goclaw/docs/12-extended-thinking.md

12 - Extended Thinking

Overview

Extended thinking allows LLM providers to "think out loud" before producing a final response. When enabled, the model generates internal reasoning tokens that improve response quality for complex tasks at the cost of additional token usage and latency. GoClaw now supports both the legacy coarse thinking_level setting and a provider-first reasoning policy for capability-aware GPT-5/Codex control.

---

1. Configuration

The reusable default now lives on the provider in settings.reasoning_defaults. Agents consume that default by inheriting it, or store a custom override in top-level reasoning_config. thinking_level remains the backward-compatible coarse shim for older builds.

Level	Behavior
off	Thinking disabled (default)
low	Minimal thinking — quick reasoning
medium	Moderate thinking — balanced reasoning
high	Maximum thinking — deep reasoning for complex tasks

Provider default

{
  "provider_type": "chatgpt_oauth",
  "settings": {
    "reasoning_defaults": {
      "effort": "high",
      "fallback": "provider_default"
    }
  }
}

Agent inherits provider default

{
  "reasoning_config": {
    "override_mode": "inherit"
  }
}

Agent custom override

{
  "thinking_level": "high",
  "reasoning_config": {
    "override_mode": "custom",
    "effort": "xhigh",
    "fallback": "downgrade"
  }
}

Rules:

• Unset provider defaults and unset agent reasoning both resolve to off.
• settings.reasoning_defaults is provider-owned and reusable across agents.
• reasoning_config.override_mode accepts inherit|custom.
• thinking_level still accepts off|low|medium|high.
• reasoning_config.effort accepts off|auto|none|minimal|low|medium|high|xhigh.
• reasoning_config.fallback accepts downgrade|off|provider_default.
• Existing legacy other_config.reasoning payloads without override_mode are treated as custom overrides for backward compatibility.
• Read path resolves provider defaults first, then applies agent inherit/custom semantics, then falls back to legacy thinking_level.
• Write path keeps a derived coarse thinking_level only for custom agent overrides so rollback to older GoClaw builds stays safe.

---

2. Provider Support

Each provider maps the normalized reasoning policy to its own implementation parameters.

flowchart TD
    CONFIG["Provider defaults +<br/>agent inherit/custom"] --> CHECK{"Provider supports<br/>thinking?"}
    CHECK -->|No| SKIP["Send request<br/>without thinking"]
    CHECK -->|Yes| MAP{"Provider type?"}

    MAP -->|Anthropic| ANTH["Budget tokens: 10,000<br/>Header: anthropic-beta<br/>Strip temperature"]
    MAP -->|OpenAI-compat| OAI["Capability-aware effort<br/>or provider default"]
    MAP -->|DashScope| DASH["enable_thinking: true<br/>Budget: 16,384 tokens<br/>⚠ Model-specific + tools limitation"]
    MAP -->|Codex| CODEX["Capability-aware effort<br/>+ trace metadata"]

    ANTH --> SEND["Send to LLM"]
    OAI --> SEND
    DASH --> SEND
    CODEX --> SEND

Anthropic (Native)

Thinking Level	Budget Tokens
low	4,096
medium	10,000
high	32,000

When thinking is enabled:

• Adds thinking: {type: "enabled", budget_tokens: N} to the request body
• Sets anthropic-beta: interleaved-thinking-2025-05-14 header
• Strips temperature parameter (Anthropic requirement — cannot use temperature with thinking)
• Auto-adjusts max_tokens to accommodate thinking budget (budget + 8,192 buffer)

OpenAI-Compatible and Codex (GPT-5 / Codex families)

Known GPT-5/Codex models use a static capability registry. The runtime resolves:

• requested effort
• actual effective effort
• fallback policy used
• whether the model default was used
• whether the source was the provider default or an agent override

If the model is known:

• supported efforts pass through unchanged
• unsupported efforts are normalized via downgrade, off, or provider_default
• auto means "use the model default effort"

If the model is unknown:

• explicit non-auto effort is passed through as requested
• auto leaves provider-default reasoning untouched

Reasoning content still streams in the provider-native format, and span metadata now records the source plus requested versus effective effort.

DashScope (Alibaba Qwen)

Thinking Level	Budget Tokens
low	4,096
medium	16,384
high	32,768

Enables thinking via enable_thinking: true plus a thinking_budget parameter.

Model-specific support: Only certain Qwen3 models accept the enable_thinking / thinking_budget parameters:

• Qwen3.5 series: qwen3.5-plus, qwen3.5-turbo (thinking + vision)
• Qwen3 hosted: qwen3-max
• Qwen3 open-weight: qwen3-235b-a22b, qwen3-32b, qwen3-14b, qwen3-8b

Other models (e.g., qwen3-plus, qwen3-turbo) silently skip thinking injection to avoid API errors.

Important limitation: DashScope does not support streaming when tools are present. When an agent has tools enabled and thinking is active, the provider automatically falls back to non-streaming mode (single Chat() call) and synthesizes chunk callbacks to maintain the event flow.

Codex (ChatGPT OAuth Responses API)

Codex natively supports extended reasoning through its Responses API. Thinking and reasoning tokens are streamed as discrete reasoning events with summary fragments.

Token tracking: Reasoning token count is exposed in response.completed / response.incomplete events as OutputTokensDetails.ReasoningTokens and accessible via ChatResponse.Usage.ThinkingTokens.

Model metadata: /v1/providers/{id}/models is now the backend source of truth for the ChatGPT OAuth model list and any known reasoning capabilities.

---

3. Streaming

When thinking is active, reasoning content streams to the client alongside regular content.

flowchart TD
    LLM["LLM generates response"] --> THINK["Thinking tokens<br/>(internal reasoning)"]
    THINK --> CONTENT["Content tokens<br/>(final response)"]

    THINK -->|Stream| CHUNK_T["StreamChunk<br/>Thinking: 'reasoning text...'"]
    CONTENT -->|Stream| CHUNK_C["StreamChunk<br/>Content: 'response text...'"]

    CHUNK_T --> CLIENT["Client receives<br/>thinking + content separately"]
    CHUNK_C --> CLIENT

Provider-Specific Streaming Events

Provider	Thinking Event	Content Event
Anthropic	thinking_delta in content blocks	text_delta in content blocks
OpenAI-compat	reasoning_content in delta	content in delta
DashScope	Same as OpenAI (when tools absent)	Same as OpenAI
Codex	reasoning items with text summaries	content items

Token Estimation

Thinking tokens are estimated as character_count / 4 for context window tracking. This rough estimate ensures the agent loop can account for thinking overhead when calculating context usage.

---

4. Tool Loop Handling

Extended thinking interacts with multi-turn tool conversations. When the LLM calls a tool and then needs to continue reasoning, thinking blocks must be preserved correctly across turns.

flowchart TD
    TURN1["Turn 1: LLM thinks + calls tool"] --> PRESERVE["Preserve thinking blocks<br/>in raw assistant content"]
    PRESERVE --> TOOL["Tool executes,<br/>result appended to history"]
    TOOL --> TURN2["Turn 2: LLM receives history<br/>including preserved thinking blocks"]
    TURN2 --> CONTINUE["LLM continues reasoning<br/>with full context"]

Anthropic Thinking Block Preservation

Anthropic requires thinking blocks (including their cryptographic signatures) to be echoed back in subsequent turns. GoClaw handles this through RawAssistantContent:

1. During streaming, raw content blocks are accumulated — including thinking type blocks with their signature fields
2. When the assistant message is appended to history, the raw blocks are preserved
3. On the next LLM call, these blocks are sent back as-is, ensuring the API can validate thinking continuity

This is critical for correctness: if thinking blocks are dropped or modified, the Anthropic API may reject the request or produce degraded responses.

Other Providers

OpenAI-compatible providers handle thinking/reasoning content as metadata. The reasoning_content is accumulated during streaming but does not require special passback handling — each turn's reasoning is independent.

---

5. Limitations

Provider	Limitation
DashScope	Cannot stream when tools are present — falls back to non-streaming mode. Only specific Qwen3 models support thinking.
Codex	Reasoning tokens tracked via API response (not in streaming chunks themselves)
Anthropic	Temperature parameter stripped when thinking is enabled
All	Thinking tokens count against the context window budget
All	Thinking increases latency and cost proportional to the budget level
GPT-5/Codex unknown models	GoClaw allows explicit effort passthrough but does not claim a capability contract

---

6. Reasoning Content Stripping (Phase 6 — OpenClaw TS port)

Some models emit chain-of-thought reasoning tokens even when effort="off" is specified. To prevent that raw CoT from reaching end users, GoClaw supports a StripThinking flag on ReasoningDecision.

Models Known to Leak CoT

Auto-flagged via modelLeaksReasoning(model) in internal/providers/reasoning_resolution.go:

• Kimi family: any model name containing kimi (case-insensitive, e.g. kimi-k2, moonshot/kimi-k2-thinking)
• DeepSeek-Reasoner: any model name containing deepseek-reasoner

The allowlist is a simple substring check — extendable as new leaky models appear.

Auto-Enable Flow

1. ResolveReasoningDecision runs its normal flow, then a defer checks: if EffectiveEffort == "off" && modelLeaksReasoning(model), it sets decision.StripThinking = true.
2. The agent loop (loop_pipeline_callbacks.go) reads decision.StripThinking and propagates it into chatReq.Options[providers.OptStripThinking] = true.
3. Each provider's streaming handler reads OptStripThinking from options at the top of ChatStream/Chat and applies guard clauses.

Implementation Per Provider

• Anthropic (anthropic_stream.go, anthropic.go): thinking_delta events skip result.Thinking accumulation AND the onChunk emit when stripping, but thinkingChars still increments so Usage.ThinkingTokens stays billable. RawAssistantContent (content blocks for tool-use passback) is never touched. Non-streaming Chat() clears resp.Thinking post-parse.
• OpenAI (openai_chat.go): streaming guards reasoning/reasoning_content delta accumulation; non-streaming clears resp.Thinking post-parse. Usage.ThinkingTokens read from the usage chunk independently.
• Codex (codex.go): processSSEEvent takes an extra stripThinking bool param; the reasoning item case skips summary text appending when set. Usage still extracted from response.completed.
• DashScope: inherits both guards via OpenAIProvider embedding — no separate implementation needed.

Invariants Preserved

Field	Stripping effect	Rationale
ChatResponse.Thinking	Cleared	User-visible output
Usage.ThinkingTokens	Unchanged	Billing accuracy (Phase 1 depends on it)
RawAssistantContent	Unchanged	Anthropic tool-use replay requires raw thinking blocks
onChunk(StreamChunk{Thinking})	Not emitted	Streaming UI display

---

7. Observability

Each LLM span can now include a metadata.reasoning section with:

• source
• requested_effort
• effective_effort
• fallback
• reason
• supported_levels
• used_provider_default

This makes silent downgrades or provider-default decisions visible in traces instead of leaving them implicit.

---

File Reference

Module	Path	Purpose
Provider types & reasoning	internal/providers/types.go, internal/providers/reasoning_capability.go, internal/providers/reasoning_resolution.go, internal/providers/reasoning_observation.go	ThinkingCapable interface, GPT-5/Codex capability registry, reasoning decision engine, trace metadata
Anthropic thinking	internal/providers/anthropic.go, internal/providers/anthropic_stream.go, internal/providers/anthropic_request.go	Budget mapping, beta header, thinking_delta streaming, block preservation for tool loops
OpenAI-compat, DashScope & Codex	internal/providers/openai.go, internal/providers/dashscope.go, internal/providers/codex.go	Reasoning effort mapping, DashScope tools+streaming fallback, Codex reasoning event streaming

Use grep or your editor's symbol search for specific files.

---

Cross-References

Document	Relevant Content
02-providers.md	Provider architecture, supported providers
01-agent-loop.md	LLM iteration loop, streaming chunk handling