feat(pipeline): agent-mode reliability improvements

[christso]

Summary

This PR addresses four issues identified from observing an agent struggle to run evals reliably via the agent-mode pipeline on Windows. All changes are in the pipeline commands and the bundled agentv-bench skill.

Issues addressed

#789 — Fix Windows subprocess in Python scripts

The bundled scripts (run_tests.py, run_code_graders.py, bench.py) used subprocess.run(["agentv", ...]) which fails on Windows because agentv resolves to a .ps1 wrapper. Changed to shutil.which("agentv") which correctly resolves the executable on all platforms.

Also added .env auto-loading to run_tests.py so target commands inherit required environment variables (e.g. SEARCH_SERVICE, OPENAI_API_KEY).

#790 — Add --llm-scores flag to pipeline bench

pipeline bench previously only accepted LLM scores via stdin. In PowerShell, the < operator is reserved, so agentv pipeline bench <dir> < llm_scores.json fails with a parse error. Added --llm-scores <path> as an optional flag. Falls back to stdin when omitted (backward compatible).

#791 — Add pipeline run combined command

Added agentv pipeline run <eval> --out <dir> which combines input extraction, CLI target invocation (parallel), and code grading into a single command. Loads .env from the eval directory automatically.

Before this PR, agent-mode required 4 steps; after, it's 2 commands plus the agent's LLM grading step:

agentv pipeline run evals/my.eval.yaml --out .agentv/results/export/run-1
# ... agent grades responses, writes llm_scores.json ...
agentv pipeline bench .agentv/results/export/run-1 --llm-scores llm_scores.json

#792 — Update agentv-bench skill docs for cross-platform use

• Changed the recommended workflow from Python wrapper scripts to direct CLI commands
• Updated the quick-reference example to show the 2-command pipeline run → pipeline bench flow
• Added --llm-scores to all bench examples
• Added guidance for non-subagent environments (VS Code Copilot, Codex) where grader subagent dispatch isn't available

Files changed

File	Change
apps/cli/src/commands/pipeline/run.ts	New pipeline run command
apps/cli/src/commands/pipeline/bench.ts	Add --llm-scores flag
apps/cli/src/commands/pipeline/index.ts	Register run subcommand
plugins/agentv-dev/skills/agentv-bench/SKILL.md	Updated instructions
plugins/agentv-dev/skills/agentv-bench/scripts/run_tests.py	shutil.which + .env loading
plugins/agentv-dev/skills/agentv-bench/scripts/run_code_graders.py	shutil.which fix
plugins/agentv-dev/skills/agentv-bench/scripts/bench.py	shutil.which fix

Closes #789
Closes #790
Closes #791
Closes #792