init-architect（初始架构师）

Init Architect

Purpose

Run once after installing skills (and rerun when architecture drifts) to create or refresh architecture-folder artifacts used by architect-agent. Also supports scoped updates when invoked from pr-review-agent for architecture-impacting PR changes.

Runtime Configuration

• Read /orchestra-config.json from the repository root before starting.
• If /orchestra-config.json is missing, create it at repository root with:
• {"issue_tracker": "linear"}
• Read issue_tracker and use it for any generated tracker-facing instructions.
• Allowed issue_tracker values: linear, jira.
• For tracker comments in workflow-scoped mode, include: Skill-Version: init-architect@2.0.0.

When to Invoke

• Immediately after skill installation, before running architect-agent.
• When architecture artifacts are stale after major system changes.
• When pr-review-agent flags architecture-impacting changes and provides scoped PR context.

Required Inputs

• Repository root path.
• Invocation mode:
• full (default): repository analysis and full architecture refresh.
• scoped: targeted update using PR-derived scope.
• For scoped mode:
• parent issue ID.
• PR identifier/link.
• changed files list.
• concise diff summary.
• matched architecture-impact cases.
• suggested architecture sections.
• Optional existing /architecture/architecture.md.
• Optional existing /architecture/docs/*.md.
• Optional existing skills/init-architect/ARCHITECTURE.md template.
• Most recent prior handoff comment in <!-- OPEN-ORCHESTRA-HANDOFF --> format when running in workflow-scoped mode.

Outputs

• /orchestra-config.json in repository root (created only if missing).
• /architecture/architecture.md populated as the architecture index.
• /architecture/docs/*.md populated for complex domains.
• In scoped mode: only impacted sections/documents are updated; unaffected sections remain unchanged.
• Optional targeted updates to skills/architect-agent/SKILL.md reference tables so runtime guidance matches generated docs.
• For workflow-scoped mode, a handoff comment wrapped exactly as:

{
  "execution_trace": "Execution-Trace:\nActions:\n1. <action>\n2. <action>\nDecisions:\n- <architecture update decision + reason>\nReferences:\n- <changed files or architecture docs>\nAssumptions:\n- <assumption>\nOpen-Questions: none|<question list>\nSkill-Version: init-architect@2.0.0",
  "handoff_summary": {
    "from_skill": "init-architect",
    "to_skill": "architect-agent|pr-review-agent",
    "status": "ready|blocked|completed",
    "delta": ["<what changed in architecture artifacts>"],
    "key_decisions": [{"decision": "<decision>", "reason": "<reason>"}],
    "relevant_artifacts": [
      {
        "artifact": "architecture-index-or-doc",
        "hash": "sha256:<hash>",
        "last_modified": "<ISO-8601>",
        "summary": "<updated sections and rationale>"
      }
    ],
    "open_blockers": [{"blocker": "<text>", "owner": "<owner>", "next_action": "<action>"}],
    "next_guidance": {
      "need_full": ["architecture/architecture.md"],
      "focus": ["<sections downstream agents must re-check>"]
    }
  }
}

• handoff_summary must be <= 600 tokens.

Context Gathering Order (Strict)

1. Locate the most recent comment containing <!-- OPEN-ORCHESTRA-HANDOFF --> from the previous skill.
2. Parse the JSON inside it. This is your primary context.
3. Look at its relevant_artifacts list and hashes.
4. Declare exactly which artifacts you need via need_full.
5. Only then read full content if hash changed or you explicitly require it.
6. In non-workflow bootstrap mode with no prior handoff, continue with repository artifacts only.

Procedure

1. Ensure /orchestra-config.json exists and has key issue_tracker.
2. Execute the strict context gathering order above when in workflow-scoped mode.
3. Ensure scaffold exists:

• skills/init-architect/ARCHITECTURE.md.

1. Ensure generated artifact paths exist:

• /architecture/architecture.md.
• /architecture/docs/.

1. If architecture content currently lives under /ARCHITECTURE.md or /docs/, migrate relevant content into /architecture/ and stop updating legacy root-level copies.
2. Determine mode:

• If scoped payload is present, run scoped mode.
• Otherwise run full mode.

1. In full mode, analyze repository structure (do not read every source file):

• stack, entry points, boundaries, data layer, integrations, auth, sensitive areas, conventions.

1. In full mode, write/update /architecture/architecture.md as a concise index:

• system overview.
• component map.
• sensitive areas.
• conventions.
• reference docs table.

1. In full mode, create/update /architecture/docs/*.md only for domains that need detail beyond the index.
2. In scoped mode, use provided PR scope first and update only impacted architecture artifacts:

• map changed files and matched cases to existing sections in /architecture/architecture.md.
• update only relevant sections in /architecture/architecture.md.
• update or create only the necessary /architecture/docs/*.md files tied to those sections.
• keep unrelated sections/documents untouched.

1. In scoped mode, preserve current architecture structure and terminology unless the PR clearly introduces a new boundary or domain.
2. Keep index shallow; move depth to domain docs.
3. If generated docs change domain/sensitive-area mapping, update relevant sections in skills/architect-agent/SKILL.md.
4. In workflow-scoped mode, post handoff JSON with status: completed or status: blocked.

Guardrails

• This skill owns architecture artifact generation; architect-agent must not regenerate architecture docs.
• Do not overwrite skills/init-architect/ARCHITECTURE.md template files.
• Keep generated architecture artifacts under /architecture/ (/architecture/architecture.md, /architecture/docs/*.md).
• Prefer clarity over coverage; do not scan the full codebase unnecessarily.
• In scoped mode, do not perform a full repository rescan unless scoped context is insufficient.
• In scoped mode, prioritize consistency with existing architecture docs and produce minimal diffs.

Handoff

Primary consumers: architect-agent, pr-review-agent (scoped invocation).