These are public records of implementation failures that changed FlowGrid's protocol or code. They are not release notes and they are not case studies written after a success.
Each entry answers five questions:
- What failed in a real or controlled project workflow?
- Why could that failure mislead a later agent or project owner?
- What did FlowGrid change?
- What evidence verifies the change?
- What does the change still not prove?
Dogfood project identities, raw conversations, and private evaluation material are deliberately omitted. The public record keeps the failure mechanism and the product boundary.
- 001 - A candidate is not a decision
- 002 - Review must happen before a candidate reaches startup context
- 003 - The ledger is the truth; the index is a cache
These are likely future topics, not commitments or claims that the problems are already solved.
- Why a complete Context Pack may still lose to short, clean raw history.
- Why a host integration needs raw-session capture, not only a better parser.
- Why an installed Agent Skill drifting from the repository version can reopen a resolved safety problem.