The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill accesses sensitive local credential files to retrieve API tokens for various AI services.
Evidence: scripts/t3-quota.mjs reads authentication data from several locations, including ~/.codex/auth.json, ~/.local/share/devin/credentials.toml, ~/.claude/settings.json, and ~/.factory/.env.
[DATA_EXFILTRATION]: The skill extracts authentication secrets from private application storage and transmits them to external service endpoints.
Evidence: scripts/t3-quota.mjs accesses the local SQLite database used by the Cursor editor (state.vscdb) to extract private cursorAuth/accessToken tokens.
Evidence: The script performs network operations to non-whitelisted domains (e.g., chatgpt.com, cursor.com, app.devin.ai, api.factory.ai, open.bigmodel.cn) to check account usage using the retrieved credentials.
[COMMAND_EXECUTION]: The skill executes local scripts and system utilities to perform its management tasks.
Evidence: SKILL.md instructs the agent to run various Node.js scripts in the scripts/ directory, such as route.mjs and t3-quota.mjs.
Evidence: scripts/t3-quota.mjs executes the sqlite3 command-line utility to query local databases if the internal SQLite library is unavailable.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to instruction poisoning from untrusted data generated by sub-agents.
Ingestion points: The skill ingests the summary and activity logs of delegated sub-tasks via t3_thread_read to inform subsequent orchestration decisions.
Boundary markers: The skill uses Markdown headers (## 背景, ## 任务) to separate context, which may not prevent the agent from obeying instructions embedded in sub-agent outputs.
Capability inventory: The skill has the capability to write files, execute shell commands through local scripts, and delegate tasks with inherited runtime permissions.
Sanitization: While the skill instructs the main agent to manually verify sub-agent outputs, it lacks explicit programmatic sanitization or escaping of the interpolated content.
[DYNAMIC_EXECUTION]: The skill generates local dashboard content using runtime data.
Evidence: scripts/route-playground.mjs generates an HTML file (route-playground.html) by injecting local routing configuration and capability data into a template for visualization purposes.

t3-orchestrator