WS0 — sea-haven-claude-plugin/ scaffold:
- CLAUDE.md: Sea Haven engineering context (pipeline overview, rules,
/new-task + /delegate usage, available phases)
- settings.template.json: UserPromptSubmit hook wiring template
- hooks/user_prompt_submit.py: standalone script (WS4)
WS2 — SlackListener /new-task intake:
- Add NewTaskCallback type alias (Callable[[str, str], str])
- Add new_task_callback param to SlackListener.__init__
- _is_new_task_command() helper for slash_commands+/new-task detection
- _handle_new_task_command() method: authorized-only, calls callback,
exception-safe (listen loop stays alive on callback errors)
- handle_event() routes /new-task BEFORE the answer path (post-AUTHZ-01)
WS4 — UserPromptSubmit auto-delegate hook:
- /delegate <text> and DELEGATE: <text> prefixes trigger delegation
- Calls POST /tasks on the agent-team HTTP API (WS1)
- Blocks the Claude Code prompt; shows thread_id + next-steps message
- Graceful degradation: missing token, HTTP error, network error all
produce a block with a human-readable reason
- run() is a pure function for testability (no stdin/stdout in tests)
Tests: 22 new tests in test_ws0_ws2_ws4_plugin_slack_hook.py.
Full suite: 1066 passed. ruff clean.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QYp761G9HojkmLqLZrASVi
3.1 KiB
Sea Haven Industries — Claude Code Engineering Plugin
This directory is the Sea Haven Claude Code plugin: context and hooks for
Claude Code sessions run against the sea-haven-industries/orchestrator repo.
Copy settings.template.json into ~/.claude/settings.json (or merge it into
an existing one) to activate the hooks.
Project overview
The orchestrator is the R720-hosted Plane-2 SDLC pipeline. It runs a
multi-model multi-agent loop — Claude Sonnet (coordinator/clarifier/planner),
GPT-4.1 (cross-reviewer), DeepSeek (fast_coder/builder), Gemini
(scanner) — gated by a CI trust boundary (agent-team-apply-verify.yml).
Key directories:
| Path | Purpose |
|---|---|
agent-team/ |
The pipeline package (agent_team.*) + tests |
agent-team/run-team.py |
CLI: serve, start, answer sub-commands |
agent-team/agent_team/coordinator.py |
Keystone: graph + transport wiring |
agent-team/agent_team/graph.py |
LangGraph state machine (P1–P3+) |
agent-team/agent_team/nodes/ |
Pipeline node implementations |
agent-team/agent_team/transport/ |
Slack / GitHub / Claude Code adapters |
.github/workflows/agent-team-apply-verify.yml |
CI trust boundary (NEVER edit directly; use a WS PR) |
sea-haven-claude-plugin/ |
This plugin |
Development rules
- Green before merge:
cd agent-team && python3 -m ruff check . && python3 -m pytest -qmust pass. - No secrets in code: all tokens go in env vars or GitHub secrets.
- No SSH/AWS/infra commands: the R720 box is VPN-only; never attempt network access from a Claude Code session.
- Draft PRs only: all automated changes go out as
--draft; humans merge. - CI YAML edits require a security review (
/sh-security-review) and a separate WS PR — never inline.
Delegating tasks to the pipeline
Via Slack
Post /new-task <description> in any channel the Sea Haven bot is in. The
listener routes it to the coordinator's start_task, which begins the
CLARIFY → PLAN → REVIEW → BUILD → VERIFY pipeline. You'll receive a clarifier
question back in Slack.
Via this Claude Code session (auto-delegate hook)
Prefix your prompt with /delegate (or DELEGATE: ) to auto-send the task
to the pipeline HTTP API instead of answering it locally:
/delegate Add OAuth2 to the admin portal — use PKCE, no client secret stored
The hook calls POST http://127.0.0.1:8765/tasks (requires
AGENT_TEAM_API_TOKEN env var) and returns the thread_id so you can track
the task in the coordinator.
Set AGENT_TEAM_API_URL to override the default http://127.0.0.1:8765.
Set AGENT_TEAM_API_TOKEN to the bearer token from run-team.py serve.
Available slash commands
These are defined in settings.template.json (copy to ~/.claude/settings.json):
| Command | Effect |
|---|---|
(hooks auto-detect /delegate prefix) |
Auto-delegate to pipeline |
Phases
- P1: INTAKE → CLARIFY (human gate) → PLAN
- P2: + adversarial REVIEW loop (GPT-4.1 cross-reviewer)
- P3: + BUILD (DeepSeek) → VERIFY (CI gate)
- P3+: + DISPATCH (push head branch + trigger
workflow_dispatch)