Context Management

Token Tracking

ContextWindowTracker (in agent-session) reads the shared agent-core context estimator. Terminal provider usage is treated as the exact post-response state. If new user or tool messages appear after the latest provider usage, the estimator uses the maximum of serialized history, latest provider usage metadata, and any caller-provided usage floor, so previous metadata cannot hide a large prompt that has not been sent yet.

const state = session.getContextState();
// { maxTokens: 200000, usedTokens: 85000, usedPercentage: 42.5, remainingPercentage: 57.5 }

Model Context Sizes

Model	Context Window
Claude Sonnet 4.6 / Opus 4.6	1,000,000 tokens
Claude Haiku 4.5	200,000 tokens

Compaction

When the context window fills up, Robota compresses the conversation by generating an LLM summary.

Auto-Compaction

Triggers at ~83.5% of the model's context window. The sequence:

1. PreCompact hook fires
2. Full conversation history sent to LLM with summarization prompt
3. History cleared and replaced with [Context Summary] message
4. Token tracking resets
5. PostCompact hook fires with the summary
6. onCompact callback notifies the UI

Core also has a last-resort hard-capacity guard before provider calls. That guard uses the same effective estimator and blocks only past 95% of the model context window, returning diagnostic values so the CLI can explain why a prompt was rejected.

Manual Compaction

await session.compact('Focus on the API design decisions');

CLI: /compact focus on API changes

Compact Instructions

CLAUDE.md can define a "Compact Instructions" section. These instructions are automatically extracted during context loading and included in the compaction prompt to preserve project-specific context.

Streaming

The text_delta event provides real-time text as the model generates it:

import { InteractiveSession } from '@robota-sdk/agent-framework';
import { AnthropicProvider } from '@robota-sdk/agent-provider/anthropic';

const provider = new AnthropicProvider({ apiKey: process.env.ANTHROPIC_API_KEY });

const session = new InteractiveSession({ cwd: process.cwd(), provider });
session.on('text_delta', (delta) => process.stdout.write(delta));

The complete event carries the final response after streaming finishes.

Web Search

Provider-side web tools are distinct from Robota local tools. Anthropic supports server-side web search (web_search_20250305) through provider-native capability configuration, and onServerToolUse fires when search executes. Qwen supports provider-side web_search and web_extractor through builtInWebTools; those hosted tools record provenance in assistant-message metadata and do not bypass local Robota permission checks. OpenAI-compatible local endpoints such as LM Studio are treated as custom function-tool capable, not provider-native web search/fetch capable, unless a concrete provider package documents and enables that hosted capability. Use Robota local WebSearch and WebFetch tools for explicit local-tool web access with those endpoints.

Abort

Cancel a running InteractiveSession.submit() via AbortSignal:

session.abort(); // Triggers AbortSignal, cancels streaming

The AbortSignal flows through the entire execution chain: InteractiveSession calls Session.run(), Session.run() passes the signal to Robota.run(), and Robota.run() passes it to the provider. AbstractAIProvider.streamWithAbort() provides a standard streaming wrapper that all providers use to handle abort — when the signal fires, the provider returns partial content with stopReason: 'aborted'. The partial response is committed to history with state: 'interrupted'.