| name | agentboard-harness |
| description | Use when the user asks Codex to operate AgentBoard, run AgentBoard slash-command-like workflows, use AgentBoard as a Codex/Claude/OpenCode harness, or move product docs through analyze, clarify, design, plan, run, QA, and change stages with the local AgentBoard CLI. |
AgentBoard Harness
Use this skill to drive the local agentboard CLI from Codex. Treat AgentBoard as the source of workflow state: projects, sources, artifacts, approvals, issues, run logs, QA, and change requests live under .agentboard/.
First Steps
- Locate the repo root. Prefer the current workspace if it contains
cmd/agentboard/main.go or an executable ./agentboard.
- Use
plugins/agentboard-harness/scripts/agentboard.sh as the CLI wrapper when possible. It uses AGENTBOARD_BIN, ./agentboard, or go run ./cmd/agentboard.
- Run
agentboard status or agentboard project create depending on whether a project already exists.
- Ask the user only when the project slug or source path cannot be inferred safely.
Slash-Style Command Map
Codex does not need native slash command registration to run these. Interpret user text like /agentboard:plan rfq-platform as the matching CLI action.
/agentboard:init -> agentboard init
/agentboard:create "Name" -> agentboard project create "Name"
/agentboard:source <path> --project <slug> -> agentboard source add <path> --project <slug>
/agentboard:analyze <slug> -> agentboard analyze <slug>
/agentboard:clarify <slug> -> agentboard clarify <slug>
/agentboard:design <slug> -> agentboard design <slug>
/agentboard:plan <slug> -> agentboard plan <slug>
/agentboard:issues <slug> -> agentboard issue list --project <slug>
/agentboard:ready <issue> -> agentboard issue ready <issue>
/agentboard:run <issue> -> agentboard run --issue <issue>
/agentboard:qa <issue> -> agentboard qa review <issue>
/agentboard:preview <slug> -> agentboard preview --project <slug> --addr 127.0.0.1:0
/agentboard:change <slug> "<summary>" -> agentboard change create --project <slug> --summary "<summary>"
/agentboard:regenerate <change-id> -> agentboard change regenerate <change-id>
/agentboard:serve -> agentboard serve --addr 127.0.0.1:8787
Stage Agent Setup
For product/design/planning stages, configure AgentBoard to use CLI agents:
agentboard config set-stage-provider cli
agentboard config set-stage-agent product "plugins/agentboard-harness/scripts/codex-stage-agent.sh"
agentboard config set-stage-agent clarify "plugins/agentboard-harness/scripts/codex-stage-agent.sh"
agentboard config set-stage-agent design "plugins/agentboard-harness/scripts/codex-stage-agent.sh"
agentboard config set-stage-agent planner "plugins/agentboard-harness/scripts/codex-stage-agent.sh"
agentboard config set-stage-agent qa "plugins/agentboard-harness/scripts/codex-stage-agent.sh"
agentboard config set-stage-agent change "plugins/agentboard-harness/scripts/codex-stage-agent.sh"
Use claude-stage-agent.sh or opencode-stage-agent.sh instead when the user wants those agents.
AgentBoard writes the generated stage prompt to AGENTBOARD_STAGE_PROMPT and expects the stage command to write the artifact body to stdout. Do not pipe long prompts through ad hoc shell quoting if a wrapper script is available.
Stage agents also receive:
AGENTBOARD_PREVIEW_HOST=127.0.0.1
AGENTBOARD_PREVIEW_PORT=<free-port>
AGENTBOARD_PREVIEW_URL=http://127.0.0.1:<free-port>
Use those variables when a spec, design, or planning agent needs to show temporary HTML prototype options for a human decision. Do not use hard-coded shared ports.
Implementation Runs
Implementation runs are separate from stage generation. Configure the default implementation runner with:
agentboard config set-default codex
agentboard config set-runner codex "codex exec"
agentboard config set-runner claude "claude"
agentboard config set-runner opencode "opencode run"
Then use agentboard run --issue ISS-0001. Check agentboard logs ISS-0001 after each run.
Operating Rules
- Prefer AgentBoard commands over manual edits to
.agentboard/ state.
- After changes, inspect artifacts or issues with
artifact list/show, issue list, logs, and qa review.
- If a command fails because a binary is missing, report the missing binary and use local template mode only when the user accepts that fallback.
- Keep stage output focused on the artifact requested by AgentBoard; avoid conversational prefaces because stdout becomes the stored artifact.