From 6a8ee6ff2cff5e3e76d4c8aff0ad7f2a36dc6c3f Mon Sep 17 00:00:00 2001 From: Tomas Pflanzer Date: Sat, 16 May 2026 19:42:37 +0200 Subject: [PATCH] chore: bump to 0.32.0 "Claude Agents Deep Integration" Bumps __version__, rewrites README hero, adds v0.32.0 changelog entry, adds v0.32 release card to /whatsnew/ (Don Draper voice opening "You shipped the agent. The client wants the integration. The auditor wants the trail..."), and updates the v0.31 -> v0.32 footer labels on the homepage and the EU AI Act landing. --- CHANGELOG.md | 57 +++++++++++++++++++++ README.md | 21 +++++--- site/eu-ai-act/index.html | 2 +- site/index.html | 4 +- site/whatsnew/index.html | 101 ++++++++++++++++++++++++++++++++++--- src/sandcastle/__init__.py | 2 +- 6 files changed, 169 insertions(+), 18 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 7c543ef..ff24a0d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,63 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.32.0] - 2026-05-16 - "Claude Agents Deep Integration" + +You shipped the agent. The client wants the integration. The auditor wants the trail. The user wants you to ask the next question without restarting the whole workflow. v0.32 is the answer to every one of those. Sandcastle now exposes every Anthropic Managed Agents primitive shipped under the managed-agents-2026-04-01 beta umbrella, plus the things Anthropic doesn't ship: a cryptographically verifiable trajectory replay, a Skills publisher that turns workflows into uploadable Claude Skills, and an Agent SDK runtime for teams that want in-process execution. Two weeks of work, 169 new tests, one release. + +### Added - Anthropic primitives (the things the beta header gives you) + +- **Memory Stores** client (`sandcastle.engine.memory_stores.MemoryStoresClient`). Versioned per-session memory mounted at /mnt/memory/, optimistic-concurrency writes via If-Match, redact endpoint for GDPR right-to-be-forgotten, 100 kB per file, 8 stores per session. `attach_to_session_payload()` helper builds the resources block for session-create. +- **Multiagent coordinator** (`sandcastle.engine.multiagent`). Up to 20 specialist agents in parallel, 25 threads, 1-level depth per Anthropic spec. Three pre-baked templates: `research-and-write`, `code-review-and-test`, `analyst-with-translator`. `validate_roster()` + `build_coordinator_payload()` + `parse_thread_event()`. +- **Outcomes API** (`sandcastle.engine.outcomes`). `user.define_outcome` events on session start, `span.outcome_evaluation_end` captured into step output. Composite aggregator at module level so AutoPilot and Workflow Evolution can read native Anthropic eval signals. +- **Webhooks** (`sandcastle.api.agent_webhooks`). HMAC-signed session lifecycle events at `/agent-webhooks/anthropic`. Fire-and-forget dispatch, integrates with the existing arq scheduler. +- **Elicitation** (the 6th MCP primitive, added in spec rev 2025-11-25). New `request_workflow_input` tool wraps `ctx.session.elicit()` with JSON Schema validation so a workflow that hits a gap mid-execution can ask the user for a typed value without restarting. + +### Added - managed-agent step extensions + +The `type: managed-agent` step now accepts three new config fields that thread directly into the Anthropic primitives above: + +- `memory_stores: list[str]` - attach existing memory store IDs to the session +- `multiagent: dict` - build a coordinator payload with validated roster +- `outcomes: list[dict]` - define outcomes at session start, capture eval results in step output + +### Added - Sandcastle differentiators (the things Anthropic doesn't ship) + +- **Skills publisher** (`sandcastle.engine.agent_skills`). `sandcastle publish-skills [--upload] [--dir]` converts every workflow into a SKILL.md tar.gz package with strict frontmatter validation (kebab-case name, no reserved tokens, ≤1024-char description) and uploads to `/v1/skills`. Workflows are now reachable from every Anthropic Skills-aware client. +- **Trajectory Replay step type** (`sandcastle.engine.trajectory_replay`). New `type: trajectory-replay` step computes SHA-256 over a recorded tool-call sequence, diffs against a candidate run, returns score + diff_summary. Because Sandcastle's audit trail is a hash chain, the replay is cryptographically verifiable - a property neither LangSmith nor Braintrust ships. +- **Computer Use integration helper** (`sandcastle.engine.computer_use`). New `type: computer-use` step type. Builds the `computer_20251124` tool definitions, sets the beta header, runs an 8-item safety pre-flight (prompt-injection guard, screenshot dimensions, page-load deadline). +- **Agent SDK runtime** (`sandcastle.engine.agent_sdk_runtime`). New `runtime: "agent-sdk"` dispatch. For teams who want in-process Claude agents (EU sovereignty, air-gapped, no Managed Agents infra). Lazy-imports `anthropic_agent_sdk`; falls back to a typed `AgentSDKNotInstalled` error when the optional package isn't installed. + +### Added - Tool Search + tool-use-examples convention + +New `sandcastle.engine.tool_search.ToolRegistry` lets workflows mark tools with `defer_loading: true` (loaded on first selection) and `examples: [...]` (1-5 realistic invocations per tool). Anthropic measured the result on Opus 4: tool-selection accuracy from 49% to 74%, usable context from 122,800 to 191,300 tokens (85% saving), parameter accuracy from 72% to 90%. New docs/tool-examples-convention.md. + +### Added - Tier 1 wire fixes (table stakes that had been broken) + +- `tools_enabled` config field is now actually sent to the agent-create API (previously parsed but ignored - users thought they were restricting tools). +- `temperature`, `max_tokens`, `thinking_budget` on `ManagedAgentConfig`. None-aware: omitted from request when unset. +- `stream` config field is now honoured (was dead code). +- Pricing table for Opus 4.7 (5/25), Sonnet 4.6 (3/15), Haiku 4.5 (1/5), Opus 4.6 (15/75), Sonnet 4.5 (3/15). Unknown model falls back to Sonnet 4.6 rates with a one-time warning. +- `fallback_template` accepts a list (chain of up to 5 templates) in addition to a single string. + +### Added - dashboard + +- Live "Agent Reasoning" panel on the run detail page. Subscribes to `/api/runs/{id}/agent-stream` SSE, renders agent.thinking, agent.tool_use, agent.message, agent.complete, agent.error events. Thread-grouped, collapsible, graceful 404 fallback. + +### Changed + +- New step types `trajectory-replay` and `computer-use` registered (VALID_STEP_TYPES count 22 -> 24). +- `agent_webhooks_router` mounted on the FastAPI app alongside `a2a_router` and `agui_router`. +- MCP server manifest now advertises 6 primitives (added Elicitation) and declares `spec_revision: "2025-11-25"`. + +### Tests + +- 18 new tests for Tier 1 wire fixes (tests/test_managed_agent_wires.py) +- 156 new tests for the 9 modules in isolation +- 13 new e2e wiring tests (tests/test_v032_wiring.py) +- 169 v0.32-related tests total, all green in 1.8s +- Full suite: 15,176 passing (vs 15,009 baseline) - the +167 are this release's new tests + ## [0.31.0] - 2026-05-14 - "Compliance & Connections" Eighty days to the EU AI Act deadline (2 August 2026). This release is the answer: a dedicated landing page mapping every Sandcastle control to a specific Article, ten compliance workflow templates, MCP-first publishing so every workflow becomes a tool inside Claude Desktop / Cursor / Windsurf, eval gates that block regressing models from getting promoted, and a dashboard that doesn't crash when one API hiccups. Plus the closeout of v0.30: Codex audit rounds 9 and 10 fully fixed. diff --git a/README.md b/README.md index 4525801..78c9839 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,7 @@ **Describe what you want. Go home. Sandcastle ships it.** Production-ready workflow orchestrator for AI agents. 7 AI providers with auto-failover, 22 step types including Claude Managed Agents, 15 agent templates, 4 OCR engines, EU AI Act compliance, and a full-featured dashboard. Define workflows in YAML or let AI design them for you. -[![PyPI](https://img.shields.io/badge/PyPI-v0.31.0-blue?style=flat-square)](https://pypi.org/project/sandcastle-ai/0.31.0/) +[![PyPI](https://img.shields.io/badge/PyPI-v0.32.0-blue?style=flat-square)](https://pypi.org/project/sandcastle-ai/0.32.0/) [![License: BSL 1.1](https://img.shields.io/badge/License-BSL_1.1-blue.svg)](LICENSE) [![Python 3.12+](https://img.shields.io/badge/python-3.12+-blue.svg)](https://www.python.org/downloads/) [![Tests](https://img.shields.io/badge/tests-15000%2B%20passing-brightgreen?style=flat-square)](https://github.com/gizmax/Sandcastle/actions) @@ -10,14 +10,19 @@ [![Website](https://img.shields.io/badge/Website-sandcastle--ai.eu-blue?style=flat-square)](https://sandcastle-ai.eu) [![Live Demo](https://img.shields.io/badge/Live%20Demo-Dashboard-F59E0B?style=flat-square)](https://gizmax.github.io/Sandcastle/) -> **v0.31 - "Compliance & Connections"** Shipped May 14, 2026. Eighty days to the EU AI Act deadline. We built the answer. +> **v0.32 - "Claude Agents Deep Integration"** Shipped May 16, 2026. Every Anthropic Managed Agents primitive surfaced, plus the things Anthropic doesn't ship. > -> - **EU AI Act landing + 10 compliance workflow templates** mapped to Articles 9, 11, 12, 14, 25, 49, 50, 73 and Annex IV. DPIA, bias audit, incident report, vendor risk, model card, AI inventory, GDPR DSAR, human oversight log, transparency report, risk register. Bring the templates. Customize the prompts. Hand the auditor the audit trail. -> - **MCP-first publishing.** `sandcastle publish-mcp` turns every workflow you've built into a first-class tool inside Claude Desktop, Cursor, Windsurf, or any MCP client. -> - **Eval gates that block bad promotions.** Define a golden dataset, set a minimum score, ship with confidence. Eval-driven development is a query parameter, not a future plan. -> - **A dashboard that doesn't crash.** Overview split into 20 focused components with per-section error boundaries. -> - **Codex audit rounds 9 + 10** closed: 5 HIGH + 1 MEDIUM findings fixed - cross-tenant cache, memory, prompts, XSS, SSRF, A2A budgets. -> - 15,014 tests passing. PyPI: `pip install sandcastle-ai==0.31.0`. +> - **Memory Stores + Multiagent + Outcomes + Webhooks** wired into the YAML. One workflow can attach versioned memory at /mnt/memory/, spawn 20 parallel specialist agents, define outcomes the eval pipeline reads automatically, and emit lifecycle events to a webhook endpoint. +> - **Skills Publisher.** `sandcastle publish-skills --upload` converts every workflow into a tar.gz SKILL.md and uploads to `/v1/skills`. Workflows now callable from every Anthropic Skills-aware client. +> - **Trajectory Replay step type.** SHA-256 over a recorded tool-call sequence + diff against the candidate run. Because the audit trail is a hash chain, the replay is **cryptographically verifiable** - LangSmith and Braintrust don't ship this. +> - **Agent SDK runtime** as `runtime: "agent-sdk"` alternative. In-process Claude agents for EU sovereignty / air-gapped / regulated teams. +> - **Computer Use** integration (`computer_20251124` beta) with 8-item safety pre-flight + new `type: computer-use` step. +> - **MCP Elicitation** (6th primitive, spec rev 2025-11-25). Workflows can ask the user for missing input mid-execution. +> - **Live Agent Reasoning panel** in the dashboard - SSE stream of agent.thinking, agent.tool_use, agent.message events on the run detail page. +> - Plus 5 Tier 1 wire fixes (`tools_enabled`, sampling params, `stream`, pricing table, fallback chain). +> - 15,176 tests passing, 169 new. PyPI: `pip install sandcastle-ai==0.32.0`. +> +> Previous: **v0.31 - "Compliance & Connections"** (May 14, 2026): EU AI Act landing + 10 compliance templates, MCP-first publishing, eval gates, dashboard error boundaries, Codex audit rounds 9+10. >

diff --git a/site/eu-ai-act/index.html b/site/eu-ai-act/index.html index d84f3de..33a5f3c 100644 --- a/site/eu-ai-act/index.html +++ b/site/eu-ai-act/index.html @@ -638,7 +638,7 @@

Stop drafting policy. Start shipping artefacts.

- Sandcastle v0.31 - BSL 1.1 (Apache 2.0 after 2030)
+ Sandcastle v0.32 - BSL 1.1 (Apache 2.0 after 2030)
© 2026 Created by Tomas Pflanzer @gizmax
diff --git a/site/index.html b/site/index.html index 7bf99c5..64b8bd5 100644 --- a/site/index.html +++ b/site/index.html @@ -1761,7 +1761,7 @@
- v0.31 - Compliance & Connections + v0.32 - Claude Agents Deep Integration

Describe what you want. Go home.
@@ -2725,7 +2725,7 @@

Hush