@steipete/oracle 0.17.0 → 0.17.1

package/dist/bin/oracle-cli.js

@@ -339,7 +339,7 @@ program
     .addOption(new Option("--browser-hide-window", "Hide the Chrome window after launch (macOS headful only).").hideHelp())
     .addOption(new Option("--browser-keep-browser", "Keep Chrome running after completion.").hideHelp())
     .addOption(new Option("--browser-model-strategy <mode>", "ChatGPT model picker strategy: select (default) switches to the requested model, current keeps the active model, ignore skips the picker entirely.").choices(["select", "current", "ignore"]))
-    .addOption(new Option("--browser-thinking-time <level>", "Thinking time intensity for Thinking/Pro models: light, standard, extended, heavy, or ChatGPT UI aliases.")
+    .addOption(new Option("--browser-thinking-time <level>", "Thinking time intensity for Thinking/Pro models: light, standard, extended, extra-high (Extra High), heavy (Pro), or ChatGPT UI aliases.")
     .argParser(parseThinkingTimeOption)
     .hideHelp())
     .addOption(new Option("--browser-research <mode>", "Browser research mode: deep activates ChatGPT Deep Research.").choices(["off", "deep"]))

package/dist/docs-site/browser-mode.html

@@ -327,7 +327,7 @@ body:not(.home) .doc>h1:first-child{display:none}
 <li>If an assistant response still times out (common with long Pro runs), Oracle marks the session as an incomplete capture, stores reattach/runtime diagnostics, and keeps enough browser metadata for <code>oracle session &lt;id&gt;</code> to recover the final answer. Visible ChatGPT rate-limit, temporary-unavailable, and authentication/challenge warnings are included in the error and session metadata instead of being reduced to a generic timeout. Increase <code>--browser-timeout</code> only when the browser session is truly unrecoverable.</li>
 <li><code>--browser-model-strategy &lt;select|current|ignore&gt;</code>: control ChatGPT model selection. <code>select</code> (default) switches to the requested model; <code>current</code> keeps the active model and logs its label; <code>ignore</code> skips the picker entirely. (Ignored for Gemini web runs.)</li>
 <li>Temporary Chat can reduce account-sidebar clutter for one-shot browser consults, but it is a different ChatGPT workflow: Oracle skips archive attempts there and the local transcript/artifacts are the durable record. Verify live behavior before relying on Project Sources, Deep Research reports, or multi-turn persistence.</li>
-<li><code>--browser-thinking-time &lt;light|standard|extended|heavy&gt;</code>: set the ChatGPT thinking-time intensity (Thinking/Pro models only). You can also set a default in <code>~/.oracle/config.json</code> via <code>browser.thinkingTime</code>.</li>
+<li><code>--browser-thinking-time &lt;light|standard|extended|extra-high|heavy&gt;</code>: set the ChatGPT thinking-time intensity (Thinking/Pro models only). On GPT-5.6 Sol, <code>extra-high</code> selects Extra High; a <code>heavy</code> request accepts an already-selected Pro pill but otherwise selects only a matching Heavy row. Effort rows are matched in English, German (<code>Sofort</code>/<code>Mittel</code>/<code>Hoch</code>/<code>Sehr hoch</code>), and Chinese; when the requested tier has no row in the current UI language, Oracle keeps the effort already selected in the tab instead of switching the model. You can also set a default in <code>~/.oracle/config.json</code> via <code>browser.thinkingTime</code>.</li>
 <li>GPT-5.5 Pro Extended is verified from the selected item in ChatGPT&#39;s standalone Pro/Thinking effort pill or compatible Intelligence/model-picker menu. A run <strong>fails closed</strong> if Extended cannot be confirmed rather than silently submitting at a weaker effort. Detection failures write a bounded, redacted model-picker diagnostic to the normal session log.</li>
 <li><code>--browser-research deep</code>: activate ChatGPT Deep Research before submitting the prompt. Use this for broad public-web research and final cited reports, not as a replacement for GPT-5.x Pro Heavy code review or pure reasoning.</li>
 <li><code>--browser-follow-up &lt;prompt&gt;</code>: submit another prompt in the same ChatGPT conversation after the initial answer. Repeat the flag for multi-turn reviews such as “challenge your recommendation”, “compare against this constraint”, then “give the final decision”. Deep Research has its own report lifecycle, so browser follow-ups are rejected when <code>--browser-research deep</code> is enabled.</li>
@@ -364,7 +364,7 @@ body:not(.home) .doc>h1:first-child{display:none}
 <span class="line"><span style="color:#6CB6FF">  -p</span><span style="color:#96D0FF"> "Research the current browser support for WebGPU in enterprise-managed Chrome and cite sources."</span></span></code></pre>
 <p>Oracle activates ChatGPT Deep Research through the composer tools menu, recognizing both the <code>Deep research</code> label and current <code>Get a detailed report</code> menu variants. It waits for the research plan to auto-confirm, logs high-level progress, then captures the final report from the Deep Research report surface instead of trusting the assistant tool-call wrapper.</p>
 <p>If ChatGPT initially exposes only <code>Called tool</code> / <code>Used tool</code>, Oracle treats that as an incomplete capture for Deep Research rather than a final answer. Reattach the existing session with <code>oracle session &lt;id&gt; --render</code> so Oracle can recover the lazy-loaded report from the existing Chrome tab; do not rerun the research unless the browser session is unrecoverable.</p>
-<p>Deep Research is browser-only. It does not use connected apps in v1; give it public-web scope, uploaded files, and any domain/source guidance in the prompt. For deep thinking over code or architecture without web search, prefer a normal browser run with a Pro/Thinking model and <code>--browser-thinking-time heavy</code>.</p>
+<p>Deep Research is browser-only. It does not use connected apps in v1; give it public-web scope, uploaded files, and any domain/source guidance in the prompt. For deep thinking over code or architecture without web search, prefer a normal browser run with GPT-5.6 Sol and <code>--browser-thinking-time extra-high</code>, or a Pro model with <code>--browser-thinking-time extended</code>.</p>
 <p>Completed browser sessions also save durable artifacts under <code>~/.oracle/sessions/&lt;id&gt;/artifacts/</code>. Deep Research writes the extracted report to <code>deep-research-report.md</code>, and every browser run writes <code>transcript.md</code> with the prompt, final answer, conversation URL, and saved artifact references. Use <code>--write-output &lt;path&gt;</code> when you also need a copy of just the final answer at a specific path.</p>
 <p>When ChatGPT generates downloadable files in the assistant response (for example a ZIP, wheel, source distribution, CSV, or PDF), Oracle saves those files beside the transcript before any archive attempt. The downloader is intentionally narrow: it only follows ChatGPT-owned file/download URLs from the assistant response and uses <code>sandbox:/mnt/data/...</code> links as source metadata and filename hints, not as arbitrary fetch targets. External links in the response are left in the transcript but are not downloaded. In bridge mode, a patched Windows host advertises artifact-transfer capability through <code>/health</code>; the Linux client then pulls each saved file over the authenticated bridge endpoint, stores it under the Linux session <code>artifacts/</code> directory, and verifies safe filename, byte size, SHA-256, and ZIP structure where applicable. If either side is older or transfer validation fails, the text response still completes and Oracle prints a manual-copy fallback instead of leaking host paths or signed download URLs.</p>
 <h3 id="conversation-archiving"><a class="anchor" href="#conversation-archiving" aria-label="Anchor link">#</a>Conversation archiving</h3>
@@ -395,7 +395,7 @@ body:not(.home) .doc>h1:first-child{display:none}
 <p>Use browser follow-ups when a one-shot review would be too easy for the model to answer shallowly. Oracle keeps the same ChatGPT conversation open, waits for each answer, then submits the next follow-up:</p>
 <pre class="shiki github-dark-dimmed" style="background-color:var(--code-bg);color:var(--code-fg)" tabindex="0"><code class="language-bash"><span class="line"><span style="color:#F69D50">oracle</span><span style="color:#6CB6FF"> --engine</span><span style="color:#96D0FF"> browser</span><span style="color:#F47067"> \</span></span>
 <span class="line"><span style="color:#6CB6FF">  --model</span><span style="color:#96D0FF"> gpt-5.5-pro</span><span style="color:#F47067"> \</span></span>
-<span class="line"><span style="color:#6CB6FF">  --browser-thinking-time</span><span style="color:#96D0FF"> heavy</span><span style="color:#F47067"> \</span></span>
+<span class="line"><span style="color:#6CB6FF">  --browser-thinking-time</span><span style="color:#96D0FF"> extended</span><span style="color:#F47067"> \</span></span>
 <span class="line"><span style="color:#6CB6FF">  -p</span><span style="color:#96D0FF"> "Review this migration plan and identify the top risks."</span><span style="color:#F47067"> \</span></span>
 <span class="line"><span style="color:#6CB6FF">  --file</span><span style="color:#96D0FF"> docs/migration-plan.md</span><span style="color:#F47067"> \</span></span>
 <span class="line"><span style="color:#6CB6FF">  --browser-follow-up</span><span style="color:#96D0FF"> "Challenge your previous recommendation. What would fail in production?"</span><span style="color:#F47067"> \</span></span>

package/dist/docs-site/cli-reference.html

@@ -264,7 +264,7 @@ body:not(.home) .doc>h1:first-child{display:none}
 <table><thead><tr><th>Flag</th><th>Purpose</th></tr></thead><tbody><tr><td><code>--base-url &lt;url&gt;</code></td><td>LiteLLM / Azure / OpenRouter / proxy.</td></tr><tr><td><code>--provider &lt;mode&gt;</code></td><td>API route: <code>auto</code>, <code>openai</code>, or <code>azure</code>.</td></tr><tr><td><code>--no-azure</code></td><td>Ignore Azure env/config for this run.</td></tr><tr><td><code>--route</code></td><td>Print redacted API route plan, then exit.</td></tr><tr><td><code>--azure-endpoint</code></td><td>Azure OpenAI endpoint.</td></tr><tr><td><code>--azure-deployment</code></td><td>Azure deployment name.</td></tr><tr><td><code>--azure-api-version</code></td><td>Azure API version.</td></tr></tbody></table>
 <p>See <a href="openai-endpoints.html">OpenAI / Azure / OpenRouter</a> and <a href="openrouter.html">OpenRouter</a>.</p>
 <h2 id="browser-mode"><a class="anchor" href="#browser-mode" aria-label="Anchor link">#</a>Browser mode</h2>
-<table><thead><tr><th>Flag</th><th>Purpose</th></tr></thead><tbody><tr><td><code>--chatgpt-url &lt;url&gt;</code></td><td>Target a ChatGPT workspace / project folder.</td></tr><tr><td><code>--browser-model-strategy &lt;select|current|ignore&gt;</code></td><td>Control ChatGPT model picker.</td></tr><tr><td><code>--browser-manual-login</code></td><td>Use persistent profile + manual login (no Keychain).</td></tr><tr><td><code>--browser-attach-running</code></td><td>Attach to your already-running Chrome via DevTools.</td></tr><tr><td><code>--browser-tab &lt;ref&gt;</code></td><td>Reuse an existing tab (<code>current</code>, id, URL, title substring).</td></tr><tr><td><code>--browser-thinking-time &lt;light|standard|extended|heavy&gt;</code></td><td>Pro / Thinking model intensity.</td></tr><tr><td><code>--browser-research deep</code></td><td>Activate Deep Research mode.</td></tr><tr><td><code>--browser-follow-up &lt;prompt&gt;</code></td><td>Multi-turn in the same ChatGPT conversation.</td></tr><tr><td><code>--browser-port &lt;port&gt;</code></td><td>Pin Chrome DevTools port.</td></tr><tr><td><code>--browser-inline-cookies[(-file)] &lt;…&gt;</code></td><td>Supply cookies inline (no Keychain / Chrome).</td></tr><tr><td><code>--browser-timeout</code>, <code>--browser-input-timeout</code>, <code>--browser-attachment-timeout</code></td><td>Overall / input / attachment readiness timeouts (h/m/s/ms).</td></tr><tr><td><code>--browser-recheck-delay</code>, <code>--browser-recheck-timeout</code></td><td>Delayed retry after a timeout.</td></tr><tr><td><code>--browser-auto-reattach-delay/-interval/-timeout</code></td><td>Poll the existing tab when ChatGPT redirects mid-load.</td></tr><tr><td><code>--browser-reuse-wait</code></td><td>Wait for shared Chrome profile before launching.</td></tr><tr><td><code>--browser-profile-lock-timeout</code></td><td>Wait for the manual-login profile lock.</td></tr><tr><td><code>--browser-max-concurrent-tabs</code></td><td>Soft limit for shared-profile parallel runs (default 3).</td></tr><tr><td><code>--browser-keep-browser</code></td><td>Keep the browser open after the run.</td></tr><tr><td><code>--browser-headless</code>, <code>--browser-hide-window</code></td><td>Visibility controls.</td></tr><tr><td><code>--browser-attachments &lt;auto|never|always&gt;</code></td><td>Attach files inline vs upload.</td></tr><tr><td><code>--browser-bundle-files</code>, <code>--browser-bundle-format &lt;auto|text|zip&gt;</code></td><td>Bundle browser uploads as text or byte-preserving ZIP.</td></tr><tr><td><code>--browser-chrome-path</code>, <code>--browser-cookie-path</code></td><td>Override Chrome / cookie store discovery (Linux / Windows).</td></tr></tbody></table>
+<table><thead><tr><th>Flag</th><th>Purpose</th></tr></thead><tbody><tr><td><code>--chatgpt-url &lt;url&gt;</code></td><td>Target a ChatGPT workspace / project folder.</td></tr><tr><td><code>--browser-model-strategy &lt;select|current|ignore&gt;</code></td><td>Control ChatGPT model picker.</td></tr><tr><td><code>--browser-manual-login</code></td><td>Use persistent profile + manual login (no Keychain).</td></tr><tr><td><code>--browser-attach-running</code></td><td>Attach to your already-running Chrome via DevTools.</td></tr><tr><td><code>--browser-tab &lt;ref&gt;</code></td><td>Reuse an existing tab (<code>current</code>, id, URL, title substring).</td></tr><tr><td><code>--browser-thinking-time &lt;light|standard|extended|extra-high|heavy&gt;</code></td><td>Effort intensity; unmatched tiers keep the current effort.</td></tr><tr><td><code>--browser-research deep</code></td><td>Activate Deep Research mode.</td></tr><tr><td><code>--browser-follow-up &lt;prompt&gt;</code></td><td>Multi-turn in the same ChatGPT conversation.</td></tr><tr><td><code>--browser-port &lt;port&gt;</code></td><td>Pin Chrome DevTools port.</td></tr><tr><td><code>--browser-inline-cookies[(-file)] &lt;…&gt;</code></td><td>Supply cookies inline (no Keychain / Chrome).</td></tr><tr><td><code>--browser-timeout</code>, <code>--browser-input-timeout</code>, <code>--browser-attachment-timeout</code></td><td>Overall / input / attachment readiness timeouts (h/m/s/ms).</td></tr><tr><td><code>--browser-recheck-delay</code>, <code>--browser-recheck-timeout</code></td><td>Delayed retry after a timeout.</td></tr><tr><td><code>--browser-auto-reattach-delay/-interval/-timeout</code></td><td>Poll the existing tab when ChatGPT redirects mid-load.</td></tr><tr><td><code>--browser-reuse-wait</code></td><td>Wait for shared Chrome profile before launching.</td></tr><tr><td><code>--browser-profile-lock-timeout</code></td><td>Wait for the manual-login profile lock.</td></tr><tr><td><code>--browser-max-concurrent-tabs</code></td><td>Soft limit for shared-profile parallel runs (default 3).</td></tr><tr><td><code>--browser-keep-browser</code></td><td>Keep the browser open after the run.</td></tr><tr><td><code>--browser-headless</code>, <code>--browser-hide-window</code></td><td>Visibility controls.</td></tr><tr><td><code>--browser-attachments &lt;auto|never|always&gt;</code></td><td>Attach files inline vs upload.</td></tr><tr><td><code>--browser-bundle-files</code>, <code>--browser-bundle-format &lt;auto|text|zip&gt;</code></td><td>Bundle browser uploads as text or byte-preserving ZIP.</td></tr><tr><td><code>--browser-chrome-path</code>, <code>--browser-cookie-path</code></td><td>Override Chrome / cookie store discovery (Linux / Windows).</td></tr></tbody></table>
 <p>See <a href="browser-mode.html">Browser Mode</a> for usage.</p>
 <h2 id="remote-browser"><a class="anchor" href="#remote-browser" aria-label="Anchor link">#</a>Remote browser</h2>
 <table><thead><tr><th>Flag</th><th>Purpose</th></tr></thead><tbody><tr><td><code>--remote-host &lt;host:port&gt;</code></td><td>Use a remote <code>oracle serve</code> host.</td></tr><tr><td><code>--remote-token &lt;secret&gt;</code></td><td>Auth for the remote host.</td></tr><tr><td><code>--remote-chrome &lt;host:port&gt;</code></td><td>Attach to an existing remote Chrome session.</td></tr></tbody></table>

package/dist/docs-site/configuration.html

@@ -283,7 +283,7 @@ body:not(.home) .doc>h1:first-child{display:none}
 <span class="line"><span style="color:#96D0FF">    autoReattachIntervalMs</span><span style="color:#ADBAC7">: </span><span style="color:#6CB6FF">0</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// interval between auto-reattach attempts (0 = disabled)</span></span>
 <span class="line"><span style="color:#96D0FF">    autoReattachTimeoutMs</span><span style="color:#ADBAC7">: </span><span style="color:#6CB6FF">120000</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// time budget per auto-reattach attempt (default: 2m)</span></span>
 <span class="line"><span style="color:#96D0FF">    modelStrategy</span><span style="color:#ADBAC7">: </span><span style="color:#96D0FF">"select"</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// select | current | ignore (ChatGPT only; ignored for Gemini web)</span></span>
-<span class="line"><span style="color:#96D0FF">    thinkingTime</span><span style="color:#ADBAC7">: </span><span style="color:#96D0FF">"extended"</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// light | standard | extended | heavy (ChatGPT Thinking/Pro models)</span></span>
+<span class="line"><span style="color:#96D0FF">    thinkingTime</span><span style="color:#ADBAC7">: </span><span style="color:#96D0FF">"extended"</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// light | standard | extended | extra-high | heavy (ChatGPT Thinking/Pro models)</span></span>
 <span class="line"><span style="color:#96D0FF">    researchMode</span><span style="color:#ADBAC7">: </span><span style="color:#96D0FF">"off"</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// off | deep (ChatGPT Deep Research; browser only)</span></span>
 <span class="line"><span style="color:#96D0FF">    manualLogin</span><span style="color:#ADBAC7">: </span><span style="color:#6CB6FF">false</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// set true to reuse a persistent automation profile and sign in once (Windows defaults to true when unset)</span></span>
 <span class="line"><span style="color:#96D0FF">    manualLoginProfileDir</span><span style="color:#ADBAC7">: </span><span style="color:#6CB6FF">null</span><span style="color:#ADBAC7">, </span><span style="color:#768390">// override profile dir (or set ORACLE_BROWSER_PROFILE_DIR)</span></span>

package/dist/docs-site/mcp.html

@@ -266,11 +266,11 @@ body:not(.home) .doc>h1:first-child{display:none}
 <ul>
 <li>Inputs: <code>prompt</code> (required), <code>files?: string[]</code> (globs), <code>model?: string</code> (defaults to CLI), <code>engine?: &quot;api&quot; | &quot;browser&quot;</code> (optional; Oracle follows CLI defaults: <code>ORACLE_ENGINE</code> and the effective config first, then API when <code>OPENAI_API_KEY</code> is set, otherwise browser), <code>slug?: string</code>.</li>
 <li>Presets: <code>preset?: &quot;chatgpt-pro-heavy&quot;</code> applies browser mode + current Pro model alias + extended thinking, unless the request overrides those fields.</li>
-<li>Browser-only extras: <code>browserAttachments?: &quot;auto&quot;|&quot;never&quot;|&quot;always&quot;</code>, <code>browserBundleFiles?: boolean</code>, <code>browserBundleFormat?: &quot;auto&quot;|&quot;text&quot;|&quot;zip&quot;</code>, <code>browserThinkingTime?: &quot;light&quot;|&quot;standard&quot;|&quot;extended&quot;|&quot;heavy&quot;</code>, <code>browserResearchMode?: &quot;deep&quot;</code>, <code>browserFollowUps?: string[]</code>, <code>browserArchive?: &quot;auto&quot;|&quot;always&quot;|&quot;never&quot;</code>, <code>browserKeepBrowser?: boolean</code>, <code>browserModelLabel?: string</code>, <code>browserModelStrategy?: &quot;select&quot;|&quot;current&quot;|&quot;ignore&quot;</code>, <code>generateImage?: string</code>, <code>outputPath?: string</code>.</li>
+<li>Browser-only extras: <code>browserAttachments?: &quot;auto&quot;|&quot;never&quot;|&quot;always&quot;</code>, <code>browserBundleFiles?: boolean</code>, <code>browserBundleFormat?: &quot;auto&quot;|&quot;text&quot;|&quot;zip&quot;</code>, <code>browserThinkingTime?: &quot;light&quot;|&quot;standard&quot;|&quot;extended&quot;|&quot;extra-high&quot;|&quot;heavy&quot;</code>, <code>browserResearchMode?: &quot;deep&quot;</code>, <code>browserFollowUps?: string[]</code>, <code>browserArchive?: &quot;auto&quot;|&quot;always&quot;|&quot;never&quot;</code>, <code>browserKeepBrowser?: boolean</code>, <code>browserModelLabel?: string</code>, <code>browserModelStrategy?: &quot;select&quot;|&quot;current&quot;|&quot;ignore&quot;</code>, <code>generateImage?: string</code>, <code>outputPath?: string</code>.</li>
 <li>Dry runs: set <code>dryRun: true</code> to preview the resolved request without creating a session or touching the browser.</li>
 <li>Behavior: starts a session, runs it with the chosen engine, returns final output + metadata. Background/foreground follows the CLI (e.g., GPT‑5 Pro detaches by default). If API mode fails because <code>OPENAI_API_KEY</code> is missing and you have ChatGPT Pro, retry with <code>engine: &quot;browser&quot;</code> or <code>preset: &quot;chatgpt-pro-heavy&quot;</code> to use your signed-in ChatGPT session instead of an API key.</li>
 <li>Logging: emits MCP logs (<code>info</code> per line, <code>debug</code> for streamed chunks with byte sizes). If browser prerequisites are missing, returns an error payload instead of running.</li>
-<li>Research mode: set <code>browserResearchMode:&quot;deep&quot;</code> for broad public-web research and cited reports. Use normal browser runs with <code>gpt-5.5-pro</code> + <code>browserThinkingTime:&quot;extended&quot;</code> for Pro Extended code review, or <code>gpt-5.5</code> + <code>browserThinkingTime:&quot;heavy&quot;</code> when you explicitly want Thinking Heavy.</li>
+<li>Research mode: set <code>browserResearchMode:&quot;deep&quot;</code> for broad public-web research and cited reports. Use normal browser runs with <code>gpt-5.5-pro</code> + <code>browserThinkingTime:&quot;extended&quot;</code> for Pro Extended code review, <code>gpt-5.6-sol</code> + <code>browserThinkingTime:&quot;extra-high&quot;</code> for Extra High, or <code>browserThinkingTime:&quot;heavy&quot;</code> when you explicitly want Pro.</li>
 <li>Multi-turn consults: set <code>browserFollowUps:[&quot;Challenge your recommendation&quot;, &quot;Give the final decision&quot;]</code> to keep one ChatGPT browser conversation open and ask sequential follow-up prompts. Use one-shot calls for narrow bugs and exact file-set reviews; use multi-turn for ambiguous architecture/product decisions where a challenge pass and final recommendation are useful; use Deep Research for broad public-web work with citations. Oracle never invents follow-ups automatically.</li>
 <li>Archiving: set <code>browserArchive:&quot;auto&quot;|&quot;always&quot;|&quot;never&quot;</code> to control ChatGPT conversation cleanup. <code>auto</code> archives only successful browser one-shots after local artifacts are saved, and skips project, Deep Research, multi-turn, failed, and incomplete sessions.</li>
 <li>ChatGPT image generation: set <code>engine:&quot;browser&quot;</code> and <code>generateImage</code> to a path under <code>ORACLE_HOME_DIR/generated</code> to use the same image-aware wait/download path as CLI <code>--generate-image</code>. Saved files are returned in <code>structuredContent.images</code> and recorded as session artifacts; multiple images save as numbered siblings. Agent-supplied <code>generateImage</code> / <code>outputPath</code> are constrained to that generated-output directory by default (set <code>ORACLE_MCP_ALLOW_EXTERNAL_OUTPUT=1</code> to allow external paths).</li>

package/dist/docs-site/mythical-pro-agents.html

@@ -278,7 +278,7 @@ body:not(.home) .doc>h1:first-child{display:none}
 <p>The captured report lands at <code>~/.oracle/sessions/&lt;id&gt;/artifacts/deep-research-report.md</code>. Deep Research can run for an hour or more — Oracle stores the session and reattaches.</p>
 <h2 id="thinking-time"><a class="anchor" href="#thinking-time" aria-label="Anchor link">#</a>Thinking time</h2>
 <p>Pro / Thinking models in browser mode accept a <code>--browser-thinking-time</code> knob:</p>
-<table><thead><tr><th>Level</th><th>What it maps to in ChatGPT</th></tr></thead><tbody><tr><td><code>light</code></td><td>Quick</td></tr><tr><td><code>standard</code></td><td>Default</td></tr><tr><td><code>extended</code></td><td>Pro Extended / Thinking Extended</td></tr><tr><td><code>heavy</code></td><td>Heavy thinking</td></tr></tbody></table>
+<table><thead><tr><th>Level</th><th>What it maps to in ChatGPT</th></tr></thead><tbody><tr><td><code>light</code></td><td>Instant / Quick</td></tr><tr><td><code>standard</code></td><td>Medium / Standard</td></tr><tr><td><code>extended</code></td><td>High / Pro Extended / Thinking Extended</td></tr><tr><td><code>extra-high</code></td><td>Extra High (GPT-5.6 Sol non-Pro max)</td></tr><tr><td><code>heavy</code></td><td>Matching Heavy row; otherwise keep the current tier</td></tr></tbody></table>
 <pre class="shiki github-dark-dimmed" style="background-color:var(--code-bg);color:var(--code-fg)" tabindex="0"><code class="language-bash"><span class="line"><span style="color:#F69D50">oracle</span><span style="color:#6CB6FF"> --engine</span><span style="color:#96D0FF"> browser</span><span style="color:#6CB6FF"> --model</span><span style="color:#96D0FF"> gpt-5.5-pro</span><span style="color:#F47067"> \</span></span>
 <span class="line"><span style="color:#6CB6FF">  --browser-thinking-time</span><span style="color:#96D0FF"> extended</span><span style="color:#F47067"> \</span></span>
 <span class="line"><span style="color:#6CB6FF">  -p</span><span style="color:#96D0FF"> "Refactor this hot path"</span><span style="color:#6CB6FF"> --file</span><span style="color:#96D0FF"> "src/render/**"</span></span></code></pre>

package/dist/src/browser/actions/thinkingTime.js

@@ -55,7 +55,9 @@ export async function ensureThinkingTime(Runtime, level, logger, desiredModel) {
             if (strictProEffort) {
                 throw new Error(`${message}; refusing to submit without confirmed Pro Extended.`);
             }
-            logger(formatBrowserThinkingLog(`${message}; continuing with ChatGPT default.`));
+            // Nothing was clicked, so the tab keeps whatever effort it already had —
+            // which is not necessarily ChatGPT's default.
+            logger(formatBrowserThinkingLog(`${message}; keeping the effort already selected in ChatGPT.`));
             return;
         }
         default: {
@@ -72,7 +74,7 @@ export async function ensureThinkingTime(Runtime, level, logger, desiredModel) {
 /**
  * Best-effort selection of a thinking time level in ChatGPT's composer pill menu.
  * Safe by default: if the pill/menu/option isn't present, we continue without throwing.
- * @param level - The thinking time intensity: 'light', 'standard', 'extended', or 'heavy'
+ * @param level - The thinking time intensity: 'light', 'standard', 'extended', 'extra-high', or 'heavy'
  */
 export async function ensureThinkingTimeIfAvailable(Runtime, level, logger, desiredModel) {
     try {
@@ -135,12 +137,13 @@ function buildThinkingTimeExpression(level, desiredModel) {
     const TARGET_MODEL_KIND = ${targetModelKindLiteral};
     const TARGET_IS_GPT56_MODEL = ${targetIsGpt56ModelLiteral};
 
-    // Bilingual matchers: English level token + observed Chinese variants.
+    // Multilingual matchers: English level token + observed German/Chinese variants.
     const LEVEL_TOKENS = {
-      light: ['light', 'instant', '轻', '极速'],
-      standard: ['standard', 'medium', '标准', '中'],
-      extended: ['extended', 'high', '扩展', '深度', '加强', '高'],
-      heavy: ['heavy', 'extra high', '重度', '加重', '极高'],
+      light: ['light', 'instant', 'sofort', 'leicht', '轻', '极速'],
+      standard: ['standard', 'medium', 'mittel', '标准', '中'],
+      extended: ['extended', 'high', 'hoch', 'erweitert', '扩展', '深度', '加强', '高'],
+      'extra-high': ['extra high', 'sehr hoch', '极高'],
+      heavy: ['heavy', 'schwer', '重度', '加重'],
     };
     const targetTokens = LEVEL_TOKENS[TARGET_LEVEL] || [TARGET_LEVEL];
 
@@ -155,19 +158,43 @@ function buildThinkingTimeExpression(level, desiredModel) {
     const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
     // Keep CJK characters so we can match Chinese labels against LEVEL_TOKENS.
     const normalize = (value) => (value || '')
+      // Compose first so NFD umlauts fold too, then map them onto ASCII before
+      // the strip below would drop them (and split the token in half).
+      .normalize('NFC')
       .toLowerCase()
+      .replace(/ä/g, 'a')
+      .replace(/ö/g, 'o')
+      .replace(/ü/g, 'u')
+      .replace(/ß/g, 'ss')
       .replace(/[^a-z0-9\\u4e00-\\u9fa5]+/g, ' ')
       .replace(/\\s+/g, ' ')
       .trim();
     const hasToken = (text, token) => normalize(text).split(' ').includes(token);
-    const matchesLevel = (text) => {
+    // Whole-word/phrase containment. Latin effort labels are short words that also
+    // occur inside unrelated UI text ("Hochladen", "Ermitteln") and inside their own
+    // row descriptions ("Hoch – für sehr komplexe Aufgaben"), so plain substring
+    // matching misclassifies rows. CJK labels have no word separators, so they keep
+    // substring semantics.
+    const hasPhrase = (text, phrase) => {
+      const haystack = ' ' + normalize(text) + ' ';
+      const needle = normalize(phrase);
+      if (!needle) return false;
+      return /^[a-z0-9 ]+$/.test(needle)
+        ? haystack.includes(' ' + needle + ' ')
+        : haystack.includes(needle);
+    };
+    // ChatGPT's Pro effort tiers are "Pro Extended"/"Pro Erweitert" per UI language.
+    const hasExtendedWord = (text) => hasPhrase(text, 'extended') || hasPhrase(text, 'erweitert');
+    const matchesTokens = (text, tokens) => {
       const t = normalize(text);
       if (!t) return false;
-      return targetTokens.some((tok) => {
+      return tokens.some((tok) => {
         const token = normalize(tok);
         if (!token) return false;
-        if (token === 'high') return hasToken(t, 'high') && !hasToken(t, 'extra');
-        if (token === 'extra high') return hasToken(t, 'extra') && hasToken(t, 'high');
+        if (token === 'high') return hasPhrase(t, 'high') && !hasPhrase(t, 'extra high');
+        if (token === 'extra high') return hasPhrase(t, 'extra high');
+        if (token === 'hoch') return hasPhrase(t, 'hoch') && !hasPhrase(t, 'sehr hoch');
+        if (token === 'sehr hoch') return hasPhrase(t, 'sehr hoch');
         if (token === '极速') {
           const suffix = t.slice(token.length);
           return t === token || hasToken(t, token) || /^[0-9]/.test(suffix);
@@ -175,27 +202,15 @@ function buildThinkingTimeExpression(level, desiredModel) {
         if (['中', '高', '极高'].includes(token)) {
           return t === token || hasToken(t, token);
         }
+        if (/^[a-z0-9 ]+$/.test(token)) {
+          return hasPhrase(t, token);
+        }
         return t === token || hasToken(t, token) || t.includes(token);
       });
     };
-    const matchesAnyEffortLevel = (text) => {
-      const normalizedText = normalize(text);
-      if (!normalizedText) return false;
-      for (const tokens of Object.values(LEVEL_TOKENS)) {
-        for (const rawToken of tokens) {
-          const token = normalize(rawToken);
-          if (!token) continue;
-          if (token.includes(' ')) {
-            if (token.split(' ').every((part) => hasToken(normalizedText, part))) return true;
-          } else if (/^[a-z0-9]+$/.test(token)) {
-            if (hasToken(normalizedText, token)) return true;
-          } else if (normalizedText.includes(token)) {
-            return true;
-          }
-        }
-      }
-      return false;
-    };
+    const matchesLevel = (text) => matchesTokens(text, targetTokens);
+    const matchesAnyEffortLevel = (text) =>
+      Object.values(LEVEL_TOKENS).some((tokens) => matchesTokens(text, tokens));
     const optionIsSelected = (node) => {
       if (!(node instanceof HTMLElement)) return false;
       const ariaChecked = node.getAttribute('aria-checked');
@@ -347,7 +362,8 @@ function buildThinkingTimeExpression(level, desiredModel) {
         return true;
       }
       const label = menu?.querySelector?.('.__menu-label, [class*="menu-label"]');
-      return normalize(label?.textContent ?? '').includes('intelligence');
+      // 'intelligen' matches both "Intelligence" and German "Intelligenz".
+      return normalize(label?.textContent ?? '').includes('intelligen');
     };
     const failure = (status, extra = {}) => ({
       status,
@@ -390,24 +406,10 @@ function buildThinkingTimeExpression(level, desiredModel) {
           return null;
         }
       }
-      if (
-        TARGET_IS_GPT56_MODEL &&
-        TARGET_LEVEL === 'heavy' &&
-        isIntelligenceEffortMenu(menu)
-      ) {
-        for (const item of items) {
-          const itemText = normalize(
-            (item.textContent ?? '') + ' ' + (item.getAttribute?.('aria-label') ?? ''),
-          );
-          if (
-            hasToken(itemText, 'pro') &&
-            !itemText.includes('gpt') &&
-            !/(?:^|\\s)5[ .-]?6(?:\\s|$)/.test(itemText)
-          ) {
-            return item;
-          }
-        }
-      }
+      // Generic effort-label match for every model/level. GPT-5.6 heavy used to
+      // short-circuit to the Pro row before reaching here; it no longer does, so
+      // a UI without a matching tier (e.g. German, which has no "heavy") falls
+      // through to null and the caller keeps the current selection.
       for (const item of items) {
         const itemText = normalize(
           (item.textContent ?? '') + ' ' + (item.getAttribute?.('aria-label') ?? ''),
@@ -422,10 +424,10 @@ function buildThinkingTimeExpression(level, desiredModel) {
           return item;
         }
       }
-      if (TARGET_LEVEL === 'heavy') {
-        // Older Chinese layouts used bare 高 for the highest effort. Keep it
-        // only as a second-pass exact fallback so a current 高 row can never
-        // win before the primary 极高 row.
+      if (TARGET_LEVEL === 'extra-high') {
+        // Older Chinese layouts used bare 高 for the highest non-Pro effort.
+        // Keep it only as a second-pass exact fallback so a current 高 row can
+        // never win before the primary 极高 row.
         for (const item of items) {
           const itemText = normalize(item.textContent ?? '');
           const ariaLabel = normalize(item.getAttribute?.('aria-label') ?? '');
@@ -434,11 +436,15 @@ function buildThinkingTimeExpression(level, desiredModel) {
       }
       return null;
     };
+    // Menu-shape heuristic only. This reads the whole menu's textContent, where
+    // adjacent row labels concatenate without a separator ("Pro StandardPro
+    // Extended"), so word-boundary matching does not apply here — substring is
+    // deliberate. Row-level classification uses matchesLevel/matchesTokens.
     const countEffortLevels = (menu) => {
       const text = normalize(menu?.textContent ?? '');
       let hits = 0;
       for (const tokens of Object.values(LEVEL_TOKENS)) {
-        if (tokens.some((token) => text.includes(String(token).toLowerCase()))) hits += 1;
+        if (tokens.some((token) => text.includes(normalize(token)))) hits += 1;
       }
       return hits;
     };
@@ -449,16 +455,22 @@ function buildThinkingTimeExpression(level, desiredModel) {
       const label = menu.querySelector?.('.__menu-label, [class*="menu-label"]');
       const labelText = normalize(label?.textContent ?? '');
       return (
-        labelText.includes('intelligence') ||
+        labelText.includes('intelligen') ||
         labelText.includes('thinking time') ||
         labelText.includes('thinking effort') ||
+        labelText.includes('denkdauer') ||
+        labelText.includes('denkzeit') ||
         countEffortLevels(menu) >= 2
       );
     };
     const isProEffortMenu = (menu) => {
       if (!isVisible(menu)) return false;
       const text = normalize(menu?.textContent ?? '');
-      return text.includes('pro standard') && text.includes('pro extended');
+      // Aggregate menu text, so plain substring only (see countEffortLevels).
+      return (
+        text.includes('pro standard') &&
+        (text.includes('pro extended') || text.includes('pro erweitert'))
+      );
     };
     const controlledMenu = (trigger) => {
       const id = trigger?.getAttribute?.('aria-controls');
@@ -493,10 +505,10 @@ function buildThinkingTimeExpression(level, desiredModel) {
         (node?.textContent ?? '') + ' ' + (node?.getAttribute?.('aria-label') ?? ''),
       );
       if (TARGET_LEVEL === 'standard') {
-        return text.includes('pro') && text.includes('standard');
+        return hasPhrase(text, 'pro') && hasPhrase(text, 'standard');
       }
       if (TARGET_LEVEL === 'extended') {
-        return text.includes('pro') && text.includes('extended');
+        return hasPhrase(text, 'pro') && hasExtendedWord(text);
       }
       return false;
     };
@@ -521,10 +533,10 @@ function buildThinkingTimeExpression(level, desiredModel) {
       }
       const label = normalize(button?.textContent ?? '');
       if (TARGET_LEVEL === 'standard') {
-        return hasToken(label, 'pro') && !hasToken(label, 'extended');
+        return hasToken(label, 'pro') && !hasExtendedWord(label);
       }
       if (TARGET_LEVEL === 'extended') {
-        return hasToken(label, 'pro') && hasToken(label, 'extended');
+        return hasToken(label, 'pro') && hasExtendedWord(label);
       }
       return false;
     };
@@ -534,13 +546,9 @@ function buildThinkingTimeExpression(level, desiredModel) {
       const normalizedLabel = normalize(
         (button?.textContent ?? '') + ' ' + (button?.getAttribute?.('aria-label') ?? ''),
       );
-      if (
-        TARGET_IS_GPT56_MODEL &&
-        TARGET_LEVEL === 'heavy' &&
-        hasToken(normalizedLabel, 'pro')
-      ) {
-        return true;
-      }
+      // No 5.6-heavy "a Pro pill counts as heavy" shortcut here: that would also
+      // make post-click verification pass on an unchanged Pro pill. selectAndVerify
+      // handles the already-on-Pro case explicitly before any click.
       if ((modelKindOverride || TARGET_MODEL_KIND || modelKindFromNode(button)) === 'pro') {
         return false;
       }
@@ -553,14 +561,21 @@ function buildThinkingTimeExpression(level, desiredModel) {
         modelKindFromNode(trigger) ||
         effectiveTargetModelKind();
       const option = findOption();
-      if (
-        !option &&
-        TARGET_IS_GPT56_MODEL &&
-        TARGET_LEVEL === 'heavy' &&
-        currentEffortPillMatchesTarget(trigger, triggerModelKind)
-      ) {
-        closeOpenMenus();
-        return { status: 'already-selected', label: trigger.textContent?.trim?.() || null };
+      if (!option && TARGET_IS_GPT56_MODEL && TARGET_LEVEL === 'heavy') {
+        // GPT-5.6 has no "heavy" tier: Pro is the closest thing. Accept a pill that
+        // is already on Pro as satisfying the request, but never click Pro to get
+        // there, and never let this stand in for post-click verification.
+        const pill = freshComposerTrigger(trigger) || findModelButton();
+        const pillLabel = normalize(
+          (pill?.textContent ?? '') + ' ' + (pill?.getAttribute?.('aria-label') ?? ''),
+        );
+        if (
+          hasToken(pillLabel, 'pro') ||
+          currentEffortPillMatchesTarget(trigger, triggerModelKind)
+        ) {
+          closeOpenMenus();
+          return { status: 'already-selected', label: trigger.textContent?.trim?.() || null };
+        }
       }
       if (!option) return failure('option-not-found', { modelKind: triggerModelKind });
       const label = option.textContent?.trim?.() || null;
@@ -857,7 +872,7 @@ function buildThinkingTimeExpression(level, desiredModel) {
         const text = normalize(
           (node?.textContent ?? '') + ' ' + (node?.getAttribute?.('aria-label') ?? ''),
         );
-        return text.includes('pro') && text.includes('extended');
+        return hasPhrase(text, 'pro') && hasExtendedWord(text);
       };
       const findProExtendedOption = () => {
         const menu = document.querySelector(INTELLIGENCE_MENU_SELECTOR);

package/dist/src/browser/chromeLifecycle.js

@@ -542,6 +542,12 @@ function buildChromeFlags(headless, debugBindAddress, hideWindow = false) {
         "--disable-features=TranslateUI,AutomationControlled",
         "--mute-audio",
         "--window-size=1280,720",
+        // Chrome that *we* launch is pinned to English, so ChatGPT renders the labels
+        // our selectors were written against. This does not make English the only case
+        // to handle: --browser-attach-running and --remote-chrome never build these
+        // flags (see controlPlan.ts), so those runs inherit the user's own Chrome
+        // locale, and a ChatGPT account language setting can localize the UI even here.
+        // That is why the model/effort matchers must stay language-tolerant.
         "--lang=en-US",
         "--accept-lang=en-US,en",
     ];

package/dist/src/cli/options.js

@@ -133,7 +133,7 @@ export function parseThinkingTimeOption(value) {
     if (normalized) {
         return normalized;
     }
-    throw new InvalidArgumentError('Thinking time must be one of "light", "standard", "extended", "heavy", or a ChatGPT UI alias like "instant", "medium", "high", or "extra-high".');
+    throw new InvalidArgumentError('Thinking time must be one of "light", "standard", "extended", "extra-high", "heavy", or a ChatGPT UI alias like "instant", "medium", "high", or "xhigh".');
 }
 export function normalizeModelOption(value) {
     return (value ?? "").trim();

package/dist/src/oracle/thinkingTime.js

@@ -1,10 +1,15 @@
-export const THINKING_TIME_LEVELS = ["light", "standard", "extended", "heavy"];
+export const THINKING_TIME_LEVELS = [
+    "light",
+    "standard",
+    "extended",
+    "extra-high",
+    "heavy",
+];
 export const THINKING_TIME_ALIASES = [
     "instant",
     "low",
     "medium",
     "high",
-    "extra-high",
     "extra high",
     "extrahigh",
     "xhigh",
@@ -29,10 +34,11 @@ export function normalizeThinkingTimeLevel(value) {
         case "extended":
         case "high":
             return "extended";
-        case "heavy":
         case "extra-high":
         case "extrahigh":
         case "xhigh":
+            return "extra-high";
+        case "heavy":
             return "heavy";
         default:
             return null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@steipete/oracle",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "description": "CLI wrapper around OpenAI Responses API with GPT-5.6 Sol, GPT-5.6, GPT-5.5 Pro, GPT-5.5, GPT-5.4, GPT-5.2, GPT-5.1, and GPT-5.1 Codex high reasoning modes.",
   "keywords": [],
   "homepage": "https://askoracle.sh",
@@ -73,11 +73,11 @@
     "inquirer": "14.0.2",
     "json5": "^2.2.3",
     "kleur": "^4.1.5",
-    "markdansi": "0.3.2",
-    "openai": "^7.3.0",
+    "markdansi": "0.3.3",
+    "openai": "^7.4.0",
     "osc-progress": "^0.3.2",
     "qs": "^6.15.3",
-    "shiki": "^4.4.1",
+    "shiki": "^4.4.2",
     "toasted-notifier": "^10.1.0",
     "tokentally": "^0.1.3",
     "zod": "^4.4.3"
@@ -88,13 +88,13 @@
     "@types/inquirer": "^9.0.10",
     "@types/node": "^26.1.2",
     "@vitest/coverage-v8": "4.1.10",
-    "devtools-protocol": "0.0.1672245",
+    "devtools-protocol": "0.0.1673900",
     "es-toolkit": "^1.50.0",
     "esbuild": "^0.28.1",
-    "oxfmt": "0.61.0",
-    "oxlint": "^1.76.0",
-    "puppeteer-core": "^25.4.0",
-    "tsx": "^4.23.5",
+    "oxfmt": "0.62.0",
+    "oxlint": "^1.77.0",
+    "puppeteer-core": "^25.5.0",
+    "tsx": "^4.23.7",
     "typescript": "^7.0.2",
     "vitest": "^4.1.10"
   },
@@ -109,5 +109,5 @@
   "engines": {
     "node": ">=24"
   },
-  "packageManager": "pnpm@10.33.2"
+  "packageManager": "pnpm@10.34.5"
 }