@yanlinglabs/winter-provider-catalog 0.0.24 → 0.0.27

package/overlay/models.json

@@ -208,7 +208,14 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "anthropic/claude-opus-5",
@@ -216,7 +223,7 @@
       "upstreamId": "claude-opus-5",
       "displayName": "Claude Opus 5",
       "aliases": [],
-      "$comment": "ADDED by Lane X, beyond T2's seed. `opus` is one of the four PINNED aliases `packages/runtime/src/provider/selection.ts` routes to the `anthropic` provider (`PINNED_ANTHROPIC_ALIASES`), and with no row carrying it the alias resolved to a typed `unknown-model` refusal — a hole the seed's `sonnet`/`haiku` rows hid. The capability facts are upstream's at the pin; the prices are the vendor's own page. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-5/overview — actual Claude API wire model ID claude-opus-5. https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — thinking defaults on; disabled allowed only at low/medium/high, rejected at xhigh/max. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor July 24, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "$comment": "ADDED by Lane X, beyond T2's seed. `opus` is one of the four PINNED aliases `packages/runtime/src/provider/selection.ts` routes to the `anthropic` provider (`PINNED_ANTHROPIC_ALIASES`), and with no row carrying it the alias resolved to a typed `unknown-model` refusal — a hole the seed's `sonnet`/`haiku` rows hid. The capability facts are upstream's at the pin; the prices are the vendor's own page. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-5/overview — actual Claude API wire model ID claude-opus-5. https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — thinking defaults on; disabled allowed only at low/medium/high, rejected at xhigh/max. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor July 24, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): `thinking.type.disabled+output_config.effort.xhigh` / `.max` -- a CONJUNCTION token (`+` joins two ordinary tokens): disabled thinking is rejected only together with those efforts (https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting; the model page's 'disabled allowed only at low/medium/high'). The Anthropic adapter's `buildThinking` rewrites such a request to adaptive thinking rather than sending a documented 400. ",
       "endpoints": [
         "chat"
       ],
@@ -356,6 +363,15 @@
           "confidence": "declared",
           "observedAt": "2026-09-25T12:30:00Z",
           "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort — the effort page's supportedModels list names claude-fable-5-1, claude-fable-5, claude-opus-5-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5-20251101, claude-sonnet-5 and claude-sonnet-4-6: effort is requested through `output_config.effort` (Claude 4.7 and later reject a manual `thinking.budget_tokens`, https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting#rejected-configurations)."
+        },
+        "perMessageEffort": {
+          "value": {
+            "beta": "mid-conversation-output-config-2026-07-01"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T18:00:00Z",
+          "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta — \"On Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5, use a per-message effort change, which keeps the prompt cache\": a role:\"system\" message with empty content and output_config.effort, beta header mid-conversation-output-config-2026-07-01; \"Models without per-message effort, including Claude Fable 5, return a 400 error\". claude 2.1.282 sends the older alias per-turn-control-2026-07-01, which the page does not document."
         }
       },
       "pricing": {
@@ -374,9 +390,50 @@
         "temperature",
         "top_p",
         "top_k",
-        "thinking.type.enabled"
+        "thinking.type.enabled",
+        "thinking.type.disabled+output_config.effort.xhigh",
+        "thinking.type.disabled+output_config.effort.max"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "anthropic/claude-sonnet-5",
@@ -546,7 +603,14 @@
         "top_k",
         "thinking.type.enabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "anthropic/claude-fable-5",
@@ -672,7 +736,46 @@
         "thinking.type.enabled",
         "thinking.type.disabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "anthropic/claude-haiku-4.5",
@@ -769,7 +872,14 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "anthropic/claude-opus-4.5",
@@ -892,15 +1002,24 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "anthropic/claude-opus-4.6",
       "providerId": "anthropic",
       "upstreamId": "claude-opus-4.6",
       "displayName": "Claude Opus 4.6",
-      "aliases": [],
-      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-opus-4.6 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-6/overview — actual Claude API wire model ID claude-opus-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 5, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-opus-4-6"
+      ],
+      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-opus-4.6 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-6/overview — actual Claude API wire model ID claude-opus-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 5, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-opus-4-6` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -1011,15 +1130,31 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "anthropic/claude-opus-4.7",
       "providerId": "anthropic",
       "upstreamId": "claude-opus-4.7",
       "displayName": "Claude Opus 4.7",
-      "aliases": [],
-      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-opus-4.7 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-7/overview — actual Claude API wire model ID claude-opus-4-7 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor April 16, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-opus-4-7"
+      ],
+      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-opus-4.7 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-7/overview — actual Claude API wire model ID claude-opus-4-7 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor April 16, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-opus-4-7` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -1136,15 +1271,31 @@
         "top_k",
         "thinking.type.enabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "anthropic/claude-opus-4.8",
       "providerId": "anthropic",
       "upstreamId": "claude-opus-4.8",
       "displayName": "Claude Opus 4.8",
-      "aliases": [],
-      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-opus-4.8 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-8/overview — actual Claude API wire model ID claude-opus-4-8 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor May 28, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-opus-4-8"
+      ],
+      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-opus-4.8 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-8/overview — actual Claude API wire model ID claude-opus-4-8 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor May 28, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-opus-4-8` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -1261,7 +1412,46 @@
         "top_k",
         "thinking.type.enabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "anthropic/claude-sonnet-4.5",
@@ -1375,15 +1565,24 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "anthropic/claude-sonnet-4.6",
       "providerId": "anthropic",
       "upstreamId": "claude-sonnet-4.6",
       "displayName": "Claude Sonnet 4.6",
-      "aliases": [],
-      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-sonnet-4.6 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/sonnet-4-6/overview — actual Claude API wire model ID claude-sonnet-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 17, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-sonnet-4-6"
+      ],
+      "$comment": "Fix wave (scoped re-review, item 3, 2026-09-17): this row is added to the overlay SOLELY to carry `promptCaching` -- I3 tightened provider-runtime's `promptCachingLayout` (packages/provider-runtime/src/adapters/anthropic/messages.ts) to require an explicit per-row `promptCaching.value === true`, and this key had none: it was pure upstream extraction, which cannot see a capability upstream's own registry never states. Every OTHER field below is that SAME upstream extraction, copied verbatim (identical source/sourceRef/confidence/observedAt to what open-sse/config/providers/registry/anthropic/index.ts#claude-sonnet-4.6 already produced) -- nothing here re-verifies context window, modalities or tool calling, and because mergeLayers merges at the ROW level (its own header: \"the overlay wins intentionally; a re-sync MUST NEVER overwrite it\"), a future upstream re-sync will no longer refresh ANY field of this key, not just `promptCaching`. `promptCaching` itself is `official-doc`/`declared`, reusing `anthropic/claude-fable-5-1`'s own already-admitted citation under its own precedent for this class of fact ('...Messages API mechanics... architectural facts of the wire protocol, shared by every current Claude model...'): prompt caching (`cache_control` marking on a Messages API request) is exactly such a wire-level mechanic, not a per-model measurement, so the SAME citation applies without a fresh per-row pricing-page confirmation like the sibling rows (haiku-4-5-20251001/opus-5/sonnet-5/fable-5-1) each cite for themselves. — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/sonnet-4-6/overview — actual Claude API wire model ID claude-sonnet-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 17, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-sonnet-4-6` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -1494,7 +1693,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "arcee-ai/trinity-large-thinking",
@@ -2005,7 +2211,21 @@
         }
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://github.com/openai/codex/blob/main/codex-rs/core/src/client.rs — Codex (the vendor's own client for this backend) sets prompt_cache_key on every Responses request, to the session id (or source:parent_thread_id for internal sessions); https://developers.openai.com/api/docs/guides/prompt-caching documents the field."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "upstream-static",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "codex-rs models-manager/models.json (cloned 25270df, 2026-09-26): supports_search_tool is true on every bundled model; core/src/tools/spec_plan.rs:653-666 turns on the client tool_search (tools/src/tool_search.rs:84-110, core/src/tools/handlers/tool_search_spec.rs:98) when the provider's namespace capability is on (default true) -- the vendor's own client sends it to chatgpt.com/backend-api/codex."
+      }
     },
     {
       "key": "codex-oauth/gpt-5.6-terra",
@@ -2111,7 +2331,21 @@
         }
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://github.com/openai/codex/blob/main/codex-rs/core/src/client.rs — Codex (the vendor's own client for this backend) sets prompt_cache_key on every Responses request, to the session id (or source:parent_thread_id for internal sessions); https://developers.openai.com/api/docs/guides/prompt-caching documents the field."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "upstream-static",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "codex-rs models-manager/models.json (cloned 25270df, 2026-09-26): supports_search_tool is true on every bundled model; core/src/tools/spec_plan.rs:653-666 turns on the client tool_search (tools/src/tool_search.rs:84-110, core/src/tools/handlers/tool_search_spec.rs:98) when the provider's namespace capability is on (default true) -- the vendor's own client sends it to chatgpt.com/backend-api/codex."
+      }
     },
     {
       "key": "codex-oauth/gpt-5.6-luna",
@@ -2217,7 +2451,21 @@
         }
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://github.com/openai/codex/blob/main/codex-rs/core/src/client.rs — Codex (the vendor's own client for this backend) sets prompt_cache_key on every Responses request, to the session id (or source:parent_thread_id for internal sessions); https://developers.openai.com/api/docs/guides/prompt-caching documents the field."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "upstream-static",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "codex-rs models-manager/models.json (cloned 25270df, 2026-09-26): supports_search_tool is true on every bundled model; core/src/tools/spec_plan.rs:653-666 turns on the client tool_search (tools/src/tool_search.rs:84-110, core/src/tools/handlers/tool_search_spec.rs:98) when the provider's namespace capability is on (default true) -- the vendor's own client sends it to chatgpt.com/backend-api/codex."
+      }
     },
     {
       "key": "deepseek-anthropic/deepseek-flash",
@@ -4169,7 +4417,14 @@
         "observedAt": "2026-09-05T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/o4-mini",
@@ -4315,7 +4570,14 @@
         "observedAt": "2026-09-05T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "opencode/big-pickle",
@@ -5041,9 +5303,10 @@
         "grok-4.20-beta-0309-non-reasoning",
         "grok-4.20-non-reasoning-gv2"
       ],
-      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. The two `targetFormat: \"openai-responses\"` siblings are omitted rather than reshaped; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.20-non-reasoning. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. Its templated detail page displays a generic Reasoning capability, conflicting with the explicit non-reasoning model name; reasoning block omitted pending probe. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25).",
+      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. Upstream's two `targetFormat: \"openai-responses\"` siblings are now representable (WS-23): `grok-4.6` is its own overlay row and `grok-4.20-multi-agent-0309` is a Responses-only row; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.20-non-reasoning. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. Its templated detail page displays a generic Reasoning capability, conflicting with the explicit non-reasoning model name; reasoning block omitted pending probe. — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions.",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 1000000,
@@ -5135,9 +5398,10 @@
         "grok-4.20-experimental-beta-latest",
         "grok-4.20-reasoning-gv2"
       ],
-      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. The two `targetFormat: \"openai-responses\"` siblings are omitted rather than reshaped; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.20. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. Model page does not document an effort vocabulary for this exact model, so efforts is empty. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25).",
+      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. Upstream's two `targetFormat: \"openai-responses\"` siblings are now representable (WS-23): `grok-4.6` is its own overlay row and `grok-4.20-multi-agent-0309` is a Responses-only row; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.20. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions. Continuation (WS-23): `opaque-provider-state` — on Responses the reasoning item carries `encrypted_content` when `include: [\"reasoning.encrypted_content\"]` is sent and is replayed unchanged in the next `input` (https://docs.x.ai/developers/model-capabilities/text/reasoning, \"Encrypted Reasoning Content\"). Effort (WS-23 fix of the \"supported, no efforts, continuation none\" incoherence): the model page says \"Reasoning: Yes\", but https://docs.x.ai/developers/model-capabilities/text/reasoning names `reasoning_effort` only for `grok-4.7`, `grok-4.6`, `grok-4.5` (and the multi-agent model's agent count), so this model reasons with NO documented effort knob and `efforts` stays empty — an effort or `thinking` request is refused before the request rather than guessed. What was incoherent was `continuation: \"none\"`, now `opaque-provider-state` per the note above. Since WS-23 fix round 1 the adapter asks for `include` on EVERY turn of an opaque-continuation row, effort or not, so the item is requested here although no effort can be named. LIVE-GATE ITEM: that xAI returns it for this model (WS-23 probe step 5b/5c); if it does not, `continuation` goes back to `none`.",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 1000000,
@@ -5202,7 +5466,7 @@
           "observedAt": "2026-09-25T08:40:00Z"
         },
         "efforts": [],
-        "continuation": "none"
+        "continuation": "opaque-provider-state"
       },
       "pricing": {
         "value": {
@@ -5218,6 +5482,110 @@
       "unsupportedParameters": [],
       "status": "candidate"
     },
+    {
+      "key": "xai/grok-4.20-multi-agent-0309",
+      "providerId": "xai",
+      "upstreamId": "grok-4.20-multi-agent-0309",
+      "displayName": "Grok 4.20 Multi-Agent Beta",
+      "aliases": [
+        "grok-4.20-multi-agent",
+        "grok-4.20-multi-agent-latest",
+        "grok-4.20-multi-agent-beta-latest",
+        "grok-4.20-multi-agent-experimental-beta-0304",
+        "grok-4.20-multi-agent-experimental-beta-latest",
+        "grok-4.20-multi-agent-beta-0309"
+      ],
+      "$comment": "WS-23 (2026-09-25), authored from the catalog-owning session's refresh row with its `notes` folded in here. Model page: https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309 — \"Model name: `grok-4.20-multi-agent-0309`\" with `grok-4.20-multi-agent` as its first alias, which is the spelling the multi-agent guide tells callers to use (https://docs.x.ai/developers/model-capabilities/text/multi-agent). The dated model name is the wire id and the undated spelling resolves through `aliases`. RESPONSES-ONLY: \"The multi-agent model does **not** work with the OpenAI Chat Completions API\" (guide, Limitations), so `endpoints` is `[\"responses\"]` and the row is legal only because the `xai` provider now routes through `winter.openai-responses` (catalog-integrity I2). TOOLS: \"Client-side tools (function calling) and custom tools are not currently supported by the multi-agent model variant\" — the model page's generic \"Function calling: Yes\" is overridden by that specific limitation, so `toolCalling` is `none` and `tools` is in `unsupportedParameters`, which the Responses adapter checks BEFORE sending (a typed `capability` refusal, never a vendor 400). OUTPUT LIMIT: \"The `max_tokens` parameter is not currently supported by the multi-agent model variant\"; on Responses that field is `max_output_tokens`, the name the adapter checks. EFFORT: `reasoning.effort` selects the AGENT COUNT (\"low\"/\"medium\" = 4 agents, \"high\"/\"xhigh\" = 16), not reasoning depth; no default is documented, so none is recorded. CONTINUATION: the guide's multi-turn path is `previous_response_id`, which Winter never uses (stateless, `store: false`); sub-agent state \"is encrypted and included in the response only when `use_encrypted_content` is set\" (the xAI SDK's name for `include: [\"reasoning.encrypted_content\"]`), so the row records `opaque-provider-state`. UNVERIFIED LIVE: whether the encrypted items come back and are accepted on a `store: false` replay is what the WS-23 probe steps 4a/4b check; if they do not, this becomes `none`. Beta: \"The API interface and behavior may change\". xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models).",
+      "endpoints": [
+        "responses"
+      ],
+      "contextWindow": {
+        "value": 1000000,
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309 — model page context window 1,000,000 tokens",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "inputModalities": {
+        "value": [
+          "text",
+          "image"
+        ],
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309 — model page Modalities: Text, Image input",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "outputModalities": {
+        "value": [
+          "text"
+        ],
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309 — model page Modalities: Text output",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "toolCalling": {
+        "value": "none",
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/model-capabilities/text/multi-agent — multi-agent limitations: client-side/custom function calling unsupported; built-in server tools only",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "nativeTools": {
+        "value": false,
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/model-capabilities/text/multi-agent — client-side/custom function tools unsupported",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "structuredOutput": {
+        "value": true,
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309 — model page Structured outputs capability",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "promptCaching": {
+        "value": true,
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309 — model page lists Cached tokens input rate; prompt caching available",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T08:40:00Z"
+      },
+      "reasoning": {
+        "supported": {
+          "value": true,
+          "source": "official-doc",
+          "sourceRef": "https://docs.x.ai/developers/model-capabilities/text/multi-agent — reasoning.effort low/medium selects 4 agents, high/xhigh selects 16; previous_response_id supports multi-turn",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T08:40:00Z"
+        },
+        "efforts": [
+          "low",
+          "medium",
+          "high",
+          "xhigh"
+        ],
+        "continuation": "opaque-provider-state"
+      },
+      "pricing": {
+        "value": {
+          "inputPerMTokUsd": 1.25,
+          "outputPerMTokUsd": 2.5,
+          "cacheReadPerMTokUsd": 0.2
+        },
+        "source": "official-doc",
+        "sourceRef": "https://docs.x.ai/developers/pricing — grok-4.20-multi-agent-0309 Standard global short-context rate (<200k prompt tokens): $1.25 input, $0.20 cached input, $2.50 output per 1M; >=200k the entire request is charged $2.50/$0.40/$5.00. US-regional 1.1x applies only to models currently on that endpoint, listed as grok-4.7 and grok-4.6; not this model (re-read 2026-09-25, WS-23).",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T17:09:48Z"
+      },
+      "unsupportedParameters": [
+        "max_output_tokens",
+        "tools"
+      ],
+      "status": "candidate"
+    },
     {
       "key": "xai/grok-4.3",
       "providerId": "xai",
@@ -5226,9 +5594,10 @@
       "aliases": [
         "grok-4.3-latest"
       ],
-      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. The two `targetFormat: \"openai-responses\"` siblings are omitted rather than reshaped; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.3. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. Detail page lists none/low/medium/high/xhigh, default low. Its prose only names none/low/medium/high; xhigh appears in the detailed effort list. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25). — 2026-09-25 audit fix (0.0.24): https://docs.x.ai/developers/pricing — US regional availability currently only grok-4.7 and grok-4.6",
+      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. Upstream's two `targetFormat: \"openai-responses\"` siblings are now representable (WS-23): `grok-4.6` is its own overlay row and `grok-4.20-multi-agent-0309` is a Responses-only row; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.3. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. Detail page lists none/low/medium/high/xhigh, default low. Its prose only names none/low/medium/high; xhigh appears in the detailed effort list. — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions. — 2026-09-25 audit fix (0.0.24): https://docs.x.ai/developers/pricing — US regional availability currently only grok-4.7 and grok-4.6 Continuation (WS-23): `opaque-provider-state` — on Responses the reasoning item carries `encrypted_content` when `include: [\"reasoning.encrypted_content\"]` is sent and is replayed unchanged in the next `input` (https://docs.x.ai/developers/model-capabilities/text/reasoning, \"Encrypted Reasoning Content\").",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 1000000,
@@ -5299,7 +5668,7 @@
           "high",
           "xhigh"
         ],
-        "continuation": "plaintext",
+        "continuation": "opaque-provider-state",
         "defaultEffort": "low"
       },
       "pricing": {
@@ -5326,9 +5695,10 @@
         "grok-code-fast",
         "grok-code-fast-1-0825"
       ],
-      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. The two `targetFormat: \"openai-responses\"` siblings are omitted rather than reshaped; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-build-0.1. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. xAI's May 15 retirement page says grok-code-fast-1 redirects to grok-4.3, whereas this model page lists it as a grok-build-0.1 alias; live resolution of that alias needs checking (https://docs.x.ai/developers/migration/may-15-retirement). Model page does not document an effort vocabulary for this exact model, so efforts is empty. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25). — 2026-09-25 audit fix (0.0.24): https://docs.x.ai/developers/pricing — US regional availability currently only grok-4.7 and grok-4.6",
+      "$comment": "Read from the pinned entry's own accepted literals — only the provider's `executor` is unrepresentable. Upstream's two `targetFormat: \"openai-responses\"` siblings are now representable (WS-23): `grok-4.6` is its own overlay row and `grok-4.20-multi-agent-0309` is a Responses-only row; see the provider row's comment. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-build-0.1. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. xAI's May 15 retirement page says grok-code-fast-1 redirects to grok-4.3, whereas this model page lists it as a grok-build-0.1 alias; live resolution of that alias needs checking (https://docs.x.ai/developers/migration/may-15-retirement). — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions. — 2026-09-25 audit fix (0.0.24): https://docs.x.ai/developers/pricing — US regional availability currently only grok-4.7 and grok-4.6 Continuation (WS-23): `opaque-provider-state` — on Responses the reasoning item carries `encrypted_content` when `include: [\"reasoning.encrypted_content\"]` is sent and is replayed unchanged in the next `input` (https://docs.x.ai/developers/model-capabilities/text/reasoning, \"Encrypted Reasoning Content\"). Effort (WS-23 fix of the \"supported, no efforts, continuation none\" incoherence): the model page says \"Reasoning: Yes\", but https://docs.x.ai/developers/model-capabilities/text/reasoning names `reasoning_effort` only for `grok-4.7`, `grok-4.6`, `grok-4.5` (and the multi-agent model's agent count), so this model reasons with NO documented effort knob and `efforts` stays empty — an effort or `thinking` request is refused before the request rather than guessed. What was incoherent was `continuation: \"none\"`, now `opaque-provider-state` per the note above. Since WS-23 fix round 1 the adapter asks for `include` on EVERY turn of an opaque-continuation row, effort or not, so the item is requested here although no effort can be named. LIVE-GATE ITEM: that xAI returns it for this model (WS-23 probe step 5b/5c); if it does not, `continuation` goes back to `none`.",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 256000,
@@ -5393,7 +5763,7 @@
           "observedAt": "2026-09-25T08:40:00Z"
         },
         "efforts": [],
-        "continuation": "none"
+        "continuation": "opaque-provider-state"
       },
       "pricing": {
         "value": {
@@ -6425,6 +6795,15 @@
           "sourceRef": "continuity report §5.1/§5.3 for own-state acceptance — the Responses surface replays its own completed `reasoning.encrypted_content` items to the same model.  WS-13 §8.2: \"OpenAI-compatible\" is never the capability test, so this key shares NO domain with any other row.",
           "confidence": "declared",
           "observedAt": "2026-09-06T00:00:00Z"
+        },
+        "perMessageEffort": {
+          "value": {
+            "item": "configuration_update"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-26T00:00:00Z",
+          "sourceRef": "https://developers.openai.com/api/docs/guides/reasoning — \"Configuration updates are supported by the GPT-6 model family in standard, single-agent mode. They change only reasoning effort.\" Shape {\"type\":\"configuration_update\",\"reasoning\":{\"effort\":…}}, placed \"before the next user message in the input array\"; \"the API rejects adjacent updates\"; \"The response's reasoning.effort continues to report the request-level setting\"; with store:false, replay updates \"in their original positions\". Retrieved 2026-09-26. NOT set on codex-oauth/gpt-6-* (codex-rs enables it only behind an off-by-default feature, backend acceptance unprobed — scripts/probe-openai-midconv.ts decides) nor on azure-openai (undocumented there)."
         }
       },
       "pricing": {
@@ -6440,7 +6819,35 @@
         "observedAt": "2026-09-06T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "codex-oauth/gpt-6-astra",
@@ -6548,7 +6955,21 @@
         }
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://github.com/openai/codex/blob/main/codex-rs/core/src/client.rs — Codex (the vendor's own client for this backend) sets prompt_cache_key on every Responses request, to the session id (or source:parent_thread_id for internal sessions); https://developers.openai.com/api/docs/guides/prompt-caching documents the field."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "upstream-static",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "codex-rs models-manager/models.json (cloned 25270df, 2026-09-26): supports_search_tool is true on every bundled model; core/src/tools/spec_plan.rs:653-666 turns on the client tool_search (tools/src/tool_search.rs:84-110, core/src/tools/handlers/tool_search_spec.rs:98) when the provider's namespace capability is on (default true) -- the vendor's own client sends it to chatgpt.com/backend-api/codex."
+      }
     },
     {
       "key": "anthropic/claude-fable-5-1",
@@ -6698,6 +7119,15 @@
           "confidence": "declared",
           "observedAt": "2026-09-25T13:00:00Z",
           "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — \"A 400 error says a thinking block signature is invalid\": on Claude Fable 5.1 and Opus 5.5 a replayed thinking block is bound to the conversation (system, tools and earlier messages) and is rejected once that prefix changes (enforced for accounts created on or after 2026-08-31); the documented escape is the `thinking-binding-controls-2026-08-01` beta with `thinking.block_binding.prefix_mismatch_behavior: \"drop_block\"`. See also https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5 (\"Thinking blocks are tied to the model and the conversation\")."
+        },
+        "perMessageEffort": {
+          "value": {
+            "beta": "mid-conversation-output-config-2026-07-01"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T18:00:00Z",
+          "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta — \"On Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5, use a per-message effort change, which keeps the prompt cache\": a role:\"system\" message with empty content and output_config.effort, beta header mid-conversation-output-config-2026-07-01; \"Models without per-message effort, including Claude Fable 5, return a 400 error\". claude 2.1.282 sends the older alias per-turn-control-2026-07-01, which the page does not document."
         }
       },
       "pricing": {
@@ -6718,7 +7148,46 @@
         "tool_choice.any",
         "tool_choice.tool"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "google/gemini-3.8-flash",
@@ -6951,9 +7420,10 @@
       "upstreamId": "grok-4.6",
       "displayName": "Grok 4.6",
       "aliases": [],
-      "$comment": "P6.6 Task 1b (WS-13c owed row): the `grok` family's `grok-4.6` slot already resolves canonicalModelId `grok-4.6` through the existing `xai-oauth/grok-4.6` (subscription) row; this row makes the SAME canonical model servable on the plain token-priced `xai` provider too, so families.json needs no canonicalModelId change here — only the slot's citation prose is updated to record that the owed row is now verified. FOUND on https://docs.x.ai/docs/models (retrieved 2026-09-07): the Text API pricing table lists `grok-4.6` at $2.00/$6.00 per 1M input/output tokens (<200k prompt tokens) and $4.00/$12.00 (>=200k), a 500k context window, beside \"For everything else, including code, use Grok 4.6. It is the most intelligent and fastest model we've built.\" No subscription/OAuth gating language sets `grok-4.6` apart from the other rows in the same plain pricing table (grok-4.5, grok-4.3, grok-4.20-*, grok-build-0.1 — the last of which the `xai` provider already carries as `xai/grok-build-0.1`), so this is the plain API-key surface, distinct from the `xai-oauth` sibling's Grok Build subscription entitlement. `reasoning.supported` mirrors `xai-oauth/grok-4.6`'s own already-admitted evidence (xAI Grok Build's own model catalogue at the pinned commit, `derived-shapes-p6b-xai.md` §6) since both rows describe the identical vendor model. `toolCalling`/`nativeTools`/`inputModalities` are ALSO inherited from that sibling row, each at `confidence: \"inferred\"` rather than the sibling's `\"declared\"` — fix r1 (I-2, I-3) found that the sibling's own citations for these three fields do not check out against a direct re-read of `https://docs.x.ai/docs/models` (no tool-calling statement anywhere on it) and `derived-shapes-p6b-xai.md` §6 (no modality data at all, and its own text hints at image input via \"web search / image description\"). Values are carried forward rather than dropped (general-API grounds for tool-calling; the sibling's admission for modalities), with the gap disclosed in each field's own `sourceRef`; the same two downgrades were applied to the `xai-oauth/grok-4.6` and `xai-oauth/grok-4.5` sibling rows in the same fix round, since the overlay is this lane's to correct during the phase. `contextWindow` and `pricing` (now including the published `cacheReadPerMTokUsd`, fix r1 I-1) are independently confirmed on this task's own fetch of the same allowed page, so they carry a fresh `sourceRef` rather than inheriting the sibling's. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.6. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. https://docs.x.ai/developers/model-capabilities/text/reasoning says presencePenalty, frequencyPenalty, and stop return an error on reasoning models (snake_case names recorded for the Chat Completions wire). Chat Completions does not carry encrypted reasoning ciphertext; Grok 4.7 Responses does, so continuation is 'none' for this chat provider. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25).",
+      "$comment": "P6.6 Task 1b (WS-13c owed row): the `grok` family's `grok-4.6` slot already resolves canonicalModelId `grok-4.6` through the existing `xai-oauth/grok-4.6` (subscription) row; this row makes the SAME canonical model servable on the plain token-priced `xai` provider too, so families.json needs no canonicalModelId change here — only the slot's citation prose is updated to record that the owed row is now verified. FOUND on https://docs.x.ai/docs/models (retrieved 2026-09-07): the Text API pricing table lists `grok-4.6` at $2.00/$6.00 per 1M input/output tokens (<200k prompt tokens) and $4.00/$12.00 (>=200k), a 500k context window, beside \"For everything else, including code, use Grok 4.6. It is the most intelligent and fastest model we've built.\" No subscription/OAuth gating language sets `grok-4.6` apart from the other rows in the same plain pricing table (grok-4.5, grok-4.3, grok-4.20-*, grok-build-0.1 — the last of which the `xai` provider already carries as `xai/grok-build-0.1`), so this is the plain API-key surface, distinct from the `xai-oauth` sibling's Grok Build subscription entitlement. `reasoning.supported` mirrors `xai-oauth/grok-4.6`'s own already-admitted evidence (xAI Grok Build's own model catalogue at the pinned commit, `derived-shapes-p6b-xai.md` §6) since both rows describe the identical vendor model. `toolCalling`/`nativeTools`/`inputModalities` are ALSO inherited from that sibling row, each at `confidence: \"inferred\"` rather than the sibling's `\"declared\"` — fix r1 (I-2, I-3) found that the sibling's own citations for these three fields do not check out against a direct re-read of `https://docs.x.ai/docs/models` (no tool-calling statement anywhere on it) and `derived-shapes-p6b-xai.md` §6 (no modality data at all, and its own text hints at image input via \"web search / image description\"). Values are carried forward rather than dropped (general-API grounds for tool-calling; the sibling's admission for modalities), with the gap disclosed in each field's own `sourceRef`; the same two downgrades were applied to the `xai-oauth/grok-4.6` and `xai-oauth/grok-4.5` sibling rows in the same fix round, since the overlay is this lane's to correct during the phase. `contextWindow` and `pricing` (now including the published `cacheReadPerMTokUsd`, fix r1 I-1) are independently confirmed on this task's own fetch of the same allowed page, so they carry a fresh `sourceRef` rather than inheriting the sibling's. — 2026-09-25 refresh: Model and aliases documented at https://docs.x.ai/developers/models/grok-4.6. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. https://docs.x.ai/developers/model-capabilities/text/reasoning says presencePenalty, frequencyPenalty, and stop return an error on reasoning models (snake_case names recorded for the Chat Completions wire). Continuation (WS-23): on Responses the reasoning item comes back with `encrypted_content` when `include: [\"reasoning.encrypted_content\"]` is sent, and is passed back unchanged in the next request's `input` (https://docs.x.ai/developers/model-capabilities/text/reasoning, \"Encrypted Reasoning Content\"), so `continuation` is `opaque-provider-state`; Chat Completions \"has no field for the ciphertext\" (same page). The penalty/stop names stay recorded but are moot on Responses, which Winter never sends them on. — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions.",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 500000,
@@ -7023,7 +7493,7 @@
           "high",
           "xhigh"
         ],
-        "continuation": "plaintext",
+        "continuation": "opaque-provider-state",
         "defaultEffort": "high"
       },
       "pricing": {
@@ -7155,7 +7625,35 @@
         "observedAt": "2026-09-25T10:21:48Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.6-terra",
@@ -7268,7 +7766,35 @@
         "observedAt": "2026-09-25T10:21:48Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.6-luna",
@@ -7381,7 +7907,35 @@
         "observedAt": "2026-09-25T10:21:48Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "console/claude-fable-5",
@@ -7507,7 +8061,46 @@
         "thinking.type.enabled",
         "thinking.type.disabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-fable-5-1",
@@ -7657,6 +8250,15 @@
           "confidence": "declared",
           "observedAt": "2026-09-25T13:00:00Z",
           "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — \"A 400 error says a thinking block signature is invalid\": on Claude Fable 5.1 and Opus 5.5 a replayed thinking block is bound to the conversation (system, tools and earlier messages) and is rejected once that prefix changes (enforced for accounts created on or after 2026-08-31); the documented escape is the `thinking-binding-controls-2026-08-01` beta with `thinking.block_binding.prefix_mismatch_behavior: \"drop_block\"`. See also https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5 (\"Thinking blocks are tied to the model and the conversation\")."
+        },
+        "perMessageEffort": {
+          "value": {
+            "beta": "mid-conversation-output-config-2026-07-01"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T18:00:00Z",
+          "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta — \"On Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5, use a per-message effort change, which keeps the prompt cache\": a role:\"system\" message with empty content and output_config.effort, beta header mid-conversation-output-config-2026-07-01; \"Models without per-message effort, including Claude Fable 5, return a 400 error\". claude 2.1.282 sends the older alias per-turn-control-2026-07-01, which the page does not document."
         }
       },
       "pricing": {
@@ -7677,7 +8279,46 @@
         "tool_choice.any",
         "tool_choice.tool"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-haiku-4-5-20251001",
@@ -7774,7 +8415,14 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "console/claude-haiku-4.5",
@@ -7871,7 +8519,14 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "console/claude-opus-4.5",
@@ -7994,15 +8649,24 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "console/claude-opus-4.6",
       "providerId": "console",
       "upstreamId": "claude-opus-4.6",
       "displayName": "Claude Opus 4.6",
-      "aliases": [],
-      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-6/overview — actual Claude API wire model ID claude-opus-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 5, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-opus-4-6"
+      ],
+      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-6/overview — actual Claude API wire model ID claude-opus-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 5, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-opus-4-6` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -8113,15 +8777,31 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-opus-4.7",
       "providerId": "console",
       "upstreamId": "claude-opus-4.7",
       "displayName": "Claude Opus 4.7",
-      "aliases": [],
-      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-7/overview — actual Claude API wire model ID claude-opus-4-7 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor April 16, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-opus-4-7"
+      ],
+      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-7/overview — actual Claude API wire model ID claude-opus-4-7 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor April 16, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-opus-4-7` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -8238,15 +8918,31 @@
         "top_k",
         "thinking.type.enabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-opus-4.8",
       "providerId": "console",
       "upstreamId": "claude-opus-4.8",
       "displayName": "Claude Opus 4.8",
-      "aliases": [],
-      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-8/overview — actual Claude API wire model ID claude-opus-4-8 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor May 28, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-opus-4-8"
+      ],
+      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-4-8/overview — actual Claude API wire model ID claude-opus-4-8 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; can set thinking.type adaptive or disabled. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor May 28, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-opus-4-8` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -8363,7 +9059,46 @@
         "top_k",
         "thinking.type.enabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-opus-5",
@@ -8371,7 +9106,7 @@
       "upstreamId": "claude-opus-5",
       "displayName": "Claude Opus 5",
       "aliases": [],
-      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-5/overview — actual Claude API wire model ID claude-opus-5. https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — thinking defaults on; disabled allowed only at low/medium/high, rejected at xhigh/max. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor July 24, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/opus-5/overview — actual Claude API wire model ID claude-opus-5. https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, xhigh, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — thinking defaults on; disabled allowed only at low/medium/high, rejected at xhigh/max. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor July 24, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): `thinking.type.disabled+output_config.effort.xhigh` / `.max` -- a CONJUNCTION token (`+` joins two ordinary tokens): disabled thinking is rejected only together with those efforts (https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting; the model page's 'disabled allowed only at low/medium/high'). The Anthropic adapter's `buildThinking` rewrites such a request to adaptive thinking rather than sending a documented 400. ",
       "endpoints": [
         "chat"
       ],
@@ -8511,6 +9246,15 @@
           "confidence": "declared",
           "observedAt": "2026-09-25T12:30:00Z",
           "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort — the effort page's supportedModels list names claude-fable-5-1, claude-fable-5, claude-opus-5-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5-20251101, claude-sonnet-5 and claude-sonnet-4-6: effort is requested through `output_config.effort` (Claude 4.7 and later reject a manual `thinking.budget_tokens`, https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting#rejected-configurations)."
+        },
+        "perMessageEffort": {
+          "value": {
+            "beta": "mid-conversation-output-config-2026-07-01"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T18:00:00Z",
+          "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta — \"On Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5, use a per-message effort change, which keeps the prompt cache\": a role:\"system\" message with empty content and output_config.effort, beta header mid-conversation-output-config-2026-07-01; \"Models without per-message effort, including Claude Fable 5, return a 400 error\". claude 2.1.282 sends the older alias per-turn-control-2026-07-01, which the page does not document."
         }
       },
       "pricing": {
@@ -8529,9 +9273,50 @@
         "temperature",
         "top_p",
         "top_k",
-        "thinking.type.enabled"
+        "thinking.type.enabled",
+        "thinking.type.disabled+output_config.effort.xhigh",
+        "thinking.type.disabled+output_config.effort.max"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-sonnet-4.5",
@@ -8645,15 +9430,24 @@
       "unsupportedParameters": [
         "thinking.type.adaptive"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "console/claude-sonnet-4.6",
       "providerId": "console",
       "upstreamId": "claude-sonnet-4.6",
       "displayName": "Claude Sonnet 4.6",
-      "aliases": [],
-      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/sonnet-4-6/overview — actual Claude API wire model ID claude-sonnet-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 17, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. ",
+      "aliases": [
+        "claude-sonnet-4-6"
+      ],
+      "$comment": "WS-20 (2026-09-16): Console-arm twin of anthropic/<id>; copied structurally, identity fields differ (see catalog-integrity 'console twin' test). Fix wave (scoped re-review, item 3, 2026-09-17): `promptCaching` backfilled here too -- see the sibling `anthropic/<id>` overlay row's own comment for the evidence tier and rationale (I3 tightened the adapter gate to require this declaration per row). — 2026-09-25 refresh: https://platform.claude.com/docs/en/models/sonnet-4-6/overview — actual Claude API wire model ID claude-sonnet-4-6 (catalog dotted family-first key is intentionally mapped by daemon). https://platform.claude.com/docs/en/build-with-claude/effort — output_config.effort values low, medium, high, max, default high. https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — adaptive thinking defaults off; enabled/budget_tokens remains accepted but deprecated; disabled turns off. https://platform.claude.com/docs/en/about-claude/model-deprecations — retirement floor February 17, 2027; a floor is not a scheduled deprecation. Thinking blocks carry opaque signatures and must be replayed unchanged; Fable 5.1 and Opus 5.5 additionally bind them to the preceding system/tools/messages prefix. WS-23 (2026-09-25): alias `claude-sonnet-4-6` is the vendor's own API id for this model (its model overview page); a transcript the claude binary wrote records that dashed id, and without the alias it resolved to no row, so an old session's thinking was stripped into `<recovered_reasoning>` text on resume. The key and upstream id stay dotted (a stored-tag migration is out of scope); the adapter already sends the dashed id on the wire. ",
       "endpoints": [
         "chat"
       ],
@@ -8764,7 +9558,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      }
     },
     {
       "key": "console/claude-sonnet-5",
@@ -8934,7 +9735,14 @@
         "top_k",
         "thinking.type.enabled"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "openai/gpt-5.6",
@@ -9047,7 +9855,35 @@
         "observedAt": "2026-09-25T10:21:48Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-4.1-mini",
@@ -9126,7 +9962,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/gpt-4.1-nano",
@@ -9212,7 +10055,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/gpt-4o",
@@ -9291,7 +10141,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/gpt-4o-2024-11-20",
@@ -9370,7 +10227,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/gpt-4o-mini",
@@ -9449,7 +10313,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/gpt-5.4",
@@ -9553,7 +10424,35 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.4-mini",
@@ -9664,7 +10563,35 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.4-nano",
@@ -9775,7 +10702,35 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.4-pro",
@@ -9861,7 +10816,35 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.5",
@@ -9965,7 +10948,35 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-5.5-pro",
@@ -10058,7 +11069,35 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/o3",
@@ -10159,7 +11198,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openai/o3-mini",
@@ -10252,7 +11298,14 @@
         "observedAt": "2026-09-19T00:00:00Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      }
     },
     {
       "key": "openrouter/openai/gpt-5.4",
@@ -20760,9 +21813,10 @@
       "upstreamId": "grok-4.7",
       "displayName": "Grok 4.7",
       "aliases": [],
-      "$comment": "Model and aliases documented at https://docs.x.ai/developers/models/grok-4.7. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. https://docs.x.ai/developers/model-capabilities/text/reasoning says presencePenalty, frequencyPenalty, and stop return an error on reasoning models (snake_case names recorded for the Chat Completions wire). Chat Completions does not carry encrypted reasoning ciphertext; Grok 4.7 Responses does, so continuation is 'none' for this chat provider. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25).",
+      "$comment": "Model and aliases documented at https://docs.x.ai/developers/models/grok-4.7. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. https://docs.x.ai/developers/model-capabilities/text/reasoning says presencePenalty, frequencyPenalty, and stop return an error on reasoning models (snake_case names recorded for the Chat Completions wire). Continuation (WS-23): on Responses the reasoning item comes back with `encrypted_content` when `include: [\"reasoning.encrypted_content\"]` is sent, and is passed back unchanged in the next request's `input` (https://docs.x.ai/developers/model-capabilities/text/reasoning, \"Encrypted Reasoning Content\"), so `continuation` is `opaque-provider-state`; Chat Completions \"has no field for the ciphertext\" (same page). The penalty/stop names stay recorded but are moot on Responses, which Winter never sends them on. — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions. Readable state (WS-23): `summary`, grok-4.7 only — xAI documents summarizations for this model and no other. `grok-4.7` also returns `encrypted_content` on every Responses response \"whether or not `include` lists it\" (reasoning page, \"Always returned for grok-4.7\"). Which SSE event carries the summary text is unconfirmed (xAI's own OpenAI-SDK example reads both `response.reasoning_text.delta` and `response.reasoning_summary_text.delta`); Winter's mapper surfaces both as this row's summary since WS-23 fix round 1, and the probe's event counts show which one xAI sends.",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 500000,
@@ -20832,8 +21886,29 @@
           "high",
           "xhigh"
         ],
-        "continuation": "plaintext",
-        "defaultEffort": "high"
+        "continuation": "opaque-provider-state",
+        "defaultEffort": "high",
+        "readableState": {
+          "value": "summary",
+          "source": "official-doc",
+          "sourceRef": "https://docs.x.ai/developers/model-capabilities/text/reasoning — \"For `grok-4.7`, we expose summarizations of the model's internal reasoning\" (Summarized Reasoning Content); https://docs.x.ai/developers/rest-api-reference/inference/responses — the example reasoning output item carries `summary: [{ type: \"summary_text\", … }]`",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T17:09:48Z"
+        },
+        "summaryRequest": {
+          "value": {
+            "field": "reasoning.summary",
+            "values": [
+              "detailed",
+              "auto",
+              "concise"
+            ]
+          },
+          "source": "official-doc",
+          "sourceRef": "https://docs.x.ai/developers/rest-api-reference/inference/responses — `reasoning.summary`: \"Possible values are `auto`, `concise` and `detailed`. Only included for compatibility. The model shall always return `detailed`.\" `detailed` is listed FIRST because the adapter sends the first value, and it is the one the model returns regardless",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T17:09:48Z"
+        }
       },
       "pricing": {
         "value": {
@@ -20862,9 +21937,10 @@
         "grok-4.5-latest",
         "grok-build-latest"
       ],
-      "$comment": "Model and aliases documented at https://docs.x.ai/developers/models/grok-4.5. The catalog's xai provider currently declares Chat Completions; xAI also offers Responses.  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. https://docs.x.ai/developers/model-capabilities/text/reasoning says presencePenalty, frequencyPenalty, and stop return an error on reasoning models (snake_case names recorded for the Chat Completions wire). Chat Completions does not carry encrypted reasoning ciphertext; Grok 4.7 Responses does, so continuation is 'none' for this chat provider. Conflict: https://docs.x.ai/developers/models/grok-4.5 lists xhigh on its detail page, but https://docs.x.ai/developers/model-capabilities/text/reasoning says xhigh on 4.5 is treated as high; efforts omit xhigh because it is not a distinct behavior. — The vendor also serves this model on its OpenAI Responses endpoint (same host and key); `endpoints` lists only what this row's Chat Completions adapter routes, so `responses` is omitted until adapter-level Responses support lands (tracked follow-up, 2026-09-25).",
+      "$comment": "Model and aliases documented at https://docs.x.ai/developers/models/grok-4.5. The catalog's xai provider declared Chat Completions until WS-23 and now speaks Responses (same host and key).  No model-specific max output token cap was found. xAI's models page says logprobs/top_logprobs on grok-4.20 and newer are silently ignored, not rejected (https://docs.x.ai/developers/models). Legacy Anthropic-compatible POST /v1/messages remains documented but fully deprecated (https://docs.x.ai/developers/rest-api-reference/inference/legacy); unauthenticated POST today returned 401 after validating the request, so route exists but successful inference was not verified. POST /anthropic returned 404. https://docs.x.ai/developers/model-capabilities/text/reasoning says presencePenalty, frequencyPenalty, and stop return an error on reasoning models (snake_case names recorded for the Chat Completions wire). Continuation (WS-23): on Responses the reasoning item comes back with `encrypted_content` when `include: [\"reasoning.encrypted_content\"]` is sent, and is passed back unchanged in the next request's `input` (https://docs.x.ai/developers/model-capabilities/text/reasoning, \"Encrypted Reasoning Content\"), so `continuation` is `opaque-provider-state`; Chat Completions \"has no field for the ciphertext\" (same page). The penalty/stop names stay recorded but are moot on Responses, which Winter never sends them on. Conflict: https://docs.x.ai/developers/models/grok-4.5 lists xhigh on its detail page, but https://docs.x.ai/developers/model-capabilities/text/reasoning says xhigh on 4.5 is treated as high; efforts omit xhigh because it is not a distinct behavior. — WS-23 (2026-09-25): the `xai` provider now routes through `winter.openai-responses`, so `endpoints` lists both surfaces the family routes (the `openai/*` convention) rather than only Chat Completions.",
       "endpoints": [
-        "chat"
+        "chat",
+        "responses"
       ],
       "contextWindow": {
         "value": 500000,
@@ -20933,7 +22009,7 @@
           "medium",
           "high"
         ],
-        "continuation": "plaintext",
+        "continuation": "opaque-provider-state",
         "defaultEffort": "high"
       },
       "pricing": {
@@ -47374,7 +48450,21 @@
         }
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://github.com/openai/codex/blob/main/codex-rs/core/src/client.rs — Codex (the vendor's own client for this backend) sets prompt_cache_key on every Responses request, to the session id (or source:parent_thread_id for internal sessions); https://developers.openai.com/api/docs/guides/prompt-caching documents the field."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "upstream-static",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "codex-rs models-manager/models.json (cloned 25270df, 2026-09-26): supports_search_tool is true on every bundled model; core/src/tools/spec_plan.rs:653-666 turns on the client tool_search (tools/src/tool_search.rs:84-110, core/src/tools/handlers/tool_search_spec.rs:98) when the provider's namespace capability is on (default true) -- the vendor's own client sends it to chatgpt.com/backend-api/codex."
+      }
     },
     {
       "key": "codex-oauth/gpt-6-luna",
@@ -47482,7 +48572,21 @@
         }
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://github.com/openai/codex/blob/main/codex-rs/core/src/client.rs — Codex (the vendor's own client for this backend) sets prompt_cache_key on every Responses request, to the session id (or source:parent_thread_id for internal sessions); https://developers.openai.com/api/docs/guides/prompt-caching documents the field."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "upstream-static",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "codex-rs models-manager/models.json (cloned 25270df, 2026-09-26): supports_search_tool is true on every bundled model; core/src/tools/spec_plan.rs:653-666 turns on the client tool_search (tools/src/tool_search.rs:84-110, core/src/tools/handlers/tool_search_spec.rs:98) when the provider's namespace capability is on (default true) -- the vendor's own client sends it to chatgpt.com/backend-api/codex."
+      }
     },
     {
       "key": "openai/gpt-6-sol",
@@ -47617,6 +48721,15 @@
           "sourceRef": "continuity report §5.1/§5.3 for own-state acceptance — the Responses surface replays its own completed `reasoning.encrypted_content` items to the same model.  WS-13 §8.2: \"OpenAI-compatible\" is never the capability test, so this key shares NO domain with any other row.",
           "confidence": "declared",
           "observedAt": "2026-09-06T00:00:00Z"
+        },
+        "perMessageEffort": {
+          "value": {
+            "item": "configuration_update"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-26T00:00:00Z",
+          "sourceRef": "https://developers.openai.com/api/docs/guides/reasoning — \"Configuration updates are supported by the GPT-6 model family in standard, single-agent mode. They change only reasoning effort.\" Shape {\"type\":\"configuration_update\",\"reasoning\":{\"effort\":…}}, placed \"before the next user message in the input array\"; \"the API rejects adjacent updates\"; \"The response's reasoning.effort continues to report the request-level setting\"; with store:false, replay updates \"in their original positions\". Retrieved 2026-09-26. NOT set on codex-oauth/gpt-6-* (codex-rs enables it only behind an off-by-default feature, backend acceptance unprobed — scripts/probe-openai-midconv.ts decides) nor on azure-openai (undocumented there)."
         }
       },
       "pricing": {
@@ -47632,7 +48745,35 @@
         "observedAt": "2026-09-25T10:21:48Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "openai/gpt-6-luna",
@@ -47767,6 +48908,15 @@
           "sourceRef": "continuity report §5.1/§5.3 for own-state acceptance — the Responses surface replays its own completed `reasoning.encrypted_content` items to the same model.  WS-13 §8.2: \"OpenAI-compatible\" is never the capability test, so this key shares NO domain with any other row.",
           "confidence": "declared",
           "observedAt": "2026-09-06T00:00:00Z"
+        },
+        "perMessageEffort": {
+          "value": {
+            "item": "configuration_update"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-26T00:00:00Z",
+          "sourceRef": "https://developers.openai.com/api/docs/guides/reasoning — \"Configuration updates are supported by the GPT-6 model family in standard, single-agent mode. They change only reasoning effort.\" Shape {\"type\":\"configuration_update\",\"reasoning\":{\"effort\":…}}, placed \"before the next user message in the input array\"; \"the API rejects adjacent updates\"; \"The response's reasoning.effort continues to report the request-level setting\"; with store:false, replay updates \"in their original positions\". Retrieved 2026-09-26. NOT set on codex-oauth/gpt-6-* (codex-rs enables it only behind an off-by-default feature, backend acceptance unprobed — scripts/probe-openai-midconv.ts decides) nor on azure-openai (undocumented there)."
         }
       },
       "pricing": {
@@ -47782,7 +48932,35 @@
         "observedAt": "2026-09-25T10:21:48Z"
       },
       "unsupportedParameters": [],
-      "status": "candidate"
+      "status": "candidate",
+      "promptCacheKey": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:30:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/prompt-caching — the Responses API takes prompt_cache_key: \"Use a stable prompt_cache_key to optimize cache routing for requests that share a reusable prefix\" (pre-GPT-5.6); on GPT-5.6+ routing is automatic and the key is \"optional for separate cache accounting\"."
+      },
+      "clientToolSearch": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — \"Only gpt-5.4 and later models support tool_search\" in the Responses API; client execution: {\"type\":\"tool_search\",\"execution\":\"client\",description,parameters}, tool_search_call -> tool_search_output {call_id, execution:\"client\", status, tools}, deferred functions keep defer_loading:true, optionally inside {\"type\":\"namespace\",name,description,tools}. Retrieved 2026-09-26."
+      },
+      "additionalToolsItem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/tools-tool-search — {\"type\":\"additional_tools\",\"role\":\"developer\",\"tools\":[...]}: \"Tools in an additional_tools item become available only after that item appears in the input\"; its position is kept on replay. The page names no model list, so it is set only on the rows the same page documents tool_search for (GPT-5.4 and later). NOT on codex-oauth: codex-rs sends one only at input[0] (core/src/client.rs:910-941), never mid-conversation -- the live probe decides. Retrieved 2026-09-26."
+      },
+      "allowedToolsChoice": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://developers.openai.com/api/docs/guides/function-calling — tool_choice {\"type\":\"allowed_tools\",\"mode\":\"auto\"|\"required\",\"tools\":[{\"type\":\"function\",\"name\":…}]}: \"make only a subset of tools available across model requests, but not modify the list of tools you pass in, so you can maximize savings from prompt caching\"; \"When you use tool search, tool_choice still applies to the tools that are currently callable in the turn.\" No model list on the page; set on the same GPT-5.4+ rows. NOT on codex-oauth: codex-rs always sends tool_choice \"auto\" (core/src/client.rs:1004) and its allowed_tools is a client-side filter (ext/extension-api/src/tool_policy.rs:14-36) -- the live probe decides. Retrieved 2026-09-26."
+      }
     },
     {
       "key": "anthropic/claude-opus-5-5",
@@ -47941,6 +49119,15 @@
           "confidence": "declared",
           "observedAt": "2026-09-25T13:00:00Z",
           "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — \"A 400 error says a thinking block signature is invalid\": on Claude Fable 5.1 and Opus 5.5 a replayed thinking block is bound to the conversation (system, tools and earlier messages) and is rejected once that prefix changes (enforced for accounts created on or after 2026-08-31); the documented escape is the `thinking-binding-controls-2026-08-01` beta with `thinking.block_binding.prefix_mismatch_behavior: \"drop_block\"`. See also https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5 (\"Thinking blocks are tied to the model and the conversation\")."
+        },
+        "perMessageEffort": {
+          "value": {
+            "beta": "mid-conversation-output-config-2026-07-01"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T18:00:00Z",
+          "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta — \"On Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5, use a per-message effort change, which keeps the prompt cache\": a role:\"system\" message with empty content and output_config.effort, beta header mid-conversation-output-config-2026-07-01; \"Models without per-message effort, including Claude Fable 5, return a 400 error\". claude 2.1.282 sends the older alias per-turn-control-2026-07-01, which the page does not document."
         }
       },
       "pricing": {
@@ -47961,7 +49148,46 @@
         "tool_choice.any",
         "tool_choice.tool"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "console/claude-opus-5-5",
@@ -48120,6 +49346,15 @@
           "confidence": "declared",
           "observedAt": "2026-09-25T13:00:00Z",
           "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting — \"A 400 error says a thinking block signature is invalid\": on Claude Fable 5.1 and Opus 5.5 a replayed thinking block is bound to the conversation (system, tools and earlier messages) and is rejected once that prefix changes (enforced for accounts created on or after 2026-08-31); the documented escape is the `thinking-binding-controls-2026-08-01` beta with `thinking.block_binding.prefix_mismatch_behavior: \"drop_block\"`. See also https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5 (\"Thinking blocks are tied to the model and the conversation\")."
+        },
+        "perMessageEffort": {
+          "value": {
+            "beta": "mid-conversation-output-config-2026-07-01"
+          },
+          "source": "official-doc",
+          "confidence": "declared",
+          "observedAt": "2026-09-25T18:00:00Z",
+          "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta — \"On Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5, use a per-message effort change, which keeps the prompt cache\": a role:\"system\" message with empty content and output_config.effort, beta header mid-conversation-output-config-2026-07-01; \"Models without per-message effort, including Claude Fable 5, return a 400 error\". claude 2.1.282 sends the older alias per-turn-control-2026-07-01, which the page does not document."
         }
       },
       "pricing": {
@@ -48140,7 +49375,46 @@
         "tool_choice.any",
         "tool_choice.tool"
       ],
-      "status": "candidate"
+      "status": "candidate",
+      "deferredToolLoading": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T18:30:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool — the Model compatibility table lists Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (not Sonnet 5); \"Custom tool search implementation\": a client tool returns a tool_result whose content carries {type:\"tool_reference\", tool_name} blocks, every referenced tool declared in tools with defer_loading: true, and the API expands them; \"The prefix is untouched, so prompt caching is preserved\" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-use-with-prompt-caching)."
+      },
+      "midConversationSystem": {
+        "value": true,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-25T19:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — \"This feature is available on Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Mythos 5, Claude Opus 5.5, Claude Opus 4.8, and Claude Opus 5. No beta header is required for mid-conversation system messages. This feature is not available on Claude Sonnet 5.\""
+      },
+      "midConversationToolChanges": {
+        "value": {
+          "beta": "mid-conversation-tool-changes-2026-07-01"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition / tool_removal blocks in a role:\"system\" message, by reference (\"Add/remove by reference: mid-conversation-tool-changes-2026-07-01 — Claude API, Amazon Bedrock, Google Cloud\"); available on Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5, \"Not available on Claude Sonnet 5\". Retrieved 2026-09-26; release note 2026-07-24 (https://platform.claude.com/docs/en/release-notes/overview)."
+      },
+      "inlineToolDefinitions": {
+        "value": {
+          "beta": "inline-tools-2026-09-15"
+        },
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages — tool_addition with a tool_definition (\"Add/remove by reference + define by value: inline-tools-2026-09-15 — Claude API\"); \"The inline-tools-2026-09-15 header covers all reference-based changes\"; redefinition: \"send a different definition under the same name ... The new definition replaces the earlier one from that position onward\" (no removal first, a tool declared in tools included); a definition reusing the name of a different type of tool is a 400 tool_name_conflict. Same model list as midConversationToolChanges. Claude API only, so set on the anthropic and console providers (both api.anthropic.com), never on a cloud reseller. Retrieved 2026-09-26; release note 2026-09-22."
+      },
+      "assistantPrefill": {
+        "value": false,
+        "source": "official-doc",
+        "confidence": "declared",
+        "observedAt": "2026-09-26T00:00:00Z",
+        "sourceRef": "https://platform.claude.com/docs/en/models/opus-5-5/migration-guide — \"Don't end messages with a prefilled assistant turn: it is rejected\"; \"Prefilling assistant messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5\"; Sonnet 5 and Opus 5: \"Manual extended thinking, non-default sampling parameters, and assistant prefill return a 400 error on both models\"; the Fable 5.1 guide: prefill \"returns a 400 error, unchanged from Claude Fable 5\". Live gate 2026-09-26 on claude-opus-5-5: \"This model does not support assistant message prefill. The conversation must end with a user message.\""
+      }
     },
     {
       "key": "mistral/mistral-large-2512",