@tachu/extensions 1.0.0-beta.1 → 1.0.0-rc.0

package/CHANGELOG.md

@@ -5,73 +5,48 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased]
+## [1.0.0-rc.0] - 2026-06-01
 
-## [1.0.0-beta.1] - 2026-04-17
+### Added
 
-First public beta of Tachu — the agentic engine Harness. All subsystems listed below are production-grade implementations; the `-beta` marker reflects that this is the first externally-consumable release prior to 1.0.0 stable. No functional gaps relative to the 1.0 architectural design.
+#### `@tachu/core`
 
-### Added
+- 9-phase execution pipeline with session management, minimum safety, intent analysis, planning, graph validation, sub-task execution, candidate-answer synthesis, result validation, and output normalization.
+- Built-in `direct-answer` and `tool-use` sub-flows, including tool-loop execution, streaming deltas, cancellation propagation, and turn retry bookkeeping.
+- Descriptor Registry for Rules, Skills, Tools, and Agents, including version-aware descriptor lookup and deterministic activation gates.
+- Prompt assembly with token accounting, context budgeting hooks, multimodal resource placeholders, and provider-bound reference materialization.
+- Result Validation contracts: `EvidenceEntry`, `ClaimEntry`, `CandidateAnswer`, `ValidationOutcome`, deterministic validation rules, optional semantic judge adapter, and degrade / handoff / retry outcomes.
+- Turn Policy support for tool include/exclude, skill pin/exclude, explicit skill priority, final-answer skill inheritance, and visualization metadata passthrough.
+- Resource Reference Pool, token-level Resource Demand routing, multimodal memory fidelity for new sessions, and provider image carrier support.
+
+#### `@tachu/extensions`
+
+- OpenAI, Anthropic, Qwen, Gemini, and Mock provider adapters with streaming, tool calling, configurable base URLs, timeouts, and embedding support where implemented.
+- Built-in tools for file, terminal, search, web, git, task tracking, and typecheck/test execution workflows.
+- MCP stdio and SSE adapters, OpenTelemetry / JSONL emitters, Qdrant and local vector index adapters, and file-backed memory with projection outbox support.
+
+#### `@tachu/host-defaults`
+
+- Shared host wiring for CLI and embedded hosts: provider inference, capability checks, semantic retrieval facade setup, semantic judge resolution, and memory projection stack helpers.
+
+#### `@tachu/cli`
+
+- `tachu init`, `tachu run`, `tachu chat`, and approval-management commands.
+- Interactive session persistence, streaming renderer, terminal Markdown rendering, configuration loading, provider setup, and opt-in real-provider end-to-end test path.
+
+#### `@tachu/web-fetch-server`
+
+- Private optional sidecar for browser-backed `web-fetch` and `web-search` workflows, with Playwright-based rendering and Docker runtime alignment.
+
+### Changed
+
+- Workspace packages now publish in lockstep as `1.0.0-rc.0`.
+- Public package publish order is `@tachu/core` → `@tachu/extensions` → `@tachu/host-defaults` → `@tachu/cli`.
+- Release artifact validation checks public package metadata, mirrored package docs, executable bins, built artifacts, private sidecar exclusion, and Web Fetch Docker / Playwright runtime alignment.
+- Package descriptions and install guidance now use release-candidate terminology.
+
+### Known Limitations
 
-- Initial production-grade release of Tachu — the agentic engine Harness
-
-- **`@tachu/core`**:
-  - `Engine` entry class with `run()`, `runStream()`, and `cancel()` APIs
-  - 9-phase execution pipeline: session management, minimum safety check, intent analysis, pre-check, task planning, DAG validation, sub-task execution, result validation, output normalization
-  - `Registry` with startup validation (name uniqueness, `requires` completeness)
-  - `PromptAssembler` with tiktoken-based exact token counting and KV-cache-friendly prompt ordering
-  - `TaskScheduler` with topological-sort-based parallel/serial dispatch (`Promise.all` for zero-indegree tasks)
-  - 8 core modules: `SessionManager`, `MemorySystem` (Head-Middle-Tail compression), `RuntimeState`, `ModelRouter`, `ProviderAdapter` (interface), `SafetyModule`, `ObservabilityEmitter`, `HookRegistry`
-  - `VectorStore` interface + built-in in-memory lightweight implementation (cosine similarity, ≤10,000 entries)
-  - Dual-plane matching model: vector semantic discovery + deterministic execution gate (scopes → whitelist → approval)
-  - Complete error taxonomy: `EngineError`, `SafetyError`, `ProviderError`, `BudgetExhaustedError`, `ValidationError`, `TimeoutError`, `RegistryError`, `PlanningError`
-  - Full TypeScript type definitions for all interfaces, descriptors, I/O structures, and configuration schema
-  - `AbortController`-based cancellation propagation (last-message-wins)
-  - Budget circuit-breaker (token + wall-time)
-  - Streaming output via `AsyncIterable<StreamChunk>`
-  - Comprehensive unit test suite (≥80% line coverage, ≥70% branch coverage)
-  - Integration test: full 9-phase pipeline with `NoopProvider`
-  - Performance benchmarks: scheduler (100 parallel tasks), vector store (10,000 entries, topK=10), prompt assembler (4K token window)
-
-- **`@tachu/extensions`**:
-  - `OpenAIProviderAdapter`: streaming, function calling, `listAvailableModels()`, standardized error handling
-  - `AnthropicProviderAdapter`: streaming, function calling, `listAvailableModels()`, standardized error handling
-  - `MockProviderAdapter`: configurable responses for testing
-  - `McpToolAdapter`: MCP stdio + SSE connection, capability negotiation, `AbortSignal`→MCP cancel propagation
-  - 7 built-in tools: file read, file write, directory list, shell command, web fetch, code search, grep
-  - `TerminalBackend`, `FileBackend`, `WebBackend` execution backends
-  - `LocalFsVectorStore`: file-backed vector store with persistence
-  - `QdrantVectorStore`: full Qdrant REST API integration via `@qdrant/js-client-rest`
-  - `OtelEmitter`: maps `EngineEvent` to OpenTelemetry spans via `@opentelemetry/api`
-  - `JsonlEmitter`: append-only JSONL event log
-  - `ConsoleEmitter`: structured development console output
-  - 4 common rules: `no-sensitive-output`, `always-cite-sources`, `concise-response`, `structured-output`
-  - `VisionTransformer`, `DocumentToTextTransformer` input transformers
-  - Unit tests for every adapter, tool, backend, and emitter (≥80% line coverage)
-  - Integration tests: OpenAI/Anthropic provider smoke tests, MCP local server connectivity, Qdrant CRUD
-
-- **`@tachu/cli`**:
-  - `tachu init`: project workspace scaffolding with `--template`, `--provider`, `--force`, `--no-examples` options; generates `.tachu/` directory, `tachu.config.ts`, and `.gitignore` entries
-  - `tachu run <prompt>`: single-shot execution with streaming render; `--session`, `--resume`, `--model`, `--provider`, `--input`, `--json`, `--output`, `--no-validation`, `--plan-mode`, `--verbose`, `--no-color`, `--timeout` options
-  - `tachu chat`: interactive multi-turn session with readline loop; `/exit`, `/reset`, `/new`, `/list`, `/load`, `/save`, `/export`, `/history`, `/stats`, `/help` built-in commands
-  - Session persistence to `.tachu/sessions/<id>.json` with auto-save after each turn
-  - `--resume` and `--session` for session continuity across invocations
-  - `--history` and `--export` non-interactive modes
-  - Ctrl+C first-press cancel, second-press (within 1s) save-and-exit behavior
-  - Hand-written ANSI color renderer with `NO_COLOR` / non-TTY / `--no-color` auto-detection
-  - Hand-written spinner for long-running phases (non-TTY auto-disabled)
-  - Plan mode interactive approval flow
-  - `citty`-based CLI framework with auto-generated `--help` and `--version`
-  - Config loader: `tachu.config.ts` discovery (CWD → git root), descriptor scanner for all 4 abstraction types
-  - End-to-end integration tests for all 3 commands using `MockProvider`
-  - Unit tests (≥80% line coverage)
-
-### Docs
-
-- English README (`README.md`) with 21 sections including architecture mermaid diagrams
-- Chinese README (`README_ZH.md`) with equivalent content and structure
-- ADR documents: `architecture-design.md`, `detailed-design.md`, `technical-design.md`
-- `CHANGELOG.md` (this file)
-
-[Unreleased]: https://github.com/dangaogit/tachu/compare/v1.0.0-beta.1...HEAD
-[1.0.0-beta.1]: https://github.com/dangaogit/tachu/releases/tag/v1.0.0-beta.1
+- Runtime provider fallback after `ProviderError` is not implemented in this release candidate.
+- Semantic judge support is available but not production-complete.
+- The Web Fetch sidecar remains private and is not published to npm.