You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Correct the DeepSeek binding terminology: it is a compatibility pairing, not DeepSeek's first-party harness #660
Part of #653. Rescoped: the schema half of the original issue is struck (see below); what remains is a terminology-accuracy fix.
Problem
The shipped tree asserts, in its most visible surfaces, that Claude Code is DeepSeek's official / supported coding-agent integration:
the root ultrafuzz.toml comments on the opt-in DeepSeek profile, which docs/config.md calls the default root config;
the class comment in packages/runtime/src/templates/smithers/agents/deepseek.tsx.
The already-merged research page under docs/explanation/ establishes the opposite: DeepSeek's first-party harness is DeepSeek Harness (dsh), listed first under Agent Integrations in DeepSeek's own API documentation, and what ultrafuzz ships is Claude Code pointed at DeepSeek's Anthropic-compatible endpoint — a compatibility pairing. The research page names this exact failure mode: the risk is that the harness and the provider get treated as a single selection. Calling one vendor's harness another vendor's official integration is that treatment.
This has to be corrected before any real dsh binding exists (#661), so the two never share a name or a claim.
Scope
Reword the DeepSeek profile comments in the root ultrafuzz.toml and the class comment in the DeepSeek adapter template to say: this runs the Claude Code harness against DeepSeek's Anthropic-compatible endpoint — a compatibility pairing, not DeepSeek's first-party coding agent.
Add one short paragraph to the DeepSeek section of docs/config.md saying the same, and noting the pairing is opt-in and non-default.
Fold in the residual editorial half of Document the provider/harness pairing policy and migration paths #664: docs/config.md should not present a single provider+harness combination as the only possible architecture, and should state default-selection status explicitly rather than by omission.
Renaming the DeepSeekAgent ref/export. It would require alias plumbing across the config package, the doctor's agent-to-binary map, Modal defaults, the workspace config, the benchmark manifest and the lane definitions — churn with no behavioural or documentation gain.
Acceptance criteria
No shipped comment, config file, or doc page calls Claude Code DeepSeek's official/supported/first-party coding-agent integration.
The DeepSeek pairing is described as a compatibility pairing wherever it is introduced.
docs/config.md states which pairing is the default and that others are opt-in.
changed the title [-]Rename the legacy DeepSeekAgent binding to Claude Code + DeepSeek[/-][+]Correct the DeepSeek binding terminology: it is a compatibility pairing, not DeepSeek's first-party harness[/+]on Aug 20, 2026
Closing as done on main via PR #695 (reached main through #744): ultrafuzz.toml (L54–56) and docs/config.md (L247) introduce the DeepSeek pairing as a compatibility pairing rather than DeepSeek's first-party harness, state that it's opt-in/non-default with [models.default] staying on CodexAgent, and the OpenRouter/Codex pairing got the same framing. Docs/comments only, zero behavior change. The deferred dsh adapter idea lives in #661.
Activity
docs: separate provider, harness, and model for #653
docs: separate provider, harness, and model for #653
docs: fix the reasoning-surface count and make cloudPortable expressi…
[-]Rename the legacy DeepSeekAgent binding to Claude Code + DeepSeek[/-][+]Correct the DeepSeek binding terminology: it is a compatibility pairing, not DeepSeek's first-party harness[/+]docs: state what actually ships instead of a provider/harness split
docs: rewrap the DeepSeek changelog entry
Release/v0.1.0 (#744)
aviggiano commentedon Aug 24, 2026
Closing as done on main via PR #695 (reached main through #744): ultrafuzz.toml (L54–56) and docs/config.md (L247) introduce the DeepSeek pairing as a compatibility pairing rather than DeepSeek's first-party harness, state that it's opt-in/non-default with
[models.default]staying on CodexAgent, and the OpenRouter/Codex pairing got the same framing. Docs/comments only, zero behavior change. The deferred dsh adapter idea lives in #661.