v0.4.0 — Staged process and AgenticLoop

What's new in v0.4.0

All features are backward compatible — no breaking changes.

Staged process

Third orchestration mode: stages run in sequence, while tasks within a single stage run concurrently. The output of each stage feeds the following ones via Task.Context. Stages marked Optional continue on failure.

crew.Process = crewai.Staged
crew.Stages = []crewai.Stage{
    {Name: "collect", Tasks: []*crewai.Task{researchA, researchB}},
    {Name: "synthesize", Tasks: []*crewai.Task{write}},
}

Agentic loop

Optional Plan → Execute → Evaluate → Refine cycle. Opt-in via Agent.Loop or Task.Loop. Supports independent evaluator, pass threshold, skip-plan, and rewrite-only refine.

agent.Loop = crewai.NewAgenticLoop(
    crewai.WithMaxRefinements(3),
    crewai.WithPassThreshold(80),
)

New sentinels

• ErrNoStages
• ErrEvaluationFailed
• ErrInvalidEvaluation

Examples

• go run ./examples/staged
• go run ./examples/agentic_loop (offline, mock LLM)

Docs

See CHANGELOG and the bilingual README / docs.

Full changelog: https://github.com/rhgs/crewai-go/blob/v0.4.0/CHANGELOG.md

v0.3.0 — Native tool calling, Web search, slog migration

What's new in v0.3.0

Native tool calling

Agent.ToolMode ("react" | "native") selects between the existing text-based ReAct loop and native function calling via the new ToolCallingLLM interface. Implemented for Ollama, OpenAI, Anthropic, and Mock.

• ToolSpec, ToolCall, ToolCallResponse, ToolTrace types
• ToolCallingLLM.CallWithTools(ctx, messages, toolSpecs) — provider-native tool dispatch
• Security limits: max args/output/response bytes, JSON depth, arg validation
• ToolTraces in TaskOutput for observability

Web search — agent-driven

WebSearcher interface + SearchWeb(ctx, llm, query, max) helper for direct web search from Go code. Implemented by:

• Ollama — POST /api/web_search (pure search, no model invocation)
• OpenAI — web_search_options with search-capable models
• Anthropic — web_search_20250305 server tool
• xAI — delegates to OpenAI-compatible client

Web search — model-driven

tools.WebSearchTool implements Tool + FactSource for the ReAct loop with pluggable SearchProvider:

Provider	API key	Free tier
Wikipedia	❌	100% free (default)
LangSearch	✅	100% free
Serpstack	✅	1000/month free
DuckDuckGo	❌	Optional, may be blocked
Google CSE	✅ + CSE ID	Paid
Brave	✅	Paid

SSRF protection: blocks non-http(s) schemes, loopback, private, link-local, and unspecified IPs. DNS rebinding prevention via net.LookupIP (fail-closed).

Logging migration to log/slog (#15)

Replaces the custom Logger interface with *slog.Logger from the standard library.

• Crew.WithLogger(*slog.Logger) *Crew and Agent.WithLogger(*slog.Logger) *Agent — fluent setters
• defaultLogger(verbose) — text handler on stderr, LevelDebug when Verbose=true, LevelError when Verbose=false (matches legacy "silent unless verbose")
• All log calls use InfoContext/DebugContext/WarnContext with structured key-value pairs
• Verbose field preserved for backward compatibility
• Subpackages (llm/*, tools/*) remain logging-free

Security

• Secret redaction in logs (redact.go): provider errors logged via WarnContext pass through redactError, which masks:
  • Long alphanumeric tokens (≥20 chars), preserving first/last 4 when ≥24
  • Bearer <token> in HTTP-style messages
  • api_key=/token=/key=/secret= query-string values
• Logging safety docs: README + doc.go warn about Debug-level logs containing full LLM output
• WithLogger thread-safety: documented as not concurrent-safe; idempotent (last wins)

Examples

• examples/logging/ — custom *slog.Logger with redaction wrapper
• examples/native_tools/ — native function calling demonstration

Stats

• 70 files changed, +9,208 / -99 lines (vs v0.2.0)
• Zero external dependencies (stdlib only)
• Test coverage: 96.3% (root package)
• All packages pass under -race

Full changelog: CHANGELOG.md

v0.2.0 — Structured Output, Guardrails, Facts & Provenance

What's New

All features are backward compatible — no breaking changes.

Structured Output

Task.Structured (*StructuredOutput) requires the model to produce JSON validated against a JSON Schema, with a bounded repair loop (RepairMax, default 2). New sentinels ErrInvalidOutput and ErrRepairBudgetExceeded. Minimal in-house schema validator (type, properties, required, enum, items) — stdlib only, no new dependencies.

Guardrails

Crew-level (Crew.Guardrails) and task-level (Task.Guardrail) post-output validation hooks that block publication of outputs violating business invariants. New sentinel ErrBlockedByGuardrail. Functional options WithGuardrails (crew) and WithGuardrail (task). Complements structured-output schema validation (shape vs. meaning).

Facts & Provenance

First-class Fact type populated only by FactSource tools, never by the LLM. Facts carry source org, source URL, collection time, and payload hash (SHA-256). NewFactSourceTool constructor, AllFactsProvenanced helper for guardrails, dedupFacts by PayloadHash. CrewOutput.Facts and TaskOutput.Facts.

Other

• gofmt applied to all Go files (CI now enforces formatting).

---

Full Changelog: v0.1.0...v0.2.0

v0.1.0 — crewai-go: idiomatic Go port of CrewAI core

First public release — an idiomatic Go port of the CrewAI framework core. Zero external dependencies (stdlib only).

🌐 Docs in English (default) and Português · CHANGELOG / CHANGELOG.pt-BR.md

✨ Added

• Core orchestration: Agent, Task, Crew, Process, Tool, Memory, and a text-based ReAct executor.
• Processes: Sequential (default) and Hierarchical (manager-driven delegation via ManagerLLM / ManagerAgent).
• Context & interpolation: chain task outputs with WithContext; inject {key} variables through Crew.Kickoff.
• LLM providers (stdlib only, no external deps):
  • OpenAI and compatible endpoints (Groq, Azure, Ollama /v1, …) — llm/openai
  • Anthropic (Claude) — llm/anthropic
  • Ollama local + Ollama Cloud — llm/ollama
  • xAI (Grok) via API key or subscription OAuth (Device Flow RFC 8628 + PKCE + refresh + persistence) — llm/xai
  • Deterministic mock for tests — llm/mock
• Built-in tools: Calculator (safe recursive-descent parser), CurrentTime, WordCount — tools package.
• Memory: concurrency-safe in-process memory with substring search.
• Docs & examples: English README + 7 guides; Portuguese mirrors (docs/pt-BR/); 7 runnable examples; hermetic tests (~90% core coverage).

🔒 Security

• Secrets never committed; .gitignore protects .claude/, .env, *token.json.
• xAI OAuth token persisted with 0600 permissions.

📦 Install

go get github.com/rhgs/crewai-go@v0.1.0

📝 Notes

• License: MIT
• Known limitations of this version: simplified hierarchical delegation (no runtime inter-agent calls), no streaming, in-process memory only, no native function calling. See PLAN.md for the full roadmap.

---

Primeira release pública — um port idiomático do núcleo do framework CrewAI para Go. Zero dependências externas (apenas stdlib).