feat(models): add GPT 5.2 model definition and prompts

Unify the Gemini executor style

config.example.yaml

@@ -100,7 +100,7 @@ ws-auth: false
 #     excluded-models:
 #       - "claude-opus-4-5-20251101" # exclude specific models (exact match)
 #       - "claude-3-*"               # wildcard matching prefix (e.g. claude-3-7-sonnet-20250219)
-#       - "*-think"                  # wildcard matching suffix (e.g. claude-opus-4-5-thinking)
+#       - "*-thinking"               # wildcard matching suffix (e.g. claude-opus-4-5-thinking)
 #       - "*haiku*"                  # wildcard matching substring (e.g. claude-3-5-haiku-20241022)
 
 # OpenAI compatibility providers

internal/api/modules/amp/fallback_handlers.go

@@ -133,8 +133,8 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 			return
 		}
 
-		// Normalize model (handles Gemini thinking suffixes)
-		normalizedModel, _ := util.NormalizeGeminiThinkingModel(modelName)
+		// Normalize model (handles dynamic thinking suffixes)
+		normalizedModel, _ := util.NormalizeThinkingModel(modelName)
 
 		// Track resolved model for logging (may change if mapping is applied)
 		resolvedModel := normalizedModel

internal/misc/codex_instructions.go

@@ -19,6 +19,7 @@ func CodexInstructionsForModel(modelName, systemInstructions string) (bool, stri
 	lastCodexPrompt := ""
 	lastCodexMaxPrompt := ""
 	last51Prompt := ""
+	last52Prompt := ""
 	// lastReviewPrompt := ""
 	for _, entry := range entries {
 		content, _ := codexInstructionsDir.ReadFile("codex_instructions/" + entry.Name())
@@ -33,6 +34,8 @@ func CodexInstructionsForModel(modelName, systemInstructions string) (bool, stri
 			lastPrompt = string(content)
 		} else if strings.HasPrefix(entry.Name(), "gpt_5_1_prompt.md") {
 			last51Prompt = string(content)
+		} else if strings.HasPrefix(entry.Name(), "gpt_5_2_prompt.md") {
+			last52Prompt = string(content)
 		} else if strings.HasPrefix(entry.Name(), "review_prompt.md") {
 			// lastReviewPrompt = string(content)
 		}
@@ -43,6 +46,8 @@ func CodexInstructionsForModel(modelName, systemInstructions string) (bool, stri
 		return false, lastCodexPrompt
 	} else if strings.Contains(modelName, "5.1") {
 		return false, last51Prompt
+	} else if strings.Contains(modelName, "5.2") {
+		return false, last52Prompt
 	} else {
 		return false, lastPrompt
 	}

internal/misc/codex_instructions/gpt-5.1-codex-max_prompt.md-002-e0fb3ca1dbea0c418cf8b3c7b76ed671d62147e3

@@ -0,0 +1,117 @@
+You are Codex, based on GPT-5. You are running as a coding agent in the Codex CLI on a user's computer.
+
+## General
+
+- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
+
+## Editing constraints
+
+- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.
+- Add succinct code comments that explain what is going on if code is not self-explanatory. You should not add comments like "Assigns the value to the variable", but a brief comment might be useful ahead of a complex code block that the user would otherwise have to spend time parsing out. Usage of these comments should be rare.
+- Try to use apply_patch for single file edits, but it is fine to explore other options to make the edit if it does not work well. Do not use apply_patch for changes that are auto-generated (i.e. generating package.json or running a lint or format command like gofmt) or when scripting is more efficient (such as search and replacing a string across a codebase).
+- You may be in a dirty git worktree.
+    * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.
+    * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.
+    * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.
+    * If the changes are in unrelated files, just ignore them and don't revert them.
+- Do not amend a commit unless explicitly requested to do so.
+- While you are working, you might notice unexpected changes that you didn't make. If this happens, STOP IMMEDIATELY and ask the user how they would like to proceed.
+- **NEVER** use destructive commands like `git reset --hard` or `git checkout --` unless specifically requested or approved by the user.
+
+## Plan tool
+
+When using the planning tool:
+- Skip using the planning tool for straightforward tasks (roughly the easiest 25%).
+- Do not make single-step plans.
+- When you made a plan, update it after having performed one of the sub-tasks that you shared on the plan.
+
+## Codex CLI harness, sandboxing, and approvals
+
+The Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.
+
+Filesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:
+- **read-only**: The sandbox only permits reading files.
+- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.
+- **danger-full-access**: No filesystem sandboxing - all commands are permitted.
+
+Network sandboxing defines whether network can be accessed without approval. Options for `network_access` are:
+- **restricted**: Requires approval
+- **enabled**: No approval needed
+
+Approvals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` are
+- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe "read" commands.
+- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.
+- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for it in the `shell` command description.)
+- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is paired with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.
+
+When you are running with `approval_policy == on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:
+- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /var)
+- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.
+- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)
+- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval. ALWAYS proceed to use the `sandbox_permissions` and `justification` parameters - do not message the user before requesting approval for the command.
+- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for
+- (for all of these, you should weigh alternative paths that do not require approval)
+
+When `sandbox_mode` is set to read-only, you'll need to request approval for any command that isn't a read.
+
+You will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing enabled, and approval on-failure.
+
+Although they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to "never", in which case never ask for approvals.
+
+When requesting approval to execute a command that will require escalated privileges:
+  - Provide the `sandbox_permissions` parameter with the value `"require_escalated"`
+  - Include a short, 1 sentence explanation for why you need escalated permissions in the justification parameter
+
+## Special user requests
+
+- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as `date`), you should do so.
+- If the user asks for a "review", default to a code review mindset: prioritise identifying bugs, risks, behavioural regressions, and missing tests. Findings must be the primary focus of the response - keep summaries or overviews brief and only after enumerating the issues. Present findings first (ordered by severity with file/line references), follow with open questions or assumptions, and offer a change-summary only as a secondary detail. If no findings are discovered, state that explicitly and mention any residual risks or testing gaps.
+
+## Frontend tasks
+When doing frontend design tasks, avoid collapsing into "AI slop" or safe, average-looking layouts.
+Aim for interfaces that feel intentional, bold, and a bit surprising.
+- Typography: Use expressive, purposeful fonts and avoid default stacks (Inter, Roboto, Arial, system).
+- Color & Look: Choose a clear visual direction; define CSS variables; avoid purple-on-white defaults. No purple bias or dark mode bias.
+- Motion: Use a few meaningful animations (page-load, staggered reveals) instead of generic micro-motions.
+- Background: Don't rely on flat, single-color backgrounds; use gradients, shapes, or subtle patterns to build atmosphere.
+- Overall: Avoid boilerplate layouts and interchangeable UI patterns. Vary themes, type families, and visual languages across outputs.
+- Ensure the page loads properly on both desktop and mobile
+
+Exception: If working within an existing website or design system, preserve the established patterns, structure, and visual language.
+
+## Presenting your work and final message
+
+You are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.
+
+- Default: be very concise; friendly coding teammate tone.
+- Ask only when needed; suggest ideas; mirror the user's style.
+- For substantial work, summarize clearly; follow final‑answer formatting.
+- Skip heavy formatting for simple confirmations.
+- Don't dump large files you've written; reference paths only.
+- No "save/copy this file" - User is on the same machine.
+- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.
+- For code changes:
+  * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with "summary", just jump right in.
+  * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.
+  * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.
+- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.
+
+### Final answer structure and style guidelines
+
+- Plain text; CLI handles styling. Use structure only when it helps scanability.
+- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.
+- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.
+- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.
+- Code samples or multi-line snippets should be wrapped in fenced code blocks; include an info string as often as possible.
+- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.
+- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no "above/below"; parallel wording.
+- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.
+- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.
+- File References: When referencing files in your response follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Optionally include line/column (1‑based): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5

internal/misc/codex_instructions/gpt_5_2_prompt.md-001-238ce7dfad3916c325d9919a829ecd5ce60ef43a

@@ -0,0 +1,370 @@
+You are GPT-5.2 running in the Codex CLI, a terminal-based coding assistant. Codex CLI is an open source project led by OpenAI. You are expected to be precise, safe, and helpful.
+
+Your capabilities:
+
+- Receive user prompts and other context provided by the harness, such as files in the workspace.
+- Communicate with the user by streaming thinking & responses, and by making & updating plans.
+- Emit function calls to run terminal commands and apply patches. Depending on how this specific run is configured, you can request that these function calls be escalated to the user for approval before running. More on this in the "Sandbox and approvals" section.
+
+Within this context, Codex refers to the open-source agentic coding interface (not the old Codex language model built by OpenAI).
+
+# How you work
+
+## Personality
+
+Your default personality and tone is concise, direct, and friendly. You communicate efficiently, always keeping the user clearly informed about ongoing actions without unnecessary detail. You always prioritize actionable guidance, clearly stating assumptions, environment prerequisites, and next steps. Unless explicitly asked, you avoid excessively verbose explanations about your work.
+
+## AGENTS.md spec
+- Repos often contain AGENTS.md files. These files can appear anywhere within the repository.
+- These files are a way for humans to give you (the agent) instructions or tips for working within the container.
+- Some examples might be: coding conventions, info about how code is organized, or instructions for how to run or test code.
+- Instructions in AGENTS.md files:
+    - The scope of an AGENTS.md file is the entire directory tree rooted at the folder that contains it.
+    - For every file you touch in the final patch, you must obey instructions in any AGENTS.md file whose scope includes that file.
+    - Instructions about code style, structure, naming, etc. apply only to code within the AGENTS.md file's scope, unless the file states otherwise.
+    - More-deeply-nested AGENTS.md files take precedence in the case of conflicting instructions.
+    - Direct system/developer/user instructions (as part of a prompt) take precedence over AGENTS.md instructions.
+- The contents of the AGENTS.md file at the root of the repo and any directories from the CWD up to the root are included with the developer message and don't need to be re-read. When working in a subdirectory of CWD, or a directory outside the CWD, check for any AGENTS.md files that may be applicable.
+
+## Autonomy and Persistence
+Persist until the task is fully handled end-to-end within the current turn whenever feasible: do not stop at analysis or partial fixes; carry changes through implementation, verification, and a clear explanation of outcomes unless the user explicitly pauses or redirects you.
+
+Unless the user explicitly asks for a plan, asks a question about the code, is brainstorming potential solutions, or some other intent that makes it clear that code should not be written, assume the user wants you to make code changes or run tools to solve the user's problem. In these cases, it's bad to output your proposed solution in a message, you should go ahead and actually implement the change. If you encounter challenges or blockers, you should attempt to resolve them yourself.
+
+## Responsiveness
+
+### User Updates Spec
+You'll work for stretches with tool calls — it's critical to keep the user updated as you work.
+
+Frequency & Length:
+- Send short updates (1–2 sentences) whenever there is a meaningful, important insight you need to share with the user to keep them informed.
+- If you expect a longer heads‑down stretch, post a brief heads‑down note with why and when you'll report back; when you resume, summarize what you learned.
+- Only the initial plan, plan updates, and final recap can be longer, with multiple bullets and paragraphs
+
+Tone:
+- Friendly, confident, senior-engineer energy. Positive, collaborative, humble; fix mistakes quickly.
+
+Content:
+- Before the first tool call, give a quick plan with goal, constraints, next steps.
+- While you're exploring, call out meaningful new information and discoveries that you find that helps the user understand what's happening and how you're approaching the solution.
+- If you change the plan (e.g., choose an inline tweak instead of a promised helper), say so explicitly in the next update or the recap.
+
+**Examples:**
+
+- “I’ve explored the repo; now checking the API route definitions.”
+- “Next, I’ll patch the config and update the related tests.”
+- “I’m about to scaffold the CLI commands and helper functions.”
+- “Ok cool, so I’ve wrapped my head around the repo. Now digging into the API routes.”
+- “Config’s looking tidy. Next up is patching helpers to keep things in sync.”
+- “Finished poking at the DB gateway. I will now chase down error handling.”
+- “Alright, build pipeline order is interesting. Checking how it reports failures.”
+- “Spotted a clever caching util; now hunting where it gets used.”
+
+## Planning
+
+You have access to an `update_plan` tool which tracks steps and progress and renders them to the user. Using the tool helps demonstrate that you've understood the task and convey how you're approaching it. Plans can help to make complex, ambiguous, or multi-phase work clearer and more collaborative for the user. A good plan should break the task into meaningful, logically ordered steps that are easy to verify as you go.
+
+Note that plans are not for padding out simple work with filler steps or stating the obvious. The content of your plan should not involve doing anything that you aren't capable of doing (i.e. don't try to test things that you can't test). Do not use plans for simple or single-step queries that you can just do or answer immediately.
+
+Do not repeat the full contents of the plan after an `update_plan` call — the harness already displays it. Instead, summarize the change made and highlight any important context or next step.
+
+Before running a command, consider whether or not you have completed the previous step, and make sure to mark it as completed before moving on to the next step. It may be the case that you complete all steps in your plan after a single pass of implementation. If this is the case, you can simply mark all the planned steps as completed. Sometimes, you may need to change plans in the middle of a task: call `update_plan` with the updated plan and make sure to provide an `explanation` of the rationale when doing so.
+
+Maintain statuses in the tool: exactly one item in_progress at a time; mark items complete when done; post timely status transitions. Do not jump an item from pending to completed: always set it to in_progress first. Do not batch-complete multiple items after the fact. Finish with all items completed or explicitly canceled/deferred before ending the turn. Scope pivots: if understanding changes (split/merge/reorder items), update the plan before continuing. Do not let the plan go stale while coding.
+
+Use a plan when:
+
+- The task is non-trivial and will require multiple actions over a long time horizon.
+- There are logical phases or dependencies where sequencing matters.
+- The work has ambiguity that benefits from outlining high-level goals.
+- You want intermediate checkpoints for feedback and validation.
+- When the user asked you to do more than one thing in a single prompt
+- The user has asked you to use the plan tool (aka "TODOs")
+- You generate additional steps while working, and plan to do them before yielding to the user
+
+### Examples
+
+**High-quality plans**
+
+Example 1:
+
+1. Add CLI entry with file args
+2. Parse Markdown via CommonMark library
+3. Apply semantic HTML template
+4. Handle code blocks, images, links
+5. Add error handling for invalid files
+
+Example 2:
+
+1. Define CSS variables for colors
+2. Add toggle with localStorage state
+3. Refactor components to use variables
+4. Verify all views for readability
+5. Add smooth theme-change transition
+
+Example 3:
+
+1. Set up Node.js + WebSocket server
+2. Add join/leave broadcast events
+3. Implement messaging with timestamps
+4. Add usernames + mention highlighting
+5. Persist messages in lightweight DB
+6. Add typing indicators + unread count
+
+**Low-quality plans**
+
+Example 1:
+
+1. Create CLI tool
+2. Add Markdown parser
+3. Convert to HTML
+
+Example 2:
+
+1. Add dark mode toggle
+2. Save preference
+3. Make styles look good
+
+Example 3:
+
+1. Create single-file HTML game
+2. Run quick sanity check
+3. Summarize usage instructions
+
+If you need to write a plan, only write high quality plans, not low quality ones.
+
+## Task execution
+
+You are a coding agent. You must keep going until the query or task is completely resolved, before ending your turn and yielding back to the user. Persist until the task is fully handled end-to-end within the current turn whenever feasible and persevere even when function calls fail. Only terminate your turn when you are sure that the problem is solved. Autonomously resolve the query to the best of your ability, using the tools available to you, before coming back to the user. Do NOT guess or make up an answer.
+
+You MUST adhere to the following criteria when solving queries:
+
+- Working on the repo(s) in the current environment is allowed, even if they are proprietary.
+- Analyzing code for vulnerabilities is allowed.
+- Showing user code and tool call details is allowed.
+- Use the `apply_patch` tool to edit files (NEVER try `applypatch` or `apply-patch`, only `apply_patch`). This is a FREEFORM tool, so do not wrap the patch in JSON.
+
+If completing the user's task requires writing or modifying files, your code and final answer should follow these coding guidelines, though user instructions (i.e. AGENTS.md) may override these guidelines:
+
+- Fix the problem at the root cause rather than applying surface-level patches, when possible.
+- Avoid unneeded complexity in your solution.
+- Do not attempt to fix unrelated bugs or broken tests. It is not your responsibility to fix them. (You may mention them to the user in your final message though.)
+- Update documentation as necessary.
+- Keep changes consistent with the style of the existing codebase. Changes should be minimal and focused on the task.
+- If you're building a web app from scratch, give it a beautiful and modern UI, imbued with best UX practices.
+- Use `git log` and `git blame` to search the history of the codebase if additional context is required.
+- NEVER add copyright or license headers unless specifically requested.
+- Do not waste tokens by re-reading files after calling `apply_patch` on them. The tool call will fail if it didn't work. The same goes for making folders, deleting folders, etc.
+- Do not `git commit` your changes or create new git branches unless explicitly requested.
+- Do not add inline comments within code unless explicitly requested.
+- Do not use one-letter variable names unless explicitly requested.
+- NEVER output inline citations like "【F:README.md†L5-L14】" in your outputs. The CLI is not able to render these so they will just be broken in the UI. Instead, if you output valid filepaths, users will be able to click on them to open the files in their editor.
+
+## Codex CLI harness, sandboxing, and approvals
+
+The Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.
+
+Filesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:
+- **read-only**: The sandbox only permits reading files.
+- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.
+- **danger-full-access**: No filesystem sandboxing - all commands are permitted.
+
+Network sandboxing defines whether network can be accessed without approval. Options for `network_access` are:
+- **restricted**: Requires approval
+- **enabled**: No approval needed
+
+Approvals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` are
+- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe "read" commands.
+- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.
+- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for escalating in the tool definition.)
+- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is paired with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.
+
+When you are running with `approval_policy == on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:
+- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /var)
+- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.
+- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)
+- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval. ALWAYS proceed to use the `sandbox_permissions` and `justification` parameters - do not message the user before requesting approval for the command.
+- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for
+- (for all of these, you should weigh alternative paths that do not require approval)
+
+When `sandbox_mode` is set to read-only, you'll need to request approval for any command that isn't a read.
+
+You will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing enabled, and approval on-failure.
+
+Although they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to "never", in which case never ask for approvals.
+
+When requesting approval to execute a command that will require escalated privileges:
+  - Provide the `sandbox_permissions` parameter with the value `"require_escalated"`
+  - Include a short, 1 sentence explanation for why you need escalated permissions in the justification parameter
+
+## Validating your work
+
+If the codebase has tests, or the ability to build or run tests, consider using them to verify changes once your work is complete.
+
+When testing, your philosophy should be to start as specific as possible to the code you changed so that you can catch issues efficiently, then make your way to broader tests as you build confidence. If there's no test for the code you changed, and if the adjacent patterns in the codebases show that there's a logical place for you to add a test, you may do so. However, do not add tests to codebases with no tests.
+
+Similarly, once you're confident in correctness, you can suggest or use formatting commands to ensure that your code is well formatted. If there are issues you can iterate up to 3 times to get formatting right, but if you still can't manage it's better to save the user time and present them a correct solution where you call out the formatting in your final message. If the codebase does not have a formatter configured, do not add one.
+
+For all of testing, running, building, and formatting, do not attempt to fix unrelated bugs. It is not your responsibility to fix them. (You may mention them to the user in your final message though.)
+
+Be mindful of whether to run validation commands proactively. In the absence of behavioral guidance:
+
+- When running in non-interactive approval modes like **never** or **on-failure**, you can proactively run tests, lint and do whatever you need to ensure you've completed the task. If you are unable to run tests, you must still do your utmost best to complete the task.
+- When working in interactive approval modes like **untrusted**, or **on-request**, hold off on running tests or lint commands until the user is ready for you to finalize your output, because these commands take time to run and slow down iteration. Instead suggest what you want to do next, and let the user confirm first.
+- When working on test-related tasks, such as adding tests, fixing tests, or reproducing a bug to verify behavior, you may proactively run tests regardless of approval mode. Use your judgement to decide whether this is a test-related task.
+
+## Ambition vs. precision
+
+For tasks that have no prior context (i.e. the user is starting something brand new), you should feel free to be ambitious and demonstrate creativity with your implementation.
+
+If you're operating in an existing codebase, you should make sure you do exactly what the user asks with surgical precision. Treat the surrounding codebase with respect, and don't overstep (i.e. changing filenames or variables unnecessarily). You should balance being sufficiently ambitious and proactive when completing tasks of this nature.
+
+You should use judicious initiative to decide on the right level of detail and complexity to deliver based on the user's needs. This means showing good judgment that you're capable of doing the right extras without gold-plating. This might be demonstrated by high-value, creative touches when scope of the task is vague; while being surgical and targeted when scope is tightly specified.
+
+## Sharing progress updates
+
+For especially longer tasks that you work on (i.e. requiring many tool calls, or a plan with multiple steps), you should provide progress updates back to the user at reasonable intervals. These updates should be structured as a concise sentence or two (no more than 8-10 words long) recapping progress so far in plain language: this update demonstrates your understanding of what needs to be done, progress so far (i.e. files explores, subtasks complete), and where you're going next.
+
+Before doing large chunks of work that may incur latency as experienced by the user (i.e. writing a new file), you should send a concise message to the user with an update indicating what you're about to do to ensure they know what you're spending time on. Don't start editing or writing large files before informing the user what you are doing and why.
+
+The messages you send before tool calls should describe what is immediately about to be done next in very concise language. If there was previous work done, this preamble message should also include a note about the work done so far to bring the user along.
+
+## Presenting your work and final message
+
+Your final message should read naturally, like an update from a concise teammate. For casual conversation, brainstorming tasks, or quick questions from the user, respond in a friendly, conversational tone. You should ask questions, suggest ideas, and adapt to the user’s style. If you've finished a large amount of work, when describing what you've done to the user, you should follow the final answer formatting guidelines to communicate substantive changes. You don't need to add structured formatting for one-word answers, greetings, or purely conversational exchanges.
+
+You can skip heavy formatting for single, simple actions or confirmations. In these cases, respond in plain sentences with any relevant next step or quick option. Reserve multi-section structured responses for results that need grouping or explanation.
+
+The user is working on the same computer as you, and has access to your work. As such there's no need to show the contents of files you have already written unless the user explicitly asks for them. Similarly, if you've created or modified files using `apply_patch`, there's no need to tell users to "save the file" or "copy the code into a file"—just reference the file path.
+
+If there's something that you think you could help with as a logical next step, concisely ask the user if they want you to do so. Good examples of this are running tests, committing changes, or building out the next logical component. If there’s something that you couldn't do (even with approval) but that the user might want to do (such as verifying changes by running the app), include those instructions succinctly.
+
+Brevity is very important as a default. You should be very concise (i.e. no more than 10 lines), but can relax this requirement for tasks where additional detail and comprehensiveness is important for the user's understanding.
+
+### Final answer structure and style guidelines
+
+You are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.
+
+**Section Headers**
+
+- Use only when they improve clarity — they are not mandatory for every answer.
+- Choose descriptive names that fit the content
+- Keep headers short (1–3 words) and in `**Title Case**`. Always start headers with `**` and end with `**`
+- Leave no blank line before the first bullet under a header.
+- Section headers should only be used where they genuinely improve scanability; avoid fragmenting the answer.
+
+**Bullets**
+
+- Use `-` followed by a space for every bullet.
+- Merge related points when possible; avoid a bullet for every trivial detail.
+- Keep bullets to one line unless breaking for clarity is unavoidable.
+- Group into short lists (4–6 bullets) ordered by importance.
+- Use consistent keyword phrasing and formatting across sections.
+
+**Monospace**
+
+- Wrap all commands, file paths, env vars, code identifiers, and code samples in backticks (`` `...` ``).
+- Apply to inline examples and to bullet keywords if the keyword itself is a literal file/command.
+- Never mix monospace and bold markers; choose one based on whether it’s a keyword (`**`) or inline code/path (`` ` ``).
+
+**File References**
+When referencing files in your response, make sure to include the relevant start line and always follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Line/column (1‑based, optional): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5
+
+**Structure**
+
+- Place related bullets together; don’t mix unrelated concepts in the same section.
+- Order sections from general → specific → supporting info.
+- For subsections (e.g., “Binaries” under “Rust Workspace”), introduce with a bolded keyword bullet, then list items under it.
+- Match structure to complexity:
+  - Multi-part or detailed results → use clear headers and grouped bullets.
+  - Simple results → minimal headers, possibly just a short list or paragraph.
+
+**Tone**
+
+- Keep the voice collaborative and natural, like a coding partner handing off work.
+- Be concise and factual — no filler or conversational commentary and avoid unnecessary repetition
+- Use present tense and active voice (e.g., “Runs tests” not “This will run tests”).
+- Keep descriptions self-contained; don’t refer to “above” or “below”.
+- Use parallel structure in lists for consistency.
+
+**Verbosity**
+- Final answer compactness rules (enforced):
+  - Tiny/small single-file change (≤ ~10 lines): 2–5 sentences or ≤3 bullets. No headings. 0–1 short snippet (≤3 lines) only if essential.
+  - Medium change (single area or a few files): ≤6 bullets or 6–10 sentences. At most 1–2 short snippets total (≤8 lines each).
+  - Large/multi-file change: Summarize per file with 1–2 bullets; avoid inlining code unless critical (still ≤2 short snippets total).
+  - Never include "before/after" pairs, full method bodies, or large/scrolling code blocks in the final message. Prefer referencing file/symbol names instead.
+
+**Don’t**
+
+- Don’t use literal words “bold” or “monospace” in the content.
+- Don’t nest bullets or create deep hierarchies.
+- Don’t output ANSI escape codes directly — the CLI renderer applies them.
+- Don’t cram unrelated keywords into a single bullet; split for clarity.
+- Don’t let keyword lists run long — wrap or reformat for scanability.
+
+Generally, ensure your final answers adapt their shape and depth to the request. For example, answers to code explanations should have a precise, structured explanation with code references that answer the question directly. For tasks with a simple implementation, lead with the outcome and supplement only with what’s needed for clarity. Larger changes can be presented as a logical walkthrough of your approach, grouping related steps, explaining rationale where it adds value, and highlighting next actions to accelerate the user. Your answers should provide the right level of detail while being easily scannable.
+
+For casual greetings, acknowledgements, or other one-off conversational messages that are not delivering substantive information or structured results, respond naturally without section headers or bullet formatting.
+
+# Tool Guidelines
+
+## Shell commands
+
+When using the shell, you must adhere to the following guidelines:
+
+- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
+- Do not use python scripts to attempt to output larger chunks of a file. Command line output will be truncated after 10 kilobytes, regardless of the command used.
+- Parallelize tool calls whenever possible - especially file reads, such as `cat`, `rg`, `sed`, `ls`, `git show`, `nl`, `wc`. Use `multi_tool_use.parallel` to parallelize tool calls and only this.
+
+## apply_patch
+
+Use the `apply_patch` tool to edit files. Your patch language is a stripped‑down, file‑oriented diff format designed to be easy to parse and safe to apply. You can think of it as a high‑level envelope:
+
+*** Begin Patch
+[ one or more file sections ]
+*** End Patch
+
+Within that envelope, you get a sequence of file operations.
+You MUST include a header to specify the action you are taking.
+Each operation starts with one of three headers:
+
+*** Add File: <path> - create a new file. Every following line is a + line (the initial contents).
+*** Delete File: <path> - remove an existing file. Nothing follows.
+*** Update File: <path> - patch an existing file in place (optionally with a rename).
+
+Example patch:
+
+```
+*** Begin Patch
+*** Add File: hello.txt
++Hello world
+*** Update File: src/app.py
+*** Move to: src/main.py
+@@ def greet():
+-print("Hi")
++print("Hello, world!")
+*** Delete File: obsolete.txt
+*** End Patch
+```
+
+It is important to remember:
+
+- You must include a header with your intended action (Add/Delete/Update)
+- You must prefix new lines with `+` even when creating a new file
+
+## `update_plan`
+
+A tool named `update_plan` is available to you. You can use it to keep an up‑to‑date, step‑by‑step plan for the task.
+
+To create a new plan, call `update_plan` with a short list of 1‑sentence steps (no more than 5-7 words each) with a `status` for each step (`pending`, `in_progress`, or `completed`).
+
+When steps have been completed, use `update_plan` to mark each finished step as `completed` and the next step you are working on as `in_progress`. There should always be exactly one `in_progress` step until everything is done. You can mark multiple items as complete in a single `update_plan` call.
+
+If all steps are complete, ensure you call `update_plan` to mark all steps as `completed`.

internal/registry/model_definitions.go

@@ -16,6 +16,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.5 Haiku",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			// Thinking: not supported for Haiku models
 		},
 		{
 			ID:                  "claude-sonnet-4-5-20250929",
@@ -26,60 +27,6 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.5 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-		},
-		{
-			ID:                  "claude-sonnet-4-5-thinking",
-			Object:              "model",
-			Created:             1759104000, // 2025-09-29
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Sonnet Thinking",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                  "claude-opus-4-5-thinking",
-			Object:              "model",
-			Created:             1761955200, // 2025-11-01
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Opus Thinking",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                  "claude-opus-4-5-thinking-low",
-			Object:              "model",
-			Created:             1761955200, // 2025-11-01
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Opus Thinking Low",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                  "claude-opus-4-5-thinking-medium",
-			Object:              "model",
-			Created:             1761955200, // 2025-11-01
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Opus Thinking Medium",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                  "claude-opus-4-5-thinking-high",
-			Object:              "model",
-			Created:             1761955200, // 2025-11-01
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Opus Thinking High",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
@@ -92,6 +39,7 @@ func GetClaudeModels() []*ModelInfo {
 			Description:         "Premium model combining maximum intelligence with practical performance",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                  "claude-opus-4-1-20250805",
@@ -102,6 +50,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.1 Opus",
 			ContextLength:       200000,
 			MaxCompletionTokens: 32000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                  "claude-opus-4-20250514",
@@ -112,6 +61,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4 Opus",
 			ContextLength:       200000,
 			MaxCompletionTokens: 32000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                  "claude-sonnet-4-20250514",
@@ -122,6 +72,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                  "claude-3-7-sonnet-20250219",
@@ -132,6 +83,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 3.7 Sonnet",
 			ContextLength:       128000,
 			MaxCompletionTokens: 8192,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                  "claude-3-5-haiku-20241022",
@@ -142,6 +94,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 3.5 Haiku",
 			ContextLength:       128000,
 			MaxCompletionTokens: 8192,
+			// Thinking: not supported for Haiku models
 		},
 	}
 }
@@ -529,58 +482,7 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-minimal",
-			Object:              "model",
-			Created:             1754524800,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-08-07",
-			DisplayName:         "GPT 5 Minimal",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-low",
-			Object:              "model",
-			Created:             1754524800,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-08-07",
-			DisplayName:         "GPT 5 Low",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-medium",
-			Object:              "model",
-			Created:             1754524800,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-08-07",
-			DisplayName:         "GPT 5 Medium",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-high",
-			Object:              "model",
-			Created:             1754524800,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-08-07",
-			DisplayName:         "GPT 5 High",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"minimal", "low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5-codex",
@@ -594,45 +496,7 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-codex-low",
-			Object:              "model",
-			Created:             1757894400,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-09-15",
-			DisplayName:         "GPT 5 Codex Low",
-			Description:         "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-codex-medium",
-			Object:              "model",
-			Created:             1757894400,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-09-15",
-			DisplayName:         "GPT 5 Codex Medium",
-			Description:         "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-codex-high",
-			Object:              "model",
-			Created:             1757894400,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-09-15",
-			DisplayName:         "GPT 5 Codex High",
-			Description:         "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5-codex-mini",
@@ -646,32 +510,7 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-codex-mini-medium",
-			Object:              "model",
-			Created:             1762473600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-11-07",
-			DisplayName:         "GPT 5 Codex Mini Medium",
-			Description:         "Stable version of GPT 5 Codex Mini: cheaper, faster, but less capable version of GPT 5 Codex.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5-codex-mini-high",
-			Object:              "model",
-			Created:             1762473600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-11-07",
-			DisplayName:         "GPT 5 Codex Mini High",
-			Description:         "Stable version of GPT 5 Codex Mini: cheaper, faster, but less capable version of GPT 5 Codex.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1",
@@ -685,58 +524,7 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-none",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Nothink",
-			Description:         "Stable version of GPT 5.1, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-low",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5 Low",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-medium",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Medium",
-			Description:         "Stable version of GPT 5.1, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-high",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 High",
-			Description:         "Stable version of GPT 5.1, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1-codex",
@@ -750,45 +538,7 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-low",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex Low",
-			Description:         "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-medium",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex Medium",
-			Description:         "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-high",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex High",
-			Description:         "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1-codex-mini",
@@ -802,34 +552,8 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
-		{
-			ID:                  "gpt-5.1-codex-mini-medium",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex Mini Medium",
-			Description:         "Stable version of GPT 5.1 Codex Mini: cheaper, faster, but less capable version of GPT 5.1 Codex.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-mini-high",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex Mini High",
-			Description:         "Stable version of GPT 5.1 Codex Mini: cheaper, faster, but less capable version of GPT 5.1 Codex.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-
 		{
 			ID:                  "gpt-5.1-codex-max",
 			Object:              "model",
@@ -842,58 +566,21 @@ func GetOpenAIModels() []*ModelInfo {
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
 		},
 		{
-			ID:                  "gpt-5.1-codex-max-low",
+			ID:                  "gpt-5.2",
 			Object:              "model",
-			Created:             1763424000,
+			Created:             1765440000,
 			OwnedBy:             "openai",
 			Type:                "openai",
-			Version:             "gpt-5.1-max",
-			DisplayName:         "GPT 5.1 Codex Max Low",
-			Description:         "Stable version of GPT 5.1 Codex Max Low",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-max-medium",
-			Object:              "model",
-			Created:             1763424000,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-max",
-			DisplayName:         "GPT 5.1 Codex Max Medium",
-			Description:         "Stable version of GPT 5.1 Codex Max Medium",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-max-high",
-			Object:              "model",
-			Created:             1763424000,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-max",
-			DisplayName:         "GPT 5.1 Codex Max High",
-			Description:         "Stable version of GPT 5.1 Codex Max High",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-		},
-		{
-			ID:                  "gpt-5.1-codex-max-xhigh",
-			Object:              "model",
-			Created:             1763424000,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-max",
-			DisplayName:         "GPT 5.1 Codex Max XHigh",
-			Description:         "Stable version of GPT 5.1 Codex Max XHigh",
+			Version:             "gpt-5.2",
+			DisplayName:         "GPT 5.2",
+			Description:         "Stable version of GPT 5.2",
 			ContextLength:       400000,
 			MaxCompletionTokens: 128000,
 			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
 		},
 	}
 }
@@ -950,6 +637,7 @@ func GetIFlowModels() []*ModelInfo {
 		DisplayName string
 		Description string
 		Created     int64
+		Thinking    *ThinkingSupport
 	}{
 		{ID: "tstars2.0", DisplayName: "TStars-2.0", Description: "iFlow TStars-2.0 multimodal assistant", Created: 1746489600},
 		{ID: "qwen3-coder-plus", DisplayName: "Qwen3-Coder-Plus", Description: "Qwen3 Coder Plus code generation", Created: 1753228800},
@@ -959,17 +647,17 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "kimi-k2-0905", DisplayName: "Kimi-K2-Instruct-0905", Description: "Moonshot Kimi K2 instruct 0905", Created: 1757030400},
 		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400},
 		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
-		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 general model", Created: 1762387200},
+		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
 		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2", Created: 1764576000},
 		{ID: "deepseek-v3.2", DisplayName: "DeepSeek-V3.2-Exp", Description: "DeepSeek V3.2 experimental", Created: 1759104000},
 		{ID: "deepseek-v3.1", DisplayName: "DeepSeek-V3.1-Terminus", Description: "DeepSeek V3.1 Terminus", Created: 1756339200},
-		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200},
+		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
 		{ID: "deepseek-v3", DisplayName: "DeepSeek-V3-671B", Description: "DeepSeek V3 671B", Created: 1734307200},
 		{ID: "qwen3-32b", DisplayName: "Qwen3-32B", Description: "Qwen3 32B", Created: 1747094400},
-		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600},
+		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
 		{ID: "qwen3-235b-a22b-instruct", DisplayName: "Qwen3-235B-A22B-Instruct", Description: "Qwen3 235B A22B Instruct", Created: 1753401600},
 		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
-		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000},
+		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {
@@ -981,6 +669,7 @@ func GetIFlowModels() []*ModelInfo {
 			Type:        "iflow",
 			DisplayName: entry.DisplayName,
 			Description: entry.Description,
+			Thinking:    entry.Thinking,
 		})
 	}
 	return models

internal/registry/model_registry.go

@@ -63,6 +63,9 @@ type ThinkingSupport struct {
 	ZeroAllowed bool `json:"zero_allowed,omitempty"`
 	// DynamicAllowed indicates whether -1 is a valid value (dynamic thinking budget).
 	DynamicAllowed bool `json:"dynamic_allowed,omitempty"`
+	// Levels defines discrete reasoning effort levels (e.g., "low", "medium", "high").
+	// When set, the model uses level-based reasoning instead of token budgets.
+	Levels []string `json:"levels,omitempty"`
 }
 
 // ModelRegistration tracks a model's availability

internal/runtime/executor/aistudio_executor.go

@@ -1,3 +1,6 @@
+// Package executor provides runtime execution capabilities for various AI service providers.
+// This file implements the AI Studio executor that routes requests through a websocket-backed
+// transport for the AI Studio provider.
 package executor
 
 import (
@@ -26,19 +29,28 @@ type AIStudioExecutor struct {
 	cfg      *config.Config
 }
 
-// NewAIStudioExecutor constructs a websocket executor for the provider name.
+// NewAIStudioExecutor creates a new AI Studio executor instance.
+//
+// Parameters:
+//   - cfg: The application configuration
+//   - provider: The provider name
+//   - relay: The websocket relay manager
+//
+// Returns:
+//   - *AIStudioExecutor: A new AI Studio executor instance
 func NewAIStudioExecutor(cfg *config.Config, provider string, relay *wsrelay.Manager) *AIStudioExecutor {
 	return &AIStudioExecutor{provider: strings.ToLower(provider), relay: relay, cfg: cfg}
 }
 
-// Identifier returns the logical provider key for routing.
+// Identifier returns the executor identifier.
 func (e *AIStudioExecutor) Identifier() string { return "aistudio" }
 
-// PrepareRequest is a no-op because websocket transport already injects headers.
+// PrepareRequest prepares the HTTP request for execution (no-op for AI Studio).
 func (e *AIStudioExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error {
 	return nil
 }
 
+// Execute performs a non-streaming request to the AI Studio API.
 func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
@@ -92,6 +104,7 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 	return resp, nil
 }
 
+// ExecuteStream performs a streaming request to the AI Studio API.
 func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
@@ -239,6 +252,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 	return stream, nil
 }
 
+// CountTokens counts tokens for the given request using the AI Studio API.
 func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
 	_, body, err := e.translateRequest(req, opts, false)
 	if err != nil {
@@ -293,8 +307,8 @@ func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.A
 	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }
 
-func (e *AIStudioExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
-	_ = ctx
+// Refresh refreshes the authentication credentials (no-op for AI Studio).
+func (e *AIStudioExecutor) Refresh(_ context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
 	return auth, nil
 }
 

internal/runtime/executor/antigravity_executor.go

@@ -1,3 +1,6 @@
+// Package executor provides runtime execution capabilities for various AI service providers.
+// This file implements the Antigravity executor that proxies requests to the antigravity
+// upstream using OAuth credentials.
 package executor
 
 import (
@@ -29,16 +32,15 @@ import (
 const (
 	antigravityBaseURLDaily = "https://daily-cloudcode-pa.sandbox.googleapis.com"
 	// antigravityBaseURLAutopush     = "https://autopush-cloudcode-pa.sandbox.googleapis.com"
-	antigravityBaseURLProd      = "https://cloudcode-pa.googleapis.com"
-	antigravityStreamPath       = "/v1internal:streamGenerateContent"
-	antigravityGeneratePath     = "/v1internal:generateContent"
-	antigravityModelsPath       = "/v1internal:fetchAvailableModels"
-	antigravityClientID         = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
-	antigravityClientSecret     = "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"
-	defaultAntigravityAgent     = "antigravity/1.11.5 windows/amd64"
-	antigravityAuthType         = "antigravity"
-	refreshSkew                 = 3000 * time.Second
-	streamScannerBuffer     int = 20_971_520
+	antigravityBaseURLProd  = "https://cloudcode-pa.googleapis.com"
+	antigravityStreamPath   = "/v1internal:streamGenerateContent"
+	antigravityGeneratePath = "/v1internal:generateContent"
+	antigravityModelsPath   = "/v1internal:fetchAvailableModels"
+	antigravityClientID     = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
+	antigravityClientSecret = "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"
+	defaultAntigravityAgent = "antigravity/1.11.5 windows/amd64"
+	antigravityAuthType     = "antigravity"
+	refreshSkew             = 3000 * time.Second
 )
 
 var randSource = rand.New(rand.NewSource(time.Now().UnixNano()))
@@ -48,18 +50,24 @@ type AntigravityExecutor struct {
 	cfg *config.Config
 }
 
-// NewAntigravityExecutor constructs a new executor instance.
+// NewAntigravityExecutor creates a new Antigravity executor instance.
+//
+// Parameters:
+//   - cfg: The application configuration
+//
+// Returns:
+//   - *AntigravityExecutor: A new Antigravity executor instance
 func NewAntigravityExecutor(cfg *config.Config) *AntigravityExecutor {
 	return &AntigravityExecutor{cfg: cfg}
 }
 
-// Identifier implements ProviderExecutor.
+// Identifier returns the executor identifier.
 func (e *AntigravityExecutor) Identifier() string { return antigravityAuthType }
 
-// PrepareRequest implements ProviderExecutor.
+// PrepareRequest prepares the HTTP request for execution (no-op for Antigravity).
 func (e *AntigravityExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
 
-// Execute handles non-streaming requests via the antigravity generate endpoint.
+// Execute performs a non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
@@ -152,7 +160,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	return resp, err
 }
 
-// ExecuteStream handles streaming requests via the antigravity upstream.
+// ExecuteStream performs a streaming request to the Antigravity API.
 func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
 	ctx = context.WithValue(ctx, "alt", "")
 
@@ -292,7 +300,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	return nil, err
 }
 
-// Refresh refreshes the OAuth token using the refresh token.
+// Refresh refreshes the authentication credentials using the refresh token.
 func (e *AntigravityExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
 	if auth == nil {
 		return auth, nil
@@ -304,7 +312,7 @@ func (e *AntigravityExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Au
 	return updated, nil
 }
 
-// CountTokens is not supported for the antigravity provider.
+// CountTokens counts tokens for the given request (not supported for Antigravity).
 func (e *AntigravityExecutor) CountTokens(context.Context, *cliproxyauth.Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
 	return cliproxyexecutor.Response{}, statusErr{code: http.StatusNotImplemented, msg: "count tokens not supported"}
 }

internal/runtime/executor/claude_executor.go

@@ -54,15 +54,22 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	modelForUpstream := req.Model
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		body, _ = sjson.SetBytes(body, "model", modelOverride)
-		modelForUpstream = modelOverride
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel == "" {
+		upstreamModel = req.Model
 	}
-	// Inject thinking config based on model suffix for thinking variants
-	body = e.injectThinkingConfig(req.Model, body)
+	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
+		upstreamModel = modelOverride
+	} else if !strings.EqualFold(upstreamModel, req.Model) {
+		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+			upstreamModel = modelOverride
+		}
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	// Inject thinking config based on model metadata for thinking variants
+	body = e.injectThinkingConfig(req.Model, req.Metadata, body)
 
-	if !strings.HasPrefix(modelForUpstream, "claude-3-5-haiku") {
+	if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -161,11 +168,20 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		body, _ = sjson.SetBytes(body, "model", modelOverride)
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel == "" {
+		upstreamModel = req.Model
 	}
-	// Inject thinking config based on model suffix for thinking variants
-	body = e.injectThinkingConfig(req.Model, body)
+	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
+		upstreamModel = modelOverride
+	} else if !strings.EqualFold(upstreamModel, req.Model) {
+		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+			upstreamModel = modelOverride
+		}
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	// Inject thinking config based on model metadata for thinking variants
+	body = e.injectThinkingConfig(req.Model, req.Metadata, body)
 	body = checkSystemInstructions(body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 
@@ -238,7 +254,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		// If from == to (Claude → Claude), directly forward the SSE stream without translation
 		if from == to {
 			scanner := bufio.NewScanner(decodedBody)
-			scanner.Buffer(nil, 20_971_520)
+			scanner.Buffer(nil, 52_428_800) // 50MB
 			for scanner.Scan() {
 				line := scanner.Bytes()
 				appendAPIResponseChunk(ctx, e.cfg, line)
@@ -261,7 +277,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 
 		// For other formats, use translation
 		scanner := bufio.NewScanner(decodedBody)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, 52_428_800) // 50MB
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()
@@ -295,13 +311,20 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	modelForUpstream := req.Model
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		body, _ = sjson.SetBytes(body, "model", modelOverride)
-		modelForUpstream = modelOverride
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel == "" {
+		upstreamModel = req.Model
 	}
+	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
+		upstreamModel = modelOverride
+	} else if !strings.EqualFold(upstreamModel, req.Model) {
+		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+			upstreamModel = modelOverride
+		}
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
-	if !strings.HasPrefix(modelForUpstream, "claude-3-5-haiku") {
+	if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
 
@@ -427,31 +450,15 @@ func extractAndRemoveBetas(body []byte) ([]string, []byte) {
 	return betas, body
 }
 
-// injectThinkingConfig adds thinking configuration based on model name suffix
-func (e *ClaudeExecutor) injectThinkingConfig(modelName string, body []byte) []byte {
-	// Only inject if thinking config is not already present
-	if gjson.GetBytes(body, "thinking").Exists() {
+// injectThinkingConfig adds thinking configuration based on metadata using the unified flow.
+// It uses util.ResolveClaudeThinkingConfig which internally calls ResolveThinkingConfigFromMetadata
+// and NormalizeThinkingBudget, ensuring consistency with other executors like Gemini.
+func (e *ClaudeExecutor) injectThinkingConfig(modelName string, metadata map[string]any, body []byte) []byte {
+	budget, ok := util.ResolveClaudeThinkingConfig(modelName, metadata)
+	if !ok {
 		return body
 	}
-
-	var budgetTokens int
-	switch {
-	case strings.HasSuffix(modelName, "-thinking-low"):
-		budgetTokens = 1024
-	case strings.HasSuffix(modelName, "-thinking-medium"):
-		budgetTokens = 8192
-	case strings.HasSuffix(modelName, "-thinking-high"):
-		budgetTokens = 24576
-	case strings.HasSuffix(modelName, "-thinking"):
-		// Default thinking without suffix uses medium budget
-		budgetTokens = 8192
-	default:
-		return body
-	}
-
-	body, _ = sjson.SetBytes(body, "thinking.type", "enabled")
-	body, _ = sjson.SetBytes(body, "thinking.budget_tokens", budgetTokens)
-	return body
+	return util.ApplyClaudeThinkingConfig(body, budget)
 }
 
 // ensureMaxTokensForThinking ensures max_tokens > thinking.budget_tokens when thinking is enabled.
@@ -491,35 +498,45 @@ func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
 }
 
 func (e *ClaudeExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	if alias == "" {
+	trimmed := strings.TrimSpace(alias)
+	if trimmed == "" {
 		return ""
 	}
-	// Hardcoded mappings for thinking models to actual Claude model names
-	switch alias {
-	case "claude-opus-4-5-thinking", "claude-opus-4-5-thinking-low", "claude-opus-4-5-thinking-medium", "claude-opus-4-5-thinking-high":
-		return "claude-opus-4-5-20251101"
-	case "claude-sonnet-4-5-thinking":
-		return "claude-sonnet-4-5-20250929"
-	}
+
 	entry := e.resolveClaudeConfig(auth)
 	if entry == nil {
 		return ""
 	}
+
+	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
+
+	// Candidate names to match against configured aliases/names.
+	candidates := []string{strings.TrimSpace(normalizedModel)}
+	if !strings.EqualFold(normalizedModel, trimmed) {
+		candidates = append(candidates, trimmed)
+	}
+	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
+		candidates = append(candidates, original)
+	}
+
 	for i := range entry.Models {
 		model := entry.Models[i]
 		name := strings.TrimSpace(model.Name)
 		modelAlias := strings.TrimSpace(model.Alias)
-		if modelAlias != "" {
-			if strings.EqualFold(modelAlias, alias) {
+
+		for _, candidate := range candidates {
+			if candidate == "" {
+				continue
+			}
+			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
 				if name != "" {
 					return name
 				}
-				return alias
+				return candidate
+			}
+			if name != "" && strings.EqualFold(name, candidate) {
+				return name
 			}
-			continue
-		}
-		if name != "" && strings.EqualFold(name, alias) {
-			return name
 		}
 	}
 	return ""

internal/runtime/executor/codex_executor.go

@@ -49,14 +49,18 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-
-	body = e.setReasoningEffortByAlias(req.Model, body)
-
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
+	body = normalizeThinkingConfig(body, upstreamModel)
+	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+		return resp, errValidate
+	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
-
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.SetBytes(body, "stream", true)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 
@@ -142,13 +146,20 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	body = e.setReasoningEffortByAlias(req.Model, body)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
+	body = normalizeThinkingConfig(body, upstreamModel)
+	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+		return nil, errValidate
+	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/responses"
 	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -205,7 +216,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			}
 		}()
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, 52_428_800) // 50MB
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()
@@ -235,14 +246,16 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }
 
 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
 
 	modelForCounting := req.Model
 
-	body = e.setReasoningEffortByAlias(req.Model, body)
-
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 	body, _ = sjson.SetBytes(body, "stream", false)
 
@@ -261,83 +274,6 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }
 
-func (e *CodexExecutor) setReasoningEffortByAlias(modelName string, payload []byte) []byte {
-	if util.InArray([]string{"gpt-5", "gpt-5-minimal", "gpt-5-low", "gpt-5-medium", "gpt-5-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5")
-		switch modelName {
-		case "gpt-5-minimal":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "minimal")
-		case "gpt-5-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5-codex", "gpt-5-codex-low", "gpt-5-codex-medium", "gpt-5-codex-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5-codex")
-		switch modelName {
-		case "gpt-5-codex-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5-codex-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5-codex-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5-codex-mini", "gpt-5-codex-mini-medium", "gpt-5-codex-mini-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5-codex-mini")
-		switch modelName {
-		case "gpt-5-codex-mini-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5-codex-mini-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1", "gpt-5.1-none", "gpt-5.1-low", "gpt-5.1-medium", "gpt-5.1-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1")
-		switch modelName {
-		case "gpt-5.1-none":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "none")
-		case "gpt-5.1-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5.1-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1-codex", "gpt-5.1-codex-low", "gpt-5.1-codex-medium", "gpt-5.1-codex-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex")
-		switch modelName {
-		case "gpt-5.1-codex-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5.1-codex-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-codex-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1-codex-mini", "gpt-5.1-codex-mini-medium", "gpt-5.1-codex-mini-high"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex-mini")
-		switch modelName {
-		case "gpt-5.1-codex-mini-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-codex-mini-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		}
-	} else if util.InArray([]string{"gpt-5.1-codex-max", "gpt-5.1-codex-max-low", "gpt-5.1-codex-max-medium", "gpt-5.1-codex-max-high", "gpt-5.1-codex-max-xhigh"}, modelName) {
-		payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex-max")
-		switch modelName {
-		case "gpt-5.1-codex-max-low":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
-		case "gpt-5.1-codex-max-medium":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
-		case "gpt-5.1-codex-max-high":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
-		case "gpt-5.1-codex-max-xhigh":
-			payload, _ = sjson.SetBytes(payload, "reasoning.effort", "xhigh")
-		}
-	}
-	return payload
-}
-
 func tokenizerForCodexModel(model string) (tokenizer.Codec, error) {
 	sanitized := strings.ToLower(strings.TrimSpace(model))
 	switch {

internal/runtime/executor/gemini_cli_executor.go

@@ -1,3 +1,6 @@
+// Package executor provides runtime execution capabilities for various AI service providers.
+// This file implements the Gemini CLI executor that talks to Cloud Code Assist endpoints
+// using OAuth credentials from auth metadata.
 package executor
 
 import (
@@ -29,11 +32,11 @@ import (
 const (
 	codeAssistEndpoint      = "https://cloudcode-pa.googleapis.com"
 	codeAssistVersion       = "v1internal"
-	geminiOauthClientID     = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
-	geminiOauthClientSecret = "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"
+	geminiOAuthClientID     = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
+	geminiOAuthClientSecret = "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"
 )
 
-var geminiOauthScopes = []string{
+var geminiOAuthScopes = []string{
 	"https://www.googleapis.com/auth/cloud-platform",
 	"https://www.googleapis.com/auth/userinfo.email",
 	"https://www.googleapis.com/auth/userinfo.profile",
@@ -44,14 +47,24 @@ type GeminiCLIExecutor struct {
 	cfg *config.Config
 }
 
+// NewGeminiCLIExecutor creates a new Gemini CLI executor instance.
+//
+// Parameters:
+//   - cfg: The application configuration
+//
+// Returns:
+//   - *GeminiCLIExecutor: A new Gemini CLI executor instance
 func NewGeminiCLIExecutor(cfg *config.Config) *GeminiCLIExecutor {
 	return &GeminiCLIExecutor{cfg: cfg}
 }
 
+// Identifier returns the executor identifier.
 func (e *GeminiCLIExecutor) Identifier() string { return "gemini-cli" }
 
+// PrepareRequest prepares the HTTP request for execution (no-op for Gemini CLI).
 func (e *GeminiCLIExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
 
+// Execute performs a non-streaming request to the Gemini CLI API.
 func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
@@ -189,6 +202,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	return resp, err
 }
 
+// ExecuteStream performs a streaming request to the Gemini CLI API.
 func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
@@ -309,7 +323,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			}()
 			if opts.Alt == "" {
 				scanner := bufio.NewScanner(resp.Body)
-				scanner.Buffer(nil, 20_971_520)
+				scanner.Buffer(nil, streamScannerBuffer)
 				var param any
 				for scanner.Scan() {
 					line := scanner.Bytes()
@@ -371,6 +385,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	return nil, err
 }
 
+// CountTokens counts tokens for the given request using the Gemini CLI API.
 func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
@@ -471,9 +486,8 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	return cliproxyexecutor.Response{}, newGeminiStatusErr(lastStatus, lastBody)
 }
 
-func (e *GeminiCLIExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
-	log.Debugf("gemini cli executor: refresh called")
-	_ = ctx
+// Refresh refreshes the authentication credentials (no-op for Gemini CLI).
+func (e *GeminiCLIExecutor) Refresh(_ context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
 	return auth, nil
 }
 
@@ -515,9 +529,9 @@ func prepareGeminiCLITokenSource(ctx context.Context, cfg *config.Config, auth *
 	}
 
 	conf := &oauth2.Config{
-		ClientID:     geminiOauthClientID,
-		ClientSecret: geminiOauthClientSecret,
-		Scopes:       geminiOauthScopes,
+		ClientID:     geminiOAuthClientID,
+		ClientSecret: geminiOAuthClientSecret,
+		Scopes:       geminiOAuthScopes,
 		Endpoint:     google.Endpoint,
 	}
 

internal/runtime/executor/gemini_executor.go

@@ -11,7 +11,6 @@ import (
 	"io"
 	"net/http"
 	"strings"
-	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
@@ -21,8 +20,6 @@ import (
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
-	"golang.org/x/oauth2"
-	"golang.org/x/oauth2/google"
 )
 
 const (
@@ -31,6 +28,9 @@ const (
 
 	// glAPIVersion is the API version used for Gemini requests.
 	glAPIVersion = "v1beta"
+
+	// streamScannerBuffer is the buffer size for SSE stream scanning.
+	streamScannerBuffer = 52_428_800
 )
 
 // GeminiExecutor is a stateless executor for the official Gemini API using API keys.
@@ -48,9 +48,11 @@ type GeminiExecutor struct {
 //
 // Returns:
 //   - *GeminiExecutor: A new Gemini executor instance
-func NewGeminiExecutor(cfg *config.Config) *GeminiExecutor { return &GeminiExecutor{cfg: cfg} }
+func NewGeminiExecutor(cfg *config.Config) *GeminiExecutor {
+	return &GeminiExecutor{cfg: cfg}
+}
 
-// Identifier returns the executor identifier for Gemini.
+// Identifier returns the executor identifier.
 func (e *GeminiExecutor) Identifier() string { return "gemini" }
 
 // PrepareRequest prepares the HTTP request for execution (no-op for Gemini).
@@ -75,6 +77,8 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	// Official Gemini API via API key or OAuth bearer
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
@@ -85,6 +89,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -93,7 +98,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		}
 	}
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, action)
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -161,12 +166,15 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	return resp, nil
 }
 
+// ExecuteStream performs a streaming request to the Gemini API.
 func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
 	apiKey, bearer := geminiCreds(auth)
 
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
@@ -176,9 +184,10 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -243,7 +252,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			}
 		}()
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, streamScannerBuffer)
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()
@@ -274,6 +283,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	return stream, nil
 }
 
+// CountTokens counts tokens for the given request using the Gemini API.
 func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
 	apiKey, bearer := geminiCreds(auth)
 
@@ -347,106 +357,8 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }
 
-func (e *GeminiExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
-	log.Debugf("gemini executor: refresh called")
-	// OAuth bearer token refresh for official Gemini API.
-	if auth == nil {
-		return nil, fmt.Errorf("gemini executor: auth is nil")
-	}
-	if auth.Metadata == nil {
-		return auth, nil
-	}
-	// Token data is typically nested under "token" map in Gemini files.
-	tokenMap, _ := auth.Metadata["token"].(map[string]any)
-	var refreshToken, accessToken, clientID, clientSecret, tokenURI, expiryStr string
-	if tokenMap != nil {
-		if v, ok := tokenMap["refresh_token"].(string); ok {
-			refreshToken = v
-		}
-		if v, ok := tokenMap["access_token"].(string); ok {
-			accessToken = v
-		}
-		if v, ok := tokenMap["client_id"].(string); ok {
-			clientID = v
-		}
-		if v, ok := tokenMap["client_secret"].(string); ok {
-			clientSecret = v
-		}
-		if v, ok := tokenMap["token_uri"].(string); ok {
-			tokenURI = v
-		}
-		if v, ok := tokenMap["expiry"].(string); ok {
-			expiryStr = v
-		}
-	} else {
-		// Fallback to top-level keys if present
-		if v, ok := auth.Metadata["refresh_token"].(string); ok {
-			refreshToken = v
-		}
-		if v, ok := auth.Metadata["access_token"].(string); ok {
-			accessToken = v
-		}
-		if v, ok := auth.Metadata["client_id"].(string); ok {
-			clientID = v
-		}
-		if v, ok := auth.Metadata["client_secret"].(string); ok {
-			clientSecret = v
-		}
-		if v, ok := auth.Metadata["token_uri"].(string); ok {
-			tokenURI = v
-		}
-		if v, ok := auth.Metadata["expiry"].(string); ok {
-			expiryStr = v
-		}
-	}
-	if refreshToken == "" {
-		// Nothing to do for API key or cookie based entries
-		return auth, nil
-	}
-
-	// Prepare oauth2 config; default to Google endpoints
-	endpoint := google.Endpoint
-	if tokenURI != "" {
-		endpoint.TokenURL = tokenURI
-	}
-	conf := &oauth2.Config{ClientID: clientID, ClientSecret: clientSecret, Endpoint: endpoint}
-
-	// Ensure proxy-aware HTTP client for token refresh
-	httpClient := util.SetProxy(&e.cfg.SDKConfig, &http.Client{})
-	ctx = context.WithValue(ctx, oauth2.HTTPClient, httpClient)
-
-	// Build base token
-	tok := &oauth2.Token{AccessToken: accessToken, RefreshToken: refreshToken}
-	if t, err := time.Parse(time.RFC3339, expiryStr); err == nil {
-		tok.Expiry = t
-	}
-	newTok, err := conf.TokenSource(ctx, tok).Token()
-	if err != nil {
-		return nil, err
-	}
-
-	// Persist back to metadata; prefer nested token map if present
-	if tokenMap == nil {
-		tokenMap = make(map[string]any)
-	}
-	tokenMap["access_token"] = newTok.AccessToken
-	tokenMap["refresh_token"] = newTok.RefreshToken
-	tokenMap["expiry"] = newTok.Expiry.Format(time.RFC3339)
-	if clientID != "" {
-		tokenMap["client_id"] = clientID
-	}
-	if clientSecret != "" {
-		tokenMap["client_secret"] = clientSecret
-	}
-	if tokenURI != "" {
-		tokenMap["token_uri"] = tokenURI
-	}
-	auth.Metadata["token"] = tokenMap
-
-	// Also mirror top-level access_token for compatibility if previously present
-	if _, ok := auth.Metadata["access_token"]; ok {
-		auth.Metadata["access_token"] = newTok.AccessToken
-	}
+// Refresh refreshes the authentication credentials (no-op for Gemini API key).
+func (e *GeminiExecutor) Refresh(_ context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
 	return auth, nil
 }
 

internal/runtime/executor/gemini_vertex_executor.go

@@ -1,6 +1,6 @@
-// Package executor contains provider executors. This file implements the Vertex AI
-// Gemini executor that talks to Google Vertex AI endpoints using service account
-// credentials imported by the CLI.
+// Package executor provides runtime execution capabilities for various AI service providers.
+// This file implements the Vertex AI Gemini executor that talks to Google Vertex AI
+// endpoints using service account credentials or API keys.
 package executor
 
 import (
@@ -36,20 +36,26 @@ type GeminiVertexExecutor struct {
 	cfg *config.Config
 }
 
-// NewGeminiVertexExecutor constructs the Vertex executor.
+// NewGeminiVertexExecutor creates a new Vertex AI Gemini executor instance.
+//
+// Parameters:
+//   - cfg: The application configuration
+//
+// Returns:
+//   - *GeminiVertexExecutor: A new Vertex AI Gemini executor instance
 func NewGeminiVertexExecutor(cfg *config.Config) *GeminiVertexExecutor {
 	return &GeminiVertexExecutor{cfg: cfg}
 }
 
-// Identifier returns provider key for manager routing.
+// Identifier returns the executor identifier.
 func (e *GeminiVertexExecutor) Identifier() string { return "vertex" }
 
-// PrepareRequest is a no-op for Vertex.
+// PrepareRequest prepares the HTTP request for execution (no-op for Vertex).
 func (e *GeminiVertexExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error {
 	return nil
 }
 
-// Execute handles non-streaming requests.
+// Execute performs a non-streaming request to the Vertex AI API.
 func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	// Try API key authentication first
 	apiKey, baseURL := vertexAPICreds(auth)
@@ -67,7 +73,7 @@ func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	return e.executeWithAPIKey(ctx, auth, req, opts, apiKey, baseURL)
 }
 
-// ExecuteStream handles SSE streaming for Vertex.
+// ExecuteStream performs a streaming request to the Vertex AI API.
 func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
 	// Try API key authentication first
 	apiKey, baseURL := vertexAPICreds(auth)
@@ -85,7 +91,7 @@ func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	return e.executeStreamWithAPIKey(ctx, auth, req, opts, apiKey, baseURL)
 }
 
-// CountTokens calls Vertex countTokens endpoint.
+// CountTokens counts tokens for the given request using the Vertex AI API.
 func (e *GeminiVertexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
 	// Try API key authentication first
 	apiKey, baseURL := vertexAPICreds(auth)
@@ -103,179 +109,7 @@ func (e *GeminiVertexExecutor) CountTokens(ctx context.Context, auth *cliproxyau
 	return e.countTokensWithAPIKey(ctx, auth, req, opts, apiKey, baseURL)
 }
 
-// countTokensWithServiceAccount handles token counting using service account credentials.
-func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (cliproxyexecutor.Response, error) {
-	from := opts.SourceFormat
-	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
-	}
-	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
-	respCtx := context.WithValue(ctx, "alt", opts.Alt)
-	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
-	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
-	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
-
-	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "countTokens")
-
-	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
-	if errNewReq != nil {
-		return cliproxyexecutor.Response{}, errNewReq
-	}
-	httpReq.Header.Set("Content-Type", "application/json")
-	if token, errTok := vertexAccessToken(ctx, e.cfg, auth, saJSON); errTok == nil && token != "" {
-		httpReq.Header.Set("Authorization", "Bearer "+token)
-	} else if errTok != nil {
-		log.Errorf("vertex executor: access token error: %v", errTok)
-		return cliproxyexecutor.Response{}, statusErr{code: 500, msg: "internal server error"}
-	}
-	applyGeminiHeaders(httpReq, auth)
-
-	var authID, authLabel, authType, authValue string
-	if auth != nil {
-		authID = auth.ID
-		authLabel = auth.Label
-		authType, authValue = auth.AccountInfo()
-	}
-	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
-		URL:       url,
-		Method:    http.MethodPost,
-		Headers:   httpReq.Header.Clone(),
-		Body:      translatedReq,
-		Provider:  e.Identifier(),
-		AuthID:    authID,
-		AuthLabel: authLabel,
-		AuthType:  authType,
-		AuthValue: authValue,
-	})
-
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
-	httpResp, errDo := httpClient.Do(httpReq)
-	if errDo != nil {
-		recordAPIResponseError(ctx, e.cfg, errDo)
-		return cliproxyexecutor.Response{}, errDo
-	}
-	defer func() {
-		if errClose := httpResp.Body.Close(); errClose != nil {
-			log.Errorf("vertex executor: close response body error: %v", errClose)
-		}
-	}()
-	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		b, _ := io.ReadAll(httpResp.Body)
-		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(b)}
-	}
-	data, errRead := io.ReadAll(httpResp.Body)
-	if errRead != nil {
-		recordAPIResponseError(ctx, e.cfg, errRead)
-		return cliproxyexecutor.Response{}, errRead
-	}
-	appendAPIResponseChunk(ctx, e.cfg, data)
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
-	}
-	count := gjson.GetBytes(data, "totalTokens").Int()
-	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
-}
-
-// countTokensWithAPIKey handles token counting using API key credentials.
-func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (cliproxyexecutor.Response, error) {
-	from := opts.SourceFormat
-	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
-	}
-	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
-	respCtx := context.WithValue(ctx, "alt", opts.Alt)
-	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
-	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
-	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
-
-	// For API key auth, use simpler URL format without project/location
-	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
-	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, "countTokens")
-
-	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
-	if errNewReq != nil {
-		return cliproxyexecutor.Response{}, errNewReq
-	}
-	httpReq.Header.Set("Content-Type", "application/json")
-	if apiKey != "" {
-		httpReq.Header.Set("x-goog-api-key", apiKey)
-	}
-	applyGeminiHeaders(httpReq, auth)
-
-	var authID, authLabel, authType, authValue string
-	if auth != nil {
-		authID = auth.ID
-		authLabel = auth.Label
-		authType, authValue = auth.AccountInfo()
-	}
-	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
-		URL:       url,
-		Method:    http.MethodPost,
-		Headers:   httpReq.Header.Clone(),
-		Body:      translatedReq,
-		Provider:  e.Identifier(),
-		AuthID:    authID,
-		AuthLabel: authLabel,
-		AuthType:  authType,
-		AuthValue: authValue,
-	})
-
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
-	httpResp, errDo := httpClient.Do(httpReq)
-	if errDo != nil {
-		recordAPIResponseError(ctx, e.cfg, errDo)
-		return cliproxyexecutor.Response{}, errDo
-	}
-	defer func() {
-		if errClose := httpResp.Body.Close(); errClose != nil {
-			log.Errorf("vertex executor: close response body error: %v", errClose)
-		}
-	}()
-	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		b, _ := io.ReadAll(httpResp.Body)
-		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(b)}
-	}
-	data, errRead := io.ReadAll(httpResp.Body)
-	if errRead != nil {
-		recordAPIResponseError(ctx, e.cfg, errRead)
-		return cliproxyexecutor.Response{}, errRead
-	}
-	appendAPIResponseChunk(ctx, e.cfg, data)
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
-	}
-	count := gjson.GetBytes(data, "totalTokens").Int()
-	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
-}
-
-// Refresh is a no-op for service account based credentials.
+// Refresh refreshes the authentication credentials (no-op for Vertex).
 func (e *GeminiVertexExecutor) Refresh(_ context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
 	return auth, nil
 }
@@ -286,10 +120,12 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -301,6 +137,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -309,7 +146,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 		}
 	}
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, action)
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -383,10 +220,12 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -398,6 +237,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -410,7 +250,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, action)
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -481,10 +321,12 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -496,9 +338,10 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -564,7 +407,7 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 			}
 		}()
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, streamScannerBuffer)
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()
@@ -595,10 +438,12 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)
 
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
 		if budgetOverride != nil {
 			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
 			budgetOverride = &norm
@@ -610,12 +455,13 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
 	body = fixGeminiImageAspectRatio(req.Model, body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -678,7 +524,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 			}
 		}()
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, streamScannerBuffer)
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()
@@ -704,6 +550,184 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	return stream, nil
 }
 
+// countTokensWithServiceAccount counts tokens using service account credentials.
+func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (cliproxyexecutor.Response, error) {
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("gemini")
+	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+		if budgetOverride != nil {
+			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
+			budgetOverride = &norm
+		}
+		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
+	}
+	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
+	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
+	respCtx := context.WithValue(ctx, "alt", opts.Alt)
+	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
+	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
+	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
+
+	baseURL := vertexBaseURL(location)
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "countTokens")
+
+	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
+	if errNewReq != nil {
+		return cliproxyexecutor.Response{}, errNewReq
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	if token, errTok := vertexAccessToken(ctx, e.cfg, auth, saJSON); errTok == nil && token != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+token)
+	} else if errTok != nil {
+		log.Errorf("vertex executor: access token error: %v", errTok)
+		return cliproxyexecutor.Response{}, statusErr{code: 500, msg: "internal server error"}
+	}
+	applyGeminiHeaders(httpReq, auth)
+
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      translatedReq,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, errDo := httpClient.Do(httpReq)
+	if errDo != nil {
+		recordAPIResponseError(ctx, e.cfg, errDo)
+		return cliproxyexecutor.Response{}, errDo
+	}
+	defer func() {
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("vertex executor: close response body error: %v", errClose)
+		}
+	}()
+	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		b, _ := io.ReadAll(httpResp.Body)
+		appendAPIResponseChunk(ctx, e.cfg, b)
+		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(b)}
+	}
+	data, errRead := io.ReadAll(httpResp.Body)
+	if errRead != nil {
+		recordAPIResponseError(ctx, e.cfg, errRead)
+		return cliproxyexecutor.Response{}, errRead
+	}
+	appendAPIResponseChunk(ctx, e.cfg, data)
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
+		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
+	}
+	count := gjson.GetBytes(data, "totalTokens").Int()
+	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
+	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+}
+
+// countTokensWithAPIKey handles token counting using API key credentials.
+func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (cliproxyexecutor.Response, error) {
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("gemini")
+	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
+		if budgetOverride != nil {
+			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
+			budgetOverride = &norm
+		}
+		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
+	}
+	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
+	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
+	respCtx := context.WithValue(ctx, "alt", opts.Alt)
+	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
+	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
+	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
+
+	// For API key auth, use simpler URL format without project/location
+	if baseURL == "" {
+		baseURL = "https://generativelanguage.googleapis.com"
+	}
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, "countTokens")
+
+	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
+	if errNewReq != nil {
+		return cliproxyexecutor.Response{}, errNewReq
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	if apiKey != "" {
+		httpReq.Header.Set("x-goog-api-key", apiKey)
+	}
+	applyGeminiHeaders(httpReq, auth)
+
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      translatedReq,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, errDo := httpClient.Do(httpReq)
+	if errDo != nil {
+		recordAPIResponseError(ctx, e.cfg, errDo)
+		return cliproxyexecutor.Response{}, errDo
+	}
+	defer func() {
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("vertex executor: close response body error: %v", errClose)
+		}
+	}()
+	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		b, _ := io.ReadAll(httpResp.Body)
+		appendAPIResponseChunk(ctx, e.cfg, b)
+		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(b)}
+	}
+	data, errRead := io.ReadAll(httpResp.Body)
+	if errRead != nil {
+		recordAPIResponseError(ctx, e.cfg, errRead)
+		return cliproxyexecutor.Response{}, errRead
+	}
+	appendAPIResponseChunk(ctx, e.cfg, data)
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
+		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
+	}
+	count := gjson.GetBytes(data, "totalTokens").Int()
+	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
+	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+}
+
 // vertexCreds extracts project, location and raw service account JSON from auth metadata.
 func vertexCreds(a *cliproxyauth.Auth) (projectID, location string, serviceAccountJSON []byte, err error) {
 	if a == nil || a.Metadata == nil {

internal/runtime/executor/iflow_executor.go

@@ -57,6 +57,15 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
+	body = normalizeThinkingConfig(body, upstreamModel)
+	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+		return resp, errValidate
+	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
@@ -139,6 +148,15 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
+	body = normalizeThinkingConfig(body, upstreamModel)
+	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+		return nil, errValidate
+	}
 	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.
 	toolsResult := gjson.GetBytes(body, "tools")
 	if toolsResult.Exists() && toolsResult.IsArray() && len(toolsResult.Array()) == 0 {
@@ -201,7 +219,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		}()
 
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, 52_428_800) // 50MB 
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()

internal/runtime/executor/logging_helpers.go

@@ -157,7 +157,7 @@ func appendAPIResponseChunk(ctx context.Context, cfg *config.Config, chunk []byt
 	if ginCtx == nil {
 		return
 	}
-	_, attempt := ensureAttempt(ginCtx)
+	attempts, attempt := ensureAttempt(ginCtx)
 	ensureResponseIntro(attempt)
 
 	if !attempt.headersWritten {
@@ -175,6 +175,8 @@ func appendAPIResponseChunk(ctx context.Context, cfg *config.Config, chunk []byt
 	}
 	attempt.response.WriteString(string(data))
 	attempt.bodyHasContent = true
+
+	updateAggregatedResponse(ginCtx, attempts)
 }
 
 func ginContextFrom(ctx context.Context) *gin.Context {

internal/runtime/executor/openai_compat_executor.go

@@ -58,6 +58,15 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel != "" {
+		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
+	}
+	translated = normalizeThinkingConfig(translated, upstreamModel)
+	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
+		return resp, errValidate
+	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
@@ -143,6 +152,15 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel != "" {
+		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
+	}
+	translated = normalizeThinkingConfig(translated, upstreamModel)
+	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
+		return nil, errValidate
+	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
@@ -206,7 +224,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 			}
 		}()
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, 52_428_800) // 50MB
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()

internal/runtime/executor/payload_helpers.go

@@ -1,6 +1,8 @@
 package executor
 
 import (
+	"fmt"
+	"net/http"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
@@ -9,11 +11,11 @@ import (
 	"github.com/tidwall/sjson"
 )
 
-// applyThinkingMetadata applies thinking config from model suffix metadata (e.g., -reasoning, -thinking-N)
+// applyThinkingMetadata applies thinking config from model suffix metadata (e.g., (high), (8192))
 // for standard Gemini format payloads. It normalizes the budget when the model supports thinking.
 func applyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(metadata)
-	if !ok {
+	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
+	if !ok || (budgetOverride == nil && includeOverride == nil) {
 		return payload
 	}
 	if !util.ModelSupportsThinking(model) {
@@ -26,20 +28,44 @@ func applyThinkingMetadata(payload []byte, metadata map[string]any, model string
 	return util.ApplyGeminiThinkingConfig(payload, budgetOverride, includeOverride)
 }
 
-// applyThinkingMetadataCLI applies thinking config from model suffix metadata (e.g., -reasoning, -thinking-N)
+// applyThinkingMetadataCLI applies thinking config from model suffix metadata (e.g., (high), (8192))
 // for Gemini CLI format payloads (nested under "request"). It normalizes the budget when the model supports thinking.
 func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(metadata)
-	if !ok {
+	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
+	if !ok || (budgetOverride == nil && includeOverride == nil) {
 		return payload
 	}
-	if budgetOverride != nil && util.ModelSupportsThinking(model) {
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if budgetOverride != nil {
 		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
 		budgetOverride = &norm
 	}
 	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
 }
 
+// applyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
+// Metadata values take precedence over any existing field when the model supports thinking, intentionally
+// overwriting caller-provided values to honor suffix/default metadata priority.
+func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string) []byte {
+	if len(metadata) == 0 {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if field == "" {
+		return payload
+	}
+	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
+		if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+			return updated
+		}
+	}
+	return payload
+}
+
 // applyPayloadConfig applies payload default and override rules from configuration
 // to the given JSON payload for the specified model.
 // Defaults only fill missing fields, while overrides always overwrite existing values.
@@ -189,3 +215,93 @@ func matchModelPattern(pattern, model string) bool {
 	}
 	return pi == len(pattern)
 }
+
+// normalizeThinkingConfig normalizes thinking-related fields in the payload
+// based on model capabilities. For models without thinking support, it strips
+// reasoning fields. For models with level-based thinking, it validates and
+// normalizes the reasoning effort level.
+func normalizeThinkingConfig(payload []byte, model string) []byte {
+	if len(payload) == 0 || model == "" {
+		return payload
+	}
+
+	if !util.ModelSupportsThinking(model) {
+		return stripThinkingFields(payload)
+	}
+
+	if util.ModelUsesThinkingLevels(model) {
+		return normalizeReasoningEffortLevel(payload, model)
+	}
+
+	return payload
+}
+
+// stripThinkingFields removes thinking-related fields from the payload for
+// models that do not support thinking.
+func stripThinkingFields(payload []byte) []byte {
+	fieldsToRemove := []string{
+		"reasoning",
+		"reasoning_effort",
+		"reasoning.effort",
+	}
+	out := payload
+	for _, field := range fieldsToRemove {
+		if gjson.GetBytes(out, field).Exists() {
+			out, _ = sjson.DeleteBytes(out, field)
+		}
+	}
+	return out
+}
+
+// normalizeReasoningEffortLevel validates and normalizes the reasoning_effort
+// or reasoning.effort field for level-based thinking models.
+func normalizeReasoningEffortLevel(payload []byte, model string) []byte {
+	out := payload
+
+	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
+		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
+			out, _ = sjson.SetBytes(out, "reasoning_effort", normalized)
+		}
+	}
+
+	if effort := gjson.GetBytes(out, "reasoning.effort"); effort.Exists() {
+		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
+			out, _ = sjson.SetBytes(out, "reasoning.effort", normalized)
+		}
+	}
+
+	return out
+}
+
+// validateThinkingConfig checks for unsupported reasoning levels on level-based models.
+// Returns a statusErr with 400 when an unsupported level is supplied to avoid silently
+// downgrading requests.
+func validateThinkingConfig(payload []byte, model string) error {
+	if len(payload) == 0 || model == "" {
+		return nil
+	}
+	if !util.ModelSupportsThinking(model) || !util.ModelUsesThinkingLevels(model) {
+		return nil
+	}
+
+	levels := util.GetModelThinkingLevels(model)
+	checkField := func(path string) error {
+		if effort := gjson.GetBytes(payload, path); effort.Exists() {
+			if _, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); !ok {
+				return statusErr{
+					code: http.StatusBadRequest,
+					msg:  fmt.Sprintf("unsupported reasoning effort level %q for model %s (supported: %s)", effort.String(), model, strings.Join(levels, ", ")),
+				}
+			}
+		}
+		return nil
+	}
+
+	if err := checkField("reasoning_effort"); err != nil {
+		return err
+	}
+	if err := checkField("reasoning.effort"); err != nil {
+		return err
+	}
+	return nil
+}

internal/runtime/executor/qwen_executor.go

@@ -12,6 +12,7 @@ import (
 
 	qwenauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -50,6 +51,15 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
+	body = normalizeThinkingConfig(body, upstreamModel)
+	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+		return resp, errValidate
+	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -121,6 +131,15 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	if upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
+	body = normalizeThinkingConfig(body, upstreamModel)
+	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+		return nil, errValidate
+	}
 	toolsResult := gjson.GetBytes(body, "tools")
 	// I'm addressing the Qwen3 "poisoning" issue, which is caused by the model needing a tool to be defined. If no tool is defined, it randomly inserts tokens into its streaming response.
 	// This will have no real consequences. It's just to scare Qwen3.
@@ -181,7 +200,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			}
 		}()
 		scanner := bufio.NewScanner(httpResp.Body)
-		scanner.Buffer(nil, 20_971_520)
+		scanner.Buffer(nil, 52_428_800) // 50MB
 		var param any
 		for scanner.Scan() {
 			line := scanner.Bytes()

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -123,6 +123,15 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 							functionResponse := client.FunctionResponse{ID: toolCallID, Name: funcName, Response: map[string]interface{}{"result": responseData}}
 							clientContent.Parts = append(clientContent.Parts, client.Part{FunctionResponse: &functionResponse})
 						}
+					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "image" {
+						sourceResult := contentResult.Get("source")
+						if sourceResult.Get("type").String() == "base64" {
+							inlineData := &client.InlineData{
+								MimeType: sourceResult.Get("media_type").String(),
+								Data:     sourceResult.Get("data").String(),
+							}
+							clientContent.Parts = append(clientContent.Parts, client.Part{InlineData: inlineData})
+						}
 					}
 				}
 				contents = append(contents, clientContent)

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -12,6 +12,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -34,8 +35,12 @@ type Params struct {
 	TotalTokenCount      int64  // Cached total token count from usage metadata
 	HasSentFinalEvents   bool   // Indicates if final content/message events have been sent
 	HasToolUse           bool   // Indicates if tool use was observed in the stream
+	HasContent           bool   // Tracks whether any content (text, thinking, or tool use) has been output
 }
 
+// toolUseIDCounter provides a process-wide unique counter for tool use identifiers.
+var toolUseIDCounter uint64
+
 // ConvertAntigravityResponseToClaude performs sophisticated streaming response format conversion.
 // This function implements a complex state machine that translates backend client responses
 // into Claude Code-compatible Server-Sent Events (SSE) format. It manages different response types
@@ -65,11 +70,14 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 
 	if bytes.Equal(rawJSON, []byte("[DONE]")) {
 		output := ""
-		appendFinalEvents(params, &output, true)
-
-		return []string{
-			output + "event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n\n",
+		// Only send final events if we have actually output content
+		if params.HasContent {
+			appendFinalEvents(params, &output, true)
+			return []string{
+				output + "event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n\n",
+			}
 		}
+		return []string{}
 	}
 
 	output := ""
@@ -115,10 +123,12 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":""}}`, params.ResponseIndex), "delta.signature", thoughtSignature.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						params.HasContent = true
 					} else if params.ResponseType == 2 { // Continue existing thinking block if already in thinking state
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, params.ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						params.HasContent = true
 					} else {
 						// Transition from another state to thinking
 						// First, close any existing content block
@@ -142,6 +152,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, params.ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						params.ResponseType = 2 // Set state to thinking
+						params.HasContent = true
 					}
 				} else {
 					finishReasonResult := gjson.GetBytes(rawJSON, "response.candidates.0.finishReason")
@@ -152,6 +163,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 							output = output + "event: content_block_delta\n"
 							data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, params.ResponseIndex), "delta.text", partTextResult.String())
 							output = output + fmt.Sprintf("data: %s\n\n\n", data)
+							params.HasContent = true
 						} else {
 							// Transition from another state to text content
 							// First, close any existing content block
@@ -175,6 +187,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 								data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, params.ResponseIndex), "delta.text", partTextResult.String())
 								output = output + fmt.Sprintf("data: %s\n\n\n", data)
 								params.ResponseType = 1 // Set state to content
+								params.HasContent = true
 							}
 						}
 					}
@@ -216,7 +229,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, params.ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 
@@ -226,6 +239,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 					output = output + fmt.Sprintf("data: %s\n\n\n", data)
 				}
 				params.ResponseType = 3
+				params.HasContent = true
 			}
 		}
 	}
@@ -265,6 +279,11 @@ func appendFinalEvents(params *Params, output *string, force bool) {
 		return
 	}
 
+	// Only send final events if we have actually output content
+	if !params.HasContent {
+		return
+	}
+
 	if params.ResponseType != 0 {
 		*output = *output + "event: content_block_stop\n"
 		*output = *output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, params.ResponseIndex)

internal/translator/antigravity/openai/chat-completions/antigravity_openai_response.go

@@ -11,6 +11,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/chat-completions"
@@ -24,6 +25,9 @@ type convertCliResponseToOpenAIChatParams struct {
 	FunctionIndex int
 }
 
+// functionCallIDCounter provides a process-wide unique counter for function call identifiers.
+var functionCallIDCounter uint64
+
 // ConvertAntigravityResponseToOpenAI translates a single chunk of a streaming response from the
 // Gemini CLI API format to the OpenAI Chat Completions streaming format.
 // It processes various Gemini CLI event types and transforms them into OpenAI-compatible JSON responses.
@@ -146,7 +150,7 @@ func ConvertAntigravityResponseToOpenAI(_ context.Context, _ string, originalReq
 
 				functionCallTemplate := `{"id": "","index": 0,"type": "function","function": {"name": "","arguments": ""}}`
 				fcName := functionCallResult.Get("name").String()
-				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
 				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "index", functionCallIndex)
 				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.name", fcName)
 				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {

internal/translator/claude/gemini/claude_gemini_response.go

@@ -331,8 +331,8 @@ func ConvertClaudeResponseToGeminiNonStream(_ context.Context, modelName string,
 	streamingEvents := make([][]byte, 0)
 
 	scanner := bufio.NewScanner(bytes.NewReader(rawJSON))
-	buffer := make([]byte, 20_971_520)
-	scanner.Buffer(buffer, 20_971_520)
+	buffer := make([]byte, 52_428_800) // 50MB
+	scanner.Buffer(buffer, 52_428_800)
 	for scanner.Scan() {
 		line := scanner.Bytes()
 		// log.Debug(string(line))

internal/translator/claude/openai/responses/claude_openai-responses_response.go

@@ -445,8 +445,8 @@ func ConvertClaudeResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 		// Use a simple scanner to iterate through raw bytes
 		// Note: extremely large responses may require increasing the buffer
 		scanner := bufio.NewScanner(bytes.NewReader(rawJSON))
-		buf := make([]byte, 20_971_520)
-		scanner.Buffer(buf, 20_971_520)
+		buf := make([]byte, 52_428_800) // 50MB
+		scanner.Buffer(buf, 52_428_800)
 		for scanner.Scan() {
 			line := scanner.Bytes()
 			if !bytes.HasPrefix(line, dataTag) {

internal/translator/gemini-cli/claude/gemini-cli_claude_response.go

@@ -12,6 +12,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -25,8 +26,12 @@ type Params struct {
 	HasFirstResponse bool // Indicates if the initial message_start event has been sent
 	ResponseType     int  // Current response type: 0=none, 1=content, 2=thinking, 3=function
 	ResponseIndex    int  // Index counter for content blocks in the streaming response
+	HasContent       bool // Tracks whether any content (text, thinking, or tool use) has been output
 }
 
+// toolUseIDCounter provides a process-wide unique counter for tool use identifiers.
+var toolUseIDCounter uint64
+
 // ConvertGeminiCLIResponseToClaude performs sophisticated streaming response format conversion.
 // This function implements a complex state machine that translates backend client responses
 // into Claude Code-compatible Server-Sent Events (SSE) format. It manages different response types
@@ -53,9 +58,13 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 	}
 
 	if bytes.Equal(rawJSON, []byte("[DONE]")) {
-		return []string{
-			"event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n\n",
+		// Only send message_stop if we have actually output content
+		if (*param).(*Params).HasContent {
+			return []string{
+				"event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n\n",
+			}
 		}
+		return []string{}
 	}
 
 	// Track whether tools are being used in this response chunk
@@ -104,6 +113,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, (*param).(*Params).ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						(*param).(*Params).HasContent = true
 					} else {
 						// Transition from another state to thinking
 						// First, close any existing content block
@@ -127,6 +137,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, (*param).(*Params).ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						(*param).(*Params).ResponseType = 2 // Set state to thinking
+						(*param).(*Params).HasContent = true
 					}
 				} else {
 					// Process regular text content (user-visible output)
@@ -135,6 +146,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, (*param).(*Params).ResponseIndex), "delta.text", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						(*param).(*Params).HasContent = true
 					} else {
 						// Transition from another state to text content
 						// First, close any existing content block
@@ -158,6 +170,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, (*param).(*Params).ResponseIndex), "delta.text", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						(*param).(*Params).ResponseType = 1 // Set state to content
+						(*param).(*Params).HasContent = true
 					}
 				}
 			} else if functionCallResult.Exists() {
@@ -197,7 +210,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, (*param).(*Params).ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 
@@ -207,6 +220,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 					output = output + fmt.Sprintf("data: %s\n\n\n", data)
 				}
 				(*param).(*Params).ResponseType = 3
+				(*param).(*Params).HasContent = true
 			}
 		}
 	}
@@ -215,28 +229,31 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 	// Process usage metadata and finish reason when present in the response
 	if usageResult.Exists() && bytes.Contains(rawJSON, []byte(`"finishReason"`)) {
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
-			// Close the final content block
-			output = output + "event: content_block_stop\n"
-			output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, (*param).(*Params).ResponseIndex)
-			output = output + "\n\n\n"
+			// Only send final events if we have actually output content
+			if (*param).(*Params).HasContent {
+				// Close the final content block
+				output = output + "event: content_block_stop\n"
+				output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, (*param).(*Params).ResponseIndex)
+				output = output + "\n\n\n"
 
-			// Send the final message delta with usage information and stop reason
-			output = output + "event: message_delta\n"
-			output = output + `data: `
+				// Send the final message delta with usage information and stop reason
+				output = output + "event: message_delta\n"
+				output = output + `data: `
 
-			// Create the message delta template with appropriate stop reason
-			template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-			// Set tool_use stop reason if tools were used in this response
-			if usedTool {
-				template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				// Create the message delta template with appropriate stop reason
+				template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				// Set tool_use stop reason if tools were used in this response
+				if usedTool {
+					template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				}
+
+				// Include thinking tokens in output token count if present
+				thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
+				template, _ = sjson.Set(template, "usage.output_tokens", candidatesTokenCountResult.Int()+thoughtsTokenCount)
+				template, _ = sjson.Set(template, "usage.input_tokens", usageResult.Get("promptTokenCount").Int())
+
+				output = output + template + "\n\n\n"
 			}
-
-			// Include thinking tokens in output token count if present
-			thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
-			template, _ = sjson.Set(template, "usage.output_tokens", candidatesTokenCountResult.Int()+thoughtsTokenCount)
-			template, _ = sjson.Set(template, "usage.input_tokens", usageResult.Get("promptTokenCount").Int())
-
-			output = output + template + "\n\n\n"
 		}
 	}
 

internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_response.go

@@ -11,6 +11,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/chat-completions"
@@ -24,6 +25,9 @@ type convertCliResponseToOpenAIChatParams struct {
 	FunctionIndex int
 }
 
+// functionCallIDCounter provides a process-wide unique counter for function call identifiers.
+var functionCallIDCounter uint64
+
 // ConvertCliResponseToOpenAI translates a single chunk of a streaming response from the
 // Gemini CLI API format to the OpenAI Chat Completions streaming format.
 // It processes various Gemini CLI event types and transforms them into OpenAI-compatible JSON responses.
@@ -146,7 +150,7 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 
 				functionCallTemplate := `{"id": "","index": 0,"type": "function","function": {"name": "","arguments": ""}}`
 				fcName := functionCallResult.Get("name").String()
-				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
 				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "index", functionCallIndex)
 				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.name", fcName)
 				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {

internal/translator/gemini/claude/gemini_claude_response.go

@@ -12,6 +12,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -24,8 +25,12 @@ type Params struct {
 	HasFirstResponse bool
 	ResponseType     int
 	ResponseIndex    int
+	HasContent       bool // Tracks whether any content (text, thinking, or tool use) has been output
 }
 
+// toolUseIDCounter provides a process-wide unique counter for tool use identifiers.
+var toolUseIDCounter uint64
+
 // ConvertGeminiResponseToClaude performs sophisticated streaming response format conversion.
 // This function implements a complex state machine that translates backend client responses
 // into Claude-compatible Server-Sent Events (SSE) format. It manages different response types
@@ -53,9 +58,13 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 	}
 
 	if bytes.Equal(rawJSON, []byte("[DONE]")) {
-		return []string{
-			"event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n\n",
+		// Only send message_stop if we have actually output content
+		if (*param).(*Params).HasContent {
+			return []string{
+				"event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n\n",
+			}
 		}
+		return []string{}
 	}
 
 	// Track whether tools are being used in this response chunk
@@ -104,6 +113,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, (*param).(*Params).ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						(*param).(*Params).HasContent = true
 					} else {
 						// Transition from another state to thinking
 						// First, close any existing content block
@@ -127,6 +137,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, (*param).(*Params).ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						(*param).(*Params).ResponseType = 2 // Set state to thinking
+						(*param).(*Params).HasContent = true
 					}
 				} else {
 					// Process regular text content (user-visible output)
@@ -135,6 +146,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, (*param).(*Params).ResponseIndex), "delta.text", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						(*param).(*Params).HasContent = true
 					} else {
 						// Transition from another state to text content
 						// First, close any existing content block
@@ -158,6 +170,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, (*param).(*Params).ResponseIndex), "delta.text", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						(*param).(*Params).ResponseType = 1 // Set state to content
+						(*param).(*Params).HasContent = true
 					}
 				}
 			} else if functionCallResult.Exists() {
@@ -197,7 +210,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, (*param).(*Params).ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 
@@ -207,6 +220,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 					output = output + fmt.Sprintf("data: %s\n\n\n", data)
 				}
 				(*param).(*Params).ResponseType = 3
+				(*param).(*Params).HasContent = true
 			}
 		}
 	}
@@ -214,23 +228,26 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 	usageResult := gjson.GetBytes(rawJSON, "usageMetadata")
 	if usageResult.Exists() && bytes.Contains(rawJSON, []byte(`"finishReason"`)) {
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
-			output = output + "event: content_block_stop\n"
-			output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, (*param).(*Params).ResponseIndex)
-			output = output + "\n\n\n"
+			// Only send final events if we have actually output content
+			if (*param).(*Params).HasContent {
+				output = output + "event: content_block_stop\n"
+				output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, (*param).(*Params).ResponseIndex)
+				output = output + "\n\n\n"
 
-			output = output + "event: message_delta\n"
-			output = output + `data: `
+				output = output + "event: message_delta\n"
+				output = output + `data: `
 
-			template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-			if usedTool {
-				template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				if usedTool {
+					template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				}
+
+				thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
+				template, _ = sjson.Set(template, "usage.output_tokens", candidatesTokenCountResult.Int()+thoughtsTokenCount)
+				template, _ = sjson.Set(template, "usage.input_tokens", usageResult.Get("promptTokenCount").Int())
+
+				output = output + template + "\n\n\n"
 			}
-
-			thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
-			template, _ = sjson.Set(template, "usage.output_tokens", candidatesTokenCountResult.Int()+thoughtsTokenCount)
-			template, _ = sjson.Set(template, "usage.input_tokens", usageResult.Get("promptTokenCount").Int())
-
-			output = output + template + "\n\n\n"
 		}
 	}
 

internal/translator/gemini/openai/chat-completions/gemini_openai_response.go

@@ -11,6 +11,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -23,6 +24,9 @@ type convertGeminiResponseToOpenAIChatParams struct {
 	FunctionIndex int
 }
 
+// functionCallIDCounter provides a process-wide unique counter for function call identifiers.
+var functionCallIDCounter uint64
+
 // ConvertGeminiResponseToOpenAI translates a single chunk of a streaming response from the
 // Gemini API format to the OpenAI Chat Completions streaming format.
 // It processes various Gemini event types and transforms them into OpenAI-compatible JSON responses.
@@ -148,7 +152,7 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 
 				functionCallTemplate := `{"id": "","index": 0,"type": "function","function": {"name": "","arguments": ""}}`
 				fcName := functionCallResult.Get("name").String()
-				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
 				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "index", functionCallIndex)
 				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.name", fcName)
 				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
@@ -281,7 +285,7 @@ func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, origina
 				}
 				functionCallItemTemplate := `{"id": "","type": "function","function": {"name": "","arguments": ""}}`
 				fcName := functionCallResult.Get("name").String()
-				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
 				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", fcName)
 				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
 					functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", fcArgsResult.Raw)

internal/translator/gemini/openai/responses/gemini_openai-responses_response.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -37,6 +38,12 @@ type geminiToResponsesState struct {
 	FuncCallIDs map[int]string
 }
 
+// responseIDCounter provides a process-wide unique counter for synthesized response identifiers.
+var responseIDCounter uint64
+
+// funcCallIDCounter provides a process-wide unique counter for function call identifiers.
+var funcCallIDCounter uint64
+
 func emitEvent(event string, payload string) string {
 	return fmt.Sprintf("event: %s\ndata: %s", event, payload)
 }
@@ -205,7 +212,7 @@ func ConvertGeminiResponseToOpenAIResponses(_ context.Context, modelName string,
 					st.FuncArgsBuf[idx] = &strings.Builder{}
 				}
 				if st.FuncCallIDs[idx] == "" {
-					st.FuncCallIDs[idx] = fmt.Sprintf("call_%d", time.Now().UnixNano())
+					st.FuncCallIDs[idx] = fmt.Sprintf("call_%d_%d", time.Now().UnixNano(), atomic.AddUint64(&funcCallIDCounter, 1))
 				}
 				st.FuncNames[idx] = name
 
@@ -464,7 +471,7 @@ func ConvertGeminiResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 	// id: prefer provider responseId, otherwise synthesize
 	id := root.Get("responseId").String()
 	if id == "" {
-		id = fmt.Sprintf("resp_%x", time.Now().UnixNano())
+		id = fmt.Sprintf("resp_%x_%d", time.Now().UnixNano(), atomic.AddUint64(&responseIDCounter, 1))
 	}
 	// Normalize to response-style id (prefix resp_ if missing)
 	if !strings.HasPrefix(id, "resp_") {
@@ -575,7 +582,7 @@ func ConvertGeminiResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 			if fc := p.Get("functionCall"); fc.Exists() {
 				name := fc.Get("name").String()
 				args := fc.Get("args")
-				callID := fmt.Sprintf("call_%x", time.Now().UnixNano())
+				callID := fmt.Sprintf("call_%x_%d", time.Now().UnixNano(), atomic.AddUint64(&funcCallIDCounter, 1))
 				outputs = append(outputs, map[string]interface{}{
 					"id":     fmt.Sprintf("fc_%s", callID),
 					"type":   "function_call",

internal/translator/openai/openai/responses/openai_openai-responses_response.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -41,6 +42,9 @@ type oaiToResponsesState struct {
 	UsageSeen        bool
 }
 
+// responseIDCounter provides a process-wide unique counter for synthesized response identifiers.
+var responseIDCounter uint64
+
 func emitRespEvent(event string, payload string) string {
 	return fmt.Sprintf("event: %s\ndata: %s", event, payload)
 }
@@ -590,7 +594,7 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(_ context.Co
 	// id: use provider id if present, otherwise synthesize
 	id := root.Get("id").String()
 	if id == "" {
-		id = fmt.Sprintf("resp_%x", time.Now().UnixNano())
+		id = fmt.Sprintf("resp_%x_%d", time.Now().UnixNano(), atomic.AddUint64(&responseIDCounter, 1))
 	}
 	resp, _ = sjson.Set(resp, "id", id)
 

internal/util/claude_thinking.go

@@ -0,0 +1,46 @@
+package util
+
+import (
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// ApplyClaudeThinkingConfig applies thinking configuration to a Claude API request payload.
+// It sets the thinking.type to "enabled" and thinking.budget_tokens to the specified budget.
+// If budget is nil or the payload already has thinking config, it returns the payload unchanged.
+func ApplyClaudeThinkingConfig(body []byte, budget *int) []byte {
+	if budget == nil {
+		return body
+	}
+	if gjson.GetBytes(body, "thinking").Exists() {
+		return body
+	}
+	if *budget <= 0 {
+		return body
+	}
+	updated := body
+	updated, _ = sjson.SetBytes(updated, "thinking.type", "enabled")
+	updated, _ = sjson.SetBytes(updated, "thinking.budget_tokens", *budget)
+	return updated
+}
+
+// ResolveClaudeThinkingConfig resolves thinking configuration from metadata for Claude models.
+// It uses the unified ResolveThinkingConfigFromMetadata and normalizes the budget.
+// Returns the normalized budget (nil if thinking should not be enabled) and whether it matched.
+func ResolveClaudeThinkingConfig(modelName string, metadata map[string]any) (*int, bool) {
+	budget, include, matched := ResolveThinkingConfigFromMetadata(modelName, metadata)
+	if !matched {
+		return nil, false
+	}
+	if include != nil && !*include {
+		return nil, true
+	}
+	if budget == nil {
+		return nil, true
+	}
+	normalized := NormalizeThinkingBudget(modelName, *budget)
+	if normalized <= 0 {
+		return nil, true
+	}
+	return &normalized, true
+}

internal/util/gemini_thinking.go

@@ -1,8 +1,6 @@
 package util
 
 import (
-	"encoding/json"
-	"strconv"
 	"strings"
 
 	"github.com/tidwall/gjson"
@@ -15,80 +13,6 @@ const (
 	GeminiOriginalModelMetadataKey   = "gemini_original_model"
 )
 
-func ParseGeminiThinkingSuffix(model string) (string, *int, *bool, bool) {
-	if model == "" {
-		return model, nil, nil, false
-	}
-	lower := strings.ToLower(model)
-	if !strings.HasPrefix(lower, "gemini-") {
-		return model, nil, nil, false
-	}
-
-	if strings.HasSuffix(lower, "-nothinking") {
-		base := model[:len(model)-len("-nothinking")]
-		budgetValue := 0
-		if strings.HasPrefix(lower, "gemini-2.5-pro") {
-			budgetValue = 128
-		}
-		include := false
-		return base, &budgetValue, &include, true
-	}
-
-	// Handle "-reasoning" suffix: enables thinking with dynamic budget (-1)
-	// Maps: gemini-2.5-flash-reasoning -> gemini-2.5-flash with thinkingBudget=-1
-	if strings.HasSuffix(lower, "-reasoning") {
-		base := model[:len(model)-len("-reasoning")]
-		budgetValue := -1 // Dynamic budget
-		include := true
-		return base, &budgetValue, &include, true
-	}
-
-	idx := strings.LastIndex(lower, "-thinking-")
-	if idx == -1 {
-		return model, nil, nil, false
-	}
-
-	digits := model[idx+len("-thinking-"):]
-	if digits == "" {
-		return model, nil, nil, false
-	}
-	end := len(digits)
-	for i := 0; i < len(digits); i++ {
-		if digits[i] < '0' || digits[i] > '9' {
-			end = i
-			break
-		}
-	}
-	if end == 0 {
-		return model, nil, nil, false
-	}
-	valueStr := digits[:end]
-	value, err := strconv.Atoi(valueStr)
-	if err != nil {
-		return model, nil, nil, false
-	}
-	base := model[:idx]
-	budgetValue := value
-	return base, &budgetValue, nil, true
-}
-
-func NormalizeGeminiThinkingModel(modelName string) (string, map[string]any) {
-	baseModel, budget, include, matched := ParseGeminiThinkingSuffix(modelName)
-	if !matched {
-		return baseModel, nil
-	}
-	metadata := map[string]any{
-		GeminiOriginalModelMetadataKey: modelName,
-	}
-	if budget != nil {
-		metadata[GeminiThinkingBudgetMetadataKey] = *budget
-	}
-	if include != nil {
-		metadata[GeminiIncludeThoughtsMetadataKey] = *include
-	}
-	return baseModel, metadata
-}
-
 func ApplyGeminiThinkingConfig(body []byte, budget *int, includeThoughts *bool) []byte {
 	if budget == nil && includeThoughts == nil {
 		return body
@@ -133,80 +57,6 @@ func ApplyGeminiCLIThinkingConfig(body []byte, budget *int, includeThoughts *boo
 	return updated
 }
 
-func GeminiThinkingFromMetadata(metadata map[string]any) (*int, *bool, bool) {
-	if len(metadata) == 0 {
-		return nil, nil, false
-	}
-	var (
-		budgetPtr  *int
-		includePtr *bool
-		matched    bool
-	)
-	if rawBudget, ok := metadata[GeminiThinkingBudgetMetadataKey]; ok {
-		switch v := rawBudget.(type) {
-		case int:
-			budget := v
-			budgetPtr = &budget
-			matched = true
-		case int32:
-			budget := int(v)
-			budgetPtr = &budget
-			matched = true
-		case int64:
-			budget := int(v)
-			budgetPtr = &budget
-			matched = true
-		case float64:
-			budget := int(v)
-			budgetPtr = &budget
-			matched = true
-		case json.Number:
-			if val, err := v.Int64(); err == nil {
-				budget := int(val)
-				budgetPtr = &budget
-				matched = true
-			}
-		}
-	}
-	if rawInclude, ok := metadata[GeminiIncludeThoughtsMetadataKey]; ok {
-		switch v := rawInclude.(type) {
-		case bool:
-			include := v
-			includePtr = &include
-			matched = true
-		case string:
-			if parsed, err := strconv.ParseBool(v); err == nil {
-				include := parsed
-				includePtr = &include
-				matched = true
-			}
-		case json.Number:
-			if val, err := v.Int64(); err == nil {
-				include := val != 0
-				includePtr = &include
-				matched = true
-			}
-		case int:
-			include := v != 0
-			includePtr = &include
-			matched = true
-		case int32:
-			include := v != 0
-			includePtr = &include
-			matched = true
-		case int64:
-			include := v != 0
-			includePtr = &include
-			matched = true
-		case float64:
-			include := v != 0
-			includePtr = &include
-			matched = true
-		}
-	}
-	return budgetPtr, includePtr, matched
-}
-
 // modelsWithDefaultThinking lists models that should have thinking enabled by default
 // when no explicit thinkingConfig is provided.
 var modelsWithDefaultThinking = map[string]bool{

internal/util/thinking.go

@@ -1,6 +1,8 @@
 package util
 
 import (
+	"strings"
+
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 )
 
@@ -67,3 +69,39 @@ func thinkingRangeFromRegistry(model string) (found bool, min int, max int, zero
 	}
 	return true, info.Thinking.Min, info.Thinking.Max, info.Thinking.ZeroAllowed, info.Thinking.DynamicAllowed
 }
+
+// GetModelThinkingLevels returns the discrete reasoning effort levels for the model.
+// Returns nil if the model has no thinking support or no levels defined.
+func GetModelThinkingLevels(model string) []string {
+	if model == "" {
+		return nil
+	}
+	info := registry.GetGlobalRegistry().GetModelInfo(model)
+	if info == nil || info.Thinking == nil {
+		return nil
+	}
+	return info.Thinking.Levels
+}
+
+// ModelUsesThinkingLevels reports whether the model uses discrete reasoning
+// effort levels instead of numeric budgets.
+func ModelUsesThinkingLevels(model string) bool {
+	levels := GetModelThinkingLevels(model)
+	return len(levels) > 0
+}
+
+// NormalizeReasoningEffortLevel validates and normalizes a reasoning effort
+// level for the given model. Returns false when the level is not supported.
+func NormalizeReasoningEffortLevel(model, effort string) (string, bool) {
+	levels := GetModelThinkingLevels(model)
+	if len(levels) == 0 {
+		return "", false
+	}
+	loweredEffort := strings.ToLower(strings.TrimSpace(effort))
+	for _, lvl := range levels {
+		if strings.ToLower(lvl) == loweredEffort {
+			return lvl, true
+		}
+	}
+	return "", false
+}

internal/util/thinking_suffix.go

@@ -0,0 +1,313 @@
+package util
+
+import (
+	"encoding/json"
+	"strconv"
+	"strings"
+)
+
+const (
+	ThinkingBudgetMetadataKey          = "thinking_budget"
+	ThinkingIncludeThoughtsMetadataKey = "thinking_include_thoughts"
+	ReasoningEffortMetadataKey         = "reasoning_effort"
+	ThinkingOriginalModelMetadataKey   = "thinking_original_model"
+)
+
+// NormalizeThinkingModel parses dynamic thinking suffixes on model names and returns
+// the normalized base model with extracted metadata. Supported pattern:
+//   - "(<value>)" where value can be:
+//   - A numeric budget (e.g., "(8192)", "(16384)")
+//   - A reasoning effort level (e.g., "(high)", "(medium)", "(low)")
+//
+// Examples:
+//   - "claude-sonnet-4-5-20250929(16384)" → budget=16384
+//   - "gpt-5.1(high)" → reasoning_effort="high"
+//   - "gemini-2.5-pro(32768)" → budget=32768
+//
+// Note: Empty parentheses "()" are not supported and will be ignored.
+func NormalizeThinkingModel(modelName string) (string, map[string]any) {
+	if modelName == "" {
+		return modelName, nil
+	}
+
+	baseModel := modelName
+
+	var (
+		budgetOverride  *int
+		reasoningEffort *string
+		matched         bool
+	)
+
+	// Match "(<value>)" pattern at the end of the model name
+	if idx := strings.LastIndex(modelName, "("); idx != -1 {
+		if !strings.HasSuffix(modelName, ")") {
+			// Incomplete parenthesis, ignore
+			return baseModel, nil
+		}
+
+		value := modelName[idx+1 : len(modelName)-1] // Extract content between ( and )
+		if value == "" {
+			// Empty parentheses not supported
+			return baseModel, nil
+		}
+
+		candidateBase := modelName[:idx]
+
+		// Auto-detect: pure numeric → budget, string → reasoning effort level
+		if parsed, ok := parseIntPrefix(value); ok {
+			// Numeric value: treat as thinking budget
+			baseModel = candidateBase
+			budgetOverride = &parsed
+			matched = true
+		} else {
+			// String value: treat as reasoning effort level
+			baseModel = candidateBase
+			raw := strings.ToLower(strings.TrimSpace(value))
+			if raw != "" {
+				reasoningEffort = &raw
+				matched = true
+			}
+		}
+	}
+
+	if !matched {
+		return baseModel, nil
+	}
+
+	metadata := map[string]any{
+		ThinkingOriginalModelMetadataKey: modelName,
+	}
+	if budgetOverride != nil {
+		metadata[ThinkingBudgetMetadataKey] = *budgetOverride
+	}
+	if reasoningEffort != nil {
+		metadata[ReasoningEffortMetadataKey] = *reasoningEffort
+	}
+	return baseModel, metadata
+}
+
+// ThinkingFromMetadata extracts thinking overrides from metadata produced by NormalizeThinkingModel.
+// It accepts both the new generic keys and legacy Gemini-specific keys.
+func ThinkingFromMetadata(metadata map[string]any) (*int, *bool, *string, bool) {
+	if len(metadata) == 0 {
+		return nil, nil, nil, false
+	}
+
+	var (
+		budgetPtr  *int
+		includePtr *bool
+		effortPtr  *string
+		matched    bool
+	)
+
+	readBudget := func(key string) {
+		if budgetPtr != nil {
+			return
+		}
+		if raw, ok := metadata[key]; ok {
+			if v, okNumber := parseNumberToInt(raw); okNumber {
+				budget := v
+				budgetPtr = &budget
+				matched = true
+			}
+		}
+	}
+
+	readInclude := func(key string) {
+		if includePtr != nil {
+			return
+		}
+		if raw, ok := metadata[key]; ok {
+			switch v := raw.(type) {
+			case bool:
+				val := v
+				includePtr = &val
+				matched = true
+			case *bool:
+				if v != nil {
+					val := *v
+					includePtr = &val
+					matched = true
+				}
+			}
+		}
+	}
+
+	readEffort := func(key string) {
+		if effortPtr != nil {
+			return
+		}
+		if raw, ok := metadata[key]; ok {
+			if val, okStr := raw.(string); okStr && strings.TrimSpace(val) != "" {
+				normalized := strings.ToLower(strings.TrimSpace(val))
+				effortPtr = &normalized
+				matched = true
+			}
+		}
+	}
+
+	readBudget(ThinkingBudgetMetadataKey)
+	readBudget(GeminiThinkingBudgetMetadataKey)
+	readInclude(ThinkingIncludeThoughtsMetadataKey)
+	readInclude(GeminiIncludeThoughtsMetadataKey)
+	readEffort(ReasoningEffortMetadataKey)
+	readEffort("reasoning.effort")
+
+	return budgetPtr, includePtr, effortPtr, matched
+}
+
+// ResolveThinkingConfigFromMetadata derives thinking budget/include overrides,
+// converting reasoning effort strings into budgets when possible.
+func ResolveThinkingConfigFromMetadata(model string, metadata map[string]any) (*int, *bool, bool) {
+	budget, include, effort, matched := ThinkingFromMetadata(metadata)
+	if !matched {
+		return nil, nil, false
+	}
+
+	if budget == nil && effort != nil {
+		if derived, ok := ThinkingEffortToBudget(model, *effort); ok {
+			budget = &derived
+		}
+	}
+	return budget, include, budget != nil || include != nil || effort != nil
+}
+
+// ReasoningEffortFromMetadata resolves a reasoning effort string from metadata,
+// inferring "auto" and "none" when budgets request dynamic or disabled thinking.
+func ReasoningEffortFromMetadata(metadata map[string]any) (string, bool) {
+	budget, include, effort, matched := ThinkingFromMetadata(metadata)
+	if !matched {
+		return "", false
+	}
+	if effort != nil && *effort != "" {
+		return strings.ToLower(strings.TrimSpace(*effort)), true
+	}
+	if budget != nil {
+		switch *budget {
+		case -1:
+			return "auto", true
+		case 0:
+			return "none", true
+		}
+	}
+	if include != nil && !*include {
+		return "none", true
+	}
+	return "", true
+}
+
+// ThinkingEffortToBudget maps reasoning effort levels to approximate budgets,
+// clamping the result to the model's supported range.
+func ThinkingEffortToBudget(model, effort string) (int, bool) {
+	if effort == "" {
+		return 0, false
+	}
+	normalized, ok := NormalizeReasoningEffortLevel(model, effort)
+	if !ok {
+		normalized = strings.ToLower(strings.TrimSpace(effort))
+	}
+	switch normalized {
+	case "none":
+		return 0, true
+	case "auto":
+		return NormalizeThinkingBudget(model, -1), true
+	case "minimal":
+		return NormalizeThinkingBudget(model, 512), true
+	case "low":
+		return NormalizeThinkingBudget(model, 1024), true
+	case "medium":
+		return NormalizeThinkingBudget(model, 8192), true
+	case "high":
+		return NormalizeThinkingBudget(model, 24576), true
+	case "xhigh":
+		return NormalizeThinkingBudget(model, 32768), true
+	default:
+		return 0, false
+	}
+}
+
+// ResolveOriginalModel returns the original model name stored in metadata (if present),
+// otherwise falls back to the provided model.
+func ResolveOriginalModel(model string, metadata map[string]any) string {
+	normalize := func(name string) string {
+		if name == "" {
+			return ""
+		}
+		if base, _ := NormalizeThinkingModel(name); base != "" {
+			return base
+		}
+		return strings.TrimSpace(name)
+	}
+
+	if metadata != nil {
+		if v, ok := metadata[ThinkingOriginalModelMetadataKey]; ok {
+			if s, okStr := v.(string); okStr && strings.TrimSpace(s) != "" {
+				if base := normalize(s); base != "" {
+					return base
+				}
+			}
+		}
+		if v, ok := metadata[GeminiOriginalModelMetadataKey]; ok {
+			if s, okStr := v.(string); okStr && strings.TrimSpace(s) != "" {
+				if base := normalize(s); base != "" {
+					return base
+				}
+			}
+		}
+	}
+	// Fallback: try to re-normalize the model name when metadata was dropped.
+	if base := normalize(model); base != "" {
+		return base
+	}
+	return model
+}
+
+func parseIntPrefix(value string) (int, bool) {
+	if value == "" {
+		return 0, false
+	}
+	digits := strings.TrimLeft(value, "-")
+	if digits == "" {
+		return 0, false
+	}
+	end := len(digits)
+	for i := 0; i < len(digits); i++ {
+		if digits[i] < '0' || digits[i] > '9' {
+			end = i
+			break
+		}
+	}
+	if end == 0 {
+		return 0, false
+	}
+	val, err := strconv.Atoi(digits[:end])
+	if err != nil {
+		return 0, false
+	}
+	return val, true
+}
+
+func parseNumberToInt(raw any) (int, bool) {
+	switch v := raw.(type) {
+	case int:
+		return v, true
+	case int32:
+		return int(v), true
+	case int64:
+		return int(v), true
+	case float64:
+		return int(v), true
+	case json.Number:
+		if val, err := v.Int64(); err == nil {
+			return int(val), true
+		}
+	case string:
+		if strings.TrimSpace(v) == "" {
+			return 0, false
+		}
+		if parsed, err := strconv.Atoi(strings.TrimSpace(v)); err == nil {
+			return parsed, true
+		}
+	}
+	return 0, false
+}

sdk/api/handlers/claude/code_handlers.go

@@ -271,6 +271,11 @@ func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.
 				continue
 			}
 			if errMsg != nil {
+				status := http.StatusInternalServerError
+				if errMsg.StatusCode > 0 {
+					status = errMsg.StatusCode
+				}
+				c.Status(status)
 				// An error occurred: emit as a proper SSE error event
 				errorBytes, _ := json.Marshal(h.toClaudeError(errMsg))
 				_, _ = writer.WriteString("event: error\n")
@@ -278,6 +283,7 @@ func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.
 				_, _ = writer.Write(errorBytes)
 				_, _ = writer.WriteString("\n\n")
 				_ = writer.Flush()
+				flusher.Flush()
 			}
 			var execErr error
 			if errMsg != nil {

sdk/api/handlers/handlers.go

@@ -323,18 +323,32 @@ func (h *BaseAPIHandler) getRequestDetails(modelName string) (providers []string
 
 	providerName, extractedModelName, isDynamic := h.parseDynamicModel(resolvedModelName)
 
-	// First, normalize the model name to handle suffixes like "-thinking-128"
-	// This needs to happen before determining the provider for non-dynamic models.
-	normalizedModel, metadata = normalizeModelMetadata(resolvedModelName)
+	targetModelName := resolvedModelName
+	if isDynamic {
+		targetModelName = extractedModelName
+	}
+
+	// Normalize the model name to handle dynamic thinking suffixes before determining the provider.
+	normalizedModel, metadata = normalizeModelMetadata(targetModelName)
 
 	if isDynamic {
 		providers = []string{providerName}
-		// For dynamic models, the extractedModelName is already normalized by parseDynamicModel
-		// so we use it as the final normalizedModel.
-		normalizedModel = extractedModelName
 	} else {
 		// For non-dynamic models, use the normalizedModel to get the provider name.
 		providers = util.GetProviderName(normalizedModel)
+		if len(providers) == 0 && metadata != nil {
+			if originalRaw, ok := metadata[util.ThinkingOriginalModelMetadataKey]; ok {
+				if originalModel, okStr := originalRaw.(string); okStr {
+					originalModel = strings.TrimSpace(originalModel)
+					if originalModel != "" && !strings.EqualFold(originalModel, normalizedModel) {
+						if altProviders := util.GetProviderName(originalModel); len(altProviders) > 0 {
+							providers = altProviders
+							normalizedModel = originalModel
+						}
+					}
+				}
+			}
+		}
 	}
 
 	if len(providers) == 0 {
@@ -382,7 +396,7 @@ func cloneBytes(src []byte) []byte {
 }
 
 func normalizeModelMetadata(modelName string) (string, map[string]any) {
-	return util.NormalizeGeminiThinkingModel(modelName)
+	return util.NormalizeThinkingModel(modelName)
 }
 
 func cloneMetadata(src map[string]any) map[string]any {