FAQ¶
Why not let one coding agent do everything?¶
ReasonFirst separates high-value reasoning from high-volume execution. Architecture, ambiguous diagnosis, scope control, and final review benefit from the strongest reasoning interface. File edits and test iterations are delegated to replaceable coding workers.
Is ChatGPT required for every command?¶
No. ActualCoder can run directly from a terminal, CI job, IDE integration, or another client. That capability is useful operationally, but the reference ReasonFirst workflow keeps ChatGPT at the reasoning forefront.
Does ReasonFirst call the OpenAI model API?¶
No direct model-inference API calls are made by ReasonFirst. Coding backends use their own supported authentication and entitlement. A Tunnel runtime key, when used, is for the connection path rather than ReasonFirst model inference.
Do I need the Tunnel?¶
You need the Tunnel for the documented normal-ChatGPT read connection to a self-hosted GitLab through the read-only MCP. The execution engine itself can operate without that connection.
Which coding backend should I use?¶
ReasonFirst supports explicit codex-cli, copilot-cli, and codex-desktop selection. Backend choice should be a user decision; TaskSpec, workspace identity, validation, review gates, and publication policy should not depend on which worker is selected.
What is TaskSpec?¶
TaskSpec is the durable task contract: original goal, acceptance criteria, non-goals, project/base identity, and requested backend. It prevents a multi-attempt implementation from silently losing the approved task boundary.
What is EvidencePack?¶
EvidencePack is a bounded, recursively redacted, read-only evidence bundle for a managed workspace. It can include task/attempt metadata, project policy, changed paths, reviewability, diff, and optional CI evidence. Missing, stale, incomplete, or truncated evidence is reported explicitly.
Is a green CI run enough to merge?¶
Not necessarily. Check that the CI belongs to the intended HEAD and covers the relevant validation. Also review the diff, protected paths, secret/history scans, and task acceptance criteria.
Is a worktree a security sandbox?¶
No. ReasonFirst adds controls and review gates, but a Git worktree is not an OS security sandbox. Use trusted development hosts and trusted repositories, and read the security policy.
Why GitLab if the project is hosted on GitHub?¶
GitHub hosts the ReasonFirst source repository. GitLab is currently the implemented target SCM/CI adapter for managed project work. Source hosting does not imply a GitHub-target execution adapter already exists.
Where should a new user begin?¶
Use First 10 minutes if your connection and worker are already available. Otherwise start with First-time setup, then use the Practice lab to rehearse the complete reasoning → worker → MR/CI → review loop.
How should I bring worker results back to ChatGPT?¶
Prefer bounded evidence rather than raw transcripts. Generate actual-coder evidence WORKSPACE (optionally --from-ci) and provide the non-secret EvidencePack to the reasoning session. See Evidence & review.
What prompts should I use with ChatGPT?¶
Use prompts that keep facts, task definition, implementation, and review separate. The Working with ChatGPT page contains copyable prompt patterns for repository inspection, TaskSpec definition, implementation review, and CI repair.