Envoy

Triage

Envoy polls IMAP on a configurable interval (default: 5 minutes). For each new message it runs a triage pass: classify, prioritize, and decide whether to draft a reply or flag for direct human attention.

Drafting

For messages that clear triage, the agent drafts a reply. The draft is written to match the voice and style defined in the personality file — the agent is answering on your behalf, so it writes as you, not as an assistant.

The agent can also use its project context (via GitHub tap) to enrich replies. "Is issue X fixed?" → check GitHub, include the PR link and the version it shipped in.

Sign-off — Shared Dashboard

Envoy shares the sign-off mechanism with Evangelist. Both agents route outbound actions through the same tap human-review pattern, and both land in the same dashboard queue. The dashboard distinguishes them by type (email-draft vs post-draft) but uses a single approval flow.

Configuration

Tap middleware

tap:
  routes:
    # Outbound email always goes to human-review queue
    - pattern: "tools/email/send"
      action: human-review
      queue: /srv/data/kanipi/email/pending/
      timeout: 72h   # longer window than posts

    # SMTP credentials injected only after approval
    - pattern: "tools/email/send"
      action: inject-credential
      credential: SMTP_PASSWORD
      as: env
      name: SMTP_PASSWORD

    # IMAP read is allowed without approval (read-only)
    - pattern: "tools/email/read"
      action: inject-credential
      credential: IMAP_PASSWORD
      as: env
      name: IMAP_PASSWORD

    # GitHub access for enriching support replies
    - pattern: "tools/github/*"
      action: inject-credential
      credential: GITHUB_TOKEN
      as: header
      name: Authorization
      prefix: "token "

    # Hard blocks
    - pattern: "tools/twitter/*"
      action: reject
      reason: "Envoy agent does not have Twitter access"

    - pattern: "tools/telegram/*"
      action: reject
      reason: "Envoy agent does not run in Telegram"

config.yaml — agent personality

agents:
  email:
    personality: |
      You manage incoming email for [project-name].
      You read, triage, and draft replies. Nothing sends without human approval.

      ## Triage Rules
      Auto-draft:
      - User support and how-to questions
      - Contributor questions about architecture, process, or roadmap
      - Feature requests (acknowledge + log to GitHub issues if appropriate)
      - Thank-you replies for contributions

      Escalate (no draft):
      - Pricing, licensing, legal, contract discussions
      - Media or press inquiries
      - Anything requiring a decision you can't make from context

      Ignore (mark read, skip queue):
      - CI/CD notifications, automated reports
      - Bounce messages
      - Known mailing lists not relevant to the project

      ## Voice
      Write as [owner-name], not as an assistant.
      - Direct and brief
      - Warm but not sycophantic
      - Include specific technical detail when helpful
      - Never use: "Hope this helps!", "Feel free to", "Don't hesitate"

      ## Context sources
      GitHub issues and PRs available via tools/github/*
      Use them to answer status questions with specifics.

      ## State
      /srv/data/kanipi/email/state.md:
      - Triage log (sender, classification, draft or escalate)
      - Pending queue size
      - Rejected draft patterns (learn from what gets edited/rejected)
    image: kanipi/agent:latest
    network: true

Channel config

channels:
  email:
    enabled: true
    imap:
      host: ${IMAP_HOST}
      port: 993
      user: ${EMAIL_USER}
      poll_interval: 5m
    smtp:
      host: ${SMTP_HOST}
      port: 587
      user: ${EMAIL_USER}

What Kanipi provides