@adaptcom/core

The Adapt agent runtime: harnesses, tools, channels, connections, persistent conversations, and deployment APIs. Requires Node 24 or later. The HTTP API and its browser client live in @adaptcom/api.

pnpm add @adaptcom/core
pnpm add -D @adaptcom/cli

import { defineAgent, createHttpChannel } from "@adaptcom/core";

Channel and connection subpaths are also available, for example @adaptcom/core/channels/slack. @adaptcom/api/client is the browser-safe API client.

Use @adaptcom/cli for the adapt command. Core does not depend on the CLI. See the repository guides and examples.

Compaction

Configure compaction on the harness, not on defineAgent. The provider receives active ModelMessage[] history and returns a smaller array of the same type. Preserve current intent, required context, and complete tool-call/result groups. Return the input unchanged if no safe reduction is possible. Honor cancellation.

import { createAiSdkHarness, defineAgent, type CompactionProvider } from "@adaptcom/core";
import type { LanguageModel } from "ai";

export function agent(model: Exclude<LanguageModel, string>, compaction: CompactionProvider) {
  return defineAgent({
    harness: createAiSdkHarness({
      model,
      contextWindow: 128_000, // Use your model's actual capacity.
      compaction, // Optional compaction.threshold defaults to 0.8.
    }),
  });
}

CompactionProvider.threshold is optional, defaults to 0.8, and must be between 0 and 1 inclusive. The AI SDK harness checks before each request, including after completed tools, and compacts when the request token estimate reaches contextWindow * threshold. Zero attempts compaction before every request; one waits until the estimated full context window. A no-op reduction proceeds without looping. Compaction stays opt-in; a context window alone does not enable it.

The harness estimates tokens from serialized messages, instructions (including skills), and tool definitions at roughly three UTF-8 bytes per token. Within a turn, reported input usage corrects underestimates, with new history added to that count. This is approximate, not a model-specific tokenizer. contextWindow stays local to the harness and is required when configuring compaction; it is not passed through project or runtime configuration.

A recognized context-overflow rejection still allows one compaction-and-retry of that request, without rerunning completed tools or consuming another step. No reduction rethrows the original overflow. Other errors are not retried. Invalid/failed/aborted replacements leave the original history intact.

createGitHubAiSdkHarness accepts the same settings and preserves them when tools are injected. For an environment-selected model, use createEnvironmentHarness(process.env, { compaction, contextWindow: 128_000 }). A project harness factory can resolve provider credentials at startup. Codex harnesses own native conversation state and do not implement this compaction loop.

A smaller serialized history is not proof it will fit the model's limit. A summarizing provider must also be able to process the history it receives, using its own chunking or a larger-context model when necessary.

Successful reductions append a session.compacted event through the ordinary store commit. On replay, the harness replaces its messages with that event's history. Revisions and retry receipts keep their existing append-only semantics; runtime state, resource handles, and other sessions are unchanged. Compaction rejects unresolved tool calls and clears completed tool state.

The Next.js chat example replaces earlier messages with a "Conversation compacted" marker and the replacement's public text. System messages, reasoning, tool payloads, and provider metadata are not projected. Custom chat clients should reset their transcript on a compaction event.

This compacts model context, not disk usage. Original events and receipts remain in the journal for idempotent retries, including sensitive original history; there is no automatic retention policy. No archives or special store methods are needed. Custom stores must support the session.compacted event and its tool-state invariant, as exercised by the shared store contract. Custom harnesses expose their provider as AgentHarness.compaction, reset history on that event when reading context.events(), and invoke context.compaction.compact() only at a safe boundary. The runtime validates and checkpoints before returning.