Coding-agent governance at runtime

Coding agents can read source, run commands, call APIs, and change remote state. Runtime governance needs decision before tool execution. AegisFlow applies policy at configured protocol boundary.

Action model

AegisFlow normalizes routed call into ActionEnvelope with fields such as:

• actor and task
• protocol and tool
• target and arguments
• requested capability
• session context

Policy returns allow, review, or block.

allow sends call upstream. review creates pending approval. block returns structured error without calling upstream.

Credentials

Credential brokers can issue short-lived access after policy and approval checks. GitHub broker can request installation token with selected repository and permissions. AWS broker can apply task-specific STS session policy.

Broker support depends on deployment configuration. AegisFlow does not reduce scope of arbitrary static token automatically.

Evidence

Each governed session can record policy decision, approval event, credential issue, and upstream result. Records are hash-linked and signed with HMAC key. See tamper-evident audit guide for verification and key handling.

Start with narrow workflow

Use one policy pack and one upstream before broad rollout. PR-writer starter kit covers repository read, test, pull request review, and destructive action block. Run governed PR proof, inspect evidence, then tune target patterns for your repository.