6.5 KiB
AirCoding
Deterministic Multi-Agent AI Coding System
Overview
AirCoding is a semi-automated development agent I built based on my own workflow. Its main goal is to solve Vibe Coding's core problems — poor code quality and high model requirements — through an engineering-driven approach. The aim is to use low-cost models concurrently to complete small to medium-sized projects.
This idea has just barely gotten off the ground. Due to time constraints, the Alpha release is a fork of opencode with direct modifications. It's still constrained by opencode's framework and far from the ideal design. Bugs are plenty — I'll fix them as I find them. It works well enough for now.
When I have more free time, I'll need to build an agent from scratch. The current framework imposes too many limitations.
Acknowledgments
Thanks to Player Xiao Zhang for calling me while I was on the toilet, which cleared my mind.
Thanks to Mr. Ran, Mr. Ze, Mr. Ding, and Mr. Zhe for enduring a 10-game losing streak in League of Legends while playing with me. This is the power of friendship and bonds — I'm going to find an Evolution Crystal to super-evolve as repayment.
Thanks in advance to Mr. Fan Xiaowen, who is currently researching Go for building agent frameworks. I plan to steal his work once he's done (doge).
The following is AI-generated.
## OOvveerrvviieeww
Core design principles:
- Tool whitelist as hard gate: Agent capabilities are enforced at the code level, not by prompt suggestion
- Deterministic scheduling first: Normal flow runs through a DAG state machine; LLM is consulted only for exceptions and edge cases
- Two-layer review: Worker self-verification + Reviewer Code-to-Design audit
- Evidence-gated debugging: cppcheck mandatory (C++ projects); must collect evidence before modifying code (DEBUG mode)
- Anti-fallback: Scheduler and Executor are forbidden from using "for now", "temporary solution", or any downgrade pattern to replace the design spec
- Mandatory Reviewer: Reviewer is not optional — every Worker output goes through a three-layer review (Code-to-Design → Static Analysis → Test Verification) before it can pass
Architecture
User → Main Agent (aircoding) → dispatches
│
├─→ Architect (architecture planner)
│ Outputs: plan.md + task-graph.json + ADR + C4 docs
│
└─→ Scheduler (scheduling engine)
│ coordinator_tick: deterministic DAG scheduler
│
├─→ Worker (executor / debugger)
│ EXECUTE: write code → compile → test → cppcheck
│ DEBUG: collect evidence → record → fix → verify
│
└─→ Reviewer (code reviewer)
Code-to-Design review against plan.md
Three mandatory layers:
1. Line-by-line Code-to-Design table
2. Static analysis (security / correctness / compliance)
3. Test / build verification
Agent Permission Matrix
| Agent | Allowed Tools | Denied Tools |
|---|---|---|
| aircoding (Main) | read, glob, grep, task, question, web, coordinator_status, coordinator_tick | write, edit, bash |
| Scheduler | read, glob, grep, task, coordinator_* | write, edit, bash |
| Worker | read, write, edit, bash, glob, grep | task |
| Architect | read, glob, grep, task, edit/write (.air/shared/plan/**) | bash, source file write |
| Reviewer | read, glob, grep | write, edit, bash, task |
Install
Binary (Linux x64)
# Download release
wget http://git.airlongdian.fun/admin/AirCoding/releases/download/0.1.0/AirCoding-Alpha-0.1.0-linux-x64.tar.gz
tar xzf AirCoding-Alpha-0.1.0-linux-x64.tar.gz
cd AirCoding-Alpha-0.1.0 && ./install.sh
aircoding --version # → 0.1.0
The binary installs to ~/.aircoding/ and does not conflict with opencode (~/.opencode/).
Build from Source
bun install # install dependencies
bun typecheck # type check (29 packages, all pass)
cd packages/opencode && OPENCODE_VERSION="0.1.0" OPENCODE_CHANNEL="aircoding" bun run script/build.ts --single --skip-embed-web-ui
Directory Convention
.air/shared/plan/plan.md # architecture plan (Architect output)
.air/shared/plan/task-graph.json # task graph (Architect output, Scheduler reads)
.air/shared/plan/docs/ADR-*.md # architecture decision records
.air/shared/plan/docs/c4/ # C4 model docs
.air/local/state/scheduler-state.json # scheduler state (realtime persistence)
.air/local/debug/debug-log.md # debug log
Design Docs
Detailed design documents are in docs/:
- Architecture Design MVP
- V2 Implementation Plan
- V2 Design (detailed)
- V1 Baseline
- Agent Constraints
- Integration Notes
Key Constraints (Iron Rules)
See CLAUDE.md for the full constraint specification. Summary:
- Scheduler: No fallback implementations. Must fully follow design specs. If the design doesn't cover a scenario, dispatch Architect to update the design first — never self-adjudicate in code.
- Executor (Worker): No "get it working first / fix later" downgrades. Compile → test → cppcheck all mandatory. Fallback language in output = auto-FAIL.
- Reviewer: Mandatory, not optional. "Tests pass" / "function exists" / "build passed" are not valid reasons to PASS. Must produce a line-by-line Code-to-Design table. Coordinator code gates enforce this deterministically.
- Architect loop prevention: Milestone review has a 2-per-phase / 5-global dispatch budget with
milestone_satisfiedgate to prevent infinite architect dispatch loops.
Acknowledgments
Forked from OpenCode v1.17.4.