Merge upstream 8a0b439 (14 commits)

.github/ISSUE_TEMPLATE/01-bug-report.yml

@@ -1,7 +1,7 @@
 name: Bug report
 description: Report a reproducible problem with the Desktop app, CLI, or source build.
 title: "[Bug]: "
-labels: ["bug"]
+labels: ["bug", "needs-triage"]
 body:
   - type: markdown
     attributes:

.github/ISSUE_TEMPLATE/02-feature-request.yml

@@ -1,7 +1,7 @@
 name: Feature request
 description: Suggest an improvement to the Desktop app, CLI, or source distribution.
 title: "[Feature]: "
-labels: ["enhancement"]
+labels: ["enhancement", "needs-triage"]
 body:
   - type: dropdown
     id: product

.github/ISSUE_TEMPLATE/03-question.yml

@@ -1,20 +1,21 @@
-name: Usage question
-description: Ask about the Desktop app, CLI, or building from source.
+name: Usage or repository policy question
+description: Ask about the Desktop app, CLI, source builds, or contribution policy.
 title: "[Question]: "
-labels: ["question"]
+labels: ["question", "needs-triage"]
 body:
   - type: dropdown
     id: product
     attributes:
       label: Product or interface
+      description: Leave blank if unsure or asking about repository policy.
       options:
         - Desktop app
         - CLI - interactive TUI
         - CLI - headless
         - CLI - ACP
         - Source build or repository tooling
     validations:
-      required: true
+      required: false
   - type: textarea
     id: question
     attributes:
@@ -25,18 +26,19 @@ body:
     id: platform
     attributes:
       label: Platform
+      description: Leave blank if unknown or not applicable.
       options:
         - macOS
         - Windows
         - Linux
         - WSL
         - Multiple platforms
     validations:
-      required: true
+      required: false
   - type: textarea
     id: context
     attributes:
-      label: Version and what you have tried
-      description: Include documentation you followed and relevant redacted output. Do not include keys or account data.
+      label: Version or policy context
+      description: For usage questions, include your version and what you tried. For policy questions, link the relevant policy. Do not include keys or account data.
     validations:
       required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -1,5 +1,8 @@
 blank_issues_enabled: false
 contact_links:
+  - name: Contribution policy
+    url: https://github.com/MiniMax-AI/minimax-code/blob/main/CONTRIBUTING.md
+    about: Everyone may report issues and suggest improvements. Pull requests are accepted only from repository collaborators; external PRs are not accepted.
   - name: CLI quick start and troubleshooting
     url: https://agent.minimax.io/docs/cli/quick-start
     about: Installation, account regions, first tasks, and links to troubleshooting.

.github/ISSUE_TEMPLATE/docs.yml

@@ -1,6 +1,7 @@
 name: Documentation issue
 description: Report missing, incorrect, or unclear documentation.
 title: "[Docs] "
+labels: ["documentation", "needs-triage"]
 body:
   - type: input
     id: location

README.md

@@ -98,6 +98,8 @@ mcode
 
 Supported API formats: `openai-completions`, `openai-responses`, and `anthropic-messages`. See the [model examples](docs/examples.md#2-choose-your-own-model) for environment variable setup, connection checks, and model overrides for a single run.
 
+Providers added this way are stored under `custom_provider` in the active profile's `config.yaml`; this is the supported path for third-party or self-hosted endpoints, while `minimax_api` stays reserved for the official MiniMax API. Relays that require `Authorization: Bearer` on an Anthropic-compatible endpoint can set custom headers in `config.yaml`; see [Third-party relays and custom auth headers](docs/examples.md#third-party-relays-and-custom-auth-headers).
+
 </details>
 
 ### 3. Run your first task

README_ZH.md

@@ -98,6 +98,8 @@ mcode
 
 支持 `openai-completions`、`openai-responses` 和 `anthropic-messages`。连接测试、单次模型切换及环境变量设置见 [模型示例](docs/examples.md#2-choose-your-own-model)。
 
+通过该命令添加的提供方保存在当前 profile `config.yaml` 的 `custom_provider` 下；第三方或自建端点一律走这条路径，`minimax_api` 保留给官方 MiniMax API。若中转端点只提供 Anthropic 兼容接口且要求 `Authorization: Bearer` 鉴权，可在 `config.yaml` 中为提供方配置自定义 headers；见[第三方中转与自定义鉴权头](docs/examples.md#third-party-relays-and-custom-auth-headers)。
+
 </details>
 
 ### 3. 完成第一个任务

docs/examples.md

@@ -66,10 +66,65 @@ pnpm mcode provider list --json
 
 `--api-key-env` reads the current environment variable value and stores that value in the active profile's `config.yaml`; it does not save an environment-variable reference. The file still contains plaintext credentials. On POSIX systems, config writes and temporary copies use `0600`. When loading existing files, MCode removes group/other access while preserving the owner's permissions; already-private files such as `0400` or `0600` do not require a permission change. Loading fails if an unsafe main config cannot be restricted. Older migration backups are also checked, but inspection or repair failures produce a warning identifying the directory or backup that needs manual attention rather than preventing the main config from loading. Windows file modes do not provide equivalent ACL protection; restrict access to the profile directory using Windows permissions.
 
+### Third-party relays and custom auth headers
+
+Providers created by `mcode provider add` are stored under `custom_provider` in the active profile's `config.yaml`. That tree is the supported home for any endpoint other than the official MiniMax API; `minimax_api` is reserved for the official endpoint and keeps its fixed authentication scheme.
+
+Authentication follows the API format: `openai-completions` and `openai-responses` send `Authorization: Bearer <key>`, and `anthropic-messages` sends the key as `x-api-key`. A relay that expects Bearer authentication on an OpenAI-compatible endpoint therefore needs no extra configuration. If a relay only exposes an Anthropic-compatible endpoint (`/v1/messages`) and requires Bearer authentication, add explicit headers to the provider entry in `config.yaml`:
+
+```yaml
+custom_provider:
+  my-relay:
+    options:
+      apiKey: sk-relay-key
+      baseURL: https://relay.example.com
+      headers:
+        Authorization: Bearer sk-relay-key
+    models:
+      MiniMax-M2: {}
+```
+
+`api` can be omitted and defaults to `anthropic-messages`. Configured headers are sent on both connection tests and conversation requests, and they are part of the connection-test fingerprint, so editing them retires a cached test verdict. The default `x-api-key` header is still sent alongside; the relay must tolerate rather than reject it. Header values are stored in plaintext like `apiKey` and are reported by name only in provider views. After editing the file, verify with `mcode provider list` and `mcode provider test <provider-id>`.
+
 [Live acceptance](verification.md) separately verified MiniMax Token Plan and one configured BYOK provider. This is not a guarantee for every compatible service.
 
 ## 3. Search and image input
 
+For a custom BYOK model, declare image input support explicitly when adding the
+provider. Use this only if the selected provider endpoint and model accept images:
+
+```bash
+pnpm mcode provider add --name my-vision-provider --base-url https://example.com/v1 \
+  --api-format openai-completions --model my-vision-model \
+  --api-key-env MCODE_PROVIDER_API_KEY --support-image --use
+```
+
+`--support-image` applies to every repeated `--model` and saves
+`capabilities.support_image: true`. Without it, adding an unknown custom model
+does not infer image support from its name. `--use` and `provider test` check
+connectivity with a text request; they do not verify vision support.
+
+For an existing provider, close MCode and add the capability to the matching model
+in the active profile's `config.yaml` (normally `~/.minimax/config.yaml`; see
+[Accounts and data](installation.md#accounts-and-data) for profiles and overrides).
+Merge this fragment into the existing provider and keep its other settings:
+
+```yaml
+custom_provider:
+  my-vision-provider:
+    models:
+      my-vision-model:
+        capabilities:
+          support_image: true
+```
+
+The existing `modalities: { input: [text, image], output: [text] }` model setting
+also enables image input. Either declaration is sufficient. To make a model
+text-only again, remove `image` from `modalities.input` and remove or set
+`capabilities.support_image` to `false`. Generic `attachment: true` alone does not
+declare image input support. Restart MCode after editing configuration and select
+the configured model with `/provider`, or use `exec --model` for a single run.
+
 After signing in to MiniMax, try a task that explicitly requires search:
 
 > Use web_search to find the official Node.js test runner documentation. Summarize how to run tests and include the source URL. If the tool is unavailable, say so.

docs/installation.md

@@ -60,6 +60,52 @@ To locate data safely:
 
 For tests, explicitly set `MINIMAX_DATA_DIR` to a temporary directory to keep normal sessions separate. Use `$env:MINIMAX_DATA_DIR = 'C:\path\to\test-profile'` in PowerShell or `export MINIMAX_DATA_DIR=/path/to/test-profile` in a POSIX shell.
 
+## macOS terminal shortcuts: Ghostty Option+M
+
+In the composer, `Alt+M` (`Option+M` on macOS) cycles permission modes through Ask, Auto, and Full access. Its binding ID is `composer.cycle-permission`. `Shift+Tab` toggles **Plan mode**, a separate setting. You can also use `/permission` to choose a permission mode and `/permission status` to inspect it without an Option shortcut.
+
+If `Option+M` inserts `µ` instead, the terminal is sending text rather than the expected Alt shortcut. MCode preserves literal `µ` as text; it cannot safely infer a permission change from that character. The terminal must forward the modifier, or you can choose another binding.
+
+### Configure Ghostty's Option key
+
+Ghostty's [`macos-option-as-alt`](https://ghostty.org/docs/config/reference#macos-option-as-alt) default is **keyboard-layout dependent** when unset, not always `false`. The current reference lists U.S. Standard and U.S. International as defaulting to `true`; other layouts default to `false`. Check the active macOS input source and the configuration loaded by your installed Ghostty version.
+
+Set this in your Ghostty configuration to send both Option keys as Alt:
+
+```ini
+macos-option-as-alt = true
+```
+
+Use `left` or `right` instead of `true` to reserve only that Option key for shortcuts and retain Unicode composition on the other side. Use the configured side when testing `Option+M`.
+
+Follow Ghostty's [configuration locations and precedence](https://ghostty.org/docs/config#file-location): macOS-specific files load after XDG files, and later settings can override earlier ones. Inspect the loaded setting from the same Ghostty installation, without `--default`:
+
+```bash
+ghostty +show-config | awk '/^macos-option-as-alt[[:space:]]*=/'
+```
+
+If `ghostty` is not on `PATH`, use the executable inside your installed app bundle, typically `/Applications/Ghostty.app/Contents/MacOS/ghostty`. An empty or omitted value is not proof of `false`; unset behavior depends on the layout. After editing, reload Ghostty's configuration from its menu (default `Cmd+Shift+,`) and open a new terminal. This is separate from MCode's `/reload` command.
+
+### Keep Option for text and customize MCode instead
+
+While idle, open `/hotkeys`, select `composer.cycle-permission`, press `Enter`, press `Ctrl+X`, then press `Enter` to save. `Ctrl+X` is unused by the current default MCode bindings; check any custom terminal or multiplexer bindings too. Do not choose `Ctrl+M`: terminals commonly encode it identically to `Enter`.
+
+Alternatively, merge this entry into `<data-dir>/tui/keybindings.json`, preserving any existing entries:
+
+```json
+{
+  "composer.cycle-permission": "ctrl+x"
+}
+```
+
+For the default profile this is `~/.minimax/tui/keybindings.json`; [profiles and data-directory overrides](#accounts-and-data) change the path. Create the `tui` directory if needed. After editing the file, run `/reload` while idle with no pending interaction or queued message, or restart MCode. This replaces `Alt+M`; use `["alt+m", "ctrl+x"]` as the value to retain both bindings. `/hotkeys` shows the effective binding.
+
+### Verify the result
+
+With an idle composer and no open picker or permission prompt, note `/permission status`, press the configured shortcut, and check the status again. The mode should advance without inserting text. Restore your intended permission mode through `/permission` afterward. Confirm that `Shift+Tab` still toggles Plan mode and that pasting `µ` still inserts text.
+
+If it still fails, record `mcode --version`, Ghostty's version, active keyboard layout, which Option key you pressed, the setting above, the effective `/hotkeys` entry, and whether tmux or SSH is involved. Compare with a direct Ghostty session. MCode supports both legacy ESC-prefixed Alt input and extended keyboard reports; a shell's raw-key display alone does not establish what Ghostty sends after the TUI negotiates its keyboard protocol.
+
 ## Update or remove
 
 Save your changes, fetch a reviewed revision with Git, then repeat the frozen install and build. A source installation does not automatically become an official npm installation.

docs/maintainers.md

@@ -11,6 +11,29 @@
 
 These are repository maintenance responsibilities, not an assertion that an account owns the product's copyright. Every PR that requires approval needs a reviewer other than the author, and security reports need a working private response path. Review maintainer additions and policy changes in a PR; do not silently grant repository access.
 
+## Issue triage
+
+Published source covers the standalone CLI (interactive TUI, headless and ACP). The issue tracker also collects Desktop app feedback; accepting those reports does not expand the published source scope.
+
+Everyone may report bugs, request features, ask questions, and flag documentation problems. Code and documentation PRs are accepted only from repository collaborators. Opening an issue or having a proposal accepted does not grant collaborator access or permission to submit a PR. Do not use `good first issue` or `help wanted` while external PRs are closed.
+
+Use these labels independently:
+
+| Category | Labels and meaning |
+| --- | --- |
+| Type | `bug`: reported malfunction; `enhancement`: requested improvement; `documentation`: missing or incorrect docs; `question`: usage or project-policy question |
+| Product | `desktop`: Desktop app; `cli`: standalone mcode, including TUI, headless, ACP and its source builds/tooling |
+| Triage | `needs-triage`: awaiting maintainer assessment; `needs-info`: a maintainer has asked the reporter for specific missing information |
+
+1. Start with [needs-triage](https://github.com/MiniMax-AI/minimax-code/issues?q=is%3Aissue+is%3Aopen+label%3Aneeds-triage) and [unlabelled issues](https://github.com/MiniMax-AI/minimax-code/issues?q=is%3Aissue+is%3Aopen+no%3Alabel). Templates add a type and `needs-triage`; reports created through other routes may need both added manually. A type label describes the report and does not confirm a defect or commit to implementation.
+2. Confirm one primary type and apply `desktop` or `cli` from the report. Product dropdowns stay in the body; maintainers apply product labels manually. Use both only when the report establishes that both are affected. Leave product labels unset for unclear reports or repository-policy questions. The existing `tui` label may supplement `cli` for interactive-terminal-specific issues; it does not cover headless or ACP.
+3. If information is missing, ask for the specific version, reproduction, or expected behavior, replace `needs-triage` with `needs-info`, and retain any known type/product labels. When the reporter replies, remove `needs-info` and reassess. Do not automatically close issues for inactivity.
+4. Once assessed, remove `needs-triage` and record the next step in a comment. Assign a collaborator when someone takes responsibility; classification alone does not promise a fix or release date. For duplicates, link the original issue before closing with `duplicate`. For declined requests, explain the reason before closing with `wontfix`.
+
+Route security reports through [Security](../SECURITY.md). Do not ask for secrets or unredacted logs in public issues. Label setup is a separate repository operation: ensure the labels above exist before merging template changes, since forms cannot create missing labels. The Feishu workflow sends notifications only; it does not perform triage.
+
+This small label set draws on the type and information-request handling in [VS Code's triage guide](https://github.com/microsoft/vscode/wiki/Issues-Triaging) and the `needs-triage`/`needs-info` states used by [Ruff](https://github.com/astral-sh/ruff/labels). Contribution eligibility follows this repository's policy.
+
 ## Review and merge
 
 1. Repository collaborators open an issue for substantial scope or submit a focused feature-branch PR. Other users may open issues to discuss ideas and proposals. Security details follow the private reporting process.

docs/tui-capabilities.md

@@ -46,4 +46,20 @@ Every collected diagnostic artifact, including prioritized session artifacts, is
 
 There is no raw-attachment upload option in this flow. Prompts, conversation text, tool arguments/results, command output, workspace excerpts, raw errors and unknown fields are excluded even if they contain no recognizable credential pattern. This intentionally reduces diagnostic detail: reproducing an exact response or inspecting an original stack is not possible from these uploads. Future raw attachments would require a separate, explicit review and consent surface describing their contents and scope.
 
-The regression tests use temporary synthetic files and intercepted HTTP only. They decrypt the final automatic-report request as a receiver would, and unzip the actual feedback PUT body after real session-report collection. They do not validate production ingestion, retention policies, live services, or other platforms.
+## Interactive startup model
+
+Use `mcode -m <provider-id>/<model-id>` or
+`mcode "Fix the failing tests" --model <provider-id>/<model-id>` to select the
+model for the startup Session. References use the same syntax as `exec --model`,
+including `custom_provider:<id>/<model-id>`, model IDs containing `/`, and the
+optional `#variant` suffix. The provider must already be configured and the model
+must be available to Runtime; invalid references fail before the initial prompt runs.
+
+Without a resume option, this creates a new Session even when no prompt is given.
+With `--session <id>` or `--continue`, it updates the opened Session's model before
+submitting the prompt. The selection is saved with that Session, so later turns
+and resumes keep it. It does not change the global default, and `/new` returns to
+the configured default. Omitting `--model` preserves existing startup behavior.
+The untargeted `--session` picker cannot be combined with `--model`; provide an ID
+or use `--continue` instead. If opening or continuing a Session fails, the override
+and initial prompt are not applied.

packages/agent-modules/permission/src/tools/fs-permission.ts

@@ -687,8 +687,10 @@ function isProtectedRuntimeRead(
     // at credentials must not turn its destination into a trusted asset root.
     const readableAsset =
       within(canonicalPath, path.join(canonicalRoot, 'skills')) ||
+      within(canonicalPath, path.join(canonicalRoot, '.builtin-skills')) ||
       (context.agentName != null &&
-        within(canonicalPath, path.join(canonicalRoot, 'agents', context.agentName, 'workspace'))) ||
+        (within(canonicalPath, path.join(canonicalRoot, 'agents', context.agentName, 'workspace')) ||
+          within(canonicalPath, path.join(canonicalRoot, 'agents', context.agentName, 'skills')))) ||
       isInternalWhitelistedPath(canonicalPath, {
         ...context,
         dataDir: canonicalRoot,