Implement: Add native Anthropic provider via @ai-sdk/anthropic, not Op

[BabyKoan]

Summary

Implements #30

• docs(readme): note native Anthropic provider support
• fix(streaming): discard Anthropic reasoning deltas from assistant stream
• feat(llm): instantiate native Anthropic client and expose capabilities
• chore(provider): add @ai-sdk/anthropic and make preset baseUrl optional

Closes #30

---

Quality Report

Changes: 11 files changed, 162 insertions(+), 53 deletions(-)

Code scan: clean

Tests: failed (1 failed, 0 test)

Branch hygiene: clean

Generated by Kōan

[atoomic]

@BabyKoan rebase and fix failures exposed by CI

[BabyKoan]

Simple rebase

Branch koan/implement-30 was rebased onto main — no additional changes were needed.

Stats

11 files changed, 168 insertions(+), 58 deletions(-)

Actions performed

• Already-solved check: negative (confidence=high, reasoning=No commit in main history or PR diff matches adding @ai-sdk/anthropic native provider, making preset)
• Rebased koan/implement-30 onto upstream/main
• Pre-push CI check: previous run #28293690375 failed
• Pre-push CI fix: no changes needed or Claude found nothing to fix
• Force-pushed koan/implement-30 to origin
• Private review gate skipped: disabled by config
• CI check enqueued in ## CI (async)

CI status

CI will be checked asynchronously.

---

Automated by Kōan

[Koan-Bot]

@atoomic: @BabyKoan rebase and fix failures exposed by CI

Reviewed the diff. No CI-failing change is identifiable from it — the new supportsExtendedThinking field is covered by the updated assertions in tests/llm/client.test.ts (both full-shape spots were updated; the rest use per-field .toBe()). Recheck CI post-rebase. The substantive issues are design-level, not test failures: the extended-thinking regex matches none of the preset's recommended models, and createAnthropic drops the configured baseURL. See inline findings.

[Koan-Bot]

PR Review — Implement: Add native Anthropic provider via @ai-sdk/anthropic, not Op

Native Anthropic client lands cleanly and tests are solid, but extended thinking is inert for every recommended model and the configured baseURL is silently dropped.

The PR does the hard part well: createAnthropic is wired through the existing ProviderCapabilities shape, baseUrl is made optional without breaking other presets, reasoning stream parts are explicitly discarded, and the stream/capabilities paths get good test coverage (on/off thinking, native-client instantiation). All stated PR goals are delivered.

• Extended thinking only activates for claude-3-7-sonnet-*, which is not among the Anthropic preset's recommended models (opus-4-8, fable-5, sonnet-4-6, haiku-4-5) — the feature is dead code in practice.
• createAnthropic({apiKey}) ignores the user-configured baseURL, so proxy/regional Anthropic endpoints are silently bypassed.
• Anthropic detection is inconsistent: capabilities() checks name || URL, modelWithConfig() checks name only.
• chat.tsx hardcodes the Anthropic URL as the generic baseUrl fallback, which will mislabel any future native provider.

---

🟡 Important

1. Extended-thinking guard matches none of the preset's recommended models
src/llm/client.ts:90-92

The extended-thinking provider option is gated on /^claude-3-7-sonnet-/i, but the Anthropic preset's suggestedModels are claude-opus-4-8, claude-fable-5, claude-sonnet-4-6, claude-haiku-4-5 — none match the regex. So providerOptions.anthropic.thinking is never emitted for any model a user reaches through normal /provider → Anthropic setup.

Why it matters: the headline capability this PR exposes (native Anthropic extended thinking) is effectively dead code in practice. Every current Claude model supports extended thinking, so the restriction reads like a leftover from an older model line. The test does not enable extended thinking for other Anthropic models even codifies that claude-opus-4-8 gets no thinking — the opposite of what users on the recommended models would expect.

How to fix: broaden the guard to the models the preset actually recommends (e.g. /^claude-(opus|sonnet|haiku|fable)-[3-5]/i), or — preferably — drive the decision off a capability/model flag instead of a name regex. If the narrow scope is deliberate (e.g. token-cost caution), say so in a comment and at least cover one reachable model.

const isExtendedThinkingModel =
  config.capabilities.supportsExtendedThinking &&
  /^claude-3-7-sonnet-/i.test(config.modelName);

2. createAnthropic drops the user-configured baseURL
src/llm/client.ts:67-69

modelWithConfig computes baseURL from selection.provider.url but passes only {apiKey} to createAnthropic. The @ai-sdk/anthropic factory accepts a baseURL option (same signature as createOpenAI two lines below), so the configured URL is discarded.

Why it matters: for the default preset this is harmless — createAnthropic defaults to api.anthropic.com, matching the URL chat.tsx stores. But if a user edits the provider URL to an Anthropic-compatible proxy, gateway, or regional endpoint, that URL is silently ignored and traffic still hits api.anthropic.com; the stored URL becomes misleading. The test locks the no-baseURL behavior in (toHaveBeenCalledWith({apiKey: 'sk-ant-test'})), so the gap is easy to miss.

How to fix: createAnthropic({apiKey, baseURL}).chat(name), and relax the assertion to permit a baseURL argument. I could not open @ai-sdk/anthropic's .d.ts to confirm (the package is added by this PR and isn't in node_modules on main), but it follows the shared provider-factory signature.

model: useAnthropic
  ? createAnthropic({apiKey}).chat(name)
  : createOpenAI({apiKey, baseURL}).chat(name),

🟢 Suggestions

1. Anthropic detection diverges between capabilities() and modelWithConfig()
src/llm/client.ts:65

capabilities() treats a provider as Anthropic on name OR URL (/api\.anthropic\.com/i.test(baseURL), line 37), but modelWithConfig decides useAnthropic on name alone (line 65).

Why it matters: a custom-named provider pointed at api.anthropic.com (user enters name MyAnthropic, URL https://api.anthropic.com/v1) gets Anthropic capabilities flagged (supportsExtendedThinking, reportsCacheUsage) but is still driven through the OpenAI client. The capability flags then mis-describe what the client will actually honor.

How to fix: compute the anthropic flag once (name || URL) and reuse it for both client selection and capability detection — thread the same predicate through both sites.

// modelWithConfig (line 65): name only
const useAnthropic = isAnthropicProvider(selection.provider.name);
// capabilities (line 37): name OR URL
const anthropic = isAnthropicProvider(providerName) || /api\.anthropic\.com/i.test(baseURL);

2. Anthropic-specific URL hardcoded in the generic preset handler
src/cli/commands/chat.tsx:410

preset.baseUrl ?? 'https://api.anthropic.com/v1' bakes an Anthropic-specific URL into the generic preset-selection path. wizardSuggestions.ts uses the generic 'native provider' fallback for the same case, so the two paths disagree.

Why it matters: today only the Anthropic preset lacks baseUrl, so it works. But the moment a second native (non-OpenAI) provider is added without baseUrl (e.g. a native Google Gemini provider via @ai-sdk/google), this will silently store and display the Anthropic URL for it. It also means the stored url for the Anthropic provider is a cosmetic string that createAnthropic ignores (see the client.ts baseURL finding).

How to fix: give ProviderPreset an optional field the preset owns (e.g. defaultBaseUrl or a nativeProvider marker), or reuse the generic 'native provider' label already used in wizardSuggestions.ts.

const urlLabel = preset.baseUrl ?? 'https://api.anthropic.com/v1';

3. Reasoning cases are a no-op vs. default; verify persisted conversation doesn't leak reasoning
src/cli/commands/streaming.ts:337-342

These explicit reasoning-* cases are functionally identical to the existing default: break — assistant text only accumulates from text-delta (line 241), so reasoning deltas were already discarded before this change.

Why it matters: it's fine as documentation of intent, but be aware it doesn't alter behavior. If the PR's goal is "reasoning must never reach the user," also check the persisted-conversation path: response.messages (line 344) for a thinking turn includes reasoning content parts, and a grep finds nothing in the codebase that filters reasoning content before setConversation. I could not confirm from the render path whether those parts are surfaced as visible text — worth a quick manual check. (Largely moot today only because the narrow model guard in client.ts means thinking rarely activates.)

case 'reasoning-start':
case 'reasoning-delta':
case 'reasoning-end':
  // ... intentionally discarded ...
  break;

---

Checklist

• No hardcoded secrets or credentials
• Backward-compatible config changes (baseUrl made optional; existing Anthropic configs still resolve)
• New capability code paths are reachable, not dead code — warning ＃1
• Configured endpoint URLs are honored by the client — warning ＃2
• Provider-type detection is consistent across call sites — suggestion ＃1
• Tests cover new branches and edge cases — warning ＃1

---

To rebase specific severity levels, mention me: @Koan-Bot rebase critical (fixes 🔴 only), @Koan-Bot rebase important (fixes 🔴 + 🟡), or just @Koan-Bot rebase for all.

---

Automated review by Kōan (Claude · model opus) HEAD=ab4f3d9 11 min 22s

[Koan-Bot]

Blocking issues found.

• Extended-thinking guard matches none of the preset's recommended models
• createAnthropic drops the user-configured baseURL

[atoomic]

@BabyKoan rebase

[BabyKoan]

❌ Permission denied. Only users with write access can trigger bot commands.

[BabyKoan]

Rebase completed; review feedback not applied

Branch koan/implement-30 was rebased onto main, but review feedback could not be applied automatically.

Warning

Review feedback was NOT applied — the feedback step errored. The reviewer comments above still need to be addressed: re-run /rebase or apply them manually.

Stats

11 files changed, 168 insertions(+), 58 deletions(-)

Actions performed

• Already-solved check: skipped (Claude call failed)

• Rebased koan/implement-30 onto upstream/main

• Review feedback step failed: Exit code 1: OCGO model mapping enabled for claude (/home/baby/.config/ocgo/model-mapping.json)
  haiku -> deepseek-v4-flash
  opus -> glm-5.2
  sonnet -> kimi-k2.7-code
  Error: exit status 1
  Usage:

• Review feedback failed (continuing with rebase)

• Review feedback step errored (Exit code 1: OCGO model mapping enabled for claude (/home/baby/.config/ocgo/model-mapping.json)
  haiku -> deepseek-v4-flash
  opus -> glm-5.2
  sonnet -> kimi-k2.7-code
  Error: exit status 1
  Usage:
  ocgo launch claude [-- claude args...] [flags]

Flags:
-h, --help help for claude
--model string OpenCode Go model ID
--yes Allow Claude Code to skip permission prompts); pushing rebase without feedback changes

• Pre-push CI check: previous run passed
• Force-pushed koan/implement-30 to origin
• Private review gate skipped: disabled by config
• CI check enqueued in ## CI (async)

CI status

CI will be checked asynchronously.

---

Automated by Kōan

[atoomic]

@BabyKoan rebase

[BabyKoan]

Rebase with requested adjustments

Branch koan/implement-30 was rebased onto main and review feedback was applied.

Changes applied

• Implemented review feedback on src/llm/client.ts and tests/llm/client.test.ts:
• createAnthropic now receives baseURL so user-configured Anthropic endpoints are honored.
• Extended-thinking model guard expanded from claude-3-7-sonnet-* to also match claude-opus-4-*, claude-sonnet-4-*, and claude-fable-*, matching the preset's suggested models.
• Updated tests to cover the new thinking-eligible models and assert baseURL is passed to createAnthropic.
• Lint, typecheck, and tests/llm/client.test.ts pass.

Stats

11 files changed, 175 insertions(+), 58 deletions(-)

Actions performed

• Already-solved check: negative (confidence=high, reasoning=main lacks @ai-sdk/anthropic, keeps Anthropic preset as OpenAI-compatible URL, and has no reasoning-)
• Rebased koan/implement-30 onto upstream/main
• Applied review feedback
• Pre-push CI check: previous run passed
• Force-pushed koan/implement-30 to origin
• Private review gate skipped: disabled by config
• CI check enqueued in ## CI (async)

CI status

CI will be checked asynchronously.

---

Automated by Kōan