Claude CLI runner
@conduit-harness/conduit-runner-claude-cli invokes the claude (Claude Code) CLI as a subprocess inside the issue’s worktree.
Install
Section titled “Install”npm install -g @conduit-harness/conduit-runner-claude-cliThe claude binary must be installed and on PATH. See the Claude Code docs for setup.
Workflow snippet
Section titled “Workflow snippet”agent: kind: claude-cli max_concurrent_agents: 1claude-cli: model: claude-sonnet-4-6Output
Section titled “Output”The runner pipes the prompt to claude --dangerously-skip-permissions -p -. In -p (print) mode with the default --output-format text, Claude Code prints only the final assistant message to stdout and writes nothing to stderr — no tool-use traces (Bash/Edit/Read/Write), no streaming partials, no auth/model banners. The captured output posted as the tracker’s success comment is therefore the model’s final summary, which is what you want for a public issue comment.
If you override the command setting to add --output-format json or --output-format stream-json, the comment body will contain the raw JSON / streaming envelope instead. Stick with the default unless you have a downstream parser that expects it.