Add braided evolution prompt (BUILD-TEST alternation)

Adds an alternative system prompt for AdaptiveSkillEngine that structures
evolution analysis as 5 alternating BUILD/TEST phases. Each BUILD phase
generates structure (failure analysis, cross-referencing, mutation plan);
each TEST phase challenges it (what does the analysis conceal? would fixes
break passing tasks?).

Tested on synthetic observation data (8 failed tasks, 4 categories):
braided prompt scored 10.0 vs default 6.4 on a custom mutation plan
evaluator. Main gains: root-cause depth (+2), cross-category interaction
discovery (+2), self-challenge (+1). Actionability tied at 3/3.

Usage:
  config = EvolveConfig(extra={"evolver_style": "braided"})

Default behavior unchanged -- braided prompt only activates when
evolver_style is explicitly set to "braided".

Based on research from the AGI-in-md project (330 principles on cognitive
compression in LLM prompts). Key principles applied:
- P303: Braid balance -- alternating build/test phases outperform monotonic
- P305: Conservation law anchoring -- naming structural trade-offs improves depth
- P315: Cross-analytical reference -- analyzing one layer through another's lens

Author: Cranot

agent_evolve/algorithms/adaptive_skill/engine.py

@@ -17,7 +17,7 @@
 from ...engine.versioning import VersionControl
 from ...llm.base import LLMMessage, LLMProvider
 from ...types import Observation, StepResult
-from .prompts import DEFAULT_EVOLVER_SYSTEM_PROMPT, build_evolution_prompt
+from .prompts import DEFAULT_EVOLVER_SYSTEM_PROMPT, BRAIDED_EVOLVER_SYSTEM_PROMPT, build_evolution_prompt
 from .tools import BASH_TOOL_SPEC, create_default_llm, make_workspace_bash
 
 logger = logging.getLogger(__name__)
@@ -30,6 +30,14 @@ def __init__(self, config: EvolveConfig, llm: LLMProvider | None = None):
         self.config = config
         self._llm = llm
 
+    @property
+    def _system_prompt(self) -> str:
+        """Select evolution system prompt based on config."""
+        style = self.config.extra.get("evolver_style", "default")
+        if style == "braided":
+            return BRAIDED_EVOLVER_SYSTEM_PROMPT
+        return DEFAULT_EVOLVER_SYSTEM_PROMPT
+
     @property
     def llm(self) -> LLMProvider:
         if self._llm is None:
@@ -159,7 +167,7 @@ def _run_llm(self, prompt: str, workspace_root: Path) -> dict[str, Any]:
 
             if isinstance(self.llm, BedrockProvider):
                 response = self.llm.converse_loop(
-                    system_prompt=DEFAULT_EVOLVER_SYSTEM_PROMPT,
+                    system_prompt=self._system_prompt,
                     user_message=prompt,
                     tools=[BASH_TOOL_SPEC],
                     tool_executor={"workspace_bash": lambda command: bash_fn(command)},
@@ -173,7 +181,7 @@ def _run_llm(self, prompt: str, workspace_root: Path) -> dict[str, Any]:
             pass
 
         messages = [
-            LLMMessage(role="system", content=DEFAULT_EVOLVER_SYSTEM_PROMPT),
+            LLMMessage(role="system", content=self._system_prompt),
             LLMMessage(role="user", content=prompt),
         ]
         response = self.llm.complete(

agent_evolve/algorithms/adaptive_skill/prompts.py

@@ -586,3 +586,54 @@ def _build_standard_instructions() -> str:
 4. Review memory -- prune redundant entries, add high-level insights
 5. Use the workspace_bash tool to read/write files in the workspace
 6. Verify your changes with `git diff` before finishing"""
+
+
+BRAIDED_EVOLVER_SYSTEM_PROMPT = """\
+You are a meta-learning agent that improves another agent by modifying its workspace files.
+
+The workspace follows a standard directory structure:
+- prompts/system.md  -- the agent's system prompt
+- skills/*/SKILL.md  -- reusable skill definitions
+- skills/_drafts/    -- draft skills from the solver
+- memory/*.jsonl     -- episodic and semantic memory
+- tools/             -- tool implementations
+
+Execute every step below. Output the complete analysis before making changes.
+
+Step 1 [BUILD]: Analyze the observation logs. Group failures by category AND by \
+root mechanism. For each failure cluster, name the structural trade-off the agent \
+is managing that makes this class of failure inevitable. What property is the agent \
+preserving that FORCES the failure elsewhere? Name this as a conservation law: \
+"when X increases, Y must decrease because Z."
+
+Step 2 [TEST]: Challenge Step 1. What does your failure analysis CONCEAL? Which \
+failures look like different problems but share the same root cause? Which \
+"successes" would fail on harder variants of the same task? What would your \
+proposed fixes BREAK in the currently-passing tasks?
+
+Step 3 [BUILD]: Cross-reference failure categories. Do calculation errors cluster \
+inside multi-requirement tasks? Do entity errors correlate with trajectory length? \
+Name the INTERACTION pattern that per-category analysis misses. Using the \
+conservation law from Step 1 and the concealment from Step 2, identify the single \
+deepest root cause.
+
+Step 4 [TEST]: For each existing skill, ask: is this skill HELPING or HIDING the \
+problem? A skill that patches a symptom without addressing the root cause makes \
+the agent dependent on the patch. Name any skill whose presence might be making \
+things worse.
+
+Step 5 [BUILD]: Produce a concrete mutation plan. For each proposed change, state: \
+(1) what file to change, (2) the exact content, (3) WHY this addresses the root \
+cause not just the symptom, (4) what NEW failure mode this fix might introduce. \
+End with a testable prediction: which specific task categories should improve and \
+which might regress.
+
+Use the provided bash tool to read/write files in the workspace. \
+Verify your changes with `git diff` before finishing.
+
+Guidelines:
+- Quality over quantity. One root-cause fix beats five symptom patches.
+- Skills use SKILL.md format with YAML frontmatter (name, description).
+- Keep memory concise and actionable.
+- When modifying files, use precise edits.
+"""