---
title: Pipeline Overview
description: Architecture of the NIKI multi-agent pipeline and artifact contracts.
sidebar:
  order: 1
  label: Pipeline Overview
---

## Multi-Agent Pipeline Overview

At the heart of NIKI is an autonomous pipeline where each stage is handled by a specialized LLM agent.

```mermaid
sequenceDiagram
    autonumber
    actor User
    participant P as ◈ Planner
    participant C as ⟠ Coder
    participant T as ◉ Tester
    participant R as ◆ Reviewer
    participant Git as Git Repo

    User->>P: niki run "Task Description"
    Note over P: Inspects workspace & files
    P->>C: TaskSpec (JSON)
    Note over C: Writes unified diff
    C->>T: CodeDiff (JSON)
    Note over T: Generates & executes tests in container
    T->>R: TestReport (JSON)
    Note over R: Audits correctness, quality, coverage
    alt Revision Needed
        R-->>C: ReviewFeedback (JSON)
        Note over C: Adjusts diff based on reviewer feedback
    else Approved
        R->>Git: Commit to niki/<id> branch
        Git-->>User: Report, Patch & Artifacts
    end
```

---

## Artifact Contract

Agents never exchange unstructured conversation logs. Every stage takes a validated JSON input and produces a validated JSON output:

| Producer | Consumer | Artifact | Schema / Purpose |
|---|---|---|---|
| **Planner** | Coder | `TaskSpec` | Problem summary, files to touch, approach, acceptance criteria, constraints |
| **Coder** | Tester, Sandbox | `CodeDiff` | Search/replace hunk edits, rationale, spec adherence verification |
| **Tester** | Reviewer | `TestReport` | Synthesized tests, test execution results, pass/fail status, coverage summary |
| **Reviewer** | Orchestrator, Coder | `ReviewVerdict` | Approved / RevisionNeeded / Rejected, quality scores, granular feedback |

---

## Topology Modes

In `niki.toml`, you can configure how tasks route through the pipeline:

* **`Auto` (Default):** Evaluates task complexity during the planning stage. Low-complexity, single-line or documentation fixes take a fast-path single-agent route, while multi-file or architectural changes run the full 4-agent pipeline.
* **`MultiAgent`:** Strictly forces all tasks through the full Planner → Coder → Tester → Reviewer sequence.
* **`SingleAgent`:** Runs Coder solo without formal Tester/Reviewer revision loops.