Add reflector instrumentation, pluggable payload mapper, v2 templates

[jnzs1836]

Supersedes #2.

PR #2 was branched before the package was renamed to strands_harness_optimizer (and before CI was added), so its changes landed on now-nonexistent paths and it couldn't be updated in place due to branch-protection rules. This branch rebases that work onto current main:

• BaseAgenticOptimizer: adds _invoke_agent helper that times the agent invocation and captures Strands metrics (tokens, cycles, tool calls, latency) onto last_wall_clock_s / last_metrics. ContrastiveReflectionOptimizer routes through it.
• AgentCoreRolloutEngine: new optional payload_mapper: Callable[[dict], dict] init kwarg that transforms the canonical {"data_sample", "params"} payload before the HTTP call. Default None (no transform) — fully backward-compatible.
• New built-in templates contrastive_reflection_v2/{system_prompt,task_message_system_prompt}.jinja.

Original work by @hanDing (commit authorship preserved). Verified locally: pytest -m "not integration" → 39 passed.

[Copilot]

Pull request overview

This PR rebases and reintroduces prior work to add instrumentation around agent invocation, make AgentCore payloads customizable without subclassing, and ship updated “contrastive_reflection_v2” built-in templates.

Changes:

• Add BaseAgenticOptimizer._invoke_agent() to time agent calls and capture Strands usage/latency/tool metrics into last_wall_clock_s and last_metrics; route ContrastiveReflectionOptimizer through it.
• Add optional payload_mapper: Callable[[dict], dict] to AgentCoreRolloutEngine to transform the canonical {"data_sample", "params"} payload before invoking the runtime.
• Add new built-in templates under templates/contrastive_reflection_v2/.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
strands_harness_optimizer/optimizers/system_prompt/base_agentic_optimizer.py	Adds _invoke_agent() plus last_wall_clock_s / last_metrics fields to expose per-invocation cost/latency.
strands_harness_optimizer/optimizers/system_prompt/contrastive_reflection.py	Routes reflection invocation through _invoke_agent() to enable instrumentation.
strands_harness_optimizer/rollout_engines/agentcore_engine.py	Adds payload_mapper hook to reshape outgoing runtime payloads without subclassing.
strands_harness_optimizer/templates/contrastive_reflection_v2/system_prompt.jinja	New v2 system prompt template focusing on integrated edits vs appendices.
strands_harness_optimizer/templates/contrastive_reflection_v2/task_message_system_prompt.jinja	New v2 task message template enforcing structural preservation, safety invariant, and length guardrail.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jnzs1836]

I think adding try-catch may silently drop the agent failure, causing the further issue for the agent execution. Since we are running an LLM agent experiments, the failure in the agent execution should be visible to the users and they should figure out by themselves, otherwise the experiment results may be biased.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.