>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3590963. --> Reported by: [camoa](https://www.drupal.org/user/2448054) Related to !28 >>> <h3>Summary</h3> <p>Add two native API features to the Anthropic provider:</p> <ol> <li><strong>Prompt caching admin UI</strong> &mdash; surface the <code>CacheControlEphemeral</code> wiring already shipped in <a href="https://www.drupal.org/project/ai_provider_anthropic/issues/3572402">#3572402</a>; per-block cache breakpoints; 1-hour TTL toggle; report <code>cache_creation_input_tokens</code> alongside the existing <code>cache_read_input_tokens</code>.</li> <li><strong>PDF input</strong> &mdash; wire <code>Base64PDFSource</code> / <code>URLPDFSource</code> block params for models with the <code>pdfInput</code> capability (Opus 4.x, Sonnet 4.x, Haiku 4.x per the live capability API).</li> </ol> <h3>Features</h3> <h4>1. Prompt caching UI</h4> <p>Phase 1 shipped <code>CacheControlEphemeral::with()</code> wiring on the top-level message block (toggled via <code>configuration['prompt_cache']</code> boolean) plus <code>Usage-&amp;gt;cacheReadInputTokens</code> to <code>TokenUsageDto-&amp;gt;cached</code> reporting. Phase 2 adds the admin surface:</p> <ul> <li><strong>Admin toggle:</strong> "Enable prompt caching" checkbox in <code>getModelSettings()</code>, gated by the typed <code>ModelCapabilities</code></li> <li><strong>TTL selector:</strong> dropdown for <code>ephemeral</code> (5-min, default) vs <code>ephemeral_1h</code> (1-hour TTL, Anthropic's extended cache tier). SDK supports both via <code>CacheControlEphemeral::with(ttl: ...)</code></li> <li><strong>Per-block markers:</strong> optional pattern for attaching <code>cache_control</code> to specific message blocks via a <code>cache_breakpoint</code> marker in <code>ChatMessage</code> metadata. Falls back to top-level caching when no marker is present.</li> <li><strong>Reporting parity:</strong> <code>TokenUsageDto</code> already has <code>cached</code>; add a parallel surface for <code>cache_creation_input_tokens</code> so AI Logging surfaces both "cache hit" and "cache write" tokens.</li> </ul> <h4>2. PDF input</h4> <p>Mirrors Phase 1's image-input wiring:</p> <ul> <li><strong>Capability gating:</strong> <code>$client-&amp;gt;models-&amp;gt;retrieve($id)-&amp;gt;capabilities-&amp;gt;pdfInput-&amp;gt;supported</code> drives whether the operation is offered for a given model</li> <li><strong><code>buildMessageContent()</code> extension:</strong> detect PDF attachments in <code>ChatMessage</code>; construct <code>DocumentBlockParam</code> with <code>Base64PDFSource</code> or <code>URLPDFSource</code> and append to the message content list</li> <li><strong>Capability declaration:</strong> add an <code>AiModelCapability::ChatWithPdfInput</code> filter in <code>getConfiguredModels()</code> mirroring the existing <code>ChatWithImageVision</code> branch</li> <li><strong>AI core coordination:</strong> if <code>ChatWithPdfInput</code> capability doesn't exist in AI core 1.3.x yet, propose it upstream (one-line enum case + interface contract)</li> </ul> <h3>Out of scope (later issues)</h3> <ul> <li><strong>Compaction</strong> follows in Phase 3 alongside token counting. The cross-provider <code>CompactionInterface</code> in AI core was closed won't-fix (<a href="https://www.drupal.org/project/ai/issues/3573087">#3573087</a>, April 2026), but Marcus's reasoning <em>"compaction works differently for each provider ... most likely each provider solves this as they wish"</em> is an explicit invitation to ship it provider-internal. Phase 3 wires <code>MessageCreateParams::with(contextManagement: ...)</code> through the same capability-driven UI pattern Phase 1 established, gated by <code>ContextManagementCapability</code> (clearThinking20251015, clearToolUses20250919, compact20260112).</li> <li><strong>Token counting</strong> (<code>/v1/messages/count_tokens</code>) &mdash; Phase 3</li> <li><strong>Citations</strong> &mdash; Phase 3</li> <li><strong>Data residency headers, web fetch, code execution</strong> &mdash; Phase 4</li> </ul> <h3>Implementation approach</h3> <ul> <li><strong>TDD:</strong> tests for cache UI rendering (capability-gated), cache header attachment (per-block + top-level), PDF block construction, capability filtering</li> <li><strong>Live end-to-end:</strong> cache hit/miss demo (D1/D2 scenarios deferred from <a href="https://www.drupal.org/project/ai_provider_anthropic/issues/3572402">#3572402</a>), PDF round-trip with both base64 and URL sources, <code>cache_creation</code> vs <code>cache_read</code> accounting in AI Log</li> </ul> <h3>Depends on</h3> <ul> <li><a href="https://www.drupal.org/project/ai_provider_anthropic/issues/3572402">#3572402</a> &mdash; Phase 1 (merged in <code>d1e078a1</code>)</li> </ul> <h3>Related</h3> <ul> <li><a href="https://www.drupal.org/project/ai/issues/3573087">#3573087</a> &mdash; Add Compaction OperationType (closed won't-fix; compaction lands in our Phase 3 as a provider-internal feature)</li> <li><a href="https://www.drupal.org/project/ai/issues/3538499">#3538499</a> &mdash; Meta: Use Symfony AI (Initiative migration tracker)</li> </ul>