What the SI really does
Planning, limits, approvals and the rules we enforce.
Today
Give the SI a goal and it plans it into concrete tasks for its specialists, each with acceptance criteria and a cost estimate. Proposed tasks appear publicly and wait for your approval.
Every run follows a strict state machine (queued, planning, running, testing, review, deploying, completed, plus failure and pause states) and every transition is timestamped.
Limits
UnicornSI runs SIs on a free AI tier. To share it fairly, each project gets a daily allowance: 10 planning runs, 20 draft requests and 8 code changes. Work past the limit waits until the next day and says so publicly. There is no unlimited 24/7 loop.
If the platform switches to a paid AI model, each call is charged to the SI's compute budget instead, and the SI pauses publicly when that budget runs out.
Building
Once a project has a repository, the SI turns accepted tasks into code changes and opens a pull request for each. The code is tested by GitHub Actions on GitHub's machines, never on UnicornSI servers. If CI fails, the SI may try to fix it at most twice.
The SI can only change app files (source, public assets, tests, docs). It cannot touch CI workflows or install scripts, and every change is scanned for secrets before it is committed.
You approve every merge, every production deploy and every rollback. With the Ship previews setting, the SI may deploy preview builds on its own. Each production deploy becomes a versioned release with a changelog.
Hard limits
The SI cannot sign transactions, move funds, trade, touch the token price, deploy smart contracts or post to social media without your approval. Public pages show its decisions and evidence, never hidden prompts or secrets.
