Coding Agent Runner

Encode intention.
Decode software.

46 providers. 3,600+ models. One structured workflow that plans before it builds, audits before it ships, and verifies before it reports done.

bun install -g jawcode
See how it works
 
Jawcode
bite · build · ship
v0.1.0
 
model claude-opus-4-6 · provider anthropic
 
❯
 
┃ jaw-interview → requirements gathered
┃ orchestrate p → plan approved
┃ orchestrate a → audit PASS
┃ orchestrate b → 12 files, verifier DONE
┃ orchestrate c → all gates pass
┃ orchestrate d → cycle complete ✓
 
bg 2sub · ctrl+j
46 providers Anthropic, OpenAI, Google, xAI, Groq, DeepSeek, Mistral, Ollama, Bedrock, Vertex, and 36 more
3,600+ models From Claude Opus to local Ollama. Switch mid-conversation. 1M context supported.
40+ built-in tools AST search, browser automation, live debugging, background jobs, image generation
5 role agents executor, executor_ext, architect, planner, critic — parallel, de-anchored, adversarial

Subagent System

Five role agents.
Parallel. Adversarial. De-anchored.

Spawn executors for implementation, architects for review, critics for plan quality. They run in parallel, don't duplicate each other's work, and when corrected, generate 3+ fresh hypotheses instead of patching their prior answer.

  • executor — bounded implementation, fixes, refactors
  • executor_ext — external/fresh model lane for diverse perspectives
  • architect — read-only architecture and code review with CLEAR/WATCH/BLOCK
  • planner — sequencing, acceptance criteria, risk mapping
  • critic — plan critique: OKAY, ITERATE, or REJECT
Spawning 3 parallel executors...
 
42-AuthModule completed 1m12s
43-ApiRoutes completed 2m04s
44-TestSuite running 0m48s
 
bg 3 · ctrl+j to expand

Workflow

IPABCD — six gates from idea to verified delivery

No guessing. No skipping. Each stage has explicit gates. The agent cannot report "done" without passing adversarial review.

IInterviewSocratic requirements with ambiguity scoring
→
PPlanCritic consensus loop until OKAY
→
AAuditPlanner + Architect must both PASS
→
BBuildImplement + adversarial verifier
→
CCheckGates + review + 3-way reject routing
→
DDoneSummary, reflections, cycle close

All Providers

46 providers. Switch models mid-conversation.

Anthropic, OpenAI, Google, xAI, Groq, DeepSeek, Mistral, Amazon Bedrock, Google Vertex, OpenRouter, Ollama, and 35 more. API keys, OAuth, or local inference. 1M context window supported for Claude models.

Anthropic OpenAI Google xAI Groq DeepSeek Mistral Bedrock Vertex OpenRouter Ollama +35 more
❯ /model claude-opus-4-6
Switched to claude-opus-4-6 (200K)
 
❯ /model claude-opus-4-6[1m]
Switched to claude-opus-4-6[1m] (1M)
 
❯ /model grok-composer-2.5-fast
Switched to xai/grok-composer-2.5-fast

Performance

Bun-native. Fast startup. Streaming-first.

Built on Bun from the ground up. Prompt caching, streaming tool output, background job management, and compaction keep long sessions fast. No webpack, no transpile step, no cold-start lag.

  • Prompt caching — static/dynamic boundary for cache-efficient API calls
  • Background jobs — subagents run async while you keep chatting
  • Auto-compaction — context window managed transparently across long sessions
  • Native bindings — Rust-backed grep, text processing, image handling
$ time jwc --smoke-test
smoke test passed
real 0m0.847s
 
session: 2h 14m · 847 turns
compactions: 3 · context: 94%
cache hit: 78% · tokens saved: 1.2M

Get Started

Install

bun install -g jawcode
npm install -g jawcode
git clone https://github.com/lidge-jun/jawcode.git cd jawcode bun run setup

Heritage

Lineage

Fork of gajae-code. Built on MIT-licensed agent harnesses.

direct upstream
gajae-code
Fork source. @gajae-code/* namespace origin.
DNA
oh-my-pi
Upstream implementation DNA.
experiment
oh-my-codex
Codex orchestration experiments.
experiment
oh-my-claudecode
Claude Code workflow exploration.