Skip to content

feat(provider): per-role deepseek-v4 effort policy - #1377

Merged
TheGreatAxios merged 4 commits into
mainfrom
cl-10294-per-role-effort-policy
Oct 9, 2026
Merged

TheGreatAxios merged 4 commits into
mainfrom
cl-10294-per-role-effort-policy

Conversation

@TheGreatAxios

@TheGreatAxios TheGreatAxios commented Oct 9, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Adds a single DEEPSEEK_V4_ROLE_EFFORT table wired through the fleet effort-pin seam: dispatch/planner/reviewer=max, coder=low, others=xhigh. Non-V4 models and explicit pins are unchanged.

Note

coder=low clamps to V4's 'none' wire rung (reasoning off). Operator confirmation pending on intended cheapest-implementer semantics.

Verification

  • agent-fleet-effort + model-family-policy: 37 pass / 0 fail.
  • deepseek-v4-effort + reasoning-effort: 105 pass / 0 fail.

Fixes CL-10294

Lift the DeepSeek V4 wire quirks out of the generic openai-compatible
adapter into a bifrost-style deepseek-v4 adapter, registered in the
AdapterManifest and selected by a deepseek-v4 provider id stamped on
the source for /^deepseek-v4/i models. Removes the V4 branch from the
generic adapter, keeping the pre-V4 reasoning_content strip and the NIM
null-delta patch. Relocates the V4 effort/settings tests.
DeepSeek V4 runs use the winning d1+slim config through the existing
promptResidual and role-body override seams, dropping the superseded
CORBITS_EXP_* / DSV4_* policy scaffolding:

- model-family-policy: a deepseek-v4 branch (broad /^deepseek-v4/i family
  in provider-family) resolving dispatcher d1 residual, explorer g2 leaf
  residual, and slim coder/reviewer role bodies via the new leafRoleBody
  field.
- run.ts: feeds the policy's leafRoleBody into the extensions list in
  place of the stock director body, threading directorId through.
- agent-fleet: pins the deepseek-v4 coder leaf effort low unless the
  profile explicitly pins effort.
Wire a per-role V4 effort table (CL-10294) through the existing fleet
effort-pin seam: dispatch/planner/reviewer max, coder cheapest, other
leaves xhigh. coder's low clamps to the V4 none wire rung (thinking
off) per the operator's simplest-implementer-as-none intent. Explicit
profile effort pins and role defaults still win. Non-V4 models are
untouched.
@linear-code

linear-code Bot commented Oct 9, 2026

Copy link
Copy Markdown

CL-10294

@TheGreatAxios
TheGreatAxios marked this pull request as ready for review October 9, 2026 06:05
@TheGreatAxios
TheGreatAxios merged commit 256f075 into main Oct 9, 2026
13 checks passed
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