claude-code-workarounds

Unofficial community workarounds for Claude Code. Each entry below is an independent fix, delivered through one env-toggled launcher per platform plus standalone per-fix tools.

Not affiliated with or endorsed by Anthropic. A future Claude Code update could make any of the included workarounds obsolete. Use them at your own discretion.

Workarounds

Empty thinking summaries (Opus 4.7 / 4.8) [updated 2026-07-07]. Thinking summaries render empty in the VS Code extension and headless -p/SDK paths, even with showThinkingSummaries enabled. Fix via the launcher (recommended), a one-line extension patch, or a local proxy. -> details

2026-07-07 - new Opus 4.8 regression, upstream and server-side. On claude-opus-4-8 only, the API now returns an empty summary even when the request correctly asks for one, so the launcher's flag is no longer sufficient on that single model. It is gated on an experiment-assignment header the CLI sends. Every other model is unaffected and the launcher remains the fix for them. There is a mitigation, and the launcher can now apply it for you: set CC_ATIS_OPTOUT=1 (off by default - it has caveats). -> 2026-07-07 update
Missing context-usage icon (1M context window) [updated 2026-06-10]. The context-usage pie in the chat input is hidden until you have used more than 50% of the context window. With the 1M window that is about 500,000 tokens, so it is effectively never shown. Fix via the launcher (re-patches the webview on each launch), or a standalone patcher script. -> details
No markdown copy / export of chat [updated 2026-06-10]. The chat cannot copy a whole message or the whole conversation as Markdown, and has no transcript export. Fix via the launcher (adds copy controls, re-applied each launch), a standalone patcher, or a standalone session exporter CLI. -> details

Quick start

Two downloads, one per platform:

Linux / macOS: launcher/claudemax - the bash launcher, no build needed.
Windows: claudemax.exe from the Releases page.

Put it on your PATH, point the VS Code "Claude Code" extension's claudeCode.claudeProcessWrapper setting at its full path, and reload the window. That enables every fix above. Terminal use, per-fix toggles, and the full wiring are under The launcher and each workaround's section.

The launcher

The recommended fix for everything is one small launcher that wraps the real claude binary. It is a drop-in process wrapper carrying every fix in this repo; each fix is on by default and independently switchable with an environment variable, so the same artifact serves "I want everything" and "I want only X" without editing code and without recompiling.

Linux / macOS: the bash script launcher/claudemax.
Windows: the compiled claudemax.exe on the Releases page, built from launcher/claudemax.win.js.

Toggles (set in the environment where Claude Code launches, then reload):

Env var	Default	Effect
`CC_WORKAROUNDS`	`1`	Master switch. `0` disables every fix (argument injection and bundle patches) and reverts the webview to a clean bundle on launch.
`CC_RECONCILE`	`1`	`0` = do not read or write the webview bundle this launch (emergency bypass). Argument injection still runs.
`CC_THINKING_DISPLAY`	`summarized`	`summarized` shows extended-thinking summaries; `omitted` hides them (no injection).
`CC_PATCH_CONTEXT_ICON`	`1`	`0` leaves the context-usage icon unpatched (and reverts ours on the next launch).
`CC_PATCH_MD_COPY`	`1`	`0` leaves the webview without the markdown copy/export controls (and reverts ours on the next launch).
`CC_ATIS_OPTOUT`	`0`	`1` opts out of the server-side experiment that blanks Opus 4.8 thinking summaries: purges the cached `x-cc-atis` assignment from `~/.claude.json` (one-time backup, atomic write, idempotent) and launches with `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1` (2026-07-08 correction - `DISABLE_GROWTHBOOK` alone no longer prevents re-enrollment). Off by default - it edits the CLI's config file and restricts the CLI to essential traffic (also disables claude.ai Projects sync, DesignSync, `/feedback`, live preview, and telemetry; local sessions are unaffected); see the 2026-07-07 update.

See launcher/README.md for wiring details, the VS Code env-setting how-to, and the build command.

Note: The interactive terminal (claude in a shell) already shows thinking summaries through the showThinkingSummaries setting and always shows the context icon. Both issues affect the VS Code extension and the headless -p / SDK paths.

Requirement: The real Claude Code CLI must already be installed and working. If claude --version prints a version, this requirement is met.

The thinking fix edits nothing (it injects a launch flag). The context-icon fix does edit the extension's webview bundle on disk - idempotently, with an ownership marker, an atomic write, a one-time pristine snapshot, and a toggle. See Workaround 2.

Migration from the old launchers (clean break)

The three bash launchers and three Windows launchers are gone. There is now one launcher per platform: launcher/claudemax and launcher/claudemax.win.js (claudemax.exe). Update your wrapper path and, if you want a subset of fixes, set the matching CC_* environment variable.

Old launcher	New equivalent
`claudemax` (both fixes)	`launcher/claudemax` - all fixes on (same behavior)
`claude-think` (thinking only)	`launcher/claudemax` with `CC_PATCH_CONTEXT_ICON=0` (and `CC_PATCH_MD_COPY=0` if you do not want the copy UI)
`claude-context` (context icon only)	`launcher/claudemax` with `CC_THINKING_DISPLAY=omitted` (and `CC_PATCH_MD_COPY=0` if you do not want the copy UI)
any `.exe`	the single `claudemax.exe`; scope features via `CC_*` (VS Code `claudeCode.environmentVariables`)

The unified launcher enables every fix by default, including the new markdown copy/export controls in the chat UI. If you do not want those controls, set CC_PATCH_MD_COPY=0 (the webview is left untouched and any prior install is reverted on the next launch).

Old release assets remain available for anyone pinned to a previous version.

Workaround 1: thinking summaries

Extended-thinking summaries stopped appearing with Opus 4.7 and remain unavailable in the VS Code extension and headless paths, even when showThinkingSummaries is enabled. For background, see these GitHub issue threads:

2026-07-07 update: Opus 4.8 and the experiment header

What changed. On 2026-07-07 thinking summaries on claude-opus-4-8 went from populated to empty mid-day with no client change involved: the same CLI binary produced multi-thousand-character summaries in one session and 0-char summaries in the next. The cause is server-side and request-provable. Claude Code's built-in feature-flag client cached a new experiment assignment in ~/.claude.json:

"experimentKey": "claude_code_attic_parcel_experiment",
"atis": "attic-parcel-meridian"

The CLI attaches that value to every first-party API request as an x-cc-atis header. With the header present, the API returns Opus 4.8 thinking blocks with a valid signature and an empty thinking string - even when the request explicitly carries thinking: {type: "adaptive", display: "summarized"}. Strip the header (or purge the cached assignment) and the byte-identical request returns populated summaries. Full evidence and the single-variable A/B runs are in TECHNICAL.md.

Scope.

Affected model: claude-opus-4-8 only (all alias spellings, with and without [1m]). In the same test matrix, Opus 4.7, Sonnet 5, Haiku 4.5, and Fable 5 returned populated summaries.
Affected surfaces: all of them - the header is attached in the CLI's shared fetch layer (verified on headless -p and the VS Code extension path; the interactive TUI uses the same layer).
Affected installs: only those whose ~/.claude.json carries the assignment. No atis entry in the config means not enrolled (yet); assignment arrives through the client's periodic remote feature refresh.

Effect on this repo's fix. The launcher's --thinking-display summarized injection addresses the original, client-side gap and still does exactly that. It cannot help here: the request already asks for summaries correctly and the server withholds the text anyway. On Opus 4.8, with the assignment cached, summaries stay empty with or without the launcher. (Separately, extension 2.1.202+ now maps showThinkingSummaries into the launch flags itself, so on current builds the launcher's thinking injection is a redundant no-op there; its no-double-inject guard makes that harmless, and the launcher's other fixes are unaffected.)

What to expect. Populated summaries on every model except Opus 4.8; on Opus 4.8, empty "Thinking" sections whenever the assignment is cached. The reasoning still runs and still consumes tokens - only the human-readable text is withheld. Because assignment is remote and server-evaluated, Anthropic can widen, narrow, or end this at any time without any version changing on your machine.

Mitigation (verified 2026-07-07). Two steps, both required - the env var alone is not enough (the cached value is still sent; measured), and the purge alone lasts only until the next feature refresh re-enrolls you:

# 0) Diagnose: is this install enrolled? (empty output = not enrolled)
jq '.clientDataCacheSlots // {} | to_entries
    | map(select(.value.data.atis != null) | .value.data | {experimentKey, atis})' \
  ~/.claude.json

# 1) Purge the cached assignment (backup first)
cp ~/.claude.json ~/.claude.json.bak-cc-workarounds
jq '.clientDataCacheSlots |= (if . then with_entries(select(.value.data.atis == null)) else . end)
    | del(.clientDataCache.atis)' \
  ~/.claude.json > ~/.claude.json.tmp && mv ~/.claude.json.tmp ~/.claude.json

# 2) Prevent re-assignment (2026-07-08 correction: DISABLE_GROWTHBOOK no longer
#    works - the assignment arrives via the CLI's startup bootstrap fetch, which
#    only the essential-traffic-only switch blocks; measured)
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
# set wherever Claude Code launches: shell profile, ~/.claude/settings.json "env",
# or VS Code's claudeCode.environmentVariables

Then reload the VS Code window. Caveats: CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 restricts the CLI to essential traffic entirely - it also disables claude.ai Projects sync, DesignSync, /feedback, --enable-live-preview, and telemetry (local sessions, transcripts, and resume are unaffected); on Windows the file is %USERPROFILE%\.claude.json - if jq is unavailable, hand-edit the slot entries containing "atis" out of clientDataCacheSlots.

Launcher automation (added 2026-07-07). The launcher can apply both steps on every launch: set CC_ATIS_OPTOUT=1 in the launcher's environment (e.g. VS Code's claudeCode.environmentVariables, or your shell profile). When enabled it purges any cached assignment (idempotent - no rewrite when nothing is cached; one-time backup to ~/.claude.json.bak-cc-workarounds; atomic temp-then-rename write; best-effort, never blocks the launch) and sets CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 for the launched process, so the CLI cannot re-enroll itself. It is off by default, unlike the launcher's other fixes, because it is not free: it edits the CLI's own config file, and the env var restricts the CLI to essential traffic entirely (the caveats above)

enable it deliberately. The bash launcher needs jq for the purge half (without it, it warns and only sets the env vars, which is NOT sufficient while an assignment is already cached); the Windows launcher has no external dependency. The manual recipe above remains for non-launcher users.

2026-07-08 correction. The originally published second step (DISABLE_GROWTHBOOK=1) stopped working within a day: with the cache purged and that variable set, the CLI re-fetched and re-cached the assignment about one second after launch and Opus 4.8 summaries stayed empty (thinking length 0). The assignment is delivered by the CLI's startup bootstrap data fetch ("[Bootstrap] Cache updated, persisting to disk" in the 2.1.204 binary), which ignores DISABLE_GROWTHBOOK; the re-fetched slot also carries a new payload field (convolute_arcades), so the experiment is actively evolving server-side. Re-measured single-variable: purge + DISABLE_GROWTHBOOK=1 -> re-enrolled, len 0; purge + CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 -> no re-enrollment, len 169/234/populated. The launcher's CC_ATIS_OPTOUT=1 now sets the working variable (and keeps DISABLE_GROWTHBOOK as harmless defense in depth).

Status. Undocumented upstream: no CHANGELOG entry through 2.1.204 mentions it, and there was no notice, no opt-in, and no documented opt-out. The documented showThinkingSummaries setting is silently overridden while the assignment is active. Background thread: anthropics/claude-code#63358. Dedicated upstream issue for this finding: anthropics/claude-code#75607.

There are technically three workarounds; only the launcher is maintained:

Launcher (recommended). A small wrapper that launches Claude Code and adds the missing flag. It fixes the VS Code extension and headless CLI, and it survives Claude Code updates.
One-line extension.js patch. A direct edit to one line of the VS Code extension; see TECHNICAL.md. VS Code only, and must be re-applied after each extension update. Maintained through 2.1.172; not maintained after.
Local proxy (advanced). A localhost proxy that fixes all surfaces at the wire level; see TECHNICAL.md. Powerful but untested. Not maintained.

Launcher

The launcher starts the real claude binary and appends the missing --thinking-display summarized flag. It does not modify Claude Code files, so it continues working after updates. The same wrapper fixes both the VS Code extension and headless CLI.

Linux / macOS (tested on Ubuntu 24.04)

# 1. Install the launcher
mkdir -p ~/.local/bin
cp launcher/claudemax ~/.local/bin/claudemax
chmod +x ~/.local/bin/claudemax

# 2. Sanity check. This should print normal Claude help.
~/.local/bin/claudemax --help

Use it in VS Code

No PATH changes are required.

Open the Command Palette with Ctrl/Cmd + Shift + P.
Select "Preferences: Open User Settings (JSON)".
Add this line, replacing YOUR_USERNAME. This is the official "Claude Code" extension's setting (shown in the UI as "Claude Code: Claude Process Wrapper"):
```
"claudeCode.claudeProcessWrapper": "/home/YOUR_USERNAME/.local/bin/claudemax"
```
If you use the third-party "Claude Code Chat" extension instead, set "claudeCodeChat.executable.path" to the same path.
Reload the VS Code window by opening the Command Palette and selecting "Developer: Reload Window".
To undo the change, clear or remove the claudeCode.claudeProcessWrapper setting, then reload. Do not point this setting at claude directly: under the process-wrapper convention the extension would launch claude <REAL_CLAUDE> <args...> and the real CLI would receive its own path as a stray argument. (If you used the third-party claudeCodeChat.executable.path instead, that one can simply be pointed back to your normal claude binary or removed.)

Multi-root note: claudeCode.claudeProcessWrapper is window-scoped. In a single folder, User or Workspace settings both work. In a multi-root .code-workspace, set it in the .code-workspace file's "settings" block (or User settings); VS Code ignores it in a folder's .vscode/settings.json.

Use it in a terminal

Run claudemax in place of claude.

Windows 11

The same result is achieved with the compiled .exe.

Download claudemax.exe from this repository's Releases, or build it yourself - see Building the .exe.
Put it somewhere stable, such as C:\Users\YOU\.local\bin\claudemax.exe.
Open the Command Palette and select "Preferences: Open User Settings (JSON)".
Add the following setting (the official "Claude Code" extension setting). Use double backslashes in the path.
```
"claudeCode.claudeProcessWrapper": "C:\\Users\\YOU\\.local\\bin\\claudemax.exe"
```
If you use the third-party "Claude Code Chat" extension instead, set "claudeCodeChat.executable.path" to the same path. In a multi-root .code-workspace, put claudeCode.claudeProcessWrapper in the workspace file's "settings" block or in User settings, not a folder's .vscode/settings.json.
Reload the VS Code window by opening the Command Palette and selecting "Developer: Reload Window".
To use it in a terminal, run claudemax.exe in place of claude.

The wrapper finds the real Claude binary automatically, including native installs with claude.exe and npm installs with claude.cmd. If it cannot find the binary, set CLAUDE_REAL_BIN to the full path of your claude binary.

Want only the thinking fix?

Set CC_PATCH_CONTEXT_ICON=0 and CC_PATCH_MD_COPY=0 to leave the webview untouched and inject only the thinking-display flag. The launcher reads CC_THINKING_DISPLAY:

unset or summarized: show thinking summaries, which is the default
omitted: hide thinking summaries

Set the variable in the same environment where Claude Code launches (such as your shell profile or the VS Code extension environment), then reload. To disable the launcher entirely, clear or remove the claudeCode.claudeProcessWrapper setting (do not point it at claude, for the reason in the undo note above).

What the launcher does

When Claude Code starts a real agent run it puts one of these markers on the command line: --max-thinking-tokens N (the current VS Code extension's budget thinking mode), --thinking adaptive or enabled (the SDK and older extensions), or -p / --print (headless). It does not add the matching --thinking-display flag, so the API defaults the display to "omitted" and the Thinking section comes back empty.

The launcher inspects the arguments and, when it detects a real run via any of those markers, appends --thinking-display summarized before handing off to the real claude binary. The official extension also launches the wrapper with the real CLI path as a leading argument (a "process wrapper" convention); the launcher detects and consumes that path so it is not forwarded as a stray positional. See TECHNICAL.md for more.

Why the launcher is recommended

It survives updates because it does not edit Claude Code files (for the thinking fix; the context-icon fix re-applies on each launch).
It fixes both the VS Code extension and headless claude -p or SDK runs.
It leaves the interactive TUI unchanged because that path already works.
It only injects the flag on real agent runs, not on subcommands or probes such as mcp, config, or --version.
It does not add the flag twice, so it can coexist with a patched or updated extension.
Every fix can be toggled with one environment variable.
It provides one place to configure effort level, auto mode, timeouts, or model routing. See the commented customization section in the script and the Side note.

Other approaches (not maintained)

The one-line extension.js patch and the local proxy are documented in full - with their trade-offs and the runnable scripts - in TECHNICAL.md. Neither is maintained; the launcher is the supported fix.

Option 2: extension.js patch - VS Code only, and must be re-applied after every extension update. Script: fixes/thinking-summaries/patch-extension.sh (--revert, --dry-run). Maintained through 2.1.172; not maintained after.
Option 3: local proxy - surface-agnostic (VS Code + CLI + SDK) but untested, and it sits in the path of your live auth token. Script: fixes/thinking-summaries/proxy.js. Not maintained.

Workaround 2: context-usage icon

The context-usage indicator (the small pie in the chat input that shows how full the context window is) disappeared for many users on recent extension builds. It is not actually removed: recent builds (around 2.1.165 and later) gate the indicator so it renders only after you have used more than 50% of the context window. With the 1M context window enabled, 50% is roughly 500,000 tokens, so in normal use the icon is effectively never shown.

The CLAUDE_CODE_DISABLE_1M_CONTEXT=1 workaround that circulates in issue threads is not a real fix: it only shrinks the window so 50% is reached sooner, at the cost of giving up the 1M window, and it does not touch the threshold. This workaround addresses the threshold directly.

Related GitHub issue threads (feature requests for a persistent context indicator, useful as corroboration):

There is no environment variable or CLI flag for this threshold, so the fix is a tiny edit to the extension's webview bundle. There are two ways to apply it.

Option 1: Launcher (recommended)

The launcher carries this fix on by default. Install and wire it up exactly like Workaround 1 (copy launcher/claudemax to ~/.local/bin, point claudeCode.claudeProcessWrapper at it, reload). On Windows, download claudemax.exe from Releases. To get only the context-icon fix and skip thinking injection, set CC_THINKING_DISPLAY=omitted.

On each launch the wrapper reconciles the extension's webview/index.js, removing the startup hide guard and flipping the hidden threshold so the icon shows at any usage level. Because it re-applies every launch, an extension auto-update that reinstalls a fresh bundle is re-patched on the next launch.

First-run note: the wrapper patches index.js on disk when the CLI is spawned, which can be after the webview already loaded the old bundle. The first time you enable it you may need two reloads: reload once (the spawn patches the file), then reload again (the webview loads the patched bundle). Later windows and post-update launches are already patched on disk.

What it changes

In the indicator component, the render gates are if (t === 0) return null and if (c >= 50) return null, where t is the known context window and c is the percent of context remaining. So the icon renders only after the webview knows a session and less than 50% remains (more than 50% used). The fix removes the startup guard, flips the threshold, and tags the edit with an ownership marker:

if(t===0)return null;if(c>=50)return null}   ->   if(c>=101)return null}/*ccwa-context-icon:t:c*/

c maxes at 100, so c >= 101 is never true and the gate never hides the icon. Removing if (t === 0) return null keeps the icon visible after a window reload while usage data is still being repopulated; during that gap it can briefly show 0%. The edit is anchored on the minified guard-pair shape above, not on the component name or exact minified variable names, which change between builds. The trailing /*ccwa-context-icon:<first-var>:<remaining-var>*/ marker stores the matched names so the launcher can reverse only its own change back to the same pristine variable names.

Turn the context-icon fix on or off

The launcher reads CC_PATCH_CONTEXT_ICON:

unset or 1: patch the webview so the icon is visible, which is the default
0: leave the extension webview untouched (and revert ours on the next launch)

This edits the extension (unlike the thinking fix)

Unlike Workaround 1, this fix edits the extension's bundled webview/index.js. The edit is made safe:

Idempotent - it skips a file that is already in the desired state, and skips (rather than guesses) if the combined guard shape is absent because the extension changed.
Ownership-marked - the edit carries a /*ccwa-context-icon:<first-var>:<remaining-var>*/ marker; the launcher reverses only its own marked edit and never touches upstream code that merely resembles a patched value. Older /*ccwa-context-icon*/ markers from prior versions are still recognized and normalized.
Snapshotted once - a whole-file pristine snapshot index.js.bak-cc-workarounds is written the first time the file is rewritten, for emergency manual restore only; routine reconcile never reads it.
Atomic - the change is written to a temp file and moved into place only after it is verified, so a failed or partial write leaves the original untouched.
Best-effort - every step is guarded; a read-only file, a renamed bundle, or a missing tool simply no-ops and never blocks the launch.
Reversible - set CC_PATCH_CONTEXT_ICON=0 (the launcher reverts on the next launch), set CC_WORKAROUNDS=0 to revert every fix, or just let an extension update replace the file.

Option 2: Standalone patcher script

fixes/context-icon/fix-context-icon.py applies the same change directly, without a launcher. It auto-discovers installed extensions, backs each up to .bak-context-icon, and is idempotent.

python3 fixes/context-icon/fix-context-icon.py            # auto-discover and patch all installs
python3 fixes/context-icon/fix-context-icon.py --revert   # restore from backups
python3 fixes/context-icon/fix-context-icon.py /path/to/webview/index.js   # explicit target(s)

After patching, reload the webview (Command Palette -> "Developer: Reload Window"). Because an extension update reinstalls a fresh bundle and reverts the patch, re-run the script after updates (or use the launcher, which re-applies automatically).

Known limitations

Coarse glyph. The pie is a 3-state gauge, not a continuous fill: it only changes appearance at roughly 62.5% and 87% used. The precise percentage is in the hover tooltip and in /context, not in the glyph itself. Making the pie a fine-grained gauge would require new SVG geometry, not a one-line patch, so it is out of scope.
Transient 0% right after a reload. The icon reads from a usage store that resets to zero on a window reload and is repopulated by the next assistant turn. Immediately after reloading a continued conversation, before any new turn, the tooltip can briefly read "0% context used"; it self-corrects to the true value after the next turn. (/context is unaffected - it queries the CLI directly.) This is intentional: showing a temporary 0% icon is preferable to hiding the icon for the whole reload gap.

Side note: launching Claude Code with third-party models

Because Option 1 is a launcher you control, you can also use it to launch Claude Code with any third-party, Anthropic-API-compatible model, such as DeepSeek or GLM. Set model-routing variables in the same launcher environment:

# --- Connection ---
export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="your token"
# --- Model mapping ---
export ANTHROPIC_MODEL="deepseek-v4-pro"
export ANTHROPIC_DEFAULT_OPUS_MODEL="deepseek-v4-pro"
export ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-v4-pro"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4-flash"
export CLAUDE_CODE_SUBAGENT_MODEL="deepseek-v4-flash"

For GLM 5.2, point the connection at z.ai and map every model to glm-5.2:

# --- Connection ---
export ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
export ANTHROPIC_AUTH_TOKEN="your token"
# --- Model mapping ---
export ANTHROPIC_MODEL="glm-5.2"
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.2"
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.2"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-5.2"
export CLAUDE_CODE_SUBAGENT_MODEL="glm-5.2"

For examples of the launcher with GLM 5.2 enabled, see these two Gists:

https://gist.github.com/phase3dev/b7c341d63f666479be0aff847a5bd018 - a simplified version of the current launcher containing only the GLM 5.2 launch code and related environment variables.
https://gist.github.com/phase3dev/1f7ee1cb17151c19d538e053f4f548ca - the full current launcher from the repo with all features enabled and GLM 5.2 active.

Setup is otherwise identical to Option 1. This is unrelated to the fixes above.

Troubleshooting

Summaries suddenly empty again, but only on Opus 4.8 (since 2026-07-07): the launcher is not broken and neither is your setup - your install is likely enrolled in a server-side experiment that blanks Opus 4.8 summaries at the API. No launch flag can override it. Diagnose and mitigate per the 2026-07-07 update, or set CC_ATIS_OPTOUT=1 in the launcher environment to apply the mitigation automatically (off by default; read the caveats there first).
Thinking still empty after setup: Reload the VS Code window after changing the setting. Confirm the setting points to the launcher's full absolute path. On Windows, confirm the path uses double backslashes.
Context icon still missing after setup: It may take two reloads the first time (see the first-run note above). Confirm CC_PATCH_CONTEXT_ICON is not set to 0 and CC_WORKAROUNDS is not set to 0.
could not find the real 'claude' binary: Set CLAUDE_REAL_BIN to its full path. Use which claude on Linux/macOS or where claude on Windows.
Nothing changes in a plain terminal chat: This is expected. The interactive TUI already shows summaries and the context icon, and does not need these fixes.
Summaries are short: Summary length tracks the reasoning effort level. Try a higher CLAUDE_CODE_EFFORT_LEVEL, such as xhigh, or enable auto mode with CLAUDE_CODE_ENABLE_AUTO_MODE=1. Higher effort uses more tokens.
To verify the thinking root cause: Run fixes/thinking-summaries/test-thinking-display.sh. It performs a live A/B test and uses a small number of tokens.

Files

File	Workaround	Description
`launcher/claudemax`	both	Unified launcher (Linux/macOS), env-toggled.
`launcher/claudemax.win.js`	both	Windows source for `claudemax.exe`.
`launcher/README.md`	both	Wiring, the toggle table, the VS Code env-setting how-to, the build command.
`fixes/thinking-summaries/patch-extension.sh`	thinking	Option 2 `extension.js` patch (`--revert`, `--dry-run`). Unmaintained; see TECHNICAL.md.
`fixes/thinking-summaries/proxy.js`	thinking	Option 3 localhost proxy. Advanced, untested, unmaintained.
`fixes/thinking-summaries/test-thinking-display.sh`	thinking	Live A/B test showing that the flag is the relevant lever.
`fixes/context-icon/fix-context-icon.py`	context icon	Option 2 standalone webview patcher with `--revert`.
`fixes/markdown-copy-export/add-md-copy.py`	markdown copy	Standalone webview patcher (sentinel block, reverse-transform `--revert`).
`fixes/markdown-copy-export/cc-export.py`	markdown copy	Standalone session exporter (markdown or plain text).
`fixes/markdown-copy-export/webview-inject.js`	markdown copy	Single source of the appended copy-controls IIFE.
`tools/gen-embeds`	markdown copy	Generates the embedded payload into the launcher + patcher; `--check` drift gate.
`TECHNICAL.md`	both	Full root-cause analysis, the reconcile model, and design notes.

Releases

The prebuilt Windows claudemax.exe launcher is published on the Releases page rather than committed to the repo, since it is large and reproducible from the claudemax.win.js source. Linux and macOS users run the bash script from the repo and do not need a download.

Building the .exe

The Windows launcher is built from its *.win.js source into a standalone .exe with vercel/pkg. Node.js is required.

npm i -g pkg
pkg launcher/claudemax.win.js --targets node18-win-x64 --output claudemax.exe

Compatibility

Confirmed on Opus 4.7 and Opus 4.8 with VS Code extension builds 2.1.169 through 2.1.172 (native-binary CLI), via the claudeCode.claudeProcessWrapper setting, on Windows 11 and Ubuntu 24.04.

Thinking fix: earlier builds (2.1.165 / 2.1.167) signaled thinking with --thinking adaptive; 2.1.169 and later use --max-thinking-tokens on the VS Code path. The launcher keys off either, plus -p/--print for headless. The --thinking-display flag and the request field are stable levers; the Option 2 minified strings can change between releases (the script matches generically and skips if not found). 2026-07-07: extension 2.1.202+ maps showThinkingSummaries into the launch flags itself (the launcher's injection becomes a harmless no-op there), and on claude-opus-4-8 a server-side experiment can blank summaries regardless of the flag - see the 2026-07-07 update. The flag remains correct and sufficient for other models and for un-enrolled installs.
Context-icon fix: the >50% used gate has been observed with different minified names (Z/U in 2.1.108 and 2.1.131, t/c in 2.1.170 and 2.1.172). The patch matches the guard-pair shape with variable-name captures and records the matched names in its ownership marker, so it patches and reverts cleanly regardless of the names; if a future build changes the guard shape itself, the launcher safely no-ops and the anchor needs updating.
Markdown copy/export fix: the copy controls are appended to the webview bundle and re-applied each launch. The injection is anchored on stable structural markers and skips cleanly if the bundle shape changes.

Behavior may change in future Claude Code releases.

Name		Name	Last commit message	Last commit date
Latest commit History 36 Commits
.github/workflows		.github/workflows
fixes		fixes
launcher		launcher
media		media
tests		tests
tools		tools
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
TECHNICAL.md		TECHNICAL.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

claude-code-workarounds

Workarounds

Quick start

The launcher

Migration from the old launchers (clean break)

Workaround 1: thinking summaries

2026-07-07 update: Opus 4.8 and the experiment header

Launcher

Linux / macOS (tested on Ubuntu 24.04)

Use it in VS Code

Use it in a terminal

Windows 11

Want only the thinking fix?

What the launcher does

Why the launcher is recommended

Other approaches (not maintained)

Workaround 2: context-usage icon

Option 1: Launcher (recommended)

What it changes

Turn the context-icon fix on or off

This edits the extension (unlike the thinking fix)

Option 2: Standalone patcher script

Known limitations

Side note: launching Claude Code with third-party models

Troubleshooting

Files

Releases

Building the .exe

Compatibility

About

Uh oh!

Releases

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

claude-code-workarounds

Workarounds

Quick start

The launcher

Migration from the old launchers (clean break)

Workaround 1: thinking summaries

2026-07-07 update: Opus 4.8 and the experiment header

Launcher

Linux / macOS (tested on Ubuntu 24.04)

Use it in VS Code

Use it in a terminal

Windows 11

Want only the thinking fix?

What the launcher does

Why the launcher is recommended

Other approaches (not maintained)

Workaround 2: context-usage icon

Option 1: Launcher (recommended)

What it changes

Turn the context-icon fix on or off

This edits the extension (unlike the thinking fix)

Option 2: Standalone patcher script

Known limitations

Side note: launching Claude Code with third-party models

Troubleshooting

Files

Releases

Building the .exe

Compatibility

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Contributors

Uh oh!

Languages