Aegis Guardrails
A rewritten version of Soma: a rules engine that issues rules and enforces execution gates for AI coding agents, comprising an integrity checklist, release gates and an MCP toolset.
Timeline (last 90 days)
v4 was renamed Aegis Guardrails, cutting MCP tools from 17 to 5, and stipulated that nothing at runtime may modify rule files and that repository configuration may never specify executable commands.
In Aegis v4, repository-level configuration is merely declarative data, and executable commands can exist only in the user-level ~/.aegis/config.json.
When Aegis v4.1 matches AST rules by call name, it mistakenly flags PyTorch's model.eval() and pandas' df.eval().
When an agent reports results, Aegis requires the working tree to contain a witness line recorded by aegis check --record --tests; otherwise the report is recorded only as claimed with zero weight.
Aegis's aegis sync never automatically modifies or promotes rules, and opens a PR for human review only when the sample size is sufficient (precision is not shown below n=25).
Aegis's aegis check --ci compares gate and invariant rules against the merge base, preventing PRs from deleting gates, lowering severity levels, shrinking coverage, or weakening forbid_ast patterns.
All Facts (6)
v4 was renamed Aegis Guardrails, cutting MCP tools from 17 to 5, and stipulated that nothing at runtime may modify rule files and that repository configuration may never specify executable commands.
50%UnverifiedIn Aegis v4, repository-level configuration is merely declarative data, and executable commands can exist only in the user-level ~/.aegis/config.json.
50%UnverifiedWhen Aegis v4.1 matches AST rules by call name, it mistakenly flags PyTorch's model.eval() and pandas' df.eval().
50%UnverifiedWhen an agent reports results, Aegis requires the working tree to contain a witness line recorded by aegis check --record --tests; otherwise the report is recorded only as claimed with zero weight.
50%UnverifiedAegis's aegis sync never automatically modifies or promotes rules, and opens a PR for human review only when the sample size is sufficient (precision is not shown below n=25).
50%UnverifiedAegis's aegis check --ci compares gate and invariant rules against the merge base, preventing PRs from deleting gates, lowering severity levels, shrinking coverage, or weakening forbid_ast patterns.
50%