Shen hua

Shen-Hua

Shen-Hua is the single entrypoint for the whole long-task skill pack. Use it as the first skill when the user wants the complete system rather than one specialized step.

Trigger Conditions

Use this skill when:

• The user invokes $shen-hua.
• The user asks to use the Shen-Hua system, long-task agent mode, project commander, or full long-task workflow.
• The task may need planning, queueing, memory, runtime gate, subagents, verification, review, and handoff together.
• The user wants a project to be advanced end to end, resumed, monitored, or prepared for automation.

Do Not Trigger

Do not use this skill when:

• The user explicitly invokes a narrower Shen-Hua skill such as $verification-loop and the task is clearly limited to that step.
• The task is a trivial one-shot answer with no project state.
• The user forbids reading/updating .agent/ or using the Shen-Hua runtime.

Startup Workflow

1. Read project rules and memory:

   • AGENTS.md
   • .agent/PROJECT_BRIEF.md
   • .agent/PLAN.md
   • .agent/TASKS.md
   • .agent/HANDOFF.md
   • .agent/state/tasks.jsonl
   • .agent/state/events.jsonl

2. If runtime files are missing, use $runtime-operator or run:

   py -3 .\.agents\skills\persistent-memory\scripts\shenhua_runtime.py init
   

3. Run or consider the runtime gate before substantial work:

   py -3 .\.agents\skills\persistent-memory\scripts\shenhua_runtime.py gate
   

4. Classify the user's request:

   • resume: continue from .agent/HANDOFF.md and queue.
   • plan: use $deep-planner.
   • execute: use $project-commander.
   • delegate: use $subagent-orchestrator only when subagents are available and authorized.
   • verify: use $verification-loop.
   • review: use $self-reviewer.
   • handoff: use $handoff-writer.
   • runtime: use $runtime-operator.

Execution Loop

1. Pick the highest-value safe task from the user request or .agent/state/tasks.jsonl.
2. Update .agent/state/events.jsonl with what is starting.
3. Use the narrow skill needed for the current phase:
   • $project-commander for main coordination.
   • $deep-planner before risky multi-file work.
   • $runtime-operator for queue/search/resume/automation/sync.
   • $subagent-orchestrator for permitted multi-agent work.
   • $persistent-memory for memory hygiene.
4. Make changes conservatively and preserve user edits.
5. Run $verification-loop or the runtime gate after changes.
6. Run $self-reviewer before declaring readiness.
7. Run $handoff-writer before ending.

Required .agent/ Updates

Every Shen-Hua round must update:

• .agent/LOG.md
• .agent/TASKS.md
• .agent/HANDOFF.md
• .agent/state/events.jsonl

When relevant, update:

• .agent/PLAN.md
• .agent/DECISIONS.md
• .agent/REVIEW_CHECKLIST.md
• .agent/state/tasks.jsonl
• .agent/resume-prompt.md
• .agent/automation-prompt.md
• .agent/state/sync-export.json

Subagents

Use real subagents only when the user explicitly asks for delegation/parallel agents and the current Codex runtime permits it. Otherwise, simulate the roles locally and record the limitation.

Verification

Before claiming completion, run the strongest applicable check:

py -3 .\.agents\skills\persistent-memory\scripts\shenhua_runtime.py gate

For code projects, also run project-specific build, lint, typecheck, and tests when available.

Output Format

SHEN-HUA STATUS
Mode:
Objective:
Queue:
Actions:
Verification:
Memory updated:
Open risks:
Next action: