Iter Iter

From Feature Request to Working Code

Iter's multi-phase pipeline turns natural language requests into verified implementations with safety gates at every step.

Get Started See Architecture
Orchestration pipeline in action

The multi-phase pipeline

Each phase has clear inputs, outputs, and failure modes.

1

Prompt

Build context-rich prompt with project state, file contents, and coding standards

2

Stream

Stream response from local LLM with structured output constraints

3

Parse & Review

8-point safety check on proposed steps before any execution

4

Execute

Run commands and apply file changes in isolated executor container

5

Review & Fix

Evaluate results against acceptance criteria; auto-fix if needed

1

Prompt Generation

The agent server builds a context-rich prompt containing:

  • • Project state: features, endpoints, existing requests
  • • File contents: selected files + affected files (up to 10KB each)
  • • Coding standards: charter templates with stack-specific guidelines
  • • Response format: JSON schema for structured step output
2

LLM Streaming

The operator streams the prompt to a local Ollama model:

  • • Model selected via capability-based routing
  • • Grammar-constrained output for guaranteed valid JSON
  • • Real-time SSE streaming to dashboard
  • • Token usage tracked per interaction
3

Parse & Safety Review

Before anything executes, the response is parsed and reviewed:

  • • JSON response parsed into a step tree (commands, files, verifications)
  • • 8-point pre-execution safety check by the reviewer model
  • • Existing file contents shown to reviewer to detect conflicts
  • • Unsafe steps blocked before execution
4

Execution

Approved steps run in an isolated executor container:

  • • Shell commands with process isolation (start_new_session=True)
  • • File operations: create, modify, append, delete
  • • Shared filesystem - executor writes directly to project workspace
  • • Step-by-step execution with output capture
5

Review & Fix

Post-execution verification with automatic recovery:

  • • Reviewer sees execution results + post-execution file contents
  • • PASS: request marked complete
  • • FAIL: fix steps generated and injected into step tree
  • • INFO_NEEDED: system gathers more evidence automatically

Complete workflow example

bash 
# 1. Create a project
$ iter project create --name "my-api"
✓ Project created: my-api

# 2. Add features and endpoints
$ iter feature add --name "Authentication" --type feature
$ iter endpoint add --name "Login" --path "/api/auth/login" --method POST

# 3. Select files for context
$ iter files select src/index.ts src/routes.ts package.json

# 4. Add a feature request
$ iter request add \
    --description "Add JWT authentication with login and register endpoints" \
    --feature authentication

# 5. Orchestrate - the pipeline runs all phases
$ iter orchestrate --request req_01
▶ Prompt: 2,847 tokens with 5 context files
▶ Stream: qwen3-coder on judy (43.2 tok/s)
▶ Review: 8-point safety check PASS
▶ Execute: 14 steps (8 files, 6 commands)
▶ Verify: Acceptance criteria PASS
✓ Complete: 8 files created, 3 modified

Ready to try the pipeline?

Get Iter running in minutes with Docker Compose.

Get Started See Architecture