Configuring Context Management

A long chat can outgrow the LLM's context window, and quality drops well before that as context rot sets in. CodeCompanion manages the chat buffer's context with two operations:

• Editing - replaces old tool results with a placeholder
• Compaction - summarises the conversation and replaces the message history with the summary

See In the Chat Buffer for how they work.

Context management is enabled by default for HTTP adapters. To disable it, or decide per adapter:

Boolean

require("codecompanion").setup({
  interactions = {
    chat = {
      opts = {
        context_management = {
          enabled = false,
        },
      },
    },
  },
})

Function

require("codecompanion").setup({
  interactions = {
    chat = {
      opts = {
        context_management = {
          enabled = function(adapter)
            return adapter.name ~= "ollama"
          end,
        },
      },
    },
  },
})

NOTE

The anthropic and openai adapters compact on the server where the model supports it, and CodeCompanion's own context management then stands aside. See Disabling Compaction

Triggers

Editing and compaction each have a trigger. A decimal is a share of the model's context window, and an integer is a token count:

Decimal

require("codecompanion").setup({
  interactions = {
    chat = {
      opts = {
        context_management = {
          editing = {
            trigger = 0.65, -- 65% of the context window
          },
          compaction = {
            trigger = 0.85, -- 85% of the context window
          },
        },
      },
    },
  },
})

Integer

require("codecompanion").setup({
  interactions = {
    chat = {
      opts = {
        context_management = {
          editing = {
            trigger = 80000, -- tokens
          },
          compaction = {
            trigger = 100000, -- tokens
          },
        },
      },
    },
  },
})

A decimal trigger needs the model's context window. If the adapter doesn't report one, that operation is skipped.

Editing

Editing replaces the content of older tool results with a placeholder, leaving the shape of the conversation intact. The last three cycles are kept in full, where a cycle is one of your messages plus everything the LLM did in response. Tools whose output is referenced again later can be left alone:

require("codecompanion").setup({
  interactions = {
    chat = {
      opts = {
        context_management = {
          editing = {
            trigger = 0.65,
            exclude_tools = { "memory" },
            keep_cycles = 3,
          },
        },
      },
    },
  },
})

Compaction

Compaction summarises the chat with a single LLM request and replaces the message history with the summary. It's skipped if it would save fewer than min_token_savings tokens.

To use a cheaper or faster model for the summary, set an adapter. With fallback_to_chat_adapter, a failed summary is retried with the chat's own adapter:

require("codecompanion").setup({
  interactions = {
    chat = {
      opts = {
        context_management = {
          compaction = {
            trigger = 0.85,
            min_token_savings = 10000,
            adapter = { name = "copilot", model = "gpt-4.1" }, -- Can be nil, an adapter name or { name, model }
            fallback_to_chat_adapter = false,
          },
        },
      },
    },
  },
})

The adapter defaults to the chat's own adapter.

Limitations

• Context management only runs for HTTP adapters. ACP agents manage their own context