Repository navigation
Conversation
- Menu option 7 checks npm for newer Claude Code and Agent SDK releases and restores pins if the install fails. - Map custom model IDs via modelPicker/behavesAs and declare CLAUDE_CODE_MAX_CONTEXT_TOKENS. - Hard-disable extended thinking for local adapters with CLAUDE_CODE_DISABLE_THINKING and MAX_THINKING_TOKENS=0. - Use a neutral adapter error message and load user settings in the dashboard. - Keep the newer runtime resume/logging and Windows folder picker fixes.
Show an X in the topbar for the open conversation with a green hover matching the theme toggle. Deleting removes the dashboard mirror plus any Claude Code JSONL under data/claude, supports legacy data/chats files, and refuses active runs. Stale nativeSource paths from older installs are skipped instead of blocking deletion, and files outside the history root are never unlinked.
The row trash was hidden again as soon as syncNative marked a session as an imported Claude conversation, producing a brief hover-only flash after the first turn. Deletion is permanent for all conversation types now, so every row keeps the control.
…mprovements Node.js - Bump the bundled runtime from 24.21.0 to 26.13.1 (LTS) in the Windows and POSIX launchers, the CI matrix, package.json engines, and the README. Session resume - RESUME.bat and resume.sh are standalone entry points that never install, repair, update, or reconfigure anything. - Argument parsing lives in lib/resume.mjs: positional or --resume session ID, --cwd or positional directory, --safe/--unrestricted overrides, with duplicate/missing/unknown-argument validation. Permissions - Bypass permissions is stored as preferences.bypassPermissions and defaults to ON. START.bat/start.sh "permissions on|off|status" and menu option 8 change it permanently; --safe and --unrestricted change it for one launch. - The studio follows the same stored default. Portability - Resolve Git Bash from CLAUDE_CODE_GIT_BASH_PATH, the bundled copy, standard install paths, or a git.exe on PATH (including MSYS2 layouts). PortableGit is downloaded only when no Git for Windows is installed. - Redirect HOME, USERPROFILE, APPDATA, LOCALAPPDATA, and the XDG folders of the session process into data/. - Start the bundled Ollama server and a speed proxy automatically and route the session through it. The proxy handles /v1/messages as well as /chat/completions; only processes started by that launch are stopped. Fixes - Imported native sessions lacked an events array, so the first turn failed with "Cannot read properties of undefined (reading 'push')" and left the session locked as already running. - Adapters rejected Claude Code tool-search results as "Unsupported content: tool_reference". - Install canUseTool only in permission modes that consult it, removing the CLAUDE_SDK_CAN_USE_TOOL_SHADOWED warning at its source. Tests - 54 -> 84 test cases, including regression coverage for the three fixes. Co-Authored-By: Claude Code <noreply@anthropic.com>
A session runs with HOME, USERPROFILE and APPDATA redirected into data/ so it leaves no footprint on the host. That also hid the user's ~/.gitconfig, so an agent's git commit failed with "Author identity unknown" and credential helpers were unavailable. - Resolve the host Git configuration from every available profile location: an existing GIT_CONFIG_GLOBAL, USERPROFILE, HOMEDRIVE+HOMEPATH, then HOME (~/.gitconfig or ~/.config/git/config). Trying all candidates keeps detection working when a session starts from an already redirected environment. - Pass it to the session process as GIT_CONFIG_GLOBAL. Detection runs on the unredirected parent environment, before HOME is rewritten. - PORTABLE_AI_NO_HOST_GIT_CONFIG=1 opts out and keeps Git confined to the drive. This is a deliberate exception to zero-footprint: git config --global inside a session writes to the host file. README and CHANGELOG document the trade-off. Tests: 84 -> 86. Co-Authored-By: Claude Code <noreply@anthropic.com>
The changelog compared against rainman74/ClaudeCode-Portable at 247aeb1, which is this fork, not the project it came from. Measured against the real upstream (techjarves/ClaudeCode-Portable at b3660fd) the fork has nine commits of its own, including two community pull requests and the update-check work. - State the actual baseline and note the one later upstream commit (dfdc58d, a sponsor button) that this fork does not include. - Add an "Adopted community pull requests" section crediting PR techjarves#43 (lux-liang, IME-safe composer) and PR techjarves#47 (VikasRajput-18, HTTP for private LAN hosts). Both are still open upstream; the included code matches the PR branches. - Document the runtime update check with pin rollback, custom model-window and model-picker mapping, and the dashboard conversation delete button. - Correct the test baseline from 54 to 50 tests, measured at the merge base. Co-Authored-By: Claude Code <noreply@anthropic.com>
Studio - Paste a screenshot straight into the composer (Ctrl+V / Cmd+V). Plain text pastes are left untouched. - Drop files anywhere in the studio: the composer highlights while dragging and a drop outside never navigates away from the page. - The picker, paste and drop share one implementation with one set of limits. Attachments - Detect the real image format from magic bytes (PNG, JPEG, GIF, WebP) instead of trusting the reported MIME type; pasted screenshots often arrive as application/octet-stream with no usable name. - Store the type in a metadata sidecar so it survives a restart. - Images get a 5 MB limit of their own, because they travel to the model as base64 and providers cap them below the generic 10 MB. Agent - Attached images are sent through the SDK's streamed prompt form as real image blocks, the same thing the Claude Code terminal does with a clipboard screenshot, so a vision model sees them without opening anything first. - Messages without images keep the original single-string prompt. - The read loop stops at the result event, which the streamed form requires. Verified end to end: the official runtime and SDK deliver the image block to the provider, and a real Chromium run confirms paste, text paste, drop, and the oversized-image rejection. Terminal paste (Ctrl+V) already works through the Claude Code binary; verified that this project's Git PATH and redirected HOME do not interfere with it. Tests: 86 -> 90. Co-Authored-By: Claude Code <noreply@anthropic.com>
Browser automation now lives next to the runtime and travels with the drive: the npm package in engine/playwright, the headless Chromium build in engine/playwright-browsers. Nothing is written to the user profile. - First launch installs the pinned package and browsers after the Claude Code runtime; later launches only check and fetch what is missing, so a complete setup needs no network. - Menu option 5 (install/update) repairs both the runtime and Playwright. - Menu option 7 (check-update) reports newer Playwright releases next to the Claude Code updates and installs them with the same pin rollback: a failed install restores both pins. - The version is pinned in tools/runtime-manifest.json because it decides the Chromium build revision. --only-shell keeps the download at ~300 MB instead of ~750 MB, and --no-remove protects browsers copied from another machine. - A Playwright failure is a warning, never a startup error. --offline and PORTABLE_AI_NO_PLAYWRIGHT=1 skip it, START.bat status reports its state. - Sessions receive PLAYWRIGHT_BROWSERS_PATH, so browser automation started by the agent uses the portable builds instead of downloading into the profile. Verified on this machine: the real install produced Playwright 1.65.1 with Chromium headless shell 1255, a repeated run is a 65 ms no-op without network, and a screenshot renders through the portable browsers. Tests: 90 -> 99. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Claude Code binds image paste to alt+v on Windows and ctrl+v only on macOS, Linux, and WSL. The classic Windows console also swallows Ctrl+V for its own paste, which is why pressing it with a screenshot in the clipboard did nothing. - The launcher merges ctrl+v -> chat:imagePaste into the profile's keybindings.json, additively: existing bindings and other contexts survive, an explicit user choice (including an unbind) stays, and a file with a syntax error is left alone instead of being rewritten. - Without an image the action falls back to a text paste, so the extra binding cannot break pasting text. - README documents Alt+V as the Windows shortcut that works today in the classic console, and the two ways to make Ctrl+V arrive there (Windows Terminal, or CtrlKeyShortcutsDisabled). Tests: 99 -> 106. Co-Authored-By: Claude Code <noreply@anthropic.com>
The session list was rewritten with innerHTML on every refresh. When a refresh landed between pressing and releasing the mouse - for example while a session was still loading - the button the press started on was destroyed and the browser delivered no click at all, so the trash had to be pressed twice. - Skip the write entirely when the generated markup is identical; the redundant resetSession()+refreshSessions() pair no longer touches the DOM. - Defer any re-render requested while a press is held on the list until the release, then run it right after the click. - Delegate row, project-toggle and delete handlers on the list container so a rebuilt row can never carry a stale handler or lose one. Verified in Chromium: a re-render requested during the press no longer swallows the click, while the previous behaviour (raw innerHTML replacement during the press) does drop it. Co-Authored-By: Claude Code <noreply@anthropic.com>
Stamp the changelog as 2.0 and correct the test count in the summary table (90 -> 106, matching the Tests section). Co-Authored-By: Claude Code <noreply@anthropic.com>
A session saw only its workspace and its own configuration, so nothing told the agent that browser automation ships with the drive - it would have had to stumble over engine/playwright. - Write a skill to the profile directory (skills/portable-playwright/SKILL.md) whenever Playwright is installed. Claude Code lists the name and description in the system prompt, so the agent knows the capability exists and loads the details only on demand. - The skill names the absolute CLI and browser paths (a session's working directory is usually the user's project), the screenshot invocation, script usage, the npx warning and the headless-only limitation. - Without Playwright nothing is written and a stale copy is removed; a skill the user added to the same folder is never touched. - Verified with a real session: the SDK skills dialog reports portable-playwright with source "user". Tests: 106 -> 110. Co-Authored-By: Claude Code <noreply@anthropic.com>
Studio - Paths in replies and tool output become buttons that open with the user's application: D:\..., C:/..., UNC paths and file:/// URLs. - The click goes to the token-protected POST /api/open, which accepts only absolute paths that exist and hands them to explorer.exe / open / xdg-open. The path is one argument and never goes through a shell. - A path with spaces is only recognised inside quotes; in running text a space is far more often the sentence. Trailing punctuation stays outside. - Code blocks keep plain text so copying still works. Web links open in a new tab and never navigate the studio away. Terminal - Claude Code emits OSC 8 hyperlinks only in terminals it recognises. Windows Terminal exports WT_SESSION and works automatically; the classic console does not, so nothing was clickable there. - PORTABLE_AI_HYPERLINK=1 / PORTABLE_AI_HYPERLINKS=1 exports FORCE_HYPERLINK as an opt-in for other terminals. Verified in Chromium: paths render as focusable buttons, clicking one opens the real file, a missing path, a relative path and an empty path are all refused with a clear error, and web links keep target=_blank. Tests: 110 -> 120. Co-Authored-By: Claude Code <noreply@anthropic.com>
Move the entries under [Unreleased] into a dated 2.1 section and bump the package version to match, so the changelog and the metadata agree. Co-Authored-By: Claude Code <noreply@anthropic.com>
The rule skipped every CODE element, so the most common way of writing a path — in backticks — never became a button. Only fenced blocks need that protection, because their text is meant to be copied. - Inline code is now treated as one value: the whole span becomes one button when it holds exactly one path. Backticks mark the exact value, so a path with a space (D:\Apps\Claude Code\x.jpg) stays in one piece instead of being cut at the space, which is what the running-text rule would do. - A span with two paths, or with no path at all, stays plain text. - Fenced blocks are still skipped, so the Copy button keeps working. - The button resets its own border, background and margin so it fits inside inline code without moving the line. Verified in Chromium with the reported message: both inline paths render as buttons with the full value in data-open-path, including the one with "Claude Code" in the folder name, and the fenced block stays plain. Tests: 120 -> 122. Co-Authored-By: Claude Code <noreply@anthropic.com>
The launcher has been writing ctrl+v -> chat:imagePaste into keybindings.json since the image-paste support landed. The file is correct and is loaded, yet the shortcut did nothing. The terminal was the cause: - A terminal that claims Ctrl+V for its own paste handles the key and pastes text. Claude Code never sees the keystroke, so no binding can fire. - The classic console hard-codes that and cannot be reconfigured. - Windows Terminal does not bind Ctrl+V by itself: its default for paste is Ctrl+Shift+V. A ctrl+v entry in settings.json takes the key away again. - Alt+V works because no terminal claims it. That is exactly the reported behaviour: text pastes, images do not. README now carries the table and the exact keybindings entry to remove. With Ctrl+V unbound the key reaches Claude Code, image paste works, and text paste moves to Ctrl+Shift+V. The action has no text fallback, which the README says plainly. WT.bat starts the studio inside Windows Terminal, where paths are clickable too because WT_SESSION makes Claude Code emit OSC 8 hyperlinks. Its argument handling is tested for no argument, resume, a profile name and a two-word argument; an empty argument gets its own branch so START.bat is not called with a blank profile name. Co-Authored-By: Claude Code <noreply@anthropic.com>
The terminal documentation and WT.bat were added on the assumption that changing the user's Windows Terminal settings was within scope. It was not: the only thing that belongs in this project is the studio itself, and system configuration on the machine stays untouched. - WT.bat is removed. - The "Clickable paths in the terminal" section is gone from the README. - PORTABLE_AI_HYPERLINKS is gone from README and lib/providers.mjs, so the launcher no longer touches FORCE_HYPERLINK. - The "Why Ctrl+V pasted text but not images" section is gone from both files. The studio's clickable paths stay as they were, including the inline-code fix. Co-Authored-By: Claude Code <noreply@anthropic.com>
The 2.0 changelog said chat:imagePaste "falls back to a normal text paste
when the clipboard holds no image". The binary says otherwise:
yCe(rc(st())).then(Ct => { if (Ct) c(Ct.base64, …) })
There is no else branch. With no image the key does nothing, and the only
other effect is the feedback "No image found in clipboard."
- README states plainly that there is no text fallback and points text
pasting at Ctrl+Shift+V or the context menu.
- A Fixed entry under 2.1 records the correction instead of rewriting the
released 2.0 section, so the published text stays as it was.
- The reg add "HKCU\Console" /v CtrlKeyShortcutsDisabled advice is gone. It
disables console shortcuts for every console window on the machine, which
is a system-wide change this project has no business recommending. Windows
Terminal is left as the one supported way.
Co-Authored-By: Claude Code <noreply@anthropic.com>
The previous commit left the wrong claim standing in the released 2.0 section and recorded the correction only under 2.1. The section is now fixed where the mistake is, so a reader of 2.0 sees what actually happens. - "Without an image the action falls back to a normal text paste" is replaced by what the handler does: it attaches nothing and reports "No image found in clipboard." There is no text fallback. - The CtrlKeyShortcutsDisabled mention is gone from 2.0 as well, matching the README. That registry value disables console shortcuts for every console window on the machine. - The 2.1 Fixed entry now says the 2.0 entry was corrected in place instead of claiming it was left as published. Co-Authored-By: Claude Code <noreply@anthropic.com>
The file existed to feed the GitHub release notes. Those notes are published now, and the 2.0 and 2.1 content is fully preserved on the release pages, so the file no longer earns its place. - Nothing referenced it: no import, no script, no README link, no check. - tools/check.mjs and the 122 tests pass without it. - The deleted text stays reachable in git history if it is ever needed. Co-Authored-By: Claude Code <noreply@anthropic.com>
The thumbnail was fetched from img.youtube.com and both the picture and the "watch the video" link pointed at YouTube, so rendering the README contacted a third party. - The thumbnail is now docs/images/maxresdefault.jpg, linked to the video. - The video link targets docs/videos/xPfZyV20_RY.mp4, which GitHub plays in its own player, so nothing leaves the repository. - The original upload URL is kept as plain text for anyone who prefers to stream it; plain text is not fetched. Both files are also copied into the local working folder so the two trees stay in step and the relative paths resolve there as well. Co-Authored-By: Claude Code <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fffa7d4cf6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Findings from the Codex review on the upstream pull request, five of six addressed here. Clickable paths (both from this release, both broke a documented form): - A file:// URL was turned into a button but POST /api/open passed the URL to isAbsolute(), which is false even on Windows, so every such click failed. The endpoint now decodes it with fileURLToPath, which also unescapes %20 and turns a host into a UNC path. - C:/Users/x/file.txt never matched: both scanner patterns required backslashes, and isFilePath was only consulted after a match. A forward-slash drive pattern was added, with a lookbehind so the `s:/` in `https://` is not read as a drive. Local model stack: - prepareLocalSession replaced a profile's base URL with the proxy but never told the proxy which server to forward to, so a LAN Ollama was silently replaced by 127.0.0.1:11434. The profile's host and port now travel with the proxy through PORTABLE_OLLAMA_HOST/PORT. - proxyStatus accepted any 2xx from /health, so an unrelated service on the proxy port was adopted and received every prompt. It now requires the proxy's JSON shape and, when known, a matching upstream. Launcher: - RESUME.bat hard-coded data\settings.json while lib/paths.mjs honours PORTABLE_AI_DATA_DIR, so a relocated data folder failed the preflight. Left open: the skill cleanup in lib/skills.mjs still removes the whole portable-playwright folder. A user file inside it would be lost; the lower-risk fix is a separate change. Tests: 122 -> 128. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
Thanks for the review — five of the six findings are addressed in Fixed
Not fixed
Tests: 122 → 128. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c6b76b6f74
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Findings from the second Codex pass on the pull request. Three of four are addressed; the image-payload limit is deliberately left open. - An HTTPS Ollama profile was reduced to `http://host:443` by the previous commit, so the proxy sent plaintext to a TLS listener and every turn failed. The scheme now survives into `PORTABLE_OLLAMA_SECURE`, and the proxy uses node:https for it. Covered by a test that talks to a real TLS server. - `modelPicker` was replaced by a shallow spread, dropping any options the profile already had. It is now merged: the managed entry is appended, or replaced in place on later starts so the list does not grow. - RESUME.bat paused unconditionally in its three preflight errors, ignoring PORTABLE_AI_NO_PAUSE, which the README documents for this script. Automation could hang instead of returning an error. Not addressed: the 20 MiB adapter limit can reject a turn carrying three 5 MiB images once base64 and prompt text are added. The right limit depends on how many images should be allowed at once, which is a product decision rather than a defect fix. Tests: 128 -> 131. Co-Authored-By: Claude Code <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 01b0c7f672
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
`RESUME.bat --cwd "D:\Projects\my-app\"` failed with "Working directory not found: D:\Projects\my-app"". Windows hands the argument over with a trailing quote because the backslash escapes the closing one, and that path can never exist: Windows forbids `"` in file names. The parser now drops the stray quote on win32, so the quoted-and-terminated and the plain form both work. POSIX is left alone - there a file name may legitimately end in `"`, and the artifact never appears. Also reproduces the failure for the launcher's own `--cwd "%~dp0"` call, which is why that script now passes `"%~dp0."`. Version 2.2.0. Co-Authored-By: Claude Code <noreply@anthropic.com>
… probes
An explicit `--permission-mode` now overrides the stored bypass default
instead of being combined with `--dangerously-skip-permissions`. The
documented `START.bat cli --permission-mode acceptEdits` reached Claude Code
with two contradicting modes and could run unrestricted, so the flag is
treated like `--safe`: it belongs to this launch and no bypass flag is added.
`--unrestricted` alongside it is rejected rather than silently widening the
request.
The managed model picker is withdrawn when a profile no longer qualifies.
Lowering or clearing a 1M context window left the behavesAs entry behind, so
Claude Code kept mapping the ID to the native 1M row and ignored the declared
window. Only an exact model/behavesAs pair is removed; user options survive,
and a container that existed only for our entry disappears with it.
A profile settings file that fails to parse is left untouched. Resetting it to
{} and writing only the managed key discarded every setting that failed to
parse, in a file Claude Code owns.
`ollamaUpstream` strips the brackets from an IPv6 hostname. A URL keeps them
in an authority, but node:http resolves `[::1]` to nothing, so a profile on
`http://[::1]:11434` failed with ENOTFOUND on every turn. The brackets are
preserved where the value is formatted as a URL or a Host header.
`localStatus` requires an actual `models` array before marking Ollama online.
Any other JSON 200 on port 11434 was adopted as Ollama, so the bundled server
was never started and turns went to that listener.
The cross-platform tests no longer assume Windows: the file-URL-with-host case
and the APPDATA redirection are asserted per platform, and the test that
derives bash from a system git.exe is skipped where the host path module
cannot express the Windows layout. On Ubuntu and macOS the suite failed, which
is what the P1 report described.
132 -> 137 test cases, all passing; `npm run check` is clean.
Version 2.2.0.
Co-Authored-By: Claude Code <noreply@anthropic.com>
01b0c7f to
7a58a99
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7a58a99b1c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Two findings still applied to the current head. `file:`, `file://` and `file:///` name no path. POSIX fileURLToPath() turns them into `/`, so the studio's open endpoint passed the filesystem root to the file manager and the cross-platform assertion expected "Invalid file URL" but got "That path does not exist". The URL is now rejected before conversion when its pathname holds nothing. startProxy() spawned the proxy and then called waitForProxy() without the requested upstream. When another launch already held PORTABLE_PROXY_PORT for a different Ollama server, the new child died on EADDRINUSE and the poll adopted the stranger's healthy proxy, so this profile's prompts went to the other endpoint. The poll now checks the expected upstream and stops as soon as the child exits. waitForProxy is exported for that test and accepts the status probe, so the case is covered without spawning a process. 137 -> 138 test cases, all passing; `npm run check` is clean. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
All review threads are resolved. Six findings were already fixed in
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3055e56af0
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Two findings from the review of 3055e56. A terminal session is its own process, so "I started this" lived in module state. When two sessions used the same Ollama profile, the second adopted the running server and proxy, and the first to exit stopped both - the second then talked to a dead proxy. A small registry in data/run/local-consumers.json now tracks which live processes use a role and which child process serves it. A session releases its claim on exit and only the last live consumer stops the process; the recorded child pid lets the last consumer stop a process whose spawning session already exited. Dead pids are pruned, so a killed terminal cannot keep the processes alive forever, and the record is dropped with the last consumer so a reused pid never becomes a kill target. playwrightStatus() treated any directory named chromium* as the installed browser. An interrupted upgrade leaves the pinned package next to the previous browser build, so the installation was reported ready, ensurePlaywright() skipped the repair and every launch failed. Readiness now compares against the revisions in playwright-core/browsers.json; an unreadable browsers.json falls back to the loose check so an unknown layout never triggers a re-download. 138 -> 141 test cases, all passing; `npm run check` is clean. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
Both findings from the review of Shared Ollama processes — a terminal session is its own OS process, so "I started this" lived in module state and the first session to exit stopped the server and proxy a second session was still using. Chromium revision —
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a53a5b856a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Finding 5 of the previous round was not a defect: the URL parser normalises a
`localhost` host on a file: URL to an empty string, so the empty-path guard
already covered `file://localhost/`. It is still simplified to reject an empty
path regardless of host, because keying on the host was the kind of detail the
guard should not depend on.
The other four are fixed:
- The studio's permission picker was documented as a single-turn override, but
the chosen mode stayed in `state.permissionMode` and every later run attached
the UNRESTRICTED confirmation. The choice now ends with the accepted turn. A
request the server rejected never consumed it, and a mode picked while the
turn was running is kept for the next turn - including a re-pick of the same
mode, which a value comparison alone could not tell from "nothing picked".
- `proxyStatus()` only compared upstreams when the answer contained one, so a
foreign listener returning `{ok:true}` was adopted as the speed proxy. An
expected upstream must now match exactly.
- The local-consumer registry serialises read/modify/write with a lock file.
Breaking a stale lock verifies the file identity instead of unlinking by path,
which would have let two waiters delete each other's fresh lock; a lock whose
owner is still alive is waited for, not stolen. Stopping a shared process now
happens inside the lock, so no session can register between the decision and
the kill. A serving process that exits is removed from the registry, so a
reused pid never becomes a kill target, and a child that loses the bind no
longer overwrites the pid of a healthy proxy.
- `providerEnvironment()` prepends the running interpreter's directory to PATH,
which RESUME.bat and resume.sh never exported: shell commands in a resumed
session could not find `node` on a host without a system installation. The
path is rebuilt from its sources so Git's directories stay ahead of the host
PATH, and the Windows case-variant keys (`Path` and `PATH`) collapse into one.
141 -> 149 test cases, all passing; `npm run check` is clean.
Co-Authored-By: Claude Code <noreply@anthropic.com>
…et to its turn Four findings from the review of 35c6fc3, all reproduced before fixing. The lock honours a live owner without any age limit, and the `since` field was never read. A lock whose unlink failed, or whose pid was reused after a crash, therefore stayed forever with a live pid: every registry call paid the full timeout and then ran unserialised, which re-opens the lost update the lock exists to prevent. Age is now taken from the owner record as well as the mtime, a live owner is trusted only for maxAgeMs, and the lock carries a token so the `finally` cannot delete a later lock held by the same process. Restoring a lock that was taken in between uses COPYFILE_EXCL instead of a blind rename, so a live owner is never evicted. startProxy() and startLocal() treated health as proof of their own child. When two launches race for the port, the loser exits on EADDRINUSE while the winner answers health for the same upstream, so the loser recorded its dead child as the owner and the healthy process lost its record - it then survived every session as an orphan. Only a still-running child of ours is recorded now; a child that lost the race registers as a plain consumer and the session uses the winner's process without claiming it. startLocal also waits for the server to answer before returning, instead of reporting success right after spawn. The composer's reset tracked "picked during a run" in a boolean, but a reconnected turn ends without running the composer's reset, so the flag survived and a one-off Unrestricted choice covered two turns. The picker now records the generation of the turn it was chosen during, so a choice made for an earlier turn is reset by the next turn's end, and the marker cannot cross into another session. The two new lock tests passed under the very regressions they are named for: a lock that is only waited out still writes the registry within the timeout, and "the lock is gone" also describes a stolen lock. They now assert durations well under the timeout and that a live lock still names its owner, and a new test covers the bind race. Each was mutation-checked: breaking the corresponding guard makes exactly that test fail. 149 -> 151 test cases, all passing; `npm run check` is clean. Co-Authored-By: Claude Code <noreply@anthropic.com>
Two findings from the review of dfa5270, both reproduced. A child that is still alive is not proof that it owns the port. In a bind race the loser has not reached bind() while the winner already answers, so it claimed ownership, overwrote the winner's pid and - when it died a moment later - cleared that record. The shared process then had no owner and survived every session as an orphan. The proxy now names its pid in /health, so a matching child owns the listener and a different pid is adopted without being claimed. Ollama names no pid, so a child has to outlive a short grace window while the port answers; a child that lost the bind dies inside it. The exit of our own child no longer ends the wait either. The winner may take a moment to answer - cold start, USB, antivirus - and giving up there left the session unregistered, so the last consumer would stop the shared server under a running session. Only a window without any answer after the child died gives up. A failed spawn now reaches the caller as the real error instead of falling through to "did not answer" after a pointless poll. The tests cover what earlier rounds left open: the claim path records its own child for both roles, the adopting path still counts as a consumer, a live child that is not the listener is not recorded, a winner that answers after our child died is adopted, and the dead-owner lock branch is distinguished from the live one by an age between the two thresholds. Each was mutation-checked. 151 -> 152 test cases, all passing; `npm run check` is clean. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
All five findings are fixed in
Two further defects found by my own review of the fix are also in this head: a child that had not reached
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7857373eaf
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Both threads are closed without a code change. Neither is a scenario this build needs to support.
Nothing else changed: the head is still |
…side `WebSearch` runs inside the Anthropic API. Every other provider configured here answers it with "400 Server tool web_search_20250305 is unsupported by this provider" - the local adapter rejects the server-tool block outright, and the OpenAI Chat Completions API it forwards to has no equivalent. `WebFetch` is client-side and works with every provider, so a search engine is just a URL. A new skill tells the session which URLs answer: Bing for a query, the Google News feed for headlines, and why DuckDuckGo's HTML endpoint is not used (it serves a bot check). It also tells the model to treat fetched text as data rather than instructions. The skill is written next to the Playwright one in the profile directory, so it is discovered by the same mechanism and reaches the launcher, resume and the studio through providerEnvironment. It is on by default and stored as preferences.webSearch, so an update never switches it back on after it was turned off: `START.bat websearch off` (or menu option 9) removes the file, `on` writes it again. Only this project's own skill is ever removed. providerEnvironment takes the preferences instead of reading them, because config.mjs imports providers.mjs and the other direction would make the three modules a cycle; setWebSearch imports config.mjs lazily for the same reason. Verified end to end on the DeepSeek endpoint: a real session listed portable-websearch among its skills, used it, fetched Bing and the news feed, and reported source plus timestamp - and named the pages that blocked the fetch instead of inventing content. 152 -> 159 test cases, all passing; `npm run check` is clean. Co-Authored-By: Claude Code <noreply@anthropic.com>
2.2 stays where it is: the review rounds that followed it added behaviour, not just fixes, so the web-search skill gets its own release instead of moving a tag that has already been downloaded. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
Release 2.3 is tagged as
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 36af9ba510
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Both findings from the review of
Both are left as follow-up work rather than folded into this branch. Head remains |
What this is
The complete state of
rainman74/ClaudeCode-Portableat release 2.3, offered as one reviewablebranch. It covers everything that changed since
b3660fd: the 2.0 feature set, the 2.1additions, the 2.2 fixes, and the web-search skill of 2.3.
The head is
release-2.1-prat36af9ba, the2.3tag. It merges cleanlyonto
main: the common ancestor isb3660fd, and upstream has added only.github/FUNDING.ymlsince then (verified: no overlappingfiles).
54 files, 4700 insertions, 126 deletions, 37 commits.
npm testruns 159 cases, all passing;npm run checkis clean.The demo video and its thumbnail are part of this branch (23.6 MB
docs/videos/xPfZyV20_RY.mp4and 70 KB
docs/images/maxresdefault.jpg), so the README loads nothing from YouTube. They land inthe history permanently; drop those two files if that is not wanted.
The release notes for all three versions follow.
Release 2.3
Summary
Sessions on anything but the official Anthropic API could not search the web.
WebSearchrunsinside the Anthropic API, and every other provider in this project answers it with a 400. This
release teaches the session a route that works everywhere:
WebFetchon a search engine, driven bya skill that is written on every start and can be switched off manually.
Added
Web search without the WebSearch tool
WebSearchis a server-side tool: the search runs inside the Anthropic API. A custom gateway, aDeepSeek Anthropic endpoint, OpenRouter and local models all answer it with
400 Server tool web_search_20250305 is unsupported by this provider.WebFetchis client-side — Claude Code fetches the URL from this machine — so it works withevery provider. A search engine is just a URL, and the new skill tells the session which ones answer:
https://www.bing.com/search?q=<query>&setlang=dehttps://news.google.com/rss/search?q=<query>&hl=de&gl=DE&ceid=DE:dehttps://html.duckduckgo.com/html/data/claude/<provider>-<auth>/skills/portable-websearch/SKILL.mdnext tothe Playwright one, through
providerEnvironment(), so the launcher,RESUME.batand the studioall receive it. Claude Code lists its name and description in the system prompt.
blocked page instead of inventing content.
Bing and the news feed, and reported a headline with source and timestamp.
The switch is on by default and off only on purpose
START.bat websearch on|off|status(menu option 9) toggles it. Stored aspreferences.webSearchnext to the bypass-permissions preference, so an update never turns it back on after it was switched
off. Turning it off removes only this project's skill; a user's own skill is untouched.
Tests
152 -> 159test cases for this release on its own; the table below counts from the 2.0 baseline.Release 2.2
Summary
Fixes every finding from the review passes on this branch: six from the review of
01b0c7f, aWindows argument-parsing bug in the resume path, an empty
file:URL in the studio, a proxyreadiness check that ignored the upstream, shared Ollama processes being stopped by the first
session to exit, and a Playwright readiness check that accepted any Chromium build. Nothing on the
host is reconfigured.
Fixed
An explicit
--permission-modewas overridden by the stored bypass defaultSTART.bat cli --permission-mode acceptEditsis documented, but with the stored bypass preferenceON (the fresh-install default) the launcher appended
--dangerously-skip-permissionson top. ClaudeCode received two contradicting modes and could run unrestricted instead of honouring the
edit-only policy.
--permission-modethe caller passed now belongs to that launch, exactly like--safe, and nobypass flag is added.
returns to the stored mode. A send the server rejected does not consume the choice, and a mode
picked while a turn is running is kept for the next one.
--unrestrictedtogether with it is rejected rather than silently widening the request.--permission-mode=bypassPermissionsstill requires the interactiveUNRESTRICTEDconfirmation.A managed model picker survived a smaller context window
A non-Anthropic profile that declared at least 1M tokens got a
modelPickerbehavesAsentry.Lowering or clearing that value later left the entry in the profile's
settings.json, so ClaudeCode kept mapping the ID to the native 1M row and ignored the declared window.
behavesAspair is removed; options the user added are untouched, and acontainer that existed only for our entry disappears with it.
A malformed profile settings file was replaced
An unparseable
settings.jsonwas reset to{}and overwritten with only the managed picker entry,discarding every setting that failed to parse — in a file Claude Code owns. The file is now left
untouched.
IPv6 Ollama endpoints failed with
ENOTFOUNDURL.hostnamekeeps the brackets of an IPv6 literal. That is right inside an authority, butnode:httpresolves[::1]to nothing, sohttp://[::1]:11434failed on every proxied turn.Hostheader.
::1listener.Any service on port 11434 counted as Ollama
localStatus()marked Ollama online on any 2xx from/api/tags. A stranger's JSON service on theport therefore suppressed starting the bundled server, and turns went to that listener. The response
must now carry a
modelsarray.The same check already existed for the speed proxy and its
/healthresponse.The cross-platform matrix failed on Linux and macOS
tests/backend.test.mjsasserted thatfile://portable-nowhere/share/x.txtis a UNC path. Thatholds only on Windows; elsewhere
fileURLToPathrejects a non-local file host, so the Ubuntu andmacOS legs failed.
tests/local.test.mjsassertedAPPDATAredirection unconditionally, and deriveda bash path from a Windows
git.exeusing the host path module.APPDATAcases are asserted per platform.A working directory ending in a backslash was rejected
RESUME.bat --cwd "D:\Projects\my-app\"failed withWorking directory not found: D:\Projects\my-app".Windows hands the argument over with a trailing quote because the backslash escapes the closing one,
and that path can never exist: Windows forbids
"in file names.",and the artifact never appears.
--cwd "%~dp0"form works through the parser repair.Empty
file:URLs were opened as the filesystem rootfile:,file://andfile:///name no path.fileURLToPath()throws on Windows but returns/onPOSIX, so the studio's open endpoint handed the filesystem root to the file manager. A file URL whose
pathname holds nothing is now rejected before conversion; real URLs, including
file://server/share/file.pdf, keep working.A launch that lost the port no longer hijacks the winner
Two launches may point at different Ollama servers while sharing
PORTABLE_PROXY_PORT, or race forthe same Ollama port. The second spawned, lost the bind with
EADDRINUSE, and the readiness pollaccepted the first launch's healthy process as its own: it recorded the pid of its own dying child,
overwrote the winner's record and then cleared it on exit, so the shared proxy or server lost its
owner and survived every session as an orphan.
/health; a matching child owns the listener and a different pid isadopted without being claimed. A child that has not reached
bind()yet is therefore nevermistaken for the process holding the port.
child that lost the bind dies inside it.
up there left the session unregistered, so the last consumer stopped a shared server that was
still in use.
real error instead of "did not answer".
A shared Ollama server is no longer stopped by the first session to exit
Two terminal launches on the same Ollama profile share one server and one speed proxy. Each launch is
its own process, so "I started this" lived in module state and the first to exit stopped both while the
second was still using them. A registry in
data/run/local-consumers.jsontracks the live consumers ofeach role and the child process that serves it.
inside the registry lock so no session can register between the decision and the kill.
with the last consumer so a reused pid never becomes a kill target.
of unlinking by path, so two waiters cannot delete each other's fresh lock. A live owner is honoured
for a bounded age, because a lock whose unlink failed or whose pid was reused would otherwise
strand every later call on the unserialised fallback.
Any
chromium*directory counted as an installed browserAn interrupted Playwright upgrade leaves the pinned package next to the previous browser build.
Readiness accepted any
chromium*directory, soensurePlaywright()skipped the repair and everylaunch failed. Readiness now compares the installed builds against the revisions in
playwright-core/browsers.json; an unreadablebrowsers.jsonfalls back to the loose check so anunknown layout never triggers a re-download.
Tests
131 → 159 test cases, all passing, plus
npm run check. Every fix above added a regression test;tests/resume.test.mjsgrew from 10 to 11 cases for the working-directory one.Release 2.1
Summary
Paths in the studio are now clickable. A file or folder mentioned in a reply or in a tool's output
opens with your own application instead of having to be found by hand, and sessions now know that
the bundled Playwright is available.
This release touches only the studio and the launcher. Nothing on the host is reconfigured.
Added
Clickable local paths in the studio
A path in a reply or in a tool's output was plain text, and a browser will not start a local file from
an HTTP page, so a picture or a report had to be found by hand.
D:\folder\file.png,C:/folder/file.png,\\server\share\file.pdf, andfile:///D:/....POST /api/open, which accepts only absolute paths thatexist and hands them to
explorer.exe/open/xdg-open. The path is one argument and nevergoes through a shell, so
&, quotes and other metacharacters stay literal.often the sentence than part of the name. Trailing punctuation stays outside the link.
it (
D:\Apps\Claude Code\x.jpg) stays in one piece instead of being cut at the space. A span isonly accepted when it holds exactly one path; anything else stays plain text.
as well, which left the most common way of writing a path unclickable.
Sessions learn that Playwright exists
A session saw only its workspace and its own configuration, so nothing told the agent that the
bundled browser automation is there — it would have had to stumble over
engine/playwright.data/claude/<provider>-<auth>/skills/portable-playwright/SKILL.mdwhenever Playwright is installed. Claude Code lists its name and description in the system prompt,
so the agent knows the capability exists and loads the details only when a task needs them.
the user's project), the
screenshotinvocation, script usage, thenpxwarning, and theheadless-only limitation.
the same folder is never touched.
portable-playwrightwithsource: "user".settingSources: ['user'], so aCLAUDE.mdinside a project would not be loaded.Fixed
Release 2.0
Summary
START.bat resume <id>onlyRESUME.bat/resume.shwith full argument handlingHOME/AppData/XDG of the session redirected intodata/check-updatewith pinned-version rollbackengines, READMEAdopted community pull requests
Two pull requests that are still open in the official repository were merged into this fork
(commit
d1a6bf4). They are listed here because a user of the official repository will not findthem there.
PR #43 — Preserve IME composition in the dashboard (by lux-liang)
Pressing Enter while an input method editor (IME) is composing — Japanese, Chinese, Korean, and
similar — submitted the message instead of confirming the composition.
dashboard/composer.jswithisComposerSubmitKey(event, compositionActive), which treatsEnter as a submit only when Shift is not held,
event.isComposingis false, and no compositionis active.
next Enter sends.
tests/composer.test.mjs.PR #47 — Allow HTTP base URLs for private LAN hosts (by VikasRajput-18)
Fixes upstream issue #45.
The base-URL validator accepted plain HTTP only for
localhost,127.0.0.1and[::1], which madeLM Studio, Ollama, or any other local server unreachable over a LAN address (for example
http://192.168.1.50:1234/v1).isPrivateHost()inlib/config.mjsaccepts loopback, RFC 1918 (10/8,172.16/12,192.168/16), link-local (169.254/16) and IPv6 ULA/link-local (fc00::/7,fe80::/10)addresses over HTTP.
rejected.
2efb954).The code from both pull requests is included as authored —
dashboard/composer.jsand theisPrivateHost()helper match the PR branches exactly. The fork only merged them into a singlecommit and rebased them onto its own history.
Added
Runtime update check with rollback (
check-update, menu option 7)Upstream shipped a pinned runtime with no way to move it forward short of a manual reinstall.
availableUpdate()compares the pinned Claude Code and Agent SDK versions with npm(
lib/runtime.mjs).savePinnedVersions()writes the new pins before installing and restores them if the installfails, so a bad release cannot leave the launcher pointing at a broken version.
START.bat check-update/./start.sh check-updateand menu option 7.(
6e2fd1d,4b86741).Custom model windows and model-picker mapping
Claude Code only knows Anthropic model IDs; any other ID warned "Model isn't recognized", assumed a
200K window and exposed reduced capabilities.
profileSettings()writes amodelPickerrow for custom model IDs, so Claude Code treats them asa known model.
CLAUDE_CODE_MAX_CONTEXT_TOKENScarries the window declared in the dashboard, so the declaredvalue stays authoritative for IDs mapped to a different native window.
CLAUDE_CONFIG_DIRis merged, never replaced, because ClaudeCode stores its own keys there.
Dashboard conversation management
.jsonlbehind a syncedsession (
d101ffb).247aeb1).Deleting a conversation no longer needs a second click
The sidebar rebuilt its whole list with
innerHTMLon every refresh. When a refresh landed betweenpressing and releasing the mouse — for example while a session was still loading — the button the
press started on was destroyed, and the browser then delivered no click at all. The press looked
ignored, so the trash had to be pressed twice.
renders (
resetSession()followed byrefreshSessions()) leave the DOM untouched.the element it started on. The deferred render runs right after it.
a rebuilt row can never keep a stale handler or lose one.
click, while the previous behaviour (raw
innerHTMLreplacement during the press) does drop it.Image paste in the terminal: Ctrl+V added next to Alt+V
Claude Code binds image paste to
alt+von Windows —ctrl+vonly on macOS, Linux, and WSL — and theclassic Windows console swallows
Ctrl+Vfor its own paste. PressingCtrl+Vwith a screenshot in theclipboard therefore did nothing.
ctrl+v -> chat:imagePasteinto the profile'skeybindings.json. Existingbindings and every other context survive; an explicit user choice (including an unbind) is never
overwritten, and a file with a syntax error is left untouched.
Windows Terminal as the one supported way to make
Ctrl+Varrive there.clipboard actually held an image, and reports "No image found in clipboard." There is no text
fallback, so text paste stays with Ctrl+Shift+V or the context menu.
Playwright built into the install and update flow
Upstream shipped no browser automation. This fork installs Playwright next to the runtime, so
screenshots and browser checks work from the portable drive without touching the user profile.
engine/playwrightand the headless Chromium buildinto
engine/playwright-browsers(~300 MB;--only-shellavoids the ~750 MB full build).install/update) repairs the runtime and Playwright.check-update) reports newer Playwright releases next to the Claude Code updatesand installs them with the same pin rollback: a failed install restores both pins.
tools/runtime-manifest.json("playwright": "1.65.1"), because it decidesthe Chromium build revision.
--offlineandPORTABLE_AI_NO_PLAYWRIGHT=1skip it.Paste and drop attachments in the studio
Upstream only accepted files through the picker behind the + button.
(Ctrl+V / ⌘V in the composer). Text pastes are untouched.
as the drop target and a drop outside never navigates away from the page.
application/octet-stream. The server nowdetects the real format from the file's magic bytes (PNG, JPEG, GIF, WebP) and stores the type in
a metadata sidecar, so it survives a restart.
far below the generic 10 MB.
Images reach the model as image blocks
Upstream passed attachments to the agent as local paths only, so a picture had to be opened by the
agent with its file tools before the model could see it — and a model without tool support never saw
it at all.
imagecontent blocks,exactly what the Claude Code terminal does with a pasted screenshot.
path changes.
Standalone session resume (
RESUME.bat,resume.sh)Upstream only offered
START.bat resume <session-id>, which still ran the full launcher: runtimechecks, provider setup, update handling. This fork adds a resume path that does exactly one thing.
RESUME.bat(Windows) andresume.sh(Linux/macOS) are self-contained entry points.data/settings.jsonexist, and fail witha clear "run START.bat first" message instead of installing, repairing, updating or reconfiguring
anything. This is the behavioural difference from
START.bat: nothing is repaired silently.lib/resume.mjs), not in batch, so quoting, spaces and Unicode inpaths work on every platform and the rules are unit-tested:
<session-id>(positional) or--resume <session-id>--cwd <directory>or a positional directory--safe/--unrestrictedfor a one-off permission overridea warning is printed and a new session starts (upstream behaviour, preserved).
New files:
RESUME.bat,resume.sh,lib/resume.mjs,tools/resume.mjs,tests/resume.test.mjs.Bypass permissions as a stored setting (default ON)
Upstream chose the mode in an interactive menu on every launch. This fork stores it, so the
launcher, the resume path and the studio agree, and the choice survives restarts.
preferences.bypassPermissionsindata/settings.json; missing value means ON.START.bat cli,--quick,--offline,START.bat resume,RESUME.bat,./start.sh cli,./start.sh resumeand./resume.sh.START.bat permissions on|off,./start.sh permissions on|off,START.bat permissions status, or menu option 8 (toggles and returns to the menu).--safe(approvals on) or--unrestricted(approvals off).Neither switch is forwarded to Claude Code.
--dangerously-skip-permissionsor--permission-mode bypassPermissionsyourself stillrequires typing
UNRESTRICTED; the stored preference does not weaken that check.--permission-modeyou pass is treated like--safe: it wins over the stored default.Bypass permissions: ON/OFF).New files:
lib/permissions.mjs,tests/permissions.test.mjs.Git Bash handling
Claude Code on Windows needs Git Bash for its shell tool. Upstream assumed it was installed.
lib/git.mjsresolves a usablebash.exein this order:CLAUDE_CODE_GIT_BASH_PATH, the bundledengine/git-win-x64, the standard installation paths, then the directories around anygit.exeon
PATH— including MSYS2 layouts such as<root>/ucrt64/bin/git.exe, wherebash.exelivestwo levels up.
CLAUDE_CODE_GIT_BASH_PATHand Git'scmd,binandusr/binprepended to
PATH.tools/bootstrap.ps1downloads PortableGit only when no Git for Windows is present, and keepsthe download out of the startup path for everyone else. A failed download is a warning, not an
abort, so the app still starts.
New file:
lib/git.mjs, covered bytests/local.test.mjs.Zero-footprint environment
Upstream redirected only
CLAUDE_CONFIG_DIR,XDG_CONFIG_HOME,XDG_DATA_HOMEandXDG_CACHE_HOMEinside the launcher batch file.
portableEnvironment()(lib/providers.mjs) also pointsHOME,USERPROFILE,APPDATAandLOCALAPPDATAatdata/home,data/app_dataanddata/local_app_data, and creates them.environment, so browser and system integration are unaffected.
ANTHROPIC_API_KEYcannot leak into a session.HOMEhides~/.gitconfig, whichmade an agent's
git commitfail with "Author identity unknown". The session now receivesGIT_CONFIG_GLOBALpointing at the host~/.gitconfig(or~/.config/git/config), resolved fromevery available profile location — including
HOMEDRIVE+HOMEPATH, so detection still works whena session is started from an already redirected environment. This is a deliberate exception to
zero-footprint, and
git config --globalinside a session writes to that host file.PORTABLE_AI_NO_HOST_GIT_CONFIG=1disables the link.Automatic Ollama server and speed proxy
Upstream had no speed proxy and left starting the local server to the user.
data/ollama/ollama.exeand the speed proxy, then routes the session through the proxy.tools/local-proxy.mjs) trims Claude Code's system prompt from roughly 10k tokens to~300 before it reaches the model. It understands both protocols:
/v1/messages— how Claude Code reaches Ollama through its Anthropic-compatible endpoint.A proxy that only rewrites
/chat/completionsnever sees these requests./chat/completions— other OpenAI-style clients/healthendpoint; activity is written todata/logs/proxy.logonly, never to the terminal.PORTABLE_PROXY_PORTandPORTABLE_PROXY_MAX_CHARS.started elsewhere keeps running.
that cannot bind its port means the session runs untrimmed and says so.
New files:
lib/local-session.mjs,tools/local-proxy.mjs; extendedlib/local-models.mjs,tools/launcher.mjs,lib/resume.mjs.Launcher and documentation
START.batvsstart.sh).README.md: permission-mode matrix for Windows and Linux/macOS, environment-variable reference,Ollama automation, zero-footprint notes, updated project structure and troubleshooting entries.
Changed
tools/bootstrap.ps1,start.sh, theCI matrix (
node-version: '26'),package.json(engines: >=26) and the README. An existinginstallation is replaced automatically on the next launch; the previous runtime is moved to
engine/<platform>/previous(for exampleengine/win32-x64/previous), a single slot that thenext install overwrites.
canUseToolis installed only for permission modes that can actually consult it (see below).Fixed
Native sessions could not run a turn —
Cannot read properties of undefined (reading 'push')Sessions imported from the Claude history only carry
messagesandtranscript;syncNativecreated them without an
eventsarray, and the agent appended to it on the first turn.syncNativenow writesevents: []and preserves an existing log.AgentManager.startnormalisesmessagesandeventsbefore the run.A failed start blocked the session forever —
A turn is already runningstart()registered the run before doing work that could throw. A failure left the entry in the runmap, so every later attempt was rejected.
Providers rejected tool search —
Unsupported content: tool_referenceClaude Code's tool search returns
tool_referenceblocks naming deferred tools. The Anthropic APIexpands them server-side; third-party providers do not, and the adapters threw.
tool_referenceblocks. No information is lost: every tool schema —including deferred ones — is forwarded anyway, because
defer_loadingis not propagated.Tool returned no textual output.rather thanan empty body.
SDK warning
CLAUDE_SDK_CAN_USE_TOOL_SHADOWEDThe dashboard passed both
canUseToolandpermissionMode: 'bypassPermissions'. The SDK approvesevery tool call before consulting the callback in that mode, so the callback was dead code and the
SDK warned on every turn.
defaultandacceptEdits. Approval handling in thosemodes is unchanged, and the warning is gone at its source rather than suppressed.
Tests
50 → 106 test cases (
npm test) against the upstream baseline at the 2.0 release, plusnpm run check. The adopted pull requests contributedtests/composer.test.mjsand additions totests/backend.test.mjs.New suites:
tests/resume.test.mjs,tests/permissions.test.mjs,tests/local.test.mjs.Regression tests were added for the three bugs above and for the permission behaviour of both the
terminal launcher and the studio.
Upgrade notes
START.bat/./start.shonce. Node.js 26.13.1 is fetched automatically; the previousruntime is kept next to it.
engine/git-win-x64/.START.bat permissions off(or./start.sh permissions off, or menu option 8) if you want to beasked before every write or command.
RESUME.batno longer goes throughSTART.bat, so it will not repair a broken runtime. If itreports a missing runtime, run
START.batfirst.🤖 Generated with Claude Code