Skip to main content

System Prompt

For each agent run, Tyrum assembles a custom system prompt. The purpose is to provide the model with the minimum context and rules needed to act safely and effectively.

Typical sections​

  • Tooling: tool list with short descriptions
  • Tool schemas: machine-readable contracts for tool invocation (counts toward context)
  • Safety: short guardrail reminder to avoid bypassing oversight
  • Skills (when available): how to load skill instructions on demand
  • Self-update: how to apply config updates and run updates
  • Workspace: the working directory boundary for tools and file operations
  • Documentation: where local docs live and when to read them
  • Injected workspace files: bootstrap context included without explicit reads
  • Memory digest: budgeted long-term memory recall (from the StateStore), scoped to the agent
  • Sandbox: runtime constraints and whether elevated execution is available
  • Date and time: user-local time and formatting
  • Reply tags: optional provider-specific reply tags
  • Heartbeats: periodic prompt and acknowledgement behavior
  • Runtime: host/OS/node/model/runtime summary
  • Reasoning visibility: visibility level and how it can be toggled (when supported)

Context report (what the model saw)​

For observability, the gateway produces a per-run context report that captures:

  • injected workspace files (raw vs injected sizes and truncation)
  • system prompt section sizes
  • skill list overhead
  • the largest tool schema contributors

Operator clients expose this via /context list and /context detail (see Observability and Slash Commands).

Advisory vs enforcement​

System-prompt guardrails are advisory. Hard enforcement should come from:

  • Tool allowlists and parameter validation
  • Execution approvals
  • Sandboxing and environment constraints
  • Channel allowlists and connector policy

Injected bootstrap files​

These files (or equivalents) can be injected as "project context" so the model has identity and safety context without extra tool calls:

  • AGENTS.md
  • SOUL.md
  • TOOLS.md
  • IDENTITY.md
  • USER.md
  • HEARTBEAT.md
  • BOOTSTRAP.md (only for brand-new workspaces)

Long-term memory is not represented as a workspace file. It lives in the StateStore and is injected as a budgeted memory digest (agent-scoped recall) during context assembly.