Skip to content

feat(agent): bake in winning DeepSeek V4 Flash prompt/config - #1368

Closed
TheGreatAxios wants to merge 6 commits into
mainfrom
cl-10242-bake-in-winning-deepseek-v4-flash-promptconfig
Closed

TheGreatAxios wants to merge 6 commits into
mainfrom
cl-10242-bake-in-winning-deepseek-v4-flash-promptconfig

Conversation

@TheGreatAxios

Copy link
Copy Markdown
Collaborator

Summary

  • Bakes in the winning DeepSeek V4 Flash prompt/config as the default agent fleet profile: DeepSeek 1 + slim variant, coder agent set to low cost.
  • Adapts DeepSeek V4 Flash wire params and family detection in the OpenAI-compatible adapter.

Verification

  • 221 tests pass.
  • bun run typecheck, bun run build, and bun run test pass on this branch.

Fixes CL-10242

DeepSeek V4 Flash is the exception to the historical DeepSeek 400 rule:
its encoder keeps every turn's reasoning when tools are present, so
stripping it rendered prior tool turns empty and the model stopped
thinking. Add the deepseek-v4-flash family (provider-family.ts) and
apply the model-card wire shape for that family in the openai-compatible
adapter: chat_template_kwargs.thinking, effort folded onto low/high/max,
agentic temperature 1.0 / top_p 0.95, and stream_options.include_usage.
The family regex keeps other DeepSeek models on the old strip behavior.
Added 6 unit tests (15 total pass). CL-10241.
…, coder=low)

Make the winning round-4 config (c) a built-in default for the
deepseek-v4-flash family: d1_primary.txt as the orchestrator's appended
prompt residual, per-role leaf bodies (body_coder/body_reviewer) via the
new ModelFamilyPolicy.leafRoleBody field, and the explorer leaf residual
(g2_leaf). The coder sub-agent pins its effort to low as the winning
default, with an explicitly configured profile effort still winning.

CORBITS_EXP_DSV4_* eval hooks remain wired (an env override wins over a
built-in) but are commented as removed in the CL-10242 follow-up.
Added 5 family-residual tests. CL-10242.
@linear-code

linear-code Bot commented Oct 9, 2026

Copy link
Copy Markdown

CL-10242

Regression proof for the CL-10242 bake-in: a deepseek-v4-flash coder
leaf resolves to reasoningEffort=low while reviewer and explorer keep
their role defaults (high / medium).
The winning deepseek-v4-flash (d1+slim, coder=low) config is now fully
built-in; the eval-only CORBITS_EXP_DSV4_* file-path overrides and
their __expResidual reader are deleted so no env-override path remains.
@TheGreatAxios
TheGreatAxios marked this pull request as ready for review October 9, 2026 03:04
@TheGreatAxios

Copy link
Copy Markdown
Collaborator Author

Superseded by #1375 (slim deepseek-v4 bake-in via existing seams), which is the approved rework replacing this over-architected branch.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant