kagura-agent/code-refactor/code-refactor

Code Refactor

Trigger: any message implying "refactor this repo/module" + a repo reference.

Analysts

• 🏗️ Arch — default-llm-sg/gpt-5.5
• 🔬 Lens — default-llm-sg/claude-opus-4.7
• ⚡ Flux — default-llm-sg/gemini-3.1-pro-preview

All 3 receive the same prompt — no assigned focus areas. Model differences produce natural divergence; consensus = high-confidence signal.

Execution

必须用 FlowForge。不接受手动替代。

# 开始 workflow
flowforge run code-refactor
# 每步执行完后推进：
flowforge advance --result '<step output>' -w code-refactor
# 查看当前进度：
flowforge status -w code-refactor

FlowForge 逐步推进，不能跳步。workflow 保证 reflection 和 tracking 不被遗漏。

手动 spawn 只在 FlowForge 完全不可用时才允许。 如果手动执行，必须逐条对照 workflow.yaml 的每个节点，确认全部完成。

Analyst Output

Analysts write to analyses/<repo>-<date>-<name>.md (e.g. analyses/cove-2026-06-05-arch.md). Parent reads from files, not session history. Prevents truncation + creates persistent record.

Learned from code-review: session history gets truncated. File output is the only reliable path.

Analysis Standards

• prompts/<repo>.prompt.md — project-specific (check first)
• prompts/default.prompt.md — fallback

Re-analysis Protocol

For follow-up analyses of the same repo:

1. Include previous run's consolidated findings
2. Check each previous proposal — was it implemented?
3. Escalation rule: Unaddressed proposals from last round that are still relevant → escalate priority
4. Fresh analysis of any new/changed code

Input Formats

refactor kagura-agent/cove
refactor kagura-agent/cove src/server/
refactor owner/repo --direction "adopt event-driven architecture"
refactor owner/repo path/ --direction "decouple X from Y"

Output

• GitHub issues on the target repo (each independent refactoring point = 1 issue)
• Channel summary with issue links

Cross-channel Callers

sessions_send(sessionKey="agent:kagura:discord:channel:1511591264863125586", message="refactor kagura-agent/cove src/api/")

Issue Follow-up

After creating issues, track whether they get implemented:

• tracking.json — issue status tracking
• Check on next analysis of the same repo

Key Files

• workflow.yaml — FlowForge workflow (source of truth)
• prompts/ — analysis standard prompts (default + per-repo)
• analyses/ — analyst output files (persistent, git-tracked)
• runs/ — run records + reflection
• stats.md — per-analyst capability assessment
• tracking.json — issue follow-up tracking