graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
0237. DeepSeek thinks at its low level by default
Status: accepted 2026-10-02
Context
ADR 0054 turned DeepSeek's thinking off when the wire effort is low, the
flash models' default. Every other effort sent thinking enabled with the
effort as chosen, so on DeepSeek V4 Pro graff's default (medium) thought at
length. DeepSeek documents low, high and max.
On tasks that read an MCP server, the first request of a turn reasoned for thousands of characters and planned the whole job before its first tool call. Another harness that sends DeepSeek's own default reasoned for a fraction of that on the same tasks.
Sending low with thinking still enabled roughly halved graff's mean task
time on the core and MCP suites, and every task still passed. Turning thinking
off was faster again but lost a task there and missed a quarter of the
sub-agent runs.
On the sub-agent suites low was markedly faster too, most of all when
children did the work. Across two rounds it missed two runs that the default
passed: one child added up correct per-file counts wrongly, and one run
misread a route's method. The default missed as many in an earlier round.
Decision
On the DeepSeek family (native deepseek, or a deepseek* model on another
provider), graff's default effort (medium) sends reasoning_effort: low
with thinking.type=enabled. /effort low still sends thinking disabled
(ADR 0054). high and above are sent as before, and the flash models' default
stays thinking off. The picker, status line and ACP still call the default
Medium.
Consequences
A default DeepSeek turn reaches its first tool call sooner and spends less of
each request thinking. Work that needs care with exact figures, such as adding
up the counts a tool reported, gets less checking unless someone picks
/effort high.
Revisit if traces show default-effort DeepSeek runs failing where high
would have passed, or if DeepSeek documents a medium level.