Thinking & Reasoning

ELI5: every provider sells “think harder” under a different name (extended thinking, reasoning effort, thinking level). deeda gives you one workflow field plus per-provider knobs; the harness maps them.

The portable field

agent:
  thinking:
    type: enabled        # enabled | disabled | adaptive
    budget_tokens: 16000 # or "{{thinking_budget}}" template

Per-provider mapping

Provider/runtime	Knob (under `harness_config.sdk_settings.<provider>`)	Values
`anthropic-agent-sdk`	`thinking_type`, `thinking_budget_tokens`	enabled/disabled/adaptive + token cap; interleaved & adaptive thinking implemented
`openai-codex-sdk`, `openai-agents-sdk`, `openai-responses-api`	`reasoning_effort`	model-dependent (o-series / GPT-5 class)
`gemini-antigravity-cli`	`antigravity_thinking_level`	`MINIMAL` \| `LOW` \| `MEDIUM` \| `HIGH`
`gemini-genai-sdk`	thinking on/off/dynamic is implemented at the capability level; no extra knob needed	—
`local-llm`, `local-voice`	none — local models expose no thinking control	—

Caveats worth knowing: Anthropic summarized thinking is model-dependent (4-series vendor default, no deeda selector); OpenAI reasoning_effort is rejected by non-reasoning models — pin a supporting model; Gemini deep-research is policy-blocked (long-running paid surface).

Example: pay for reasoning only on the hard stage

state_machine:
  stage_order: [scan, deep_fix]
  states:
    scan:
      agent:
        thinking: { type: disabled }
    deep_fix:
      agent:
        thinking: { type: enabled, budget_tokens: 24000 }

When to use

Task	Setting
Bulk/mechanical edits, summaries	`disabled` — thinking is wasted spend
Debugging, architecture, proofs	`enabled` with 16–32k budget
Mixed/unknown difficulty	`adaptive` (Anthropic) — model decides

​The portable field

​Per-provider mapping

​Example: pay for reasoning only on the hard stage

​When to use

​See also

The portable field

Per-provider mapping

Example: pay for reasoning only on the hard stage

When to use

See also