Context Model

March treats the initial Agent Run context as reconstructed state. During an Agent Run, pi-agent advances by appending transcript entries for model output, tool calls, tool results, and hidden steer messages.

Stable instructions
  → Session identity
  → Recent chat summary
  → Recall hints
  → Tool-driven reads of current project facts

Principle

The repository is the source of truth. March reads files and terminal state when needed instead of pinning stale snapshots in the prompt.

Layers

• system_core: March behavior, safety rules, model-specific additions
• injections: explicit external instructions from MCP or extensions
• session_identity: cwd, workspace root, memory root, platform, shell
• recent_chat: recent Agent Runs plus compact recall hints

For the runtime boundary, see Runtime Core Boundary. For the full design note, see Context Core.