This skill has been retired (2026-09-17): consumer-side links have been removed (
~/.agents/skills/core-rulesdoes not exist); the source of truth is retained for traceability only. Final destinations of the four rules:
- No plaintext passwords → user-level memory
~/.commandcode/AGENTS.md"General Hard Constraints" (loaded by every project at that layer; the permanent home of hard rules)- Permission confirmation → Deprecated(each agent has its own permission control; the rule is redundant)
- Long-task feedback → Deprecated(each agent already has progress feedback capability)
- Session progress → consolidated into a single handoff document, see
project-handoff(formerlyproject-intake; no longer maintained separately.agents/session.md)The following is the original text at the time of retirement,for traceability only; do not use it directly—if it needs to be re-enabled, first verify whether the above destinations have been covered by a new single source.
The following 4 rules apply to all projects.
1. No plaintext passwords
Sensitive credentials such as passwords, API Keys, Tokens, Secrets, etc.,must not appear in plaintextin code, configuration, logs, or conversation output.
- Code reads credentials from project
config/secrets.jsonRead credentials - Use placeholders in examples/templates (
your-key-here) - Do not expose real keys in conversations
2. Permission confirmation
For write operations (file modifications, Git, script execution, network requests, remote commands, etc.),you must first explain and obtain confirmation。
- Except for read-only: pure queries, pulling logs, listing items, and generating plans can be done first
- Confirmation content: scope + action depth + impact and rollback
- Do not default to full-scale or destructive steps because of a colloquial shorthand (e.g., "clean it up")
3. Long-task feedback
For tasks estimated to exceed 60 seconds, progress must be continuously reported during execution.
- Prohibitedblind sleep / fixed countdown as the primary waiting method
- Prefer completion signals (
exit_code, progress JSON), then short polling - At least one progress update within 60 consecutive seconds; even if there is no change, report "still running"
- Failure: self-check → locate → report → wait for user decision(unless automatic retry has been authorized)
- When stable polling is needed, use
task-loop-progressSkill
4. Session progress
Each project independently maintains its own session.md, placed in <project>/.agents/session.md。
the workspace root directory's .agents/session.md stores only cross-project information (subproject list, global pending directions), and does not store the progress of any single project.
Layering
<project>/.agents/session.md: the project's active items + pending directions + key context (reference thedocs/path only; do not record it repeatedly)docs/: EVOLUTION, ROADMAP, etc. (project-built, not mandatory)
Format
### <事项名>
- 状态:进行中 / 等待 / 阻塞
- 下一步:1-3 条可执行动作
- 断点:路径、命令、ID 等硬事实
- 待决策:要用户拍板的问题(无则写"无")
## 待定方向
- [ ] 方向 — 优先级:高/中/低 — 来源:何时提出
Delete upon completion; do not keep history.
Wrap up
Updatethe current project's session.md, and fill it in first if missing. Do not wrap up empty.
Start work
Readthe current project's session.md → Recap: status, next step, breakpoint, pending decisions → Stop and wait for instructions。
For modifications, use Edit to add or delete items one by one; overwriting the entire text with Write is prohibited.