Reasoning

StreamOptions.ThinkingLevel maps a normalized level (minimal, low, medium, high, xhigh) to each provider’s reasoning configuration. Thinking blocks emerge as ai.Thinking content with an optional Signature for cross-turn passthrough.

Extended / Adaptive Thinking

Provider	API	pi-go	Notes
Anthropic	✅ Adaptive Thinking — effort low/med/high/xhigh/max; legacy token budget removed on 4.7+	✅	pi-go sends adaptive thinking with output_config.effort mapped from the thinking level; pre-4.6 models get the token-budget form
OpenAI Chat	✅ reasoning_effort on GPT-5.2 / o-series	✅	gated by compat.SupportsReasoningEffort (openai.go:308-310)
OpenAI Responses	✅ reasoning.effort (none/low/med/high/xhigh) + auto summary	✅	(convert.go:290-301, openairesponses.go:318-326)
Google Gemini	✅ thinkingConfig.thinkingBudget (Gemini 2.5+); thinking_level (Gemini 3+)	⚠️	the google provider does not set thinkingConfig
Claude CLI	✅ --effort (low/medium/high/xhigh/max, model-dependent)	✅	level → --effort; minimal floors to low, off/unknown omit the flag (claudecli claude.go:99, agent/claude config.go:71)
Codex CLI	✅ model_reasoning_effort (minimal/low/medium/high/xhigh)	✅	level → -c model_reasoning_effort; off/unknown omit the override (codexcli codex.go:112, agent/codex config.go:84)
Cursor CLI	❌ no reasoning-effort flag; reasoning bound to the model name (*-thinking)	—	ThinkingLevel ignored; pass a thinking-capable model instead (cursorcli cursor.go:129)

Encrypted Reasoning Passthrough

Providers return reasoning blocks with a signed/encrypted blob. Clients echo the blob back on subsequent turns; the model decrypts it server-side and continues its prior thinking.

Provider	API	pi-go	Notes
Anthropic	✅ thinking signatures	✅	round-tripped via Signature (convert.go:307)
OpenAI Chat	❌	—	reasoning kept server-side via stored conversation only
OpenAI Responses	✅ reasoning.encrypted_content; ZDR-compatible	✅	ResponseIncludableReasoningEncryptedContent (openairesponses.go:323-325)
Google Gemini	✅ thoughtSignature (Gemini 3+)	✅	preserved on Thinking and ToolCall blocks (google.go:187, 228-229)

Provider Documentation

• Anthropic — Extended / Adaptive thinking
• OpenAI — Reasoning models
• OpenAI — Encrypted reasoning content
• OpenAI — Reasoning best practices
• Google Gemini — Thinking
• Google — Thought signatures

pi-go Gaps

• Thinking display (summarized / omitted) — not exposed; the provider leaves the API default.
• Google HTTP provider asymmetry — pkg/ai/provider/google does not set thinkingConfig; thinking is uncontrolled there.
• Reasoning summaries are wired only for OpenAI Responses; Anthropic and Gemini do not surface summaries.
• Tools-in-reasoning chain (OpenAI o3/o4-mini calling tools mid-CoT) — no special handling.
• Disabling thinking (level none / zero budget) is not consistently expressible across providers.
• Encrypted reasoning passthrough is the strongest area of provider parity — no major gaps there. Verify callers correctly include thinking blocks before tool-result turns or risk losing reasoning state.