Multi-Agent Development Orchestrator.
AgentBridge coordinates development agents, repository evidence, reviews, tests, policy, and human decisions. Autoflow is the orchestration engine inside AgentBridge.
AgentBridge V1 is intentionally read-only against managed repositories. It is designed to automate coordination before automating authority.
V1 does not autonomously edit managed repository source files, create branches or commits in managed repositories, push changes, create/approve/merge pull requests, deploy, or mutate managed project databases.
Unknown or unclassified actions must fail closed and escalate.
This repository is in early bootstrap. PR 001 contains only development scaffolding, documentation, CI, and reproducible build/test tooling.
PR 002 adds the domain and action kernel: pure TypeScript models for requested operations and a deterministic classifier that maps them to a policy decision. It classifies only and executes nothing. V1 defaults allow a fixed read-only allowlist and escalate everything else to human review; unrecognized actions fail closed and never resolve to ALLOW. See docs/architecture/002-domain-action-kernel.md.
PR 003 adds the action request envelope and policy gate: a structured, untrusted request from an agent is evaluated into a GateDecision that answers exactly one question — may AgentBridge execute this without human approval? Agents request actions; they do not authorize them. Agent identity, provider, rationale, and metadata cannot increase authority, and human approval is a separate trust boundary. The gate evaluates only and executes nothing. See docs/architecture/003-action-request-gate.md.
PR 004 adds the commit-bound evidence and freshness kernel: evidence records are bound to a repository and a commit SHA, and are evaluated against a caller-supplied HEAD into CURRENT, STALE, or INVALID. Evidence whose SHA differs from HEAD is never current, so a new commit automatically stales prior evidence. Evidence is data, never authority, and the kernel performs no I/O. See docs/architecture/004-evidence-freshness-kernel.md.
PR 005 adds review ingestion: untrusted reviewer output is normalized into commit-bound review findings. Trusted caller-supplied context is the only source of repository, pull request, commit SHA, provider, and reviewer identity — reviewer content can never set a binding field, and a review of one commit stays bound to that commit. Ingestion is provider-neutral, decides neither freshness nor authority, and performs no I/O. See docs/architecture/005-review-ingestion.md.
PR 006 adds the provider-neutral agent invocation boundary: a commit-bound record of what AgentBridge asked which agent to do, and normalization of what that agent reported back. An agent's report is an untrusted claim — PR 006 records that an artifact was claimed, never that it exists remotely, is integrated, is validated, or is authorized. Provider identity is inert: it never implies a role or grants authority, and roles remain configurable. The layer invokes nothing, models no lifecycle transitions, and performs no I/O. See docs/architecture/006-agent-invocation-boundary.md.
- Node.js 24 LTS
- TypeScript (strict)
- npm
- ESM
- Vitest
- ESLint + typescript-eslint
npm ci
npm run verifynpm run verify runs type checking, linting, tests, and the production build.
Architecture and security decisions are documented under docs/. The V1 architecture is frozen for implementation unless a contradiction or security defect requires an explicit architecture decision.