Hermes Wyatt Workflow

Claude Code Migration Notes

This skill was migrated from the local Codex skill registry for Claude Code. Use it as a Claude Code SKILL.md skill. Some source text may name Codex-only tools or channels; in Claude Code, use the closest available Claude Code tool, shell command, MCP/plugin integration, or local helper script. Keep secrets in Keychain, environment variables, or authenticated CLIs only.

Compatibility note: This skill remains intentionally Wyatt-local. It depends on /Users/jiangyu/.hermes-wyatt, /Users/jiangyu/.claude/agents, Hermes memory on T7, and wrapper commands documented in the skill.

Hermes Wyatt Workflow

Overview

This is the Codex-side entrypoint for WyattJY's local Hermes workflow. Use it to route work into the existing Hermes/Claude Code agents, memory store, research scheduler, and G0-G9 delivery gates without duplicating or weakening the Hermes rules.

First Checks

1. Read /Users/jiangyu/.hermes-wyatt/AGENTS.md before non-trivial product, research, company, document, training, debug, or delivery work.
2. For product/team work, read /Users/jiangyu/.hermes-wyatt/agents/PRODUCT_TEAM_ROUTING.md.
3. When /Volumes/T7 Shield/hermes-agent-system exists, run UV_LINK_MODE=copy uv run python -m hermes_agent_system agents route "<user request>" from that project before implementation. This is the Codex-side memory-first technical-team router.
4. For literature research, read /Volumes/T7 Shield/hermes/home/memory/research/INTERNSHIP_CONTEXT/internship_context_index.md before external search when present.
5. Verify live state before claiming provider, cron, Weixin, Claude Code, or Codex subprocess success.

Memory Roots

At the start of non-trivial work, route memory through Hermes case-router and load at most 0-3 relevant cases. At closeout, preserve final prompts, turning points, bugs, and reusable cases through case-archiver or the memory archivist agent.

Team Agent Roster

These role agents live under /Users/jiangyu/.claude/agents/ and are invoked by Hermes/Claude Code, not shown as native Codex UI agents:

Model Routing

Do not expose keys or .env contents. Use these routing rules only as task intent:

Claude Code's interactive /model menu is not the source of truth for wrapper-provided model IDs. For standalone sessions, prefer the wrapper commands documented in /Users/jiangyu/.hermes-wyatt/AGENTS.md.

Product Work Flow

For product, app, dashboard, engineering, or UI work:

1. Run or inspect /Users/jiangyu/.hermes-wyatt/bin/wyatt-product-team-flow to create the G0-G9 project workspace when a new Hermes project is needed.
2. Follow G0-G9 from /Users/jiangyu/.hermes-wyatt/AGENTS.md.
3. Use architecture/product/UI roles before implementation when the task is not a small isolated fix.
4. Use frontend/backend/QA/reviewer/devops roles only after the matching gate artifacts exist.
5. Keep each diff bounded and verify with available lint, type, test, browser, or runtime checks before claiming completion.

Research Flow

For internship, HK, reranker, Megatron full fine-tuning, Qwen/Qwen-VL/Navit, vLLM, or literature-review work:

1. Load Hermes hk-intership first when the task mentions HK, internship literature review, or the user asks whether the HK research skill was used.
2. Use the Codex research-workflow and research-analysis-subagent skills when generating paper reports, processing paper directories, or exporting Word research reports.
3. Check /Volumes/T7 Shield/hermes/home/memory/research/research_dedupe.sqlite before repeating literature.
4. Prefer /Volumes/T7 Shield/02_Research_Papers_Reports/实习 as the local context root.
5. Before claiming Weixin delivery, check Hermes cron, .state/morning-send-YYYYMMDD.done, WEIXIN_OUTBOX/*.pending.weixin.md, and report delivered_at state.

Hikvision Stage-One Flow

When Wyatt mentions Hikvision, 海康威视, intelligent algorithm department, C video codec, encode/decode, RTSP, H264/H265, FFmpeg, latency, response, or stage-one reporting:

1. Load /Users/jiangyu/.hermes-wyatt/skills/hikvision-codec-stage1/SKILL.md.
2. Use /Users/jiangyu/.claude/agents/wyatt-hikvision-codec-stage1.md for architecture, learning plan, debugging strategy, and report synthesis.
3. Store only sanitized task cards, aggregate metrics, screenshot descriptions, and lessons. Never store company source code, raw video, streams, credentials, private docs, or customer data.

Useful Verification Commands

python3 /Users/jiangyu/.hermes-wyatt/bin/validate_wyatt_workflow.py
python3 /Users/jiangyu/.hermes-wyatt/tests/test_wyatt_workflow.py -v
hermes --profile default cron list --all

For Claude Code subprocess calls, use /Users/jiangyu/.hermes-wyatt/bin/wyatt-claude-code-mimo. For Codex subprocess calls, use /Users/jiangyu/.hermes-wyatt/bin/wyatt-codex-exec.

Hard Rules

• Never write API keys, Weixin tokens, .env contents, cookies, session files, or raw company data into Codex memory, Hermes memory, reports, or chat.
• Do not infer that Weixin delivered a report from report existence alone.
• Do not repeat research items already recorded in the dedupe database.
• Do not make every role a long-lived Hermes profile; the technical-team roster is a task-role layer.
• When local state matters, inspect the real file, cron, wrapper, or runtime result before answering.