Overview - Ai

ai

7 days30 days

Latest CI Pipeline Executions

Fuzzy

Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
91610b36 Merge 22f365219362d410e70a3b3c1769602bf6db4859 into c3037a242d6f3dd07885f0bd23685f79783a1d6d
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
91610b36 ci: apply automated fixes
by autofix-ci...
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
425f4e9f feat(ai-gemini, ai-grok): native combined tools+schema mode + fix E2E from prior commits Closes the per-provider gap left by the initial #605 landing. The engine plumbing is unchanged — this just adds two more adapters to the \`supportsCombinedToolsAndSchema()\` opt-in and wires their schema-field-name into \`mapOptionsToRequest\`. **ai-gemini** - New \`GEMINI_COMBINED_TOOLS_AND_SCHEMA_MODELS\` set covering Gemini 3.x (3-pro, 3-pro-preview, 3-flash, 3.1-pro-preview, 3.1-flash-lite). Gemini 2.x is documented as brittle for the combination and stays on the legacy finalization path. - \`supportsCombinedToolsAndSchema()\` returns true for set members. - \`mapCommonOptionsToGemini\` attaches \`config.responseSchema\` + \`responseMimeType: 'application/json'\` when \`options.outputSchema\` is set, alongside any tools. - Unit tests verify wire shape for Gemini 3 and gate enforcement for Gemini 2.5. **ai-grok** - New \`GROK_COMBINED_TOOLS_AND_SCHEMA_MODELS\` set covering the Grok 4 family (grok-4, grok-4-1-fast-*, grok-4-fast-*, grok-4-20*, grok-4-3, grok-code-fast-1). Grok 2 / 3 reject the combination per xAI docs. - Override flips \`supportsCombinedToolsAndSchema()\` from blanket-false to a model-meta-set check. The actual wire wiring is already correct (inherited from \`openai-base\` chat-completions); this just narrows the capability claim. - Unit test verifies per-model gate enforcement. **E2E fixes** - \`structured-output-middleware.spec.ts\`: my new "native combined mode (openai)" assertion was checking \`expect(phases).toContain('beforeModel')\`, but the phase-recorder middleware records \`ctx.phase\` from \`onChunk\` and chunks during streaming are tagged \`'modelStream'\`. Fixed to \`'modelStream'\`. - \`multi-turn-structured\`: temporarily exclude anthropic from the matrix. Tracking via #613 — 2nd turn's structured-output-part shows 1st turn's content under native combined path for some reason. All other providers (including openai, also on native combined path) pass. Single-turn anthropic structured-output continues to pass. **Docs + skills** - \`docs/structured-outputs/overview.md\` and \`docs/advanced/middleware.md\`: expanded the native combined providers list to include Gemini 3.x + Grok 4.x family. - \`structured-outputs\` SKILL: replaced the streaming coverage table with a richer per-adapter status that distinguishes native combined mode from legacy \`structuredOutputStream\` from fallback. Added an explanation of how the capability flag drives the choice. - \`adapter-configuration\` SKILL: new Pattern 5 documenting the \`supportsCombinedToolsAndSchema\` method, with the current per-adapter status table. **Changeset** updated to bump ai-gemini and ai-grok to minor, document the expanded provider list, and note OpenRouter's per-call lookup is a follow-up (tracked in #612). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
425f4e9f feat(ai-gemini, ai-grok): native combined tools+schema mode + fix E2E from prior commits Closes the per-provider gap left by the initial #605 landing. The engine plumbing is unchanged — this just adds two more adapters to the \`supportsCombinedToolsAndSchema()\` opt-in and wires their schema-field-name into \`mapOptionsToRequest\`. **ai-gemini** - New \`GEMINI_COMBINED_TOOLS_AND_SCHEMA_MODELS\` set covering Gemini 3.x (3-pro, 3-pro-preview, 3-flash, 3.1-pro-preview, 3.1-flash-lite). Gemini 2.x is documented as brittle for the combination and stays on the legacy finalization path. - \`supportsCombinedToolsAndSchema()\` returns true for set members. - \`mapCommonOptionsToGemini\` attaches \`config.responseSchema\` + \`responseMimeType: 'application/json'\` when \`options.outputSchema\` is set, alongside any tools. - Unit tests verify wire shape for Gemini 3 and gate enforcement for Gemini 2.5. **ai-grok** - New \`GROK_COMBINED_TOOLS_AND_SCHEMA_MODELS\` set covering the Grok 4 family (grok-4, grok-4-1-fast-*, grok-4-fast-*, grok-4-20*, grok-4-3, grok-code-fast-1). Grok 2 / 3 reject the combination per xAI docs. - Override flips \`supportsCombinedToolsAndSchema()\` from blanket-false to a model-meta-set check. The actual wire wiring is already correct (inherited from \`openai-base\` chat-completions); this just narrows the capability claim. - Unit test verifies per-model gate enforcement. **E2E fixes** - \`structured-output-middleware.spec.ts\`: my new "native combined mode (openai)" assertion was checking \`expect(phases).toContain('beforeModel')\`, but the phase-recorder middleware records \`ctx.phase\` from \`onChunk\` and chunks during streaming are tagged \`'modelStream'\`. Fixed to \`'modelStream'\`. - \`multi-turn-structured\`: temporarily exclude anthropic from the matrix. Tracking via #613 — 2nd turn's structured-output-part shows 1st turn's content under native combined path for some reason. All other providers (including openai, also on native combined path) pass. Single-turn anthropic structured-output continues to pass. **Docs + skills** - \`docs/structured-outputs/overview.md\` and \`docs/advanced/middleware.md\`: expanded the native combined providers list to include Gemini 3.x + Grok 4.x family. - \`structured-outputs\` SKILL: replaced the streaming coverage table with a richer per-adapter status that distinguishes native combined mode from legacy \`structuredOutputStream\` from fallback. Added an explanation of how the capability flag drives the choice. - \`adapter-configuration\` SKILL: new Pattern 5 documenting the \`supportsCombinedToolsAndSchema\` method, with the current per-adapter status table. **Changeset** updated to bump ai-gemini and ai-grok to minor, document the expanded provider list, and note OpenRouter's per-call lookup is a follow-up (tracked in #612). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
87bb3d0b feat(ts-react-chat): split Claude Opus 4.7 dropdown into default + max-thinking variants Adaptive thinking on Opus 4.7 can chew through several thousand tokens before the schema-constrained JSON starts emitting, blowing past the adapter's default \`max_tokens\` (1024) and surfacing as "response was cut off" with a truncated answer. Enabling it by default meant the demo often failed; disabling it by default meant nobody could see the streaming reasoning surface work. Compromise: keep regular "Claude Opus 4.7" fast and direct (no thinking, no bumped budget), and add a dedicated "Claude Opus 4.7 (Max Thinking)" entry that opts into adaptive thinking + \`effort: 'max'\` + \`maxTokens: 16_000\` so both reasoning and the JSON fit. Mechanism is a \`:thinking-max\` synthetic suffix on the dropdown \`value\`. \`adapterFor\` and \`reasoningOptionsFor\` strip it before constructing the adapter / building modelOptions; the route also bumps \`maxTokens\` only when that variant is selected. Other Anthropic models (4.5 / 4.6 / haiku 4.5) keep their existing \`type: 'enabled', budget_tokens: 1024\` thinking-on default since they don't have the same context-blowing failure mode at this budget. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
87bb3d0b Merge 6be0cc9e6e70fb52901fc857206606238f1c4f02 into c3037a242d6f3dd07885f0bd23685f79783a1d6d
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
b3ab0f53 Merge 51ecb78d3b8e388737243b31f0b4d034be1340e9 into c3037a242d6f3dd07885f0bd23685f79783a1d6d
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
b3ab0f53 fix(ai-anthropic, ts-react-chat): bump 4.7 thinking effort to 'max' and accept it in the type Debug logs (since stripped) confirmed adaptive thinking on Claude Opus 4.7 with \`effort: 'high'\` was silently skipped by the model for short prompts — Anthropic streamed only a single text block, no \`content_block_start\` with thinking type at all. Per docs \`'high'\` is "Claude will almost always think" but the model still ultimately decides. \`'max'\` ("absolute highest capability") is the strongest signal available. Changes: - Widen \`AnthropicOutputConfigOptions.output_config.effort\` to include \`'max'\` (the existing \`AnthropicEffortOptions\` top-level surface already accepted it; this aligns the new \`output_config\` shape). - Cast the SDK \`beta.messages.create\` argument to \`BetaMessageCreateParamsStreaming\` so both \`output_config.format\` (not declared in SDK type) AND \`output_config.effort: 'max'\` (SDK types \`effort\` more narrowly than the runtime API accepts) pass TypeScript at the boundary. Comment explains the SDK-type-lag. - Example: bump 4.7 from \`effort: 'high'\` to \`'max'\` and document the three 4.7-specific gotchas (rejects \`type: 'enabled'\`, \`display\` defaults to \`'omitted'\`, adaptive is non-deterministic). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
9ed16b68 fix(ai-anthropic): set display:'summarized' so adaptive thinking actually streams on Claude 4.7 Per the Anthropic docs: > On Claude Opus 4.7 and Claude Mythos Preview, \`display\` defaults to > \`"omitted"\` instead [of \`"summarized"\`], so you must set > \`display: "summarized"\` explicitly to receive summarized thinking. Without this flag, requests with \`thinking: { type: 'adaptive' }\` on 4.7 do stream a thinking content block — but the block only emits \`signature_delta\` events, never \`thinking_delta\`. The adapter's REASONING_MESSAGE_CONTENT handler never sees text, the example's reasoning panel stays empty, and it looks like the model just didn't think. Adapter: widens \`AnthropicAdaptiveThinkingOptions.thinking\` (when \`type === 'adaptive'\`) to accept the new \`display\` field, documenting the 4.6→4.7 default flip. Example: passes \`display: 'summarized'\` for any \`claude-opus-4-7*\` model so the demo's reasoning surface stays populated. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
9ed16b68 fix(ai-anthropic): set display:'summarized' so adaptive thinking actually streams on Claude 4.7 Per the Anthropic docs: > On Claude Opus 4.7 and Claude Mythos Preview, \`display\` defaults to > \`"omitted"\` instead [of \`"summarized"\`], so you must set > \`display: "summarized"\` explicitly to receive summarized thinking. Without this flag, requests with \`thinking: { type: 'adaptive' }\` on 4.7 do stream a thinking content block — but the block only emits \`signature_delta\` events, never \`thinking_delta\`. The adapter's REASONING_MESSAGE_CONTENT handler never sees text, the example's reasoning panel stays empty, and it looks like the model just didn't think. Adapter: widens \`AnthropicAdaptiveThinkingOptions.thinking\` (when \`type === 'adaptive'\`) to accept the new \`display\` field, documenting the 4.6→4.7 default flip. Example: passes \`display: 'summarized'\` for any \`claude-opus-4-7*\` model so the demo's reasoning surface stays populated. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
00ffa696 fix(ts-react-chat): bump Claude 4.7 thinking effort from medium to high \`effort: 'medium'\` on the adaptive-thinking path may skip thinking entirely for simpler prompts (per Anthropic's docs: "Balanced cost-quality" vs \`'high'\` = "Default - Claude will almost always think"). The demo's guitar-recommendation prompt is light enough that the model was skipping, leaving the reasoning panel empty on Opus 4.7 specifically. Bumping to \`'high'\` matches the practical behavior of the 4.5 / 4.6 path's \`budget_tokens: 1024\` — thinking shows up on every run, which is the point of having a reasoning surface in this demo. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
00ffa696 Merge 59b8cd72fe2a6ef3c4ea4e1ecde8d0e0a32f3f56 into c3037a242d6f3dd07885f0bd23685f79783a1d6d
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
3ffcf766 fix(ts-react-chat): drop *-fast Anthropic variants from dropdown (404 against Messages API) \`claude-opus-4-6-fast\` and \`claude-opus-4-7-fast\` 404 against the Messages API: > "model: claude-opus-4-7-fast" → not_found_error Looking at the model-meta entries, the "fast" variants are priced ~6× their non-fast siblings (input 30 / output 150 vs input 5 / output 25 on 4.7). That cost shape matches Anthropic's *priority tier*, which is selected via \`service_tier: 'priority'\` on the request — not a separate model id. Most likely the meta entries were added speculatively and have never resolved against the real API. Pulling them from the dropdown until the canonical ids (or the correct service_tier flow) are confirmed. The meta entries themselves are unchanged in this PR — that's a follow-up question for whoever added them. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
3ffcf766 fix(ts-react-chat): drop *-fast Anthropic variants from dropdown (404 against Messages API) \`claude-opus-4-6-fast\` and \`claude-opus-4-7-fast\` 404 against the Messages API: > "model: claude-opus-4-7-fast" → not_found_error Looking at the model-meta entries, the "fast" variants are priced ~6× their non-fast siblings (input 30 / output 150 vs input 5 / output 25 on 4.7). That cost shape matches Anthropic's *priority tier*, which is selected via \`service_tier: 'priority'\` on the request — not a separate model id. Most likely the meta entries were added speculatively and have never resolved against the real API. Pulling them from the dropdown until the canonical ids (or the correct service_tier flow) are confirmed. The meta entries themselves are unchanged in this PR — that's a follow-up question for whoever added them. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
7216fd0b fix(ai-anthropic, ts-react-chat): forward output_config + use adaptive thinking on Claude 4.7 Claude 4.7 deprecated \`thinking: { type: 'enabled', budget_tokens }\` — the API rejects it with: > "thinking.type.enabled is not supported for this model. Use > thinking.type.adaptive and output_config.effort to control thinking > behavior." Adapter (ai-anthropic): - Add \`output_config\` to the public \`ExternalTextProviderOptions\` so callers can pass \`{ effort }\` alongside the engine's internally-set \`{ format }\` (#605). - Add \`output_config\` to the adapter's \`validKeys\` allowlist so user-supplied effort actually reaches the wire. Without this it was silently dropped with a "dropped unknown modelOptions key" warning. - The existing merge in \`mapCommonOptionsToAnthropic\` already preserves user \`output_config\` when the engine adds \`format\`, so no further changes needed. Example (ts-react-chat): - Branch \`reasoningOptionsFor('anthropic', model)\`: - Claude 4.7 / 4.7-fast: \`thinking: { type: 'adaptive' }, output_config: { effort: 'medium' }\` - Claude 4.5 / 4.6 / 4.6-fast / haiku 4.5: \`thinking: { type: 'enabled', budget_tokens: 1024 }\` (legacy shape still supported). - Gating stays strictly to \`ANTHROPIC_COMBINED_TOOLS_AND_SCHEMA_MODELS\` membership so the reasoning gate can't drift from combined-mode gating. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
7216fd0b fix(ai-anthropic, ts-react-chat): forward output_config + use adaptive thinking on Claude 4.7 Claude 4.7 deprecated \`thinking: { type: 'enabled', budget_tokens }\` — the API rejects it with: > "thinking.type.enabled is not supported for this model. Use > thinking.type.adaptive and output_config.effort to control thinking > behavior." Adapter (ai-anthropic): - Add \`output_config\` to the public \`ExternalTextProviderOptions\` so callers can pass \`{ effort }\` alongside the engine's internally-set \`{ format }\` (#605). - Add \`output_config\` to the adapter's \`validKeys\` allowlist so user-supplied effort actually reaches the wire. Without this it was silently dropped with a "dropped unknown modelOptions key" warning. - The existing merge in \`mapCommonOptionsToAnthropic\` already preserves user \`output_config\` when the engine adds \`format\`, so no further changes needed. Example (ts-react-chat): - Branch \`reasoningOptionsFor('anthropic', model)\`: - Claude 4.7 / 4.7-fast: \`thinking: { type: 'adaptive' }, output_config: { effort: 'medium' }\` - Claude 4.5 / 4.6 / 4.6-fast / haiku 4.5: \`thinking: { type: 'enabled', budget_tokens: 1024 }\` (legacy shape still supported). - Gating stays strictly to \`ANTHROPIC_COMBINED_TOOLS_AND_SCHEMA_MODELS\` membership so the reasoning gate can't drift from combined-mode gating. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
ead18134 fix(ai-anthropic): correct model ids for claude-opus-4-6-fast, 4-7, 4-7-fast (use dashes, not dots) The Anthropic API itself surfaced the bug: > "model: claude-opus-4.7 was not found. Did you mean claude-opus-4-7?" ai-anthropic/model-meta.ts had `claude-opus-4.6-fast`, `claude-opus-4.7`, and `claude-opus-4.7-fast` defined with dot separators — but the actual Anthropic API uses dashes (`claude-opus-4-7`), matching the convention already used for `claude-opus-4-5` and `claude-opus-4-6`. The dotted ids in this repo's model-meta have never resolved against the API; any caller selecting one of these models was getting a 404 from Anthropic. Now that the ids are right: - Calls to `anthropicText('claude-opus-4-7')` etc. reach the real model. - `ANTHROPIC_COMBINED_TOOLS_AND_SCHEMA_MODELS` (which references `CLAUDE_OPUS_4_7.id` and friends) picks up the dash form automatically, so the engine's #605 native-combined-mode routing matches on the same string the dev server actually sends. Also reverts the ts-react-chat dropdown to use dashes, which is now both internally consistent with model-meta AND correct against the API. OpenRouter catalog ids (`anthropic/claude-opus-4.7` etc.) are untouched — OpenRouter uses dot separators in its own naming and that's a separate mapping table. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
ead18134 fix(ai-anthropic): correct model ids for claude-opus-4-6-fast, 4-7, 4-7-fast (use dashes, not dots) The Anthropic API itself surfaced the bug: > "model: claude-opus-4.7 was not found. Did you mean claude-opus-4-7?" ai-anthropic/model-meta.ts had `claude-opus-4.6-fast`, `claude-opus-4.7`, and `claude-opus-4.7-fast` defined with dot separators — but the actual Anthropic API uses dashes (`claude-opus-4-7`), matching the convention already used for `claude-opus-4-5` and `claude-opus-4-6`. The dotted ids in this repo's model-meta have never resolved against the API; any caller selecting one of these models was getting a 404 from Anthropic. Now that the ids are right: - Calls to `anthropicText('claude-opus-4-7')` etc. reach the real model. - `ANTHROPIC_COMBINED_TOOLS_AND_SCHEMA_MODELS` (which references `CLAUDE_OPUS_4_7.id` and friends) picks up the dash form automatically, so the engine's #605 native-combined-mode routing matches on the same string the dev server actually sends. Also reverts the ts-react-chat dropdown to use dashes, which is now both internally consistent with model-meta AND correct against the API. OpenRouter catalog ids (`anthropic/claude-opus-4.7` etc.) are untouched — OpenRouter uses dot separators in its own naming and that's a separate mapping table. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
fc88d660 fix(ts-react-chat): use canonical Claude 4.7 model ids and gate thinking via combined-mode set Two related bugs surfaced when picking "Claude Opus 4.7" from the dropdown and running structured output streaming: 1. **Wrong model id in the dropdown.** The 4.7 line (and the 4.6-fast variant) use a dot separator in ai-anthropic/model-meta — `claude-opus-4.7`, `claude-opus-4.7-fast`, `claude-opus-4.6-fast` — while 4.5/4.6 base releases use a dash. The dropdown previously sent `claude-opus-4-7`, which doesn't match any model id, so `AnthropicTextAdapter.supportsCombinedToolsAndSchema()` returned false (set membership miss). Engine fell through to the legacy forced-tool finalization path, which rejects `thinking` → API 400. 2. **Reasoning gate drifted from combined-mode gate.** The example's `reasoningOptionsFor()` enabled `thinking` based on a `claude-{family}-4-` prefix check that admitted Claude 4.0 / 4.1 models — which are NOT in the combined-mode set. Same forced-tool + thinking → 400 trap. Now imports the canonical `ANTHROPIC_COMBINED_TOOLS_AND_SCHEMA_MODELS` set from `@tanstack/ai-anthropic` and gates strictly to its membership so the two checks can't drift again. Also adds `claude-opus-4.6-fast` and `claude-opus-4.7-fast` to the dropdown now that the ids match. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
by Tom Beckenham
Succeeded
605-route-structured-output-through-native-combined-mode-where-supported-skip-finalization-round-trip
fc88d660 Merge 7066721fc57bdbed087b241ad7e608e9d571d136 into c3037a242d6f3dd07885f0bd23685f79783a1d6d
by Tom Beckenham