Overview
This series explains how Codewhale turns model suggestions into local, authorized work without conflating inference, execution, or proof. It follows one Runtime across terminal, headless, API, Fleet, and Workflow surfaces; traces a bounded turn through tools, approvals, diagnostics, and lifecycle records; separates permission from actual sandboxing; and examines provider routes, extensions, recovery, and orchestration. The result is a practical way to keep claims proportional to evidence when capabilities, platform containment, or release behavior remain unknown.
Learning outcomes
- Codewhale's Runtime, not the model response, is the authority boundary for consequential actions.
- Changing provider routes changes inference and capability evidence, not the Runtime's basic permission contract.
- Permission and actual OS containment are separate; platform and backend conditions determine the sandbox claim.
- A completed lifecycle record is not proof of correctness; relevant checks or independent verification support stronger claims.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What is Codewhale's Runtime model?
Codewhale treats the Runtime as the execution substrate that mediates model output, tools, policy, approvals, workspace scope, and lifecycle records; the model proposes actions but does not own repository authority.
Does changing the model provider change Codewhale's permissions?
No: changing Codewhale's provider changes the inference route, endpoint, protocol, credentials, and possibly capabilities, while the Runtime continues to apply the active mode, policy, approval, and sandbox boundaries.
Is Codewhale permission the same as a sandbox?
No. Permission determines whether a call may proceed under Runtime policy; actual containment depends on the selected and available OS or external backend, which varies by platform and configuration.
How can someone tell whether a Codewhale task is complete?
No single record proves completion: tool results and lifecycle records show what settled, while diagnostics or an independent verifier support narrower claims about the result; missing evidence remains unknown.