Set up Benny on Codex

Benny is a two-workflow automation pack: triage one Slack issue report, then reproduce or verify a trusted bug/performance verdict. These three explicit-only skills distribute its native instructions; distribution alone does not create, enable or schedule either workflow.

1. Prepare the complete project-owned pack

Identify the target repository and existing Benny setup. Read the native intent, configuration template and both prompt templates in the bundled references/benny/ pack. If this file is already inside a committed pack at skills/setup-benny/instructions.md, locate its root through ../../FOR_AGENTS.md instead. The complete pack includes all three operational files and their supporting references.

Only when project setup is requested, merge the complete bundled pack into <target-repository>/.agents/automations/benny/. Preserve destination-only files, inspect conflicts and retain local changes. Keep user-owned configuration, routing and feature maps outside the managed pack, for example .agents/benny/. Stop and ask when ownership of a conflicting file is ambiguous. Never install skills or write Cursor settings as an incidental setup step.

Resolve the ten shared dependencies through actual native discovery in a fresh session rooted in the target project: cursor-how, cursor-why, cursor-tdd, cursor-unslop, cursor-principle-separate-before-serializing-shared-state, cursor-principle-minimize-reader-load, cursor-principle-guard-the-context-window, cursor-principle-sequence-verifiable-units, cursor-principle-fix-root-causes, and cursor-principle-prove-it-works. Do not count this session's loaded context or assume another home has them. Native discovery differs from Cursor project plugin enablement; report the scope actually verified. Missing dependencies leave the dependent workflow incomplete.

The project-owned pack's operational files are read directly, not selected implicitly as chat skills. Confirm the pack and referenced secret-free configuration are committed in the same repository before a fresh automation checkout uses them; do not commit unless asked. Live prompts use stable project-relative paths, never installation/cache paths or this mirror's source-repository paths.

2. Fill configuration without secrets

Copy the bundled configuration, feature-map and optional routing-map examples into user-owned locations outside the pack. Never overwrite user configuration on refresh. Require explicit source channel, optional operations channel, repository/default branch, triage identity, tracker actions/team/project/labels/intake status, control adapter, complete user-facing feature map, verdict markers, Unicode status strings, draft-PR URL form, observation/repro/fix budgets and available model choices. Read the standing model-capability-router and the exact host's available models; do not guess slugs or change the current selection.

Secrets belong in the authorized environment or secret store, not YAML, prompts or committed files. Owner pings remain off unless the configured policy permits the specific feature owner or strongly evidenced regression author. Never guess routing destinations, tracker IDs or priorities.