v1.5.2

Release 1.5.2

• Enhances output-validator errors with instance values, union-error collapse, error modes, prompt examples, and required-field hints.
• Bumps package metadata to 1.5.2.

npm: https://www.npmjs.com/package/agentool/v/1.5.2

v1.5.1

Changes

• Fix agent tool JSON schema for native Anthropic tool validation.
• Keep strict runtime validation for agent actions while exposing a flat provider-facing object schema.
• Add provider compatibility functional test for OpenAI-compatible and Anthropic protocols.

Verification

• npm run test:provider-compat
• npm run typecheck
• npm run lint
• npm test

v1.5.0

What's Changed

• Add the managed agent tool with start, wait, status, result, list, and stop actions.
• Add subagent task management with timeouts, wait polling, result formatting, and concurrency limits.
• Prevent recursive subagent spawning by stripping nested agent tools from child toolsets.
• Add unit, build, and real-AI functional coverage for the agent tool.
• Export agentool/agent and bump package metadata to 1.5.0.

Verification

• npm run lint
• npm run typecheck
• npm run test
• npm run test:functional
• npm run build

v1.4.0

Changes

• Add output-validator tool for validating final JSON output against an application-configured JSON Schema.
• Export output-validator from the root package and subpath export.
• Add README documentation and unit tests for per-turn schema validation.
• Update package version to 1.4.0 and test against ai 6.0.175.

Verification

• npm run typecheck
• npm run lint
• npm test
• npm run build

v1.3.0 - compactMessages helper (breaking)

v1.3.0 — compactMessages helper (breaking)

Why this release

The 1.2.0 createContextCompaction middleware was stateless: every over-threshold turn re-summarized the older history from scratch, since the AI SDK middleware contract gives no path to mutate the caller's messages array. That made compaction expensive on every subsequent turn and busted provider prompt caches (each summary was a non-deterministic new prefix).

Breaking changes

• Removed: createContextCompaction, ContextCompactionConfig
• Added: compactMessages, CompactMessagesOptions, CompactSummarizer

The shape of the migration is one of intent — the old middleware ran on every model call automatically; the new function runs explicitly before each call and the caller assigns the result back. That's what makes the compacted state actually persist.

Migration

Before (1.2.x):

import { createContextCompaction } from 'agentool/context-compaction';
import { wrapLanguageModel, generateText } from 'ai';

const model = wrapLanguageModel({
  model: anthropic('claude-sonnet-4-20250514'),
  middleware: createContextCompaction({ maxContextTokens: 200_000 }),
});
const { text } = await generateText({ model, messages });

After (1.3.0):

import { compactMessages } from 'agentool/context-compaction';
import { generateText } from 'ai';

const model = anthropic('claude-sonnet-4-20250514');
messages = await compactMessages({
  messages,
  summaryModel: model,
  maxContextTokens: 200_000,
});
const { text } = await generateText({ model, messages });

When usage is under threshold, compactMessages returns the same messages reference (===) — second-call cost is cheap, no LLM round trip.

Provider compatibility

Operates on the AI SDK's unified ModelMessage[], so works with any provider the SDK supports — OpenAI, Anthropic, Google Gemini, Mistral, xAI, etc. summaryModel need not match the main model (use a cheap summarizer with a powerful main model).

Output shape

[ ...leadingSystemPrefix,
  { role: 'user',      content: <summary> },
  { role: 'assistant', content: 'Understood.' },
  ...lastNMessages ]

The synthetic user → assistant ack pair preserves role alternation (required by Anthropic and Google providers).

Tool-chain safety

The recent-window boundary auto-extends backwards via toolCallId and approvalId tracking, so tool-call / tool-result / tool-approval-request / tool-approval-response pairs are never split across the summarization boundary — no MissingToolResultsError from the SDK.

Other behavior

• Mid-conversation system messages stay in place (only the leading contiguous system prefix is hoisted as system).
• Multimodal content (images, files) is reduced to placeholders during summarization. Original binary data is gone from the persisted compacted history.
• Custom summarize callback is supported as an alternative to summaryModel (e.g. local model, cached, or anything else).

Options

Option	Type	Default
messages	ModelMessage[]	required
maxContextTokens	number	required
summaryModel or summarize	LanguageModelV3 or fn	required (exactly one)
autoCompactThresholdPct	number (0-1)	0.8
summaryTargetTokens	number	floor(maxContextTokens * 0.05)
reservedOutputTokens	number	16384
keepRecentMessages	number	1
estimateTokens	(msgs) => number	char/4 heuristic
onCompactionFailure	'passthrough' | 'throw'	'passthrough'

Caveats

• Default token estimation is char/4 — coarse but provider-agnostic. For accuracy pass estimateTokens with a real tokenizer (e.g. tiktoken for OpenAI, @anthropic-ai/tokenizer for Anthropic).
• Summary-of-summary degradation accumulates over very long sessions. Recommend periodic session restarts for long-lived agents.

🤖 Generated with Claude Code

v1.2.0 - Context Compaction Middleware

v1.2.0 - Context Compaction Middleware

Breaking Changes

• Removed: contextCompaction tool (old LLM-callable implementation)
• Added: createContextCompaction middleware for Vercel AI SDK
• Peer dependency: ai package bumped to >=5.0.17

New Features

The context-compaction middleware automatically summarizes conversation history when it exceeds a configurable threshold:

import { createContextCompaction } from 'agentool/context-compaction';
import { wrapLanguageModel } from 'ai';

const model = wrapLanguageModel({
  model: anthropic('claude-sonnet-4-20250514'),
  middleware: createContextCompaction({
    maxContextTokens: 200_000,
    autoCompactThresholdPct: 0.80,
    summaryTargetPct: 0.05,
  }),
});

Configuration options:

• maxContextTokens - Model's context window size (required)
• autoCompactThresholdPct - Trigger compaction at this usage % (default: 0.80)
• summaryTargetPct - Target summary size as % of max (default: 0.05)
• reservedOutputTokens - Tokens reserved for output (default: 16384)
• estimateTokens - Custom token estimator function
• summarize - Custom summarizer function
• onCompactionFailure - 'passthrough' or 'throw' (default: 'passthrough')

Technical Details

• Middleware intercepts generateText/streamText calls via wrapLanguageModel
• Preserves system messages verbatim
• Maintains recent conversation window (~20% of context)
• Handles cross-provider compatibility (OpenAI, Anthropic, Google)
• Role alternation preserved for Anthropic compatibility

Testing

• 26 unit tests for middleware logic
• 2 functional tests with real LLM
• All 437 tests passing

v1.1.1

What's Changed

Internal Refactoring

• Centralized filesystem access: All node:fs operations now route through src/shared/file.ts with 4 new utilities (readTextContent, listDirectory, removeFile, getFileStats)
• Unified error handling: Extracted extractErrorMessage() to src/shared/errors.ts, replacing 21 duplicated patterns across all tools
• Deduplicated countOccurrences: Moved to src/shared/edit-helpers.ts with explicit non-overlapping semantics, shared by both edit and multi-edit
• ESLint enforcement: Added no-restricted-imports rule banning direct node:fs imports outside of src/shared/file.ts

Testing

• Added 22 new unit tests for all new shared utilities
• All 420 unit tests and 30 functional tests passing

No public API changes. This is a purely internal quality improvement.

v1.1.0

What's New

6 New Tools

• task-create — Create tasks with subject, description, and metadata
• task-get — Retrieve a task by ID
• task-update — Update status, owner, metadata, and dependencies (blocks/blockedBy)
• task-list — List all non-deleted tasks
• web-search — Callback-based web search with domain filtering
• tool-search — Registry-based tool discovery by name/keyword

Aligned with Claude Code

• Read — file_path enforces "absolute path"; offset validates .int().nonnegative(); limit validates .int().positive()
• Write — file_path description says "absolute path, not relative"
• Edit — file_path description says "absolute path to modify"; replace_all uses .default(false).optional()
• Grep — All 14 parameter descriptions match Claude Code verbatim
• WebFetch — url uses .url() validation; unused prompt param removed
• LSP — 1-based indexing (required), prepareCallHierarchy added, query removed, boundary conversion

Breaking Changes

• agentool/task removed — replaced by agentool/task-create, agentool/task-get, agentool/task-update, agentool/task-list
• LSP line/character are now required and 1-based (previously optional, 0-based)
• LSP query parameter removed
• WebFetch prompt parameter removed
• WebFetch rejects invalid URLs at schema level

Infrastructure

• 22 tools total (was 16), 22 subpath exports
• Shared task store at src/shared/task-store.ts
• 398 unit tests, 30 functional tests (all with real AI provider)
• Full pipeline: typecheck, lint, test, functional test, build — all pass