fix(agent): append system_prompt to the Claude Code preset instead of replacing

[Mihaiii]

Why append instead of replace

ClaudeAgentOptions.system_prompt accepts either a plain string or the claude_code preset dict. A plain string replaces Claude Code's entire default system prompt — the only way to keep the default is {"type": "preset", "preset": "claude_code", "append": ...}. coder_eval passes the experiment's system_prompt straight through as a string, so any experiment that sets even a one-line prompt silently strips every behavioral instruction the harness ships with.

That is exactly what the skills-repo experiments do. The nightly config sets an innocuous sandbox guard:

https://github.com/UiPath/skills/blob/main/tests/experiments/nightly.yaml#L39-L40

system_prompt: |
  You are a coding agent. Do not access files in sibling runs/* directories. Everywhere else is permitted.

(same pattern in tests/experiments/default.yaml#L21, smoke.yaml#L39, smoke-windows.yaml#L23, and the skill-comparison templates)

One sentence of sandbox policy costs the whole Claude Code system prompt.

Observed impact (skills nightly, skill-rpa-execution-map-greenfield)

The task is a turn-budget gate (max_turns: 10, expected 6) that assumes the agent batches tool calls per turn. Every claude-sonnet-5 run exhausted the cap; pass/fail depended on where the cap happened to land. Transcript analysis across four runs (31172161551, 31174117722 ×3 attempts, 31178128981, 31179344004 ×2 attempts):

1. Zero parallel tool calls, ever. Example distribution (run 31179344004, attempt 2): 30 assistant messages — 0 multi-tool, 19 single-tool, 11 with no tool call at all. The instruction to emit independent tool calls together in one message lives in the default system prompt; with it gone, Sonnet paid one turn per call and blew the 10-turn budget in 7/7 attempts. Skill-doc prose telling the agent to batch (added and strengthened twice in UiPath/skills) was read in-transcript and changed nothing — a reference file cannot substitute for the missing system-level contract.
2. Narration bloat. 10–11 assistant messages per run contained no tool call, just interim commentary — the default prompt's conciseness/minimal-output rules were gone.
3. Tool-choice drift. Runs used Bash cat/sed/find where the default prompt directs the dedicated Read/Grep/Glob tools (e.g. cd TextReport && cat project.json && cat Main.xaml), losing the harness's file-tracking and permission integration.

Beyond the observed items, replacing the prompt also drops the default guidance on code-reference formatting, task management, professional tone, and the security guardrails — none of which an experiment author intends to disable when adding a sandbox-scoping sentence.

The change

• claude_code_agent.py: when system_prompt is configured, wrap it as SystemPromptPreset(type="preset", preset="claude_code", append=...) so the default prompt survives and the experiment text is appended. None still means the untouched SDK default.
• agent_config.py: system_prompt field description updated ("appended to the agent's default system prompt" — previously "Replaces").
• tests/test_agent.py: two tests via the existing _capture_sdk_options pattern (append wrapping; None passthrough).

Behavioral note for existing consumers

Every experiment that sets system_prompt switches from replace to append semantics with this release. For the known consumers (sandbox-scoping one-liners) this is the intended repair. An experiment that deliberately relied on full replacement to suppress default Claude Code behavior would need a different mechanism.

Judge (agent_judge.py) and user-simulator paths construct their own options and are unaffected.

🤖 Generated with Claude Code

[github-actions]

Claude finished @Mihaiii's task in 1m 46s —— View job

---

🔍 Reviewing PR #92: Fix system_prompt append behavior

Working on review...

Todo List:

• Read .github/code_review.md for review guidelines
• Read CLAUDE.md for project conventions
• Review the full diff (git diff origin/main...HEAD)
• Read each changed file in full context
• Perform cross-file consistency checks
• Check "what's missing" analysis
• Post comprehensive review feedback