v0.7.12 — coordinated stable release

Released: 2026-05-26 · Codename: Stable Foundations

Why v0.7.12 matters

v0.7.12 is the largest release since the project began — a single, coordinated promotion that takes every alpha/beta provider to its first stable line and ships the cross-cutting native-tool observability that powers it. After this release the project returns to a bug-fix / single-provider cadence.

Highlights

🟢 Five packages graduate to Stable (Development Status :: 5 - Production/Stable) in one release: nucleusiq-anthropic 0.2.0, nucleusiq-ollama 0.2.0, nucleusiq-groq 0.1.0, nucleusiq-mcp 0.1.0. Core (nucleusiq 0.7.12) and the already-stable nucleusiq-openai 0.7.0 / nucleusiq-gemini 0.3.0 ship aligned, with every provider flooring on nucleusiq>=0.7.12.
🧠 Anthropic Phase B feature-complete — AnthropicTool.web_search() / web_fetch() / code_execution(), prompt caching (cache_system / cache_tools), extended thinking (thinking="low"|"medium"|"high"|"max"), and strict_tools / disable_parallel_tool_use. Backed by 6 live integration tests against the real Anthropic API.
🔎 Native-tool observability in core — every traced tool call carries executed_by: Literal["local", "provider"]; LLMCallRecord gains provider, request_id, organization_id, stop_reason, cache_read_input_tokens, cache_creation_input_tokens, and a generic metadata dict. Populated by every provider in this release.
👁️ Ollama vision wire — multimodal content lists with image_url parts are split into Ollama's message.content + message.images fields, with data: URL decoding and HTTP-URL warnings.
✅ 3 705+ tests passing across the monorepo — including 79 Groq, 98 Ollama, 151 Anthropic + 6 live, 232 OpenAI, 292 Gemini, and 235 MCP unit tests.

Package matrix

Package	Before	Now	Status	Floor
`nucleusiq`	`0.7.11`	`0.7.12`	🟢 Core release	—
`nucleusiq-anthropic`	`0.1.0a1` (alpha)	`0.2.0`	🟢 Stable	`nucleusiq>=0.7.12`
`nucleusiq-ollama`	`0.1.0a1` (alpha)	`0.2.0`	🟢 Stable	`nucleusiq>=0.7.12`
`nucleusiq-groq`	`0.1.0b1` (beta)	`0.1.0`	🟢 Stable	`nucleusiq>=0.7.12`
`nucleusiq-mcp`	`0.1.0b1` (beta)	`0.1.0`	🟢 Stable	`nucleusiq>=0.7.12`
`nucleusiq-openai`	`0.6.4`	`0.7.0`	🟢 Stable	`nucleusiq>=0.7.12`
`nucleusiq-gemini`	`0.2.6`	`0.3.0`	🟢 Stable	`nucleusiq>=0.7.12`

Anthropic 0.2.0 Stable — Phase B feature-complete

What's new

nucleusiq-anthropic graduates from alpha to Production/Stable with full Phase B support. 151 unit tests + 6 live integration tests, 95.91% coverage (gate ≥ 95%).

Capability	API surface
Native server tools	`AnthropicTool.web_search()` · `AnthropicTool.web_fetch()` · `AnthropicTool.code_execution()`
Prompt caching	`AnthropicLLMParams(cache_system=True, cache_tools=True)`
Extended thinking	`AnthropicLLMParams(thinking="low"\\|"medium"\\|"high"\\|"max")`
Strict tools	`AnthropicLLMParams(strict_tools=True)`
Single-tool turn	`AnthropicLLMParams(disable_parallel_tool_use=True)`
Server-tool observability	`AnthropicLLMResponse.server_tool_calls: list[ServerToolCall]` (auto-emitted as `ToolCallRecord(executed_by="provider")`)
`LLMCallRecord` enrichment	`provider`, `request_id`, `organization_id`, `stop_reason`, `cache_read_input_tokens`, `cache_creation_input_tokens`

Live integration tests (6 / 6 passing)

Test	Validates
`test_live_web_search_emits_server_tool_calls`	`web_search` server-tool emission
`test_live_code_execution_emits_server_tool_calls_with_result`	`code_execution` round-trip + result parsing
`test_live_prompt_caching_reads_cache_on_second_call`	Cache write → cache read on consecutive calls
`test_live_extended_thinking_completes[low]`	Thinking budget `low` (2 000)
`test_live_extended_thinking_completes[medium]`	Thinking budget `medium` (8 000)
`test_live_disable_parallel_tool_use_emits_single_tool_call`	Single-tool turn semantics

Default model: claude-sonnet-4-5-20250929 (override with ANTHROPIC_PHASE_B_MODEL=<id>).

→ Deep-dive guides: Native server tools · Prompt caching · Extended thinking.

Ollama 0.2.0 Stable — vision wire

OpenAI-style multimodal messages now work natively

nucleusiq-ollama/_shared/wire.py adds _split_data_url and _extract_text_and_images helpers; sanitize_messages now produces:

# in:
{"role": "user", "content": [
    {"type": "text", "text": "What's in this image?"},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0K..."}},
]}
# out (Ollama wire shape):
{"role": "user", "content": "What's in this image?", "images": ["iVBORw0K..."]}

Direct images arrays pass through untouched.
Raw-string images entries are accepted.
data:image/...;base64,... URLs are decoded.
HTTP image URLs trigger a warning and are dropped — NucleusIQ does not fetch remote images on your behalf.

98 unit tests, 99.85% coverage. Vision now requires a multimodal Ollama model (llama3.2-vision, llava, bakllava, …).

→ See the Ollama provider guide.

Native-tool observability (cross-cutting)

One observability signal across every provider

Adding ToolCallRecord.executed_by makes it trivial to split locally-run vs provider-run tools — no matter which LLM the agent used:

local    = [tc for tc in result.tool_calls if tc.executed_by == "local"]
provider = [tc for tc in result.tool_calls if tc.executed_by == "provider"]

Provider	Native tool wire types surfaced
Anthropic	`server_tool_use` + per-tool `*_tool_result` (e.g. `code_execution_tool_result`, `web_search_tool_result`)
OpenAI	`web_search_call`, `code_interpreter_call`, `file_search_call`, `computer_use_call`, `image_generation_call`
Gemini	`executable_code` + `code_execution_result` (paired → `code_execution`), `grounding_metadata` (→ `google_search`)
Groq	`message.executed_tools` parsed into `GroqLLMResponse.server_tool_calls` (stub today; full hosted-tool Responses API in `0.2.x`)

→ See Native server tools and Observability.

Migration from v0.7.11

Zero breaking changes

v0.7.12 is additive only for application code. All new fields default to safe values (executed_by="local", cache_*_tokens=0, optional strings = None).

The only required action is to bump pins:

pip install -U \
    "nucleusiq>=0.7.12,<0.8" \
    "nucleusiq-openai>=0.7.0,<0.8" \
    "nucleusiq-gemini>=0.3.0,<0.4" \
    "nucleusiq-anthropic>=0.2.0,<0.3" \
    "nucleusiq-groq>=0.1.0,<0.2" \
    "nucleusiq-ollama>=0.2.0,<0.3" \
    "nucleusiq-mcp>=0.1.0,<0.2"

Test gate

Suite	Tests	Coverage
`nucleusiq` core	1 543 unit	≥ 95%
`nucleusiq-openai`	232 unit	≥ 95%
`nucleusiq-gemini`	292 unit	≥ 95%
`nucleusiq-anthropic`	151 unit + 6 live integration	95.91%
`nucleusiq-groq`	79 unit	92.51%
`nucleusiq-ollama`	98 unit	99.85%
`nucleusiq-mcp`	235 unit + 13 live integration	98.68%

→ Full per-package breakdown lives in the Changelog and the monorepo CHANGELOG.md.

What's next

After v0.7.12 the project returns to a bug-fix / single-provider cadence:

nucleusiq-anthropic 0.3.x — Phase C (Memory tool, computer_use, bash).
nucleusiq-groq 0.2.x — Groq Responses API + hosted tools.
nucleusiq-mcp 0.2.x — MCP resources & prompts, sampling, progress callbacks.
nucleusiq 0.7.13+ — operational polish, bug fixes, retry tuning.