feat: improve generated agent prompts

src/rdb_discovery/tasks.py

@@ -10,6 +10,7 @@ STATUS_RE = re.compile(r"^Status:\s*(.+)$", re.MULTILINE)
 
 GOAL_RE = re.compile(r"^Goal:\s*(.+)", re.MULTILINE)
 AC_RE = re.compile(r"^- (.+)$", re.MULTILINE)
+GAP_RE = re.compile(r"^Implementation Gap:\s*(.+)", re.MULTILINE)
 
 
 @dataclass
@@ -28,6 +29,12 @@ class Task:
     def ac_lines(self) -> list[str]:
         return AC_RE.findall(self.body)
 
+    def implementation_gap(self, root: Path) -> str | None:
+        match = GAP_RE.search(self.body)
+        if match and match.group(1).strip():
+            return match.group(1).strip()
+        return None
+
 
 def read_tasks(root: Path) -> list[Task]:
     tasks_path = root / "TASKS.md"
@@ -92,6 +99,12 @@ CONTEXT_FILES_TO_READ = [
     "context/agent-guidelines.md",
 ]
 
+TEST_CMD_SECTION_RE = re.compile(
+    r"^##\s*Test Commands\s*\n([\s\S]*?)(?=^##|\Z)",
+    re.MULTILINE,
+)
+BASH_BLOCK_RE = re.compile(r"```bash\n(.*?)```", re.DOTALL)
+
 
 def _read_file_safe(root: Path, relative: str) -> str:
     path = root / relative
@@ -100,6 +113,55 @@ def _read_file_safe(root: Path, relative: str) -> str:
     return f"# {relative} — not found"
 
 
+def _extract_test_commands_from_claude(root: Path) -> list[str]:
+    """Extract bash test commands from CLAUDE.md Test Commands section."""
+
+    def _looks_like_command(line: str) -> bool:
+        """Heuristic: keep lines that look like shell commands."""
+        if not line.strip():
+            return False
+        lower = line.lower()
+        # Skip markdown, prose, headings
+        if any(lower.startswith(p) for p in ("# ", "- ", "if ", "do not", "for ", "use ")) or lower in (
+            "validation commands",
+        ):
+            return False
+        # Skip lines that look like sentences (contain spaces followed by lowercase words)
+        # but keep things like `source .venv/bin/activate` and `pip install ...`
+        parts = line.split()
+        if len(parts) <= 1:
+            return True
+        # If the first word is a known shell builtin / command prefix, accept it
+        known_prefixes = ("source", "cd", "ls", "cp", "mv", "rm", "mkdir", "echo", "grep",
+                          "git", "pip", "python", "pytest", "rdb", "cat", "head", "tail",
+                          "sed", "awk", "find", "install")
+        return parts[0].lower() in known_prefixes
+
+    claude_path = root / "CLAUDE.md"
+    if not claude_path.exists():
+        return []
+
+    text = claude_path.read_text(encoding="utf-8")
+    section_match = TEST_CMD_SECTION_RE.search(text)
+    if not section_match:
+        return []
+
+    section_text = section_match.group(1)
+    blocks = BASH_BLOCK_RE.findall(section_text)
+    seen: set[str] = set()
+    commands: list[str] = []
+    for block in blocks:
+        for line in block.strip().splitlines():
+            stripped = line.strip()
+            if not stripped:
+                continue
+            if _looks_like_command(stripped):
+                if stripped not in seen:
+                    seen.add(stripped)
+                    commands.append(stripped)
+    return commands
+
+
 def generate_agent_prompt(root: Path) -> str:
     task = get_next_task(root)
     if not task:
@@ -128,9 +190,20 @@ def generate_agent_prompt(root: Path) -> str:
         f"Status: {task.status}",
         "",
         f"Goal:\n{task.goal(root)}",
+    ]
+
+    gap = task.implementation_gap(root)
+    if gap:
+        sections.extend([
+            "",
+            "Implementation Gap:",
+            gap,
+        ])
+
+    sections.extend([
         "",
         "Acceptance Criteria:",
-    ]
+    ])
 
     for line in task.ac_lines():
         sections.append(f"- {line}")
@@ -141,8 +214,25 @@ def generate_agent_prompt(root: Path) -> str:
         "",
         "# Constraints",
         "",
-        "- Implement ONE task only. Do not combine with other tasks.",
+        "- Inspect the codebase first, then edit.",
+        "- Do not repeatedly reread unchanged files.",
+        "- Limit your work to ONE small implementation step only. Do not combine tasks or features.",
         "- Make the smallest useful change possible.",
+    ])
+
+    test_cmds = _extract_test_commands_from_claude(root)
+    if test_cmds:
+        sections.extend([
+            "",
+            "---",
+            "",
+            "# Test Commands (from CLAUDE.md)",
+            "",
+            "Use these existing test commands:",
+            ("```bash\n" + "\n".join(test_cmds) + "\n```\n"),
+        ])
+
+    sections.extend([
         "",
         "---",
         "",