Agent approval workflow

When the user wants agents to drive work but only execute after human approval, follow this workflow. The human's job is to accept or decline a todo list; the agent then acts only on accepted items.

When to use

• •User says "put agents to work", "run the workflow", "run all agents", "give me the todo list", "agents suggest I approve", or similar.
• •User wants a single list of proposed actions and will reply with accept/decline.

Workflow

Step 1: Gather suggestions (do not execute)

Run each relevant subagent's logic in suggestion mode: determine what each would do (create/update file, run command, add section) and record it as a proposed action. Do not create or edit files yet.

Relevant agents for a full run: api-contracts, contributing, changelog, docs-steward, github-docs-sync, testing, security-audit, devops-release, dependencies, future-features, ui-steward, code-reviewer (if there are recent changes), build-restart (e.g. "Build frontend and restart backend" if frontend or backend changed).

ui-steward and future-features must always produce at least one concrete action:

• •ui-steward: e.g. "Review Dashboard/SessionView and update docs/UI_GUIDELINES.md with current suggestions, backlog, and scale-friendly improvements."
• •future-features: e.g. "Research roadmap ideas (web search if needed) and update docs/ROADMAP.md with prioritized list and last-updated date."

For each agent, output 1–5 concrete actions, e.g.:

• •"Create docs/API_AND_EVENTS.md with REST and Socket.io tables"
• •"Update README first paragraph and Key Features"
• •"Add npm audit note to security backlog"
• •"Create CONTRIBUTING.md with run instructions and where to change what"

Combine all actions into one list. Deduplicate (same file or same action = one item).

Step 2: Present the approval todo list

Output a single numbered todo list in this format:

## Approval todo list

Review the items below. Reply with one of:
- **Accept all** — I will execute every item.
- **Accept 1, 3, 5** — I will execute only those (comma-separated numbers or ranges like 1-4).
- **Decline 2, 4** — I will skip those and execute the rest.
- Or comment per item (e.g. "Accept 1 and 2, decline 3").

| # | Action | Source | What will be done |
|---|--------|--------|-------------------|
| 1 | Short title | agent-name | One line: file or command |
| 2 | ... | ... | ... |

Keep the table concise. "Source" = which agent suggested it (e.g. api-contracts, docs-steward). End with: "Reply with your choices (e.g. Accept all, or Accept 1 3 5, or Decline 2 4)."

Do not execute any action yet. Wait for the user's reply.

Step 3: Parse the user's reply

• •Accept all → accepted = all numbers.
• •Accept 1, 3, 5 or Accept 1 3 5 → accepted = {1, 3, 5}.
• •Decline 2, 4 → accepted = all except 2 and 4.
• •Ranges: "Accept 1-4" → accepted = {1, 2, 3, 4}.
• •Mixed: "Accept 1, 3, decline 2" → accepted = {1, 3}.
• •If the reply is ambiguous, ask once: "Confirm: execute items 1, 3, 5 only?"

Step 4: Execute only accepted items

For each accepted item, perform the action (create or update the file, run the command, add the section). Then report briefly: "Done: 1, 3, 5. Skipped: 2, 4." and list what was created/updated.

Rules

• •Never execute in Step 1 or Step 2. Only execute after the user has replied with accept/decline in Step 3.
• •One action = one row. If an agent suggests "update README and add API doc", split into two rows if they are independent.
• •If there are no suggestions (e.g. everything already done), say "No new actions from agents. Todo list is empty."
• •If the user says "accept all" in the same message as "put agents to work", treat as: first present the list, then if they already said accept all, execute all after presenting (so one round-trip is enough).