Context Engine

Builds the agent's system prompt from typed layers in a deterministic order. Each layer is independent and can be enabled, disabled, or modified by hooks.

The 10 layers

Top to bottom in the final prompt:

Inspecting the built context

To see what the agent will actually see:

curl -X POST http://127.0.0.1:18789/agent/bags-hunter/context \
  -d '{
    "sessionKey": "user-123",
    "recentMessages": [{"content": "what tokens look good today"}]
  }'

Returns the assembled prompt + a per-layer breakdown (name, priority, char count).

Excluding layers

For minimal prompts (testing, debugging), exclude layers:

curl -X POST http://127.0.0.1:18789/agent/auto/context \
  -d '{
    "sessionKey": "user-123",
    "excludeLayers": ["memory", "commitments", "standing"]
  }'

Resolution order — user vs bundled

For each agent, the engine first checks ~/.openvesper/agents/<mode>/. If the agent is installed there, files load from user dir. Otherwise from bundled .agents/<mode>/.

Project AGENTS.md walk

The engine walks up to 5 directory levels from process.cwd() looking for AGENTS.md. The first one found becomes the project layer. This lets projects inject team conventions automatically.

Source

Implementation: apps/gateway/src/context-engine.ts

What's next?

• Markdown templates
• Memory Engine