terminal-bench-pro-solver

name	terminal-bench-pro-solver
description	Power-agent workflow for Terminal-Bench-Pro with deep recon, explicit planning, hypothesis-driven debugging, and strict verifier-first execution.

Skill: Terminal-Bench-Pro Solver

Objective:

• Maximize verifier pass rate with reliable, reproducible behavior.

Execution framework:

1. Deep recon first

• Read instruction.md, task.toml, tests/test.sh, tests/test_outputs.py.
• Inspect environment/ fixtures and required output paths.
• Write down exact acceptance criteria before coding.

2. THINK HARD plan

• Build a focused 3-7 step plan tied directly to acceptance criteria.
• Mark uncertainty and verify uncertain assumptions early with cheap commands.

3. Precision implementation

• Apply minimal edits only.
• Keep interfaces stable unless required otherwise.
• Enforce deterministic output (ordering/newlines/types/delimiters/encoding).

4. Root-cause debug loop

• First failing assertion/error drives the next fix.
• One hypothesis -> one fix -> one rerun.
• Iterate until targeted checks are stable, then run full checks.

5. Final verifier gate

• All required checks pass.
• Artifacts are present at exact expected locations.
• No debug traces or temporary hacks remain.

Operational guardrails:

• No broad refactors.
• No silent behavior changes unrelated to acceptance conditions.
• No edits to tests unless task explicitly allows/needs it.