[bai-uipath]

The codex harness symlinks each skill into .agents/skills/, and those symlinks can leave a codex agent without its skill: codex lists the skills under the plugin's namespace, models sometimes open a path that doesn't exist, and the search they fall back on can't see symlinks.

Changes

• Copy skills instead of symlinking them. The symlinks cause two problems that compound:
  • Codex follows a symlinked skill to its source, finds the skills repo's .claude-plugin/plugin.json there ("name": "uipath"), and lists the skill as uipath:<skill>. Models sometimes turn that name into the folder .agents/skills/uipath/<skill>/.
  • The search that follows (rg --files .agents/skills, find .agents/skills -name SKILL.md) skips symlinks and returns nothing, so the agent continues without the skill.
• Copy before each turn, remove after it. The copies are also removed in stop() and kill_sync(). Removing them only in stop() is not enough, because grading runs before agent.stop(): 31 skill dirs (about 18 MB, 1,800 files) would sit in front of criteria and end up in run artifacts. OpenCode avoids writing skills into the sandbox for the same reason.
• Workspace content is left alone. A skill dir the task workspace already ships is not overwritten or removed, and files the agent writes under .agents/ stay.
• _setup_skills is split into three methods. _collect_skills resolves the skill dirs once at start(), and _install_skills / _remove_skills run around each turn.
• Docs updated. docs/agents/CODEX.md, docs/agents/ANTIGRAVITY.md, the skills section of .claude/notes/agents.md, and the wording in the task guide and test docstrings.
• Sample task tasks/agents/codex_skills_test.yaml rewritten. It asserted .agents/skills exists at grading time, which per-turn removal makes false. It now has the agent write the output of find .agents/skills -name SKILL.md | wc -l to a file, and checks that .agents is gone before grading.

Validation

Codex 0.156.1 (the version on main) debug prompt-input, run on the Docker sandbox layout with the real skills repo:

Check	Symlinks (main)	Copies (this PR)
Skill name codex shows the model	uipath:<skill>	<skill>
Skill path codex shows	r1/<skill>/SKILL.md	r1/<skill>/SKILL.md
find .agents/skills -name SKILL.md	0 files	31 files
Workspace after the turn	.agents/skills/ with 31 symlinks	empty

Live runs of tasks/agents/codex_skills_test.yaml (codex gpt-5.5, PLUGIN_PATH = the skills repo, 31 skills):

Code	Driver	find count the agent wrote	.agents in artifacts	Result
main (3ffb2ef)	tempdir	0	yes	0/1
this PR	tempdir	31	no	1/1
this PR	docker (agent image built from this branch, skills copied from /work/plugins/0)	31	no	1/1

Tests	Result
New TestSkillCopies (5 tests)	pass; the os.walk test fails when _install_skills symlinks
test_codex_agent.py + test_codex_agent_unit.py (codex extra installed)	176 passed
make verify (format, ruff, pyright, CE lint rules, docs budget, tests)	6,330 passed, 93% coverage; the 6 local failures (test_stats_nonfinite, one test_agent_context_window, two test_docker_runner_mounts) fail the same way on a clean main checkout
CI Live Integration Tests (Codex)	pass

🤖 Generated with Claude Code

[github-actions]

Claude finished @bai-uipath's task in 1m 43s —— View job

---

🔍 Code Review in Progress

Task List:

• Read .github/code_review.md for review standards and format
• Read CLAUDE.md for project conventions
• Get full diff with git diff origin/main...HEAD
• Read each changed file in full for context
• Perform cross-file consistency checks
• Analyze what's missing
• Provide structured review feedback

[github-actions]

Claude finished @bai-uipath's task in 1m 30s —— View job

---

Code Review in Progress

Todo List:

• Read code review guidelines (.github/code_review.md)
• Read project conventions (CLAUDE.md)
• Examine the full diff (git diff origin/main...HEAD)
• Review each changed file with full context
• Perform cross-file consistency checks
• Analyze what's missing
• Provide design-level scrutiny
• Post comprehensive review feedback

Starting review...