and querying
• src/tai/chroma_telemetry.py
  • No-op telemetry adapter for Chroma local usage
• src/tai/prompt_builder.py
  • Prompt assembly for full-context and retrieved-context paths
• src/tai/ai_client.py
  • OpenAI-compatible completions and embeddings client
• src/tai/ai_guardrails.py
  • Lightweight response guardrails and warnings
• src/tai/session_log.py
  • Optional JSONL event logging

Data Stores

• Runbook store (Tier 2): local ChromaDB path, default ~/.tai/runbooks
• Session logs: optional JSONL file configured by --log-file

Retrieval Layers

• Tier 1 (implemented): in-memory semantic retrieval over diagnostic chunks
• Tier 2 (implemented): persistent semantic retrieval over runbook corpus
• Tier 3 (implemented core): persistent retrieval over prior sessions (dedicated UX commands pending)

Safety Boundaries

Read-only policy is enforced before each remote command execution.

• Allowed command families are explicitly enumerated.
• Shell composition operators are blocked.
• Commands that fail execution are recorded and surfaced to the model as non-evidence.

Failure and Fallback Behavior

• If RAG indexing fails, analysis falls back to full-context prompts.
• If runbook store is unavailable, analysis proceeds without runbook context.
• If AI call fails, CLI exits with non-zero status and displays an error.

Test Coverage Highlights

• Planner behavior and service detection
• Prompt formatting and guardrail-sensitive messaging
• CLI command behavior and interactive loop controls
• Runbook store parsing/index/query behavior (with mocked Chroma)
• SSH policy validation and command execution contract