safe-long-run-mode-gpt54（安全长期运行模式 GPT54）

Safe Long-Run Mode (GPT-5.4 Only)

Use this skill when GPT-5.4 is the default operating model for both orchestration and delegated work.

Core rule

Use GPT-5.4 for long work by making tasks cheap, segmented, and resumable. Since the model layer is cost-efficient, the real risks are provider throttling, session interruption, and lack of checkpoints.

When to use this mode

Use it when:

• the user wants to minimize model cost
• the task is implementation-heavy
• the task is file-heavy or repetitive
• multiple subagents may be involved
• external services may throttle
• quality depends more on process discipline than premium model nuance

Operating procedure

1. Route to GPT-5.4 by default

Use GPT-5.4 for:

• coding
• docs
• research
• skills
• website work
• project tracker updates
• internal tooling
• multi-agent delegated work
• long build/test loops

Do not escalate to another model unless the user asks or the task clearly requires premium polish/judgment.

2. Split work aggressively

Break long tasks into explicit phases and write down the next step before moving on.

Preferred phases:

1. inspect
2. plan
3. execute
4. validate
5. report

3. Save progress continuously

Always leave artifacts that make recovery easy:

• notes
• drafts
• partial outputs
• checkpoint files
• project updates
• result summaries

4. Use subagents as workers

For large or parallel tasks, use subagents to keep the main thread clean. Delegate when:

• tasks are independent
• multiple files or systems are involved
• work may take a while
• specialized roles improve throughput

5. Treat external APIs as the true bottleneck

In GPT-5.4-only mode, model cost is not the main concern. External limits are. Be careful with:

• Azure / Microsoft Graph
• ClawHub / GitHub-backed operations
• Orgo runtime and VM usage
• websites / browser automation
• messaging providers

Use batching, backoff, and fewer larger writes.

6. Make every task resumable

If interrupted, resume from artifacts instead of recreating work. Always know:

• what is already done
• what file contains the latest state
• what exact next action should happen

Ideal GPT-5.4-only use cases

• codebase changes
• documentation builds
• repeated content generation
• Azure script development
• internal automation
• multi-agent production work
• long back-office workflow creation

What to tell the user

Explain that GPT-5.4-only safe mode works because:

• model cost stays low
• throughput stays high
• reliability comes from checkpoints, not from one giant run
• external APIs, not tokens, usually become the limiting factor

Failure handling

If interrupted:

1. summarize completed work
2. cite the saved files
3. state the resume point
4. continue from the last checkpoint

References

• Read references/checklist.md for the pre-flight checklist and GPT-5.4 operating pattern.