v0.7.0

What's New

PromptProvider — select which primitives and decompositions appear in prompts

New PromptProvider base class that controls which primitives and decompositions the LLM sees in the prompt. Subclass it and override select_primitives / select_decompositions to filter by rich metadata (PrimitiveInfo, DecompositionInfo) including docstrings, read_only, deterministic, parameters, return types, and source code.

from opensymbolicai import PromptProvider, PrimitiveInfo, PlanExecuteConfig

class ReadOnlyOnly(PromptProvider):
    def select_primitives(self, available: list[PrimitiveInfo]) -> list[str]:
        return [p.name for p in available if p.read_only]

config = PlanExecuteConfig(prompt_provider=ReadOnlyOnly())

• Only affects prompt construction — all primitives remain available for execution
• Works across all blueprints: PlanExecute, DesignExecute, GoalSeeking
• New exported models: PromptProvider, PrimitiveInfo, DecompositionInfo, ParameterInfo

Full Changelog: v0.6.2...v0.7.0

v0.6.2

What's Changed

New Features

• Anonymous usage telemetry via PostHog (feat: add anonymous usage telemetry)
• API key auth support for observability HTTP transport (feat: add API key auth support to observability HTTP transport)

Bug Fixes

• Pair LLM request/response into single deferred spans for correct trace duration (fix: pair LLM request/response into single deferred spans)
• Allow isinstance in plan execution — removed from DANGEROUS_BUILTINS and added to DEFAULT_ALLOWED_BUILTINS, with allowed_builtins now checked before the blocklist so user config can override (fix: allow isinstance in plan execution)

Full Changelog: v0.5.1...v0.6.2

v0.5.1

What's Changed

Features

• Observability framework — new tracing and event transport system (FileTransport, HttpTransport, InMemoryTransport)
• GoalSeeking observability instrumentation with session_id support
• Determinism metadata, signature hashing, and prompt splitting APIs
• Prompt section demarcation markers for splittable LLM prompts
• Auto-document Pydantic BaseModel schemas in primitive prompts

Docs & Maintenance

• Improved PyPI discoverability and added benchmark results to README
• Updated README with DesignExecute, GoalSeeking, and exception documentation

Full Changelog: v0.4.1...v0.5.1

v0.4.1

What's in this release

OpenSymbolicAI Core — a Python framework for building symbolic AI agents where LLMs plan and typed Python functions execute.

Blueprints

• PlanExecute — LLM generates a plan, runtime executes it against @primitive functions
• DesignExecute — control flow support (loops, conditionals) for multi-step workflows
• GoalSeeking — iterative goal-directed agents with built-in evaluation and retry

Highlights

• Symbolic firewall: LLM never touches raw data, only function signatures
• All outputs are validated Pydantic models
• Full execution tracing with before/after state snapshots
• Multi-provider LLM support (Ollama, OpenAI, Anthropic, Fireworks, Groq)
• Checkpoint system for distributed execution and state persistence

Examples

• Scientific calculator with 120-intent benchmark suite
• Function optimizer using GoalSeeking
• Shopping cart using DesignExecute

Benchmark Results (calculator, 120 intents)

Model	Accuracy
gpt-oss:20b	100%
qwen3:1.7b	100%
qwen3:8b	100%
gemma3:4b	94%
phi4:14b	80%

---

Full documentation: https://opensymbolicai.github.io/core-py/