convez.ai
AgenteDesenvolvimento WebAvançado
zdypro888

AIBridge

zdypro888/aibridge

Revisão de código colaborativa entre dois agentes IA com ping-pong automático

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent que implementa revisão de código mútua entre dois agentes IA (Claude e Codex) em sessões PTY interativas. Eles se revezam editando e auditando mudanças no repositório git até convergência, com dashboard web para acompanhamento e controle em tempo real.

Instalação

$npx convez-ai zdypro888/aibridge

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validação automática de código com múltiplos modelos de IA
  • Ciclos iterativos de edição e auditoria sem intervenção manual
  • Desenvolvimento assistido com feedback cruzado entre agentes

Veredito

Solução técnica sofisticada para orquestração de agentes IA em revisão de código. Implementação em Go com PTYs reais e estratégias de convergência plugáveis é robusta, mas requer infraestrutura local (Codex + Claude CLI) e é mais apropriada para equipes de desenvolvimento que buscam automação avançada.

README

A tradução para o português deste README ainda está na fila.

aibridge

A product for mutual code review between two live, interactive AI agents — Claude Code and OpenAI Codex. They run in live PTY-backed CLI sessions and take turns reviewing the shared git work tree: one edits, the other audits the diff, ping-ponging until both agree the code is clean. Driven from a web dashboard where you watch both agents work live, edit every setting, switch convergence strategies, and steer the run.

go build -o aibridge ./cmd/aibridge
./aibridge --repo /path/to/your/git/repo
# opens http://127.0.0.1:8799 — press Start

Why interactive (not SDK / -p / exec)

Both agents stay real interactive sessions, driven through pseudo-terminals with raw output streamed to the dashboard. This avoids the non-interactive paths (claude -p, codex exec, the Agent/Codex SDKs) which draw from separate, rate-limited credit pools (Anthropic split SDK/-p into its own pool on 2026-06-15; the Codex SDK is codex exec under the hood). Interactive CLI sessions use your normal subscription path. The two agents communicate only through the shared git diff.

Web dashboard

  • Live terminals: both agents' raw PTY output, streamed over WebSockets.
  • Verdict & timeline: each turn's CLEAN/FIXED/ISSUES and convergence reason.
  • Config editor (⚙): every setting below, saved to aibridge.yaml.
  • Steering: Start / Stop / Pause / Resume; restrict to one agent (codex-only / claude-only / both); skip a side's next turn; inject a message into either agent before its next turn.

Configurable (everything)

aibridge.yaml (created on first save; zero-config defaults otherwise):

repo: /path/to/repo
agents:
  codex:
    enabled: true
    command: "/Applications/Codex.app/Contents/Resources/codex"
    prompt_first: ""        # Go template; empty = built-in default
    prompt_next:  ""        # may reference {{.Handoff}} {{.Ask}} {{.Verdict}} {{.AskBlock}}
    stable_for: 10s         # screen-quiet duration that counts a turn "done"
    settle_for: 30s         # max wait for the response to start rendering
    timeout: 10m
  claude:
    enabled: true
    command: "claude --dangerously-skip-permissions"
    # ...same fields...
flow:
  first: codex              # who reviews first
  max_rounds: 8
  strategy: combined        # convergence strategy (below)
server:
  addr: 127.0.0.1:8799

Convergence strategies (pluggable, not hard-wired)

The "are we done / any more bugs?" decision is a swappable strategy:

strategyconverges when…
diff-fixpointa full round passes with no code change and both verdicts CLEAN
ask-gateboth sides explicitly answer NO_MORE_BUGS on consecutive turns
combined (default)BOTH of the above hold — strictest; the real "are you sure there are no other bugs?" gate

Agents emit AUDIT_RESULT: CLEAN|FIXED|ISSUES, and (for ask-gate/combined) NO_MORE_BUGS / MORE_BUGS, parsed from the captured screen.

Single-agent mode

Disable one agent (enabled: false) or use the dashboard's codex-only / claude-only buttons to drive just one AI.

Architecture

cmd/aibridge      CLI entry: web dashboard (default) or --headless one-shot
internal/config   YAML config + validation, human-readable durations
internal/agent    PTY process control (start/input/screen) + idle detection
internal/bridge   ping-pong loop, pluggable Strategy, prompt templates,
                  PTY driver, event bus, control surface
internal/runner   config → launch agents → drive loop, expose bus + control
internal/server   REST + SSE events + WebSocket terminals + embedded dashboard
cmd/mockagent     fake interactive agent for hermetic e2e tests

CLI flags

--config (path, default aibridge.yaml) · --repo / --addr (override config) · --headless (run one loop in the terminal, no UI) · --no-open (don't launch browser). AIBRIDGE_DEBUG=1 prints per-turn parse traces.

Verified

  • 14 unit tests: all three convergence strategies, verdict & NO_MORE_BUGS parsing (incl. ignoring the echoed prompt's own tokens), prompt template rendering (single-line, custom, ask-block, malformed), and single-agent control.
  • PTY e2e with mock agents: converges [codex:FIXED claude:CLEAN codex:CLEAN].
  • Real live run: Claude (Opus) ↔ Codex (gpt-5.5) on a buggy repo converged to double-clean; codex fixed an off-by-one and a divide-by-zero, claude audited and confirmed. Server-driven run over the web API, SSE, and WebSockets verified end to end.

Known edge cases

  • Codex's first run in an untrusted directory shows a "trust this directory?" prompt; mark the project trusted in ~/.codex/config.toml first.
  • Idle detection is heuristic (screen stability). Raise stable_for for slow / high-effort models. An unparseable turn is treated as UNKNOWN (never counts as clean), so a mis-timed cut costs an extra round, never a false convergence.
  • Claude needs --dangerously-skip-permissions (default) so it doesn't stall on per-command approval prompts the dashboard can't answer.
Ver o original no GitHub

Bom para

Relacionados