About OpenFlows
OpenFlows coordinates five specialized AI agents through a Redis-backed state machine to take GitHub issues all the way to merged, documented pull requests. GitHub identity is handled by Coder external auth - no personal access tokens. The entire pipeline runs autonomously - you wake up to completed features, reviewed PRs, and updated documentation.
The team
Polls GitHub for issues, provisions workspaces, assigns work to agents, and manages the entire pipeline lifecycle.
Receives tickets, writes PLAN.md, implements code segment by segment, and opens pull requests.
Reviews every code segment against 5 criteria: correctness, security, test coverage, standards compliance, and no regressions.
Handles CI status, resolves merge conflicts, and squash-merges approved PRs after SENTINEL approval.
Writes ADRs (Architecture Decision Records) and updates CHANGELOG.md after each successful merge.
"Coder governs where agents run. OpenFlows governs how they coordinate. Code is the output. Architecture is the product."
How it works
NEXUS polls GitHub for open issues and provisions ephemeral workspaces for each task. FORGE receives the ticket, writes a PLAN.md, and implements the code segment by segment. SENTINEL reviews every segment against five criteria before approval. VESSEL handles CI status and merges approved PRs. LORE documents everything with ADRs and CHANGELOG updates.
The entire pipeline runs without human intervention. NEXUS only reaches out when it genuinely needs you - spec ambiguity, security concerns, or resource limits. Otherwise, the team runs autonomously within your governed environment.
What we stand for
The pipeline runs without human input. NEXUS polls GitHub, assigns tickets, provisions ephemeral Coder workspaces, and manages the full lifecycle. It only escalates to AwaitingHuman for spec ambiguity, security concerns, or resource limits.
Agent = CLI backend (muscle) + configuration harness (brain). Each agent has its own persona, skills, hooks, and permissions. GitHub identity is handled by Coder external auth - agents inherit identity from the logged-in Coder user via the git-config module. No personal access tokens needed.
A five-layer crash recovery stack catches failures at every level, from malformed LLM output to crashed nodes. NEXUS reconcile() detects orphaned tickets, stale workers, and unmerged PRs on every cycle. Ephemeral Coder workspaces are torn down after merge - no lingering state, no stale credentials.
Each agent's LLM provider and model are configured independently in registry.json. FORGE and SENTINEL can use different providers - Anthropic for the builder, OpenAI for the reviewer - so the reviewer's blind spots differ from the writer's. SENTINEL spawns fresh per evaluation, read-only and bias-free.
The orchestration engine - PocketFlow core, SharedStore, agent harnesses, and TUI setup wizard - is on GitHub under the MIT license. Install via npm, cargo, Docker, or a one-line curl installer. Self-hosted on your infrastructure, with Coder AI Gateway and LiteLLM fallback for any model provider.
FORGE implements and opens PRs in 15–30 minutes. SENTINEL reviews every segment automatically. VESSEL merges when CI is green. A 2-person startup can ship 40 features in a month because the pipeline eliminates the PR review queue and founder context-switching.
When VESSEL detects a merge conflict, it does not create a new branch or lose context. It writes CONFLICT_RESOLUTION.md and re-routes the ticket back to the same FORGE worker. FORGE reworks the implementation, re-commits, and VESSEL re-monitors CI - no manual rebase.
LORE writes Architecture Decision Records and updates CHANGELOG.md after every successful merge. Technical decisions are captured automatically, not as an afterthought. The team ships code and the paper trail writes itself.
Community
OpenFlows is MIT licensed and self-hosted on your infrastructure. Contributors publish Coder Registry modules that install code agent CLIs (claude-code, codex, aider, and more) into ephemeral workspaces, extend the PocketFlow engine with new agent node types, improve the five-layer crash recovery stack, and integrate with additional LLM providers through the Coder AI Gateway and LiteLLM fallback.
The project is maintained by The AgenticFlow team. Whether you're writing agent harnesses with custom personas and permissions, improving VESSEL's merge conflict resolution, enhancing the TUI dashboard, or documenting architecture decisions, your work directly shapes how autonomous development teams operate.
