Description

Extend the agent loop to detect fenced bash code blocks in LLM output, execute them, and feed results back to the LLM for a final response.

Parent: #3

Branch

feat/m2/shell-execution

Files

• crates/zeph-core/src/context.rs — context builder (system + skills + history + message)
• crates/zeph-core/src/agent.rs — shell detection, execution, re-query loop
• crates/zeph-core/Cargo.toml — add zeph-skills dependency

Detection Logic

1. LLM returns response text
2. Scan for ```bash\n...\n``` blocks
3. If found: execute via tokio::process::Command
4. Capture stdout + stderr
5. Send back to LLM: "Command output:\n{stdout}\n\nFormat a response for the user."
6. Return final response to user

Acceptance Criteria

• Context includes system prompt + skill instructions
• Bash code blocks detected in LLM output
• Commands executed with 30s timeout
• Output fed back to LLM for final answer
• Non-bash code blocks ignored (not executed)
• Command failure (non-zero exit) included in output

Depends On

#13 (Skill registry), #14 (Prompt formatter), #15 (Bundled skills), #11 (CLI agent loop)