Extended thinking

Spettro exposes a normalized thinking level that controls how much extra reasoning compute the active model spends before answering. The setting is independent of the active provider — the same /thinking command works regardless of who serves the model — and each provider maps it to its own native parameter: a thinking token budget for Anthropic, a reasoning_effort value for OpenAI and every OpenAI-compatible backend. Models that don't reason simply ignore it.

Levels

Level	Anthropic budget tokens	OpenAI-style `reasoning_effort`
`off`	0 — sent as `thinking.type=disabled` so behaviour is explicit	`none` — sent as `reasoning_effort: "none"` so models that reason by default actually stop; a server that rejects the value gets a retry with the parameter omitted
unset (never configured)	same as `off`	parameter omitted (provider default)
`low`	~2 048	`low`
`medium`	~5 120	`medium`
`high`	~16 384	`high`
`x-high`	~32 768	`xhigh`
`max`	~100 000	`xhigh` (effort is an enum; `xhigh` is its highest value)

The exact mappings live in internal/provider/provider.go as ThinkingBudgetTokens and ReasoningEffort so they stay consistent across adapters and are easy to test.

Provider support

Provider	Native parameter	Notes
Anthropic (`claude-opus-4*`, `claude-sonnet-4-5`, etc.)	`thinking.type=enabled` + `thinking.budget_tokens`	Spettro automatically bumps `max_tokens` so it stays > the budget (Anthropic rejects requests where they meet or overlap).
OpenAI (`o3`, `o4-mini`, `gpt-5*`, …)	`reasoning.effort` (Responses API) / `reasoning_effort` (chat)	Only forwarded to models OpenAI marks as reasoning-capable, so setting a level does not break non-reasoning models like `gpt-4o`.
OpenAI-compatible (Groq, xAI, DeepSeek, Google's compat endpoint, Spettro Subscription, local Ollama/LM Studio, …)	`reasoning_effort` on chat completions	Sent for every explicit level — `off` sends `none`, so reasoning-capable models can be disabled. Servers that don't know the parameter ignore it; only the unset default (never configured) omits it.

Whether a model reasons at all comes from the catalog's reasoning flag ("reasoning": true on the model entry), and the Spettro Subscription's /v1/models response carries the same flag. Models without it never see a thinking parameter, and in the TUI the /thinking command, its completion menu, and the status-bar tag are hidden while such a model is active. The ACP Thinking selector is the exception: it is always present (showing Off when disabled) so the editor toolbar never loses the control.

Automatic fallback

If a model rejects the requested level (an effort value it doesn't define, or a thinking budget above its cap), Spettro does not surface the error — it silently retries one level lower (max → x-high → high → medium → low → off) until the request goes through, so the run keeps its continuity. On the reasoning_effort wire format, levels that serialize to the same value are skipped (e.g. max and x-high are both xhigh, so max falls straight to high). An off that a server rejects (it doesn't know the none effort) is retried with the parameter omitted entirely. The persisted setting is untouched: only the failing request is downgraded.

Slash commands

/thinking                 # report the current level
/thinking high            # switch to high-budget thinking
/thinking off             # disable extended thinking

/thinking is an instant command: it persists immediately and works even while an agent run is in flight. The setting is stored in config.json under thinking_level.

When extended thinking is enabled, the status bar shows a thinking:<level> tag next to the active model so you can always tell at a glance whether you're paying for thinking compute.

Caveats

Anthropic disallows non-default temperature, top_p, and top_k alongside thinking. Spettro doesn't currently set those, so this isn't a concern in practice — but anyone forking the Anthropic adapter should keep this in mind.
Sub-agents inherit their parent's thinking level (so agent tool calls don't quietly downgrade to off).