[Kaap10]

While exploring the agent-guard dispatcher and auditing how pre-execution hooks discover skill guards across different harnesses, I noticed that agent_guard.discover_guards() was returning only 1 guard (no_verify_commit.py from guards.d/), completely omitting the domain-specific safety guards documented in the README (mention, mark_ready, and security_language).

Tracking down the discovery logic, guard_dirs() was still scanning (root / "skills").glob("*/guards"). Following the modular plugin migration (#1087 / #1672), canonical guards live under plugins/*/skills/*/guards. On checkouts where skills/* entries are symlinks (or in adopter repositories deploying plugins directly without top-level symlink mirrors), the glob returned empty and the guards silently dropped.

Summary

• Extend agent_guard.guard_dirs() to discover contributed skill guards under plugins/*/skills/*/guards in addition to skills/*/guards (retaining backwards compatibility).
• Update framework_root() to recognize either skills/ or plugins/ sitting beside tools/agent-guard.
• Deduplicate discovered guard directories using d.resolve() so environments exposing both canonical plugin paths and skills/ symlinks do not register duplicate guard instances.
• Align tools/agent-guard/tests/test_skill_guards.py and test_discovery.py with canonical plugin directories, cross-platform path handling, and new regression unit tests.

Type of change

• Python package (tools/*/ with pyproject.toml)
• CI / dev loop (prek, workflows, validators)

Test plan

• ruff check and ruff format --check pass with zero warnings across all touched files.
• mypy --platform linux tools/agent-guard/src/agent_guard passes (Success: no issues found in 3 source files).
• pytest tools/agent-guard/tests/test_skill_guards.py tools/agent-guard/tests/test_discovery.py passes cleanly (87 passed, 2 skipped on Windows non-admin symlink privileges).
• Regression tests added:
  • test_plugin_skill_guards_are_discovered()
  • test_guard_dirs_deduplicates_symlinks()
• Verified in Python: agent_guard.discover_guards() discovers all 4 active guards (no_verify_commit, mention, mark_ready, security_language).

RFC-AI-0004 compliance

• Sandbox — ensures pre-execution action guards are reliably active across all harnesses and plugin environments.

Linked issues

Fixes #1756

[Kaap10]

Hi @potiuk - while auditing agent-guard hook discovery across runtimes, I noticed discover_guards() only returned the single builtin guard (no_verify_commit.py) and missed the canonical domain guards (mention, mark_ready, security_language).

Post-modularization (#1087 / #1672), guards live under plugins/*/skills/*/guards, but guard_dirs() was still scanning skills/*/guards. On checkouts where skills/* are symlinks (or in adopter setups without top-level symlink mirrors), the glob silently returned empty.

This PR adds plugin guard discovery, hardens directory deduplication with d.resolve(), aligns test paths, and adds regression tests.

[potiuk]

Discovering guards under plugins/*/skills/*/guards and deduplicating by resolved path is the right fix; I merged it with current main locally, the isolated-setup fingerprint still holds and the agent-guard suites pass. One follow-up outside this PR's scope: the separately published magpie-agent-guard marketplace plugin has neither skills/ nor plugins/ beside the engine, so framework_root() is still None there and the skill guards stay undiscovered on a marketplace install — worth its own issue.

---

This review was drafted by an AI-assisted tool and
confirmed by an Apache Magpie maintainer. The maintainer
approving this PR has read the findings and signed off. If
something feels off, please reply on the PR and a maintainer
will follow up.

More on how Apache Magpie handles maintainer review:
Contributing guide.