Hey! I built siaan — Skill Is All Agents Need.
In practice, the loop starts with a GitHub Issue labeled status:ready, and an agent works toward a review-ready PR.
Repo: https://github.com/Stevengre/siaan

What siaan adds

• GitHub Issues as the sole tracker
  Migrated from Linear. Labels drive the state machine.

• Security guardrails around issue-based management

  • allowlist-based access control, so only approved users can trigger agent work
  • a human triage gate (status:triage → status:ready)
  • full auditability through issues and PRs

• PR-comment driven interaction
  Leave feedback on a PR and the agent iterates asynchronously, similar to reviewing a human contributor.

• Docker-sandboxed workspaces
  Each issue runs in an isolated container with its own branch and directory. Agent permissions are scoped per issue to prevent cross-issue interference.

• Token-conscious execution
  Implementation work is dispatched to @codex to control cost. Each session tracks token usage so the cost of an issue is visible.

• Companion skill: gh-issue-triage
  Converts raw ideas into structured issues with:

  • classification
  • acceptance criteria
  • risk analysis
  • dependency mapping

Self-developing

siaan is already implementing its own issues.

So far, the self-loop has shipped one PR end-to-end.
Many small bug fixes were still done locally, because not every fix warrants a full issue.

Going forward, the plan is to drive all non-trivial work through issues.

Why not a fork?

Symphony is an orchestrator for coding agents. I'm exploring a different direction and using this project to rapidly iterate on that idea.

siaan is exploring something different: a platform where everything is a SKILL.

SKILL files are human-readable and machine-executable.
When they can embed scripts, they become Turing-complete, which makes them a new application paradigm rather than just a configuration format.

The core hypothesis is that complex agent configuration —

• hooks
• multi-agent orchestration
• tool routing
• memory systems

— can collapse into SKILL as a unified abstraction.

GitHub-based software development is the starting point.
The goal is to accelerate iteration on SKILL-driven agents through the Issue → PR development loop.

Where exactly "everything is a SKILL" or "Skill Is All Agents Need" leads — skill-as-application, agents composing skills autonomously, or something else — is what I am exploring through rapid iteration.

Because that direction diverges from Symphony, I rebuilt around this vision instead of extending upstream.

Feedback welcome

Especially interested in thoughts on:

• the SKILL-as-abstraction direction
• what would be missing for your workflows