How Intent-System runs today.
A two-region operational loop with a feedback path. Ⓐ the intent tree (curated by humans, reviewed by AI). Ⓑ the implementation cycle (an agent codes from what is embedded in the issue; a separate AI reviewer reads the implementation, the issue, and the intent tree at the same time, then a human gives final approval). Hand-off, review pass, and review reject move work between them on a cadence — and post-release defects re-enter the loop as Intent-Bug issues.
Hover over a node to isolate the surrounding hand-off, pass, and reject paths.
Curated by humans
- Inferred intent capture — agents propose; humans accept or reject.
- Tree placement — Purpose / User Context / Means.
- Clarification loop until ambiguity is resolved.
- Promote: inferred → clarified → canonical.
- Slice into an issue ready for implementation.
Driven by agents
- Issue picked up by the implementation thread — Claude on /loop 5m, launched via intent-cli.
- Agent implements from what is embedded in the issue; PR opens against ai-develop.
- AI review — a separate Claude /loop 5m thread, running in a different cwd (e.g. *Review worktree), diffs the implementation against the packet and the intent tree it has loaded. Pass continues; reject sends a clarification back to Ⓐ-③.
- Human final approval before promotion to ai-develop / main.
- Closeout writes the diff back into the tree as canonical.
Post-release defects re-enter the same loop
The implementation will not be perfect, and that is fine — Intent-System treats the post-release feedback path as a first-class part of the loop, not as an exception. Each defect or surprising behavior is filed as an Intent-Bug issue and triaged against the intent tree:
- Implementation bug. Intent was right; code did not honor it. Re-enter Ⓑ-① with the same intent slice; fix on ai-develop.
- Intent gap. Intent said nothing on this point; the implementer made a local choice; the result was not what the team wanted. Re-enter Ⓐ-③ as a clarification → promote the tree → then cut a new issue from the sharper intent.
Either way the bug exits via the same AI review → human approval → promotion path. Drawing this line correctly is what makes the tree get stronger over time instead of just bigger.
Want the implementation-level view?
This page shows the two-region operational model. The next page zooms into how the implementation cycle actually runs today — three Claude threads on /loop 5m, kicked off through intent-cli, with the impl and review threads in physically separate cwds.