TECHNAI distributes instructions and executable resources to coding agents. Treat every skill, script, reference, and operational document as code that can change agent behavior.
| Source | Trust level | Policy |
|---|---|---|
skills/ |
Owned | Review, validate, commit, and sync. |
docs/ |
Owned | Public operating guidance; review before release. |
scripts/ |
Owned executable tooling | Test and review with higher scrutiny. |
registry/reference_only |
External | Track location only; do not copy. |
| Plugin cache | External | Let plugin managers update it. |
| Runtime system skills | External | Do not edit or override. |
| Backups | Local recovery | Do not publish without review and sanitization. |
Never store:
- API keys
- GitHub tokens or PATs
- passwords
- cookies
- private keys
- production database URLs with credentials
- cloud provider credentials
- unredacted private session logs
Use placeholders such as <TOKEN>, <PROJECT_ID>, <HOST>, and <USER>.
Recommended scan:
rg -n "ghp_|github_pat_|OPENAI_API_KEY|ANTHROPIC_API_KEY|GEMINI_API_KEY|password|secret|token|private key" .Review matches manually; examples and policy text can be benign, but real credentials block release.
Do not promote content into a broader scope than the evidence supports:
- user-private content stays local unless explicitly approved and sanitized
- project content stays project-scoped
- tool content requires a tool detector or clear install guidance
- universal content must be safe across runtimes and projects
- secret content is forbidden
See reference-library/scope-model.md.
Bundled scripts must:
- use explicit arguments
- print clear failure messages
- avoid destructive operations by default
- avoid hardcoded machine-specific paths unless documented
- avoid network writes unless the skill explicitly requires them
- support dry-run or preview behavior when practical
Do not vendor plugin cache contents into skills/. If third-party content is
worth maintaining internally, fork or rewrite it intentionally, record
provenance, and review it as owned content.
Skills must not assume automatic permission to:
- run shell commands
- access network resources
- read private files
- edit files
- push commits
- contact production systems
State required tools, permissions, and safety constraints explicitly.
Future session audit and optimization features must default to:
- local-first collection
- redaction before export
- no raw prompts by default
- no raw terminal output by default
- explicit user or operator approval before sharing
- security review before propagation
Tracing, if enabled, must allow sensitive inputs and outputs to be excluded.