fix: extract agent markdown body as system instruction

[samueljklee]

Problem

When loading agent .md files via _load_agent_file_metadata(), the markdown body after frontmatter was being discarded (assigned to _body and ignored). This body contains the agent's behavioral instructions that should be injected as the system prompt when the agent is spawned via the task tool.

Fixes: microsoft/amplifier#174

Root Cause

Two issues combined to cause this bug:

1. _load_agent_file_metadata() discarded the body: The markdown content after frontmatter was parsed but assigned to _body (unused variable) and never stored.

2. load_agent_metadata() was never called: The method existed but Bundle.prepare() never invoked it before generating the mount plan.

Impact

Agents like smoke-tester that rely on their markdown instructions to guide behavior (e.g., "run this exact recipe path") would receive no system prompt and fall back to generic exploration behavior, causing unpredictable results.

Solution

1. Modified _load_agent_file_metadata() to capture the body and store it as system.instruction
2. Added call to self.load_agent_metadata() in Bundle.prepare() before to_mount_plan()

Changes

• amplifier_foundation/bundle.py:
  • prepare(): Added call to self.load_agent_metadata() before to_mount_plan()
  • _load_agent_file_metadata(): Changed _body to body, added logic to store as result["system"] = {"instruction": body.strip()}

Tests

Added 3 tests in TestAgentMetadataLoading:

• test_load_agent_metadata_extracts_system_instruction - verifies body becomes system.instruction
• test_load_agent_metadata_preserves_existing_system - verifies explicit inline instructions aren't overwritten
• test_load_agent_metadata_empty_body_no_system - verifies empty body doesn't create spurious instruction

All 21 tests pass.

---

🤖 Generated with Amplifier

Co-Authored-By: Amplifier 240397093+microsoft-amplifier@users.noreply.github.com

[samueljklee]

Fixed in #32 .