[aviggiano]

Closes #660.

What this changes

The shipped tree asserted, in its most visible surfaces, that Claude Code is DeepSeek's official / supported coding-agent integration. It is not. What ultrafuzz ships is the Claude Code harness pointed at DeepSeek's Anthropic-compatible endpoint — a compatibility pairing. DeepSeek's first-party harness is DeepSeek Harness (dsh), as recorded in docs/explanation/provider-harness-research.md (see Recommended Pairing Policy, rules 1 and 3). Fixing the wording now keeps this claim from colliding with the real dsh binding in #661.

File	Change
ultrafuzz.toml	[models.deepseek] profile comment and the [agents.DeepSeekAgent] comment reworded; both instances of "official" dropped.
packages/runtime/src/templates/smithers/agents/deepseek.tsx	createDeepSeekAgent doc comment reworded. Only the false first sentence changed; the accurate half about keeping credentials, model selection, telemetry semantics and pricing provenance off Anthropic defaults is kept verbatim.
docs/config.md	New paragraph in the DeepSeek section stating the pairing is a compatibility pairing, that it is opt-in and non-default, and how it is actually selected.
docs/config.md	"OpenRouter through Codex" now names itself a compatibility pairing and separates the two senses of "default".
CHANGELOG.md	The v0.1.0 DeepSeek entry said "DeepSeek's official Claude Code endpoint"; reworded to "Anthropic-compatible", matching the correction this PR makes everywhere else.
packages/runtime/src/init.ts	The pre-reword deepseek.tsx digest is registered as a stock hash, so ultrafuzz init upgrades an untouched generated adapter instead of treating it as customized.

Second pass: two false claims removed

The first pass added a ### Providers and harnesses are separate choices section to docs/config.md, with a five-row provider/harness/ref table. Both of its load-bearing claims were false, so the section is deleted.

1. "The two are not welded together" is false for three of the four shipped refs. deepseek.tsx hardcodes DEEPSEEK_ANTHROPIC_BASE_URL = "https://api.deepseek.com/anthropic" and injects it; claude.tsx contains no base_url/baseUrl at all, so ClaudeAgent cannot be repointed; kimi.tsx is wired to Kimi Code. Only CodexAgent can be repointed, and only through a hand-authored Codex-native config.toml that ultrafuzz neither owns nor validates. Documenting a separate provider/harness boundary as shipped fact is the prose form of the parallel surface PR feat(config): add provider harness binding validation #682 was closed for — and it is Document the provider/harness pairing policy and migration paths #664 scope, which is closed.
2. The selection rule — "Every other row has to be selected per node or group, or for a single run with --agent <Ref>" — is false for the OpenRouter row the same table introduces. The root config contains only [models.default|claude|kimi|deepseek] and [agents.CodexAgent|ClaudeAgent|KimiAgent|DeepSeekAgent]. There is no openrouter profile, no distinct ref, no OPENROUTER_API_KEY; that row is reached by editing [models.default] plus [agents.CodexAgent], not by selecting anything.

Two more corrections in the same pass:

• The DeepSeek section said --agent DeepSeekAgent "selects this one". It does not. This page already states (Claude agent section) that --agent and --model override fields of the default profile rather than selecting a profile by id, and runtime.test.ts:5163 asserts that an --agent override carries neither the previous profile's model nor its reasoning. Selection is model_profiles = ["deepseek"] in .ultrafuzz/topology.yml; --agent DeepSeekAgent carries neither deepseek-v4-pro nor reasoning = "max".
• "It is opt-in, not the default selection" on the OpenRouter section collapsed two different senses of "default". Now stated separately: the pairing is not selected by default, and once OpenRouter is the provider, Codex CLI is the harness this release ships for it.

Acceptance criteria

• No shipped comment, config file, or doc page calls Claude Code DeepSeek's official/supported/first-party coding-agent integration. Verified by grep across docs/, *.md, packages/**/src, and ultrafuzz.toml. Two remaining "first-party" hits — docs/config.md "pinned to the first-party deepseek catalog entry" and docs/reference/artifacts-reports.md:481 — are about the pricing catalog entry, not the harness, and are accurate as written.
• The pairing is described as a compatibility pairing wherever it is introduced (root TOML, adapter template, docs/config.md).
• docs/config.md states which pairing is the default and how the opt-in ones are actually selected, in the section that documents each one.
• No config schema, no new key on any table, no rename, no argv or routing change.

One behaviour change, deliberate

AGENT_TEMPLATES[deepseek].stockSha256 gains the pre-reword template digest (65bf43f3…). An earlier revision of this PR listed not doing this as a feature; that was wrong. Without the entry, every project already initialized keeps the false comment in its generated .smithers/agents/deepseek.ts until someone runs init --force, which defeats the point of the change. The stock-hash set exists for exactly this: claude, codex, and kimi each already carry several historical digests, and a digest match means the file is provably untouched, so nothing a user wrote is at risk. A customized adapter still does not match and is still preserved.

Explicitly not done

• No harness = "claude-code" + provider = "deepseek" schema. That surface was rejected in PR feat(config): add provider harness binding validation #682 and is struck from Correct the DeepSeek binding terminology: it is a compatibility pairing, not DeepSeek's first-party harness #660.
• No rename of the DeepSeekAgent ref/export, and no change to the doctor's agent-to-binary map.
• No provider/harness architecture table. That is Document the provider/harness pairing policy and migration paths #664, which is closed.

Tight-adapter ruling compliance

• No new adapter file. No adapter was added; deepseek.tsx changed by one block comment, 0 executable lines.
• No new config keys. No [providers.*] or [harnesses.*] table, no new field on any existing table. ultrafuzz.toml gains zero keys — the diff there is comment text only.
• Nothing reimplemented from Smithers. No argv construction, prompt assembly, output interpretation, usage accounting, session/resume handling, error classification, credential storage, provider routing, or model-catalogue knowledge was touched or duplicated.
• No fs, network, or env surface change. workflowControlChildEnvironment and workflowControlCredentialValue call sites are untouched.

Verification

• pnpm -w format:check — pass.
• pnpm -w lint — pass.
• pnpm -w docs:check (audit-profile-docs.mjs --check + docs-check.mjs) — pass, exit 0.
• packages/runtime node:test suite (node scripts/run-tests.mjs) — 658 tests, 640 pass, 0 fail, 18 skipped, exit 0, run against the tree at 47ee058f (which is the tree at HEAD apart from a CHANGELOG rewrap).