| instruction | Task instruction text sent to the LLM. Mutually exclusive with task_file — use one or the other.
| no | "" |
| task_file | Path to a task file (.md / .txt) whose content is used as the instruction. Useful for multi-line prompts stored as repository files.
| no | "" |
| project_path | Absolute project root for per-project session memory isolation. Defaults to $GITHUB_WORKSPACE (the checked-out repository root).
| no | "" |
| target_file | Relative path of the file the LLM should write or update (e.g. "src/utils/helpers.go"). Optional — omit for plan/analysis tasks.
| no | "" |
| context_files | Comma-separated list of relative file paths to include as read context alongside the instruction (e.g. "go.mod,internal/core/domain/task.go").
| no | "" |
| command | Task command type: "plan" (orchestration / documentation) or "execute" (code implementation). Default: execute.
| no | execute |
| model | LLM model override sent with the task (e.g. "gpt-4o", "claude-3-5-sonnet-20241022", "llama3").
| no | "" |
| provider | Provider hint for routing the task (e.g. "openai", "anthropic", "ollama", "lmstudio").
| no | "" |
| agent | Load a single specialist agent from el-j/agency-agents by slug or name. Examples: "engineering-backend-architect", "backend-architect", "Frontend Developer". Mutually exclusive with agents / agent_category. The agent's system prompt is automatically prepended to the instruction.
| no | "" |
| agents | Comma-separated list of agent slugs/names for swarm mode (e.g. "engineering-backend-architect,engineering-frontend-developer"). A combined orchestrator prompt is generated from all agents. Mutually exclusive with agent / agent_category.
| no | "" |
| agent_category | Load ALL agents from a specific category from el-j/agency-agents (e.g. "engineering", "design", "testing"). A combined swarm prompt is generated. Mutually exclusive with agent / agents.
| no | "" |
| agent_ref | Git ref (branch, tag, or SHA) of el-j/agency-agents to fetch identities from. Default: "main".
| no | main |
| system_prompt | Raw agent system prompt prepended to the instruction. Takes precedence over agent / agents / agent_category when set. Use this to chain with el-j/agency-agents manually:
- uses: el-j/agency-agents@main
id: ag
with:
agent: backend-architect
- uses: el-j/nexus-orchestrator@v1
with:
system_prompt: ${{ steps.ag.outputs.system_prompt }}
instruction: 'Refactor auth to use JWT'
| no | "" |
| daemon_url | HTTP base URL of a running nexus-daemon. Used only when start_daemon is "false".
| no | http://127.0.0.1:63987 |
| start_daemon | Download and start a local nexus-daemon for the duration of this job. Set to "false" to connect to an existing remote daemon via daemon_url.
| no | true |
| nexus_version | nexus-orchestrator release version to install (e.g. "v0.2.0"; defaults to "latest").
| no | latest |
| timeout_seconds | Maximum seconds to wait for the task to reach a terminal status before the step fails.
| no | 300 |
| openai_api_key | OpenAI API key — enables GPT models on the local daemon. | no | "" |
| openai_model | OpenAI model override (default: gpt-4o-mini). Requires openai_api_key. | no | "" |
| anthropic_api_key | Anthropic API key — enables Claude models on the local daemon. | no | "" |
| anthropic_model | Anthropic model override (default: claude-3-5-sonnet-20241022). Requires anthropic_api_key.
| no | "" |
| github_copilot_token | GitHub Copilot token — enables GPT-4o via Copilot on the local daemon. | no | "" |
| github_copilot_model | Copilot model override (default: gpt-4o). Requires github_copilot_token. | no | "" |