DAF Intent Guard — 意图漂移检测

Tracks intent drift across multi-turn tasks. Before executing any multi-step task, call this to detect if the user's new instruction is a small patch, requir...

Security Scan

The skill's code, runtime instructions, and requirements are consistent with its stated purpose (measuring intent drift) and it does not request extra credentials, network access, or persistent privileges.

Purpose & Capability

Name/description (intent-drift detection) match the included Python implementation: the script computes semantic distance, anchor rollback, and emits patch/replan/abort decisions. Required binary (python3) is appropriate and proportionate.

Instruction Scope

SKILL.md instructs only running the local Python script with JSON arguments. The script reads only its CLI args and performs local computation; it does not read unrelated files, environment variables, or send data externally.

Install Mechanism

No installation/spec downloads are present; the skill is delivered as a small local Python file and README. There are no external install URLs or archive extraction steps.

Credentials

No environment variables, credentials, or config paths are requested. The script takes inputs via CLI JSON parameters only, which is proportional to the task.

Persistence & Privilege

always is false and the skill does not request to modify other skills or agent-wide settings. It does not attempt to persist secrets or alter system configuration.

Assessment

This skill appears internally coherent and low-risk: it runs a local Python script to compute intent-drift decisions and does not contact external endpoints or ask for credentials. Before installing, verify the file origin (source is listed as unknown) and review the daf_guard.py contents yourself (they are small and readable). When integrating, pass only non-sensitive JSON (the script logs/prints decisions) and test on safe example tasks to confirm behavior meets your agent's expectations.

DAF Intent Guard

Use this skill when:

• A user modifies an in-progress task
• A user gives a new instruction that might conflict with what is already done
• You need to decide whether to continue, roll back, or restart

How to use

Call the guard before executing any new instruction on an ongoing task:

python3 {baseDir}/daf_guard.py \
  --current-anchor <index> \
  --old-constraints '<json>' \
  --new-constraints '<json>' \
  --new-action '<action>' \
  --requires '<json_array>' \
  --anchors '<json>'

Decision meanings

• patch: Small change. Update parameters and continue from current anchor.
• replan: Significant change. Roll back to anchor first_affected_anchor and re-execute from there.
• abort: Task cancelled or too much has changed. Stop and confirm with user before restarting.

Example

User was booking a flight (seat locked at anchor 1), now says "change destination to New York":

• old-constraints: {"destination":"KL","date":"today","cabin":"economy"}
• new-constraints: {"destination":"New York","date":"today","cabin":"economy"}
• Expected: replan (destination change affects anchor 0)