docs(changelog): sync 3.19.0 security section with the release notes

CHANGELOG.md

@@ -7,7 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [3.19.0] - 2026-07-30
 
-Development since v3.18.0 is headlined by a security-hardening wave and a proxy correctness fix with immediate field impact. On the security side (#5811 and follow-ups): skill installs from GitHub repositories are hardened against zip-slip and repository-coordinate path traversal with hard archive limits; a Gemini common-config credential leak is closed and a one-time startup cleanup scrubs keys that already leaked into other providers' configs; imported SQL backups now run under a SQLite authorizer that denies `ATTACH` and every other statement that can reach outside the import database; common-config snippet handling no longer follows `__proto__` into the global prototype; terminal launch escapes project paths with POSIX single quotes so a directory name can no longer inject commands; and `ccswitch://` import confirmations show every argument, env var, URL, and the complete usage-script body — nothing hidden by truncation, credential-shaped values masked — flag risky values, and import usage scripts disabled by default. On the proxy side, tool-result images are lifted out of stringified tool text and re-emitted as native media on every conversion bridge, ending the ~9,000× token inflation that pushed Codex sessions past their context window after a few screenshots (#4465, #5663). Usage statistics gain opt-in automatic price sync from models.dev with a persistent local override file (#5734), plus coverage of Grok CLI's official OAuth mode imported from session logs and SuperGrok subscription quota on provider cards. Rounding it out: in-app updates are served from a Cloudflare R2 mirror at `dl.ccswitch.io` with GitHub as fallback, Codex usage import reuses parsed parent rollouts across forked sessions (#5626), preset defaults advance to Claude Opus 5, GPT-5.6 Sol, and Gemini 3.6 Flash, the OpenClaw Kimi For Coding base URL is corrected, and the GPT-in-Claude-Code routing guide is now available in English and Japanese.
+Development since v3.18.0 is headlined by a security-hardening wave and a proxy correctness fix with immediate field impact. On the security side (#5811 and follow-ups): skill installs from GitHub repositories are hardened against zip-slip and repository-coordinate path traversal with hard archive limits; a Gemini common-config credential leak is closed and a one-time startup cleanup scrubs keys that already leaked into other providers' configs; imported SQL backups now run under a SQLite authorizer that denies `ATTACH` and every other statement that can reach outside the import database; common-config snippet handling no longer follows `__proto__` into the global prototype; terminal launch escapes project paths with POSIX single quotes so a directory name can no longer inject commands; and `ccswitch://` import confirmations show every argument, env var, URL, and the complete usage-script body — nothing hidden by truncation, nor by a Base64 variant the dialog could not decode, credential-shaped values masked — flag risky values, and import usage scripts disabled by default. On the proxy side, tool-result images are lifted out of stringified tool text and re-emitted as native media on every conversion bridge, ending the ~9,000× token inflation that pushed Codex sessions past their context window after a few screenshots (#4465, #5663). Usage statistics gain opt-in automatic price sync from models.dev with a persistent local override file (#5734), plus coverage of Grok CLI's official OAuth mode imported from session logs and SuperGrok subscription quota on provider cards. Rounding it out: in-app updates are served from a Cloudflare R2 mirror at `dl.ccswitch.io` with GitHub as fallback, Codex usage import reuses parsed parent rollouts across forked sessions (#5626), preset defaults advance to Claude Opus 5, GPT-5.6 Sol, and Gemini 3.6 Flash, the OpenClaw Kimi For Coding base URL is corrected, and the GPT-in-Claude-Code routing guide is now available in English and Japanese.
 
 **Stats**: 38 commits | 132 files changed | +14,926 insertions | -1,415 deletions
 
@@ -34,7 +34,6 @@ Development since v3.18.0 is headlined by a security-hardening wave and a proxy
 - **Grok Build Costs Overstated by the Cost Backfill**: The routine that recomputes missing costs treated only Codex and Gemini as providers whose reported input token counts already include cache reads. Grok Build uses the same convention, so any Grok Build row repaired by the backfill was priced on the full input count with cache reads billed a second time. The cache-inclusive provider set is now defined in one place and shared by the routing logger, the cost calculator, and the backfill, so the three can no longer disagree.
 - **Hand-Edited Config Files No Longer Crash the App or Silently Swallow Edits**: A `~/.codex/config.toml` in which `mcp_servers` exists but is not a table made MCP sync panic mid-switch, after the database and live-config writes had already committed — leaving a half-applied switch; non-table values are now normalised to an empty table with a logged warning, in both the Codex and GrokBuild writers. Inline-table forms (valid TOML) had the mirrored problem: MCP removal silently did nothing while the UI reported success, and a `base_url` edit was written at a level Codex never reads — both now handled. An `opencode.json` whose root, `provider`, or `mcp` section is an array or scalar no longer panics; such files are rejected with an error rather than rebuilt, so the user's own settings are not destroyed. (#5811)
 - **Proxy Transforms Survive Malformed Upstream Responses**: A non-object `message` or `content_block` in an Anthropic SSE stream, or a buffered body that is a top-level JSON array or scalar (what a gateway that ignores `stream: true` returns), could abort the local proxy instead of producing an error; the stream now bails out with a proper failed event. A malformed `content_block` header is additionally recovered as a text block — sanitising it to an empty object stopped the panic but silently dropped everything that followed — so the deltas after a garbled header come through, with a warning logged. (#5811)
-- **URL-Safe Base64 Payloads No Longer Render Blank in Deep-Link Confirmations**: Deep links whose Base64 payload used the URL-safe alphabet decoded correctly in the backend but not in the confirmation dialog, so usage scripts and system prompts showed opaque Base64 and MCP configs rendered as "0 servers" while the backend still imported the real entry. The frontend decoder now normalises the URL-safe alphabet before decoding, so what the confirmation shows always matches what will be imported.
 - **OpenClaw Kimi For Coding Base URL Corrected**: The OpenClaw preset pointed at `https://api.kimi.com/v1`, the general platform endpoint, rather than the endpoint the Kimi For Coding subscription is served from, so the preset did not work with a coding-plan key. The base URL is now `https://api.kimi.com/coding/v1`, with the form placeholder and default updated to match.
 
 ### Security
@@ -43,6 +42,7 @@ Development since v3.18.0 is headlined by a security-hardening wave and a proxy
 - **Skill Repository Install Hardening**: Installing or browsing skills from a GitHub repository could write files outside the intended directory: archive entries were joined onto the destination path without normalisation (zip-slip), and repository coordinates were never validated, so a branch like `../../../releases/download/v1/evil` retargeted the download at an arbitrary release asset — and since skill repositories can arrive via an untrusted `ccswitch://` deeplink and are enabled by default, merely opening the Skills panel could trigger the download. Skill `directory` values from restored backups, sync snapshots, and import-from-apps were likewise joined raw, letting uninstall `remove_dir_all` outside the managed skills directory. Every sink now validates the directory name, repository owner/name/branch are whitelisted at the single download convergence point, and extraction is capped (10,000 entries, 512 MB written, 128 MB download, 4 KB symlink targets, self-referential symlinks rejected), with the new errors localized in all four languages. (#5811)
 - **Deep-Link Import Confirmations Show the Whole Payload (Credentials Masked) and Flag Risky Values**: The `ccswitch://` MCP import confirmation previously rendered a single truncated `Command:` line and showed neither `args`, `url`, nor `env` — so a link carrying `command: "sh"` with `args: ["-c", "curl …|sh"]` and an `LD_PRELOAD` variable displayed as a harmless `sh`, yet was written to the live MCP files on confirm. The dialog now renders command, each argument, URL, and env vars on separate lines that wrap instead of truncating, so nothing is clipped out of view (env values whose name contains TOKEN, KEY, SECRET, or PASSWORD are shown masked, as a prefix plus asterisks), and highlights values that warrant a second look: shell interpreters invoked with inline-command flags, env vars that change how a process loads code (`LD_*`, `DYLD_*`, `NODE_OPTIONS`, `PYTHONPATH`, `PATH`, proxy vars), and endpoints pointing at loopback, private-range, or cloud-metadata addresses. The markers are advisory and never block an import — a local Ollama endpoint is ordinary usage. The provider confirmation gains the same treatment, and the "will be written immediately" warning is now shown unconditionally instead of being gated on a link-controlled field.
 - **Usage Scripts From Deep Links Arrive Disabled and Their Code Is Shown**: A usage-query script imported through a deep link is JavaScript that runs on every usage query, and it was previously possible to acquire one without ever seeing it: the backend treated the mere presence of code as a decision to enable it, and the dialog showed only an Enabled/Disabled badge. Scripts now default to disabled — a link must explicitly carry `usageEnabled=true` — and the confirmation displays the complete decoded script in a scrollable code block with a warning that it will execute once enabled. Decoding failures fall back to displaying the raw payload, so a malformed script can never read as "there is no script".
+- **URL-Safe Base64 Payloads Could Blank the Confirmation Dialog Entirely**: The two entries above make the confirmation show more; this one is about the confirmation showing *nothing*. The backend accepts four Base64 variants including the URL-safe alphabet (RFC 4648 §5), while the frontend's `atob` accepts only the standard one and, on failure, returned its input unchanged instead of signalling an error — so for the same payload the backend decoded and imported successfully while the frontend held an undecodable string. Usage scripts and system prompts rendered as opaque Base64, and MCP configs were worse: `JSON.parse` threw, the component swallowed it, and the dialog rendered **"0 servers" with an empty list while the backend wrote the real entry into the live MCP files**. Substituting a single `/` for `_` in the payload was enough — the dialog went blank, the import stayed fully functional, and the complete-payload display added above was defeated along with it. The frontend decoder now normalises the URL-safe alphabet before decoding, so the confirmation always matches what will be imported, and the shared decoder gained its first unit tests — with in-test preconditions asserting each sample really exercises the URL-safe path rather than happening to encode identically under both alphabets. Affected every release from v3.8.0 onward; see Upgrade notes for what to check.
 - **SQL Import Rejects Statements That Reach Outside the Import Database**: Importing a database backup validated only the file's header comment and then handed the entire text to `execute_batch`, so a crafted backup could `ATTACH DATABASE` and create a SQLite file anywhere the user can write — a side effect that landed even when the import as a whole failed, and one reachable through WebDAV/S3 sync snapshots too. A SQLite authorizer is now installed for the duration of the external batch only, denying `ATTACH`/`DETACH`, `VACUUM`, virtual-table creation, and any action SQLite reports as unknown (fail-closed), with PRAGMAs limited to the two the exporter actually emits.
 - **Prototype Pollution in Common-Config Snippet Handling**: The three walkers that apply, remove, and compare common-config snippets all followed `__proto__` into the global `Object.prototype`, so a malicious snippet — reachable via WebDAV/S3 sync, since the `settings` table is overwritten by whatever the remote sends — could write attacker-chosen values onto the global prototype the moment a provider form was opened. All three walkers now skip `__proto__`, `constructor`, and `prototype`; the "common config applied" comparison also now requires own properties, fixing a visible quirk where `{"__proto__":{}}` counted as a subset of every config.
 - **Command Substitution via Project Directory Names on Terminal Launch**: Resuming a session in an external terminal built the `cd` line with double-quote escaping, which stops spaces but not `$(…)`, backticks, or `$VAR` — and the value is the real project path recorded in the CLI's session history, which may legitimately contain those characters. Any project folder named that way executed the embedded command in the user's terminal on Resume. The three launchers that build a shell line — Terminal.app, iTerm, and kitty — now use POSIX single-quote escaping, where nothing expands, including through the AppleScript quoting layer; Ghostty, WezTerm/Kaku, and Alacritty were already safe because they pass the directory as its own argument.
@@ -63,6 +63,7 @@ Development since v3.18.0 is headlined by a security-hardening wave and a proxy
 
 - No database schema migration in this release — `SCHEMA_VERSION` stays at 16.
 - On first launch after upgrading, a one-time Gemini common-config cleanup runs before the usual config extraction. Some Gemini providers may afterwards report a missing API key: entries are removed by credential-style key name plus exact value match, so this is usually another provider's credential that leaked through the shared snippet (the provider's own original value was overwritten when the leak occurred and cannot be recovered) — but a key you intentionally reused across several Gemini providers is removed from all of them as well. Either way, rotate and re-enter the key: treat anything that sat in the shared Gemini snippet as exposed. An audit record listing the removed key names and affected provider ids — never the values — is stored in the `settings` table under `gemini_common_config_scrub_audit_v1`.
+- If you have ever imported an MCP server from a `ccswitch://` link, it is worth auditing once. Before this release the MCP import confirmation could fail to show what it was about to write: arguments and env vars were not rendered at all (`command: "sh"` with `args: ["-c", …]` displayed as a harmless `sh`), and a URL-safe Base64 payload made the whole list render as "0 servers" — while the backend wrote the real entry to the live MCP files in both cases. Both defects affected **every release from v3.8.0 onward** and are fixed here. Exploitation required opening an attacker-supplied link and confirming the import, so most users are unaffected; if you did open an MCP import link from a source you do not fully trust, review the entries in the MCP panel, or directly in `~/.claude.json` under `mcpServers` (Codex: `mcp_servers` in `~/.codex/config.toml`), and confirm you recognise each one — MCP servers are executed as subprocesses the next time the CLI starts.
 - Deep links carrying a usage-query script now import it disabled unless the link explicitly sets `usageEnabled=true`. Links that relied on auto-enablement (for example some partner one-click setup links) will import the script but leave usage querying off — enable it in the provider editor after reviewing the code.
 - The new default models apply only to newly added providers — providers already saved keep the model IDs they were created with. For Claude Desktop, the current opus route ID advances to `claude-opus-5` with `claude-opus-4-8` moving into the legacy slot, so opus route IDs stored in existing configs still resolve through the compatibility alias.
 - New built-in pricing rows (`claude-opus-5`, `gemini-3.6-flash`, `grok-4.5-build`) are appended insert-if-absent on next launch — this seeding never overwrites a price you edited. A `~/.cc-switch/model-pricing.json` file is created (empty) to hold manual overrides and deletions made from this version onward — earlier price edits are not migrated into it, so re-save any you want protected against a database rebuild. Automatic models.dev sync stays off until you enable it; once enabled, it is the one path that does overwrite matching prices, built-in and hand-edited alike.