Changelog

Quiz q1 asked for the 'only required file' and marked .claude-plugin/plugin.json correct, while a partial audit fix had already changed its explanation to say the manifest is optional — a contradiction where the graded answer rewarded picking a supposedly-required file the explanation called optional. Verified against the official plugins docs (manifest 'optional if components use default locations'), the question now asks what role the manifest plays, the correct answer is that it's optional with auto-discovery from default locations, and all four options and explanations are internally consistent. Also swept the rest of the quiz for the same stale wording: q3's decoy no longer claims plugins 'require the plugin manifest', and q1 option (d) now distinguishes the manifest's identity/configuration fields (name, version, author, userConfig) from the skill and hook definitions that live in separate files, rather than calling it 'only metadata'. Supersedes audit PR #433.

The plugin.json template card described the manifest as 'The required .claude-plugin/plugin.json manifest', contradicting the module prose and the official docs. Changed it to 'optional' in both languages so the template surface matches.

The module claimed agents are 'resumable by ID' (vague) and an audit fix pinned the mechanism to 'as of v2.1.76'; that version's official changelog records no Agent resume-parameter removal, and an internal note dated the change to v2.1.77 instead, so the version pin was inconsistent and unconfirmed. Reworded to the observable mechanism — SendMessage({to: agentId}) resumes an agent with its context intact, while a fresh Agent call starts a new conversation — in both languages, dropping the version pin. Supersedes audit PR #434.

The module prose and quiz q10 decoy still presented the 200-subagent-per-session cap as an active limit. The official CHANGELOG records it was added in v2.1.212 and removed in v2.1.224 (concurrency and depth limits still apply). Reworded the prose so the concurrent cap (default 20) is the remaining companion to the depth limit and the per-session cap reads as removed, and rewrote quiz q10 option (c) so its wrong-answer no longer cites a retired variable as current. Supersedes audit PR #431.

The feature catalog (EN + AR) and the English reference cheatsheet still listed CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION as an active config option with a 200 default, so the catalog, reference page, and site search advertised a variable removed in v2.1.224 — contradicting the corrected module. Removed the entry from all three surfaces.

The /review bullet dated its move to a lighter /code-review alias to v2.1.222; the official CHANGELOG records that change in v2.1.223. Fixed the version in both languages and kept the original 'reach for /code-review when you want effort levels or fixes' guidance instead of the auto-fixer's awkward meta-commentary. Supersedes audit PR #430.

v1.24.41 made the plugins module describe .claude-plugin/plugin.json as optional, but the reference cheatsheet still listed it as 'Required plugin manifest' (EN) / 'المطلوب' (AR), contradicting the module. Updated both entries to describe the manifest as optional and to note that components are auto-discovered from default locations when it is omitted.

The English terminal /effort output ends with 'Use /effort auto to reset to default'; the Arabic output omitted it. Added the Arabic equivalent so both languages document the reset behavior.

The /effort listing in the slash-commands module named low, medium, high, xhigh, and max but omitted ultracode, and marked only max as session-only. The official commands reference states /effort accepts low, medium, high, xhigh, max, or ultracode, that both max and ultracode are session-only, and that auto resets to the model default. Added ultracode, marked both max and ultracode session-only, and marked ultracode session-only in the terminal simulator's /effort output too (EN + AR). Supersedes audit PR #427, whose auto-generated wording redundantly re-listed xhigh.

The plugins module described .claude-plugin/plugin.json as 'the only required file.' The official plugins reference states the manifest is optional — if omitted, Claude Code auto-discovers components in their default locations (skills/, agents/, hooks/, .mcp.json, .lsp.json) and derives the plugin name from the directory. Reworded the prose accordingly and changed the '# Required manifest' label in the directory-tree diagram to '# Plugin manifest (optional)'. Supersedes audit PR #426.

Audit PR #425 proposed reducing the quiz q9 CLAUDE_EFFORT explanation from 'low, medium, high, xhigh, or max' to 'low, medium, high', citing the v2.1.73 CHANGELOG. That source is out of date: the current hooks doc lists the effort.level / $CLAUDE_EFFORT value as 'low, medium, high, xhigh, or max', and the reference cheatsheet documents the same five levels. Removing xhigh and max would have made the quiz wrong. No change made; #425 closed as superseded.

Audit PR #424 rewrote the Auto Mode 'older models never eligible' list but its own finding states the tier list 'should be verified against current docs' and was not. The applied diff also contradicts itself, listing Opus 4.8 among ineligible older models in the same paragraph that names Opus 4.8 as one of the supported models. Per 'if the docs don't support the claim, skip it', the content was left unchanged; #424 closed as superseded.

The Arabic reference cheatsheet described /effort max as 'لـ Opus بس' (Opus-only) while the English says 'session-only'. Per the official commands reference, max is session-only, not Opus-only. Corrected the Arabic to match.

The commands module, its templates, a quiz question, the reference cheatsheet, and the feature catalog all described CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 as populating the /model picker from a custom LLM gateway's /v1/models endpoint. The variable is in no official source — not the env-vars reference, the settings page, the CHANGELOG, or the LLM-gateway docs, which describe no gateway model-discovery feature. Removed every reference across both languages (module, template, quiz, cheatsheet, feature catalog). Supersedes audit PR #420, which only softened the module prose with hedge wording.

The getting-started module, a quiz question, the reference cheatsheet, and the feature catalog described CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 as keeping the conversation in the terminal's native scrollback. The variable is in no official source; the only documented alternate-screen variable is CLAUDE_CODE_ALT_SCREEN_FULL_REPAINT (a repaint control, unrelated to disabling the alternate screen). Removed every reference across both languages (module, quiz, cheatsheet, feature catalog). Supersedes audit PR #421, which only softened the module prose with hedge wording.

Audit PR #422 proposed pruning the subagents quiz from 10 to 5 questions ('schema mandates exactly 5') and deleting the 'ultracode sessions are exempt' clause from the CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS description. Both findings are wrong: the content schema imposes no quiz-question count (modules range from 5 to 30 questions), and the official subagents doc states verbatim that sessions with ultracode active are exempt from the concurrent cap. No change made to the subagents content; #422 closed as superseded.

The module described /review as a 'read-only deeper pass on the change set'. The official commands reference documents it as a distinct bundled skill: a fast single-pass, read-only review of a GitHub pull request that requires the gh CLI. Rewrote the entry to match, noting it's a lighter alias-style entry point to /code-review (v2.1.222) while /code-review remains the choice for effort levels or fixes. Rejects audit PR #416, which had proposed deleting the command entirely — the docs still document /review as its own command, so it was kept and corrected rather than removed. EN + AR modules.

The cheatsheet's /review line said 'Review code changes on the current branch with suggestions'. Corrected to match the official commands reference: a fast single-pass, read-only review of a GitHub pull request that needs the gh CLI. EN + AR cheatsheets.

The /doctor quiz explanation suggested '/review or /simplify' for code review. Per the module itself, /simplify became a separate cleanup-only skill in v2.1.154 that applies fixes without hunting for bugs, so it's no longer the tool for code review. Changed the recommendation to '/review or /code-review' in both quiz files; the correct answer is unaffected.

The module documented Ultraplan, which handed a plan to a Claude Code on the web session for cloud drafting. The official anthropics/claude-code CHANGELOG records 'Removed ultraplan feature' in v2.1.222, and it is the only Ultraplan entry in the changelog — a complete removal. Deleted the entire section, including its launch methods, the ◇ ultraplan / ◆ ultraplan ready status badges, the execute-on-web vs. send-to-terminal paths, and the claude.ai/GitHub availability note, from both the EN and AR advanced-features modules. Incorporates audit PR #413 and adds the version bump and changelog it was missing.

The module stated that --add-dir loads subagent definitions from .claude/agents/ in the added directory. Per the official /en/commands and /en/permissions docs, most .claude/ configuration is not discovered from an added directory — the flag extends read/edit file access only; CLAUDE.md files and rules still require CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1. Corrected in the EN and AR modules (incorporates audit PR #411).

The quiz answer explanation for the --add-dir question asserted 'the flag also loads subagent definitions from .claude/agents/ in the added tree'. Rewrote it to match the docs — the flag grants file access only and does not load subagent definitions — in both the EN and AR quiz files, so the quiz no longer contradicts the corrected module (and the terminal simulation and cheatsheet, which already stated .claude/ config is not loaded).

The prerequisites line said 'Ubuntu 20.04+ (and other modern Linux distros)'. Per the /en/setup System requirements section, added the specific distros the docs name — Debian 10+ and Alpine Linux 3.19+ — while keeping Ubuntu 20.04+ (the docs list 20.04, so the audit's proposed 22.04 change was rejected). This matches the getting-started install tabs, which already list Ubuntu 20.04+, Debian 10+, and Alpine 3.19+. Applied to EN and AR (incorporates the supported part of audit PR #410).

The Subagents module never mentioned that a subagent can spawn subagents of its own. Per the official /en/sub-agents 'Let subagents spawn their own subagents' section, added that nesting is allowed up to three layers below the main conversation by default (raised to three in v2.1.219, having been one in v2.1.217–v2.1.218), that at the depth limit the Agent tool is withheld from every subagent except a fork, and that CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH (v2.1.217+) sets the number of layers or 1 disables nesting. Applied to the EN and AR modules.

Alongside nesting depth, documented the two other limits the docs describe: CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS (default 20 running at once, v2.1.217+, ultracode sessions exempt) and CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION (default 200 total per session, v2.1.212+, reset by /clear). Added to the module prose (EN + AR), the feature catalog (EN + AR) and the reference cheatsheet (EN) so all three are discoverable through search.

Added a tenth question to the Subagents quiz testing that a subagent nests up to three layers below the main conversation by default (v2.1.219) and that CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH changes the number of layers (1 to disable), with decoys that conflate depth with the separate concurrent (20) and per-session (200) limits. Added to both the EN and AR quiz files.

Every addition was applied to the Arabic side alongside the English: the nested-subagents-and-limits paragraph in ar/subagents.md, the q10 quiz question in ar/quiz.yaml, and the three env-var entries in ar/feature-catalog.json. The reference cheatsheet is English-only (the Arabic cheatsheet has no env-vars section), so the cheatsheet entries are EN-only, consistent with prior passes.

Audit PR #400 corrected the deprecation-date wording but left the Fast Mode paragraph contradicting itself: it listed Opus 4.7 as a supported model in the same sentence that then said it had been removed in v2.1.219. The official CHANGELOG (v2.1.219) states 'Removed Opus 4.7 from fast mode; /fast now applies to Opus 5 and Opus 4.8' and that Opus 5 is 'now the default Opus model'. Rewrote the supported-models clause to Opus 5 (default Opus model since v2.1.219) and Opus 4.8, dropping Opus 4.7; changed the auto-switch default from Opus 4.8 to Opus 5; and updated the fast-mode rate-limit pool from 'shared across Opus 4.8 and 4.7' to 'shared across Opus 5 and Opus 4.8'. Incorporates and closes audit PR #400 and the remaining Codex line-70 finding it did not cover.

The q7 option-(a) explanation claimed --add-dir 'only grants file access; .claude/ configuration inside the added paths is not loaded', which directly contradicts the project-setup module body — it already states --add-dir loads subagent definitions from .claude/agents/ in the added tree, while CLAUDE.md files and rules require CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1 to take effect (per the /en/sub-agents docs). Corrected the quiz explanation to match. Incorporates and closes audit PR #401.

Both corrections were applied to the Arabic side alongside the English: the Fast Mode supported-models, default-model and rate-limit-pool wording in ar/commands.md, and the project-setup q7 quiz explanation in ar/quiz.yaml.

Our Windows setup guidance said the PowerShell tool 'is enabled by default on Windows for Bedrock, Vertex, and Foundry users — set CLAUDE_CODE_USE_POWERSHELL_TOOL=0 to opt out.' The current official /en/setup docs no longer describe any default-on provider: 'When Git for Windows is installed, the PowerShell tool is rolling out progressively as an additional option alongside Bash. Set CLAUDE_CODE_USE_POWERSHELL_TOOL=1 to opt in or 0 to opt out.' Rewrote the getting-started prose to lead with the opt-in rollout (=1 to opt in, =0 to opt out, not enabled by default for any specific provider), and applied the same correction to the advanced-features module, the advanced-features terminal-steps description, the feature catalog entry, and the reference cheatsheet entry. Incorporates audit PR #398 and extends it to the four other places the stale default-on claim survived.

Every change was applied to the Arabic side alongside the English: the opt-in progressive-rollout wording in the getting-started and advanced-features modules and the advanced-features terminal-steps description. The feature catalog and reference cheatsheet PowerShell entries are English-only content (the Arabic cheatsheet carries no CLAUDE_CODE_USE_POWERSHELL_TOOL entry), so no further Arabic parity fixes were needed there.

The module (and the parentSettingsBehavior quiz question) referenced settings the prose did not explain. Added three documented items: parentSettingsBehavior (v2.1.133, admin-tier), with 'merge' clarified as restrictive layering — parent-supplied settings sit beneath admin policy through a restrictive-only filter so they can tighten but not loosen it, and the setting has no effect without an admin tier; sandbox.network.strictAllowlist (v2.1.219), written with its full setting path, which denies all hosts not in allowedDomains without prompting; and a Hooks section for DirectoryAdded (v2.1.219), described as firing when /add-dir or the SDK register_repo_root registers an additional working directory mid-session — a directory-registration event, not a change of the current directory (distinct from the CwdChanged hook). Incorporates audit PR #394 with the CodeRabbit and Codex review refinements.

The official interactive-mode docs define Ctrl+O as 'Toggle transcript viewer', which shows detailed tool usage and execution — not a 'verbose mode'. Corrected the commands module description and, for consistency, every place the Ctrl+O = 'verbose mode / same as /debug' conflation appeared: the reference cheatsheet Ctrl+O and /debug entries, the /debug terminal simulation, and the commands, slash-commands and subagents quiz explanations. /debug is corrected to 'enable debug logging' per the /en/commands docs. The auto-fix's added claim that Ctrl+Shift+V or --include-partial-messages enables verbose mode was rejected: Ctrl+Shift+V is undocumented and --include-partial-messages is a streaming-JSON flag (requires --print and --output-format stream-json), not an interactive verbose toggle. Incorporates the verified part of audit PR #395.

Added catalog entries for sandbox.network.strictAllowlist (v2.1.219) and the DirectoryAdded hook (v2.1.219) so the two newly documented features are discoverable through the site search alongside their module coverage.

Audit PR #396 tried to change the workflows q7 explanation from v2.1.120 to v2.1.111, but that explanation is about the `claude ultrareview [target]` CLI subcommand running non-interactively in CI. The v2.1.121 CHANGELOG documents that subcommand under v2.1.120 ('Added `claude ultrareview [target]` subcommand to run /ultrareview non-interactively from CI or scripts'); v2.1.111 introduced the interactive /ultrareview slash command, a different feature. v2.1.120 is a real documented version, so the original attribution is kept and #396 is closed without applying its change.

Every change was applied to the Arabic side alongside the English: the parentSettingsBehavior / strictAllowlist / DirectoryAdded additions in the advanced-features module, the Ctrl+O transcript-viewer and /debug debug-logging corrections across the commands/slash-commands/subagents quizzes, the reference cheatsheet, and the terminal simulation.

The commands module and the /code-review quiz explanation described the command without noting that, as of v2.1.218, it runs as a background subagent so review work no longer fills the conversation. Per the Claude Code CHANGELOG v2.1.218 ('Changed /code-review to run as a background subagent, so review work no longer fills your conversation'), added that behaviour to the commands module prose and the quiz explanation. Applied to the EN + AR module and quiz. Incorporates audit PR #392.

The advanced-features headless-mode Q4 decoy and the workflows CI/CD Q1 decoy both named --no-conversation, a flag that does not exist. The official CLI reference lists --no-session-persistence as the print-mode (claude -p) flag that disables session saving and has no --no-conversation entry. Corrected both the decoy option text and its explanation in each quiz so the decoy teaches the real flag while remaining an incorrect answer (the CI/CD flag is still --permission-mode bypassPermissions). Incorporates audit PR #391 and resolves the CodeRabbit + Codex review notes that its explanation text was left stale.

Every change was applied to the Arabic side alongside the English: the v2.1.218 /code-review background-subagent note (commands module + quiz) and the --no-conversation → --no-session-persistence correction (advanced-features and workflows quizzes, option text and explanation).

The Windows setup guidance said the PowerShell tool 'is rolling out progressively' and led with CLAUDE_CODE_USE_POWERSHELL_TOOL=1 to opt in. Per the Claude Code CHANGELOG, the tool is now enabled by default on Windows for Bedrock, Vertex, and Foundry users — opt out with CLAUDE_CODE_USE_POWERSHELL_TOOL=0 — and only rolls out progressively for other providers. Rewrote the wording to lead with the default-on behavior in the getting-started module and, for consistency across the site, in the advanced-features module and the advanced-features terminal-steps description as well. Incorporates audit PR #386 and a Codex review note about the remaining stale references.

The workflows module attributed /ultrareview to 'new in v2.1.86, highlighted in v2.1.112'. The CHANGELOG's introducing entry ('Added /ultrareview for running comprehensive code review in the cloud…') is under v2.1.111, with no matching v2.1.86 or v2.1.112 addition. Corrected the attribution to 'added in v2.1.111', aligning the workflows module with the skills module's existing v2.1.111 reference. Incorporates the first finding of audit PR #389.

The historical v1.22.4 changelog entry still attributed /ultrareview to v2.1.86, which contradicted this release's correction to v2.1.111. Updated that entry's version references (EN + AR) so the public changelog is internally consistent on the official v2.1.111 attribution. Addresses a CodeRabbit data-integrity review comment.

#387 re-proposed the skill-listing budget 1% → 2% change (already rejected as #288/#290) — the /en/skills docs say 'The budget scales at 1% of the model's context window' — and tried to trim skill `effort` to low/medium/high, but /en/model-config and the skills frontmatter reference both list `low`, `medium`, `high`, `xhigh`, `max`; rejected, original kept. #388 tried to drop `max` from the /effort values, but /en/model-config confirms `max` is a valid (deepest, session-only) effort level; rejected. #389's second finding claimed `disableWorkflows` / `CLAUDE_CODE_DISABLE_WORKFLOWS=1` don't exist, but the workflows docs 'Turn workflows off' section documents both verbatim; rejected, and the original module wording plus the reference cheatsheet entry were kept. All of #386–#389 are closed in favour of this consolidated PR.

Every change in this pass was applied to the Arabic side alongside the English: the PowerShell default-on wording (getting-started module, advanced-features module, and advanced-features terminal-steps), the /ultrareview v2.1.111 attribution (workflows module), and the corrected v1.22.4 changelog entry. No Arabic quiz, template, cheatsheet, or catalog content asserted the corrected facts, so no further Arabic parity fixes were needed.

The guidance on experimental plugin-manifest fields (`monitors` and `themes` under `experimental: {}`) now dates the `monitors` field to its introducing release. Per the Claude Code CHANGELOG, background monitor support for plugins landed in v2.1.105; the `experimental` nesting requirement followed in v2.1.129. Applied to the EN + AR module. Incorporates audit PR #384.

The module and both quiz files stated the Advisor Tool is enabled via CLAUDE_CODE_ENABLE_EXPERIMENTAL_ADVISOR_TOOL. Per the official advisor docs (code.claude.com/docs/en/advisor), no such env var exists — the advisor is enabled by setting a model through the /advisor command, the advisorModel setting, or the --advisor flag, and CLAUDE_CODE_DISABLE_ADVISOR_TOOL is the only advisor env var (it disables the tool). Corrected the claim across the EN + AR module and EN + AR quizzes. Incorporates audit PR #382.

The Arabic module and quiz described CLAUDE_CODE_DISABLE_ADVISOR_TOOL as 'بينعّم الأداة' (softens the tool) rather than 'بيعطّل الأداة' (disables the tool), and the quiz carried a 'بيتتفعّل' → 'بتتفعّل' typo. Both fixed per review feedback on PR #382.

Per the plugins reference (code.claude.com/docs/en/plugins-reference), the LSP `restartOnCrash` and `shutdownTimeout` fields require v2.1.205+ (before which setting either skipped that LSP server at startup), and a plugin's `settings.json` supports only the `agent` and `subagentStatusLine` keys — all other fields are silently ignored. Added both to the plugins module. The unverified `--plugin-dir (v2.1.145+)` tag from audit PR #375 was NOT added (no doc support). Partially incorporates #375.

Per the memory docs (code.claude.com/docs/en/memory): expanded the `CLAUDE_CODE_NEW_INIT=1` description (asks which artifacts to set up, explores with a subagent, presents a reviewable proposal); noted that declining a first-time external-import dialog leaves imports disabled and the dialog doesn't reappear; added that files are concatenated in load order (managed → user → project → local, closest read last); and clarified in the quiz that rules without a `paths` field load unconditionally and apply to all files. Partially incorporates #374 — its 'whichever is smaller' change was dropped (docs say 'whichever comes first') and its '~v2.1.198' tag was dropped (misattributed to symlink matching).

The module claimed built-in subagents are 'always available and not part of the name-override priority system'. The sub-agents docs (code.claude.com/docs/en/sub-agents) say the opposite: 'A user or project subagent named Explore overrides the built-in and keeps its own model field.' Rewrote the sentence to describe the override behavior. Incorporates audit PR #378.

The `--add-dir` paragraph said the flag 'only grants file access, not the rest of .claude/ configuration'. Per the docs, `--add-dir` also scans a `.claude/agents/` folder in the added directory and loads those subagents; CLAUDE.md files and rules still require `CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1`. Corrected the sentence. Incorporates audit PR #376.

The routines availability line named the plans (Pro/Max/Team/Enterprise) but omitted the prerequisite. The routines docs (code.claude.com/docs/en/routines) state routines are available on those plans 'with Claude Code on the web enabled'. Added the clause. Incorporates audit PR #379.

#372 flipped the ultracode↔workflow rename direction, contradicting the current dynamic-workflow keyword and leaving the paragraph self-contradictory — rejected, original kept. #373 tried to reduce the three MCP scopes to two and claim local scope is stored in `.mcp.json` and private, but the MCP docs (code.claude.com/docs/en/mcp) list three scopes with Local in `~/.claude.json` and Project in the shared `.mcp.json`; only its `which channel plugins` allowedChannelPlugins wording fix was kept. #377 (skills) had an empty diff and was closed as a no-op. All of #372–#379 are closed in favor of this consolidated PR.

Mirrored all module edits (subagents, project-setup, workflows, plugins, memory, mcp) and the memory quiz clarification in the Arabic content and quiz files, matching the existing Egyptian-Arabic style.

Q2's explanation called `/branch` '(formerly /fork)', implying `/branch` was renamed from `/fork`. Per the commands and sub-agents docs, `/fork` was an alias for `/branch` only before v2.1.161; from v2.1.161 `/fork` is a distinct command that hands a side task to a background forked subagent. Reworded the explanation to describe `/branch` on its own terms and note that `/fork` is now a separate command, not an alias — consistent with the subagents module (src/content/modules/subagents.md).

Both auto-generated audit PRs contradicted the official docs and were closed. #367 would have removed the correct 'aliased autoremove' note on `claude plugin prune` — the Plugins reference lists 'Aliases: autoremove' for that command. #368 would have added a false '`/fork` is its alias' claim to the `/branch` bullet — `/fork` is a distinct background-subagent command from v2.1.161. The repo's existing module content on both topics is already accurate and left unchanged.

The Arabic slash-commands quiz Q2 already omits the '/fork' parenthetical, so the EN fix brings the two languages back into agreement; no Arabic change was needed for this correction.

Per the settings docs (code.claude.com/docs/en/settings), Claude Code exposes a screen-reader mode as the `axScreenReader` setting (v2.1.181+), the `CLAUDE_AX_SCREEN_READER=1` environment variable, and the `--ax-screen-reader` flag — the variable and flag take precedence over the setting. It renders flat text with no decorative borders or animations for assistive technology, and always uses the classic renderer so the fullscreen/tui preference has no effect while active. Added the environment variable to the advanced-features env-vars section, next to the existing accessibility-related CLAUDE_CODE_NATIVE_CURSOR entry.

Added the `--ax-screen-reader` flag to the CLI-flags cheatsheet (EN + AR), the `CLAUDE_AX_SCREEN_READER` variable to the EN env-vars cheatsheet, and an `env-ax-screen-reader` feature-catalog entry (EN + AR) so the accessibility feature is discoverable through catalog search in both languages.

Both auto-generated audit PRs were verified against the official docs, CodeRabbit-clean, and Cloudflare-green. #364 clarifies that `--channels` was added in v2.1.80 as a research preview and suggests checking `claude --version`; #365 clarifies `${CLAUDE_PLUGIN_DATA}` is a shell-expansion variable for persistent state and reframes `claude plugin tag` as 'introduced in v2.1.118'. Squash-merged with clearer commit titles (#364's original title was flagged as too vague by CodeRabbit).

Mirrored the advanced-features env-var paragraph, the CLI-flags cheatsheet entry, and the feature-catalog entry in Arabic.

Per the MCP docs (code.claude.com/docs/en/mcp), WebSocket is a fourth MCP transport for remote servers that push events unprompted over a persistent bidirectional connection. Added `ws` to the transport-selection guidance, noting it must be configured with `type: "ws"` in `.mcp.json` (not `claude mcp add --transport`, which doesn't accept `ws`) and supports neither OAuth nor `--transport`. Consolidates audit PR #360.

The WebSocket MCP Transport feature-catalog entry described the config as `transport: websocket`, which is not how it's configured. Corrected to `type: "ws"` in `.mcp.json`, matching the module and the official docs.

Per the permission-modes docs (code.claude.com/docs/en/permission-modes), Auto Mode is now available by default on the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry; the `CLAUDE_CODE_ENABLE_AUTO_MODE=1` opt-in that applied on the non-Anthropic providers in v2.1.158–v2.1.206 was removed in v2.1.207 (still accepted for compatibility, no effect). Also corrected the supported-model list on the three non-Anthropic providers to Claude Sonnet 5, Opus 4.7, and Opus 4.8 — the previous text and PR #359 both dropped Sonnet 5. Consolidates audit PR #359.

#361 proposed replacing `--max-turns` with a `--runs <n>` flag across the workflows module and templates (EN + AR), citing a third-party changelog. The official CLI reference (code.claude.com/docs/en/cli-reference) confirms `--max-turns` exists and lists no `--runs` flag, so the change would have introduced a non-existent flag. The existing `--max-turns` content is correct and left unchanged; #361 is closed.

The Arabic mcp and advanced-features modules and the Arabic feature-catalog mirror all three EN changes — the WebSocket (ws) transport guidance, the corrected WebSocket config, and the Auto Mode default-availability and supported-model updates.

Per the MCP docs (code.claude.com/docs/en/mcp), tool search is on by default, and in configurations without it — Google Cloud's Agent Platform, a custom ANTHROPIC_BASE_URL, or ENABLE_TOOL_SEARCH=false — Claude uses the WaitForMcpServers tool instead. Replaced the older 'off by default on Vertex AI / non-first-party proxy' aside with this doc-verified phrasing. Consolidates audit PR #355.

Per the memory docs (code.claude.com/docs/en/memory), the @path/to/file import syntax expands imported files and loads them into context at launch alongside the CLAUDE.md that references them — they still consume context tokens. Replaced the misleading 'reference rather than duplicate' phrasing. Also clarified that ./CLAUDE.md and ./.claude/CLAUDE.md are two alternative supported locations for the project scope, not one shared file. Consolidates the doc-supported part of audit PR #356.

Per the discover-plugins docs (code.claude.com/docs/en/discover-plugins), plugin marketplaces source from GitHub repositories (owner/repo), git URLs, local paths, and remote marketplace.json URLs — not npm. Removed the incorrect npm claim in the EN module and added the corrected sentence to the AR module, which previously lacked it. Also reordered the Arabic plugins quiz (q5 cache cleanup, q6 version tag, q7 subagent restrictions) to match the English sequence. Consolidates audit PR #357.

#356 proposed collapsing the CLAUDE.md scopes to 'three truly distinct locations: managed policy, user/project instructions, and local instructions.' The memory docs list four distinct scopes and keep user (~/.claude/CLAUDE.md, personal) separate from project (./CLAUDE.md, team-shared), so this reframing is not supported and was dropped; only the accurate clarification that ./CLAUDE.md and ./.claude/CLAUDE.md are two alternative supported locations for the project scope (not one shared file) was kept.

The Arabic mcp and memory modules mirror the WaitForMcpServers fallback and the @import context-cost correction; the Arabic plugins module gains the corrected marketplace-sources sentence; and the Arabic plugins quiz is realigned to the English question order (q5–q7).

Per the fast-mode docs (code.claude.com/docs/en/fast-mode), fast mode is supported on Opus 4.8 and Opus 4.7 only; earlier Opus models like 4.6 are no longer supported. Replaced the stale 'supports Opus 4.6, removed ~30 days after the Opus 4.8 launch' copy with the doc-supported facts (Opus 4.7 fast mode deprecated June 25, 2026, removed July 24, 2026, no fallback), and corrected the rate-limit pool to 'shared across Opus 4.8 and 4.7'. Avoided the unverified v2.1.176 version claim from audit #352.

Per the permission-modes docs (code.claude.com/docs/en/permission-modes), Shift+Tab cycles `default` → `acceptEdits` → `plan`; `auto`, `bypassPermissions` (added by `--permission-mode bypassPermissions`, `--dangerously-skip-permissions`, or `--allow-dangerously-skip-permissions`), and `dontAsk` are not in that default cycle. Rewrote the commands module prose to include the `--allow-dangerously-skip-permissions` variant (missed by audit #352) and fixed the wrong order (`default → plan → acceptEdits → auto`) in the terminal `/permissions` output, the project-setup terminal step, and the reference cheatsheet.

#351's 'managed-only, cannot be set in user or project settings' qualifier for `enforceAvailableModels` is not supported by the settings docs, which describe a setting settable in any scope; the existing prose stays. #353 re-proposed two already-rejected regressions — the `/less-permission-prompts` rename (live skill and commands reference use `/fewer-permission-prompts`) and the skills budget 1%→2% change (docs say the budget scales at 1% of the context window) — so both are declined.

The Arabic commands module, project-setup terminal step, terminal `/permissions` output, and reference cheatsheet mirror the fast-mode Opus 4.8/4.7 update and the corrected `default → acceptEdits → plan` cycle order.

Per the MCP docs (code.claude.com/docs/en/mcp), `alwaysLoad: true` doesn't only exempt a server from tool-search deferral — it also blocks startup until that server connects, capped at the standard 5-second connect timeout, because the tools must exist when the first prompt is built. Other configured servers keep connecting in the background. Added this to the MCP module, which previously described only the deferral-exemption behavior.

The Arabic MCP module mirrors the added note (alwaysLoad blocks startup on that server's connect, 5-second cap, other servers keep connecting in the background).

Per the hooks docs (code.claude.com/docs/en/hooks) and sub-agents docs (code.claude.com/docs/en/sub-agents), the hook-input `effort` object and the subagent frontmatter `effort` field each accept only `low`, `medium`, `high`, `xhigh`, `max`. Removed the invalid `auto` value from both modules. `auto` remains a valid `/effort` command argument (reset to model default) in the commands and slash-commands modules, so those are unchanged.

The Arabic hooks and subagents modules mirror the corrected effort-level list (low/medium/high/xhigh/max, no auto).

Per the headless docs (code.claude.com/docs/en/headless), `--bare` reduces startup time by skipping auto-discovery of hooks, skills, plugins, MCP servers, auto memory, and CLAUDE.md, so only explicitly-passed flags and settings take effect. Replaced the misleading 'cleanest scripted output' framing in the workflows module and the 'Cleanest output... no formatting or decoration' description in the reference cheatsheet.

The Arabic workflows module and reference cheatsheet mirror the corrected --bare description (startup speed and reproducibility over output cleanliness).

Per the permission-modes docs (code.claude.com/docs/en/permission-modes), pressing `Ctrl+G` at the plan-approval prompt opens the proposed plan in your default text editor to edit before Claude proceeds. Kept this correct content (audit PR #334 had wrongly removed it) and refined two surfaces for precision: the reference cheatsheet now separates the general behavior (opens your prompt) from the plan-mode behavior, and the terminal simulator's /plan output no longer suggests Ctrl+G exits plan mode.

The Arabic advanced-features module, reference cheatsheet, terminal simulator /plan output, and guided terminal step mirror the clarified Ctrl+G plan-edit content.

Per the official docs (code.claude.com/docs/en/desktop-linux), the Claude desktop app is in beta on Linux — Ubuntu 22.04+ or Debian 12+, x86_64/arm64, installed via Anthropic's apt repository or a downloaded .deb — with Computer Use, dictation, and Fedora/RHEL not yet included. Replaced the outdated 'not available on Linux' line in the getting-started module, using the full docs URL rather than a root-relative link that would 404 on the static site.

The Arabic getting-started module mirrors the Desktop-App-on-Linux-beta correction, including the beta's known gaps and the full docs URL.

Per the official docs (code.claude.com/docs/en/agents), as of v2.1.198 /agents no longer opens a panel — it prints a notice pointing to the subagent file locations, and custom subagents are created or edited by asking Claude or editing the files directly. Corrected across the subagents module prose, the reference cheatsheet, the terminal simulator's /agents output, the subagents guided terminal step, three subagents quiz explanations, and the /agents feature-catalog entry.

The Arabic subagents module, reference cheatsheet, terminal simulator, guided terminal step, subagents quiz, and feature catalog mirror the /agents-prints-a-notice correction.

Per the Claude Code CHANGELOG (v2.1.198), the built-in Explore agent now inherits the main session's model, capped at opus, instead of running on Haiku. Corrected across the subagents module prose, all four subagents quiz options (the correct answer and the a/c/d distractor explanations), the skills quiz explanation for `agent: Explore`, the reference cheatsheet, the terminal simulator's `claude agents` output, and the Explore feature-catalog entry.

Auto-generated PR #326 re-proposed renaming `/fewer-permission-prompts` to `/less-permission-prompts`, a change already investigated and reverted in v1.24.8. The current CHANGELOG mentions neither name and the live skill is `/fewer-permission-prompts`, so the existing content is correct and unchanged.

The Arabic subagents module, subagents and skills quizzes, reference cheatsheet, terminal simulator, and feature catalog mirror the Explore-inherits-the-session-model correction.

Added a Fallback Model Chains section covering the `--fallback-model` flag and the `fallbackModel` settings array: tried in order when the primary model is overloaded, unavailable, or returns a non-retryable server error; auth/billing/rate-limit/size/transport errors never switch; capped at three models; the switch lasts the current turn only; the flag takes precedence over the setting (code.claude.com/docs/en/model-config).

Added the `--fallback-model <list>` CLI flag to the reference cheatsheet and the feature catalog so the search palette surfaces fallback model chains.

Added quiz q30 testing what happens when the primary model returns an overloaded error mid-turn with a fallback chain configured (switch for the current turn only, then retry the primary).

Mirrored the Fallback Model Chains section, the cheatsheet and feature-catalog entries, and quiz q30 in the Arabic advanced-features module and data files.

Clarified that `CLAUDE.md` and `CLAUDE.local.md` files in the directory hierarchy above the working directory are loaded in full at launch, while subdirectory files load on demand when Claude reads files in those directories (code.claude.com/docs/en/memory).

Corrected an over-broad claim that PowerShell is the default shell tool on all Windows and that Git for Windows is no longer required. Git Bash is used when Git for Windows is installed; the PowerShell tool is the shell only when Git Bash is absent; with Git Bash present it is rolling out progressively (opt in with CLAUDE_CODE_USE_POWERSHELL_TOOL=1) and is default-on only for Bedrock, Vertex, and Foundry users (code.claude.com/docs/en/setup).

The Arabic advanced-features module and quiz claimed cleanupPeriodDays also removes ~/.claude/tasks/, shell-snapshots/, and backups/. The official settings docs list only session files and orphaned subagent worktrees, so the extra scope was removed to match the English module and the docs. The auto-generated PR that tried to add the same claim to the English module was not merged.

The Arabic memory and getting-started modules mirror the English wording changes, and the Arabic advanced-features module and quiz q7 were corrected back to the documented cleanupPeriodDays scope (session files plus orphaned worktrees).

`claude mcp login <name>` runs a configured server's OAuth flow from the shell without opening the in-session `/mcp` panel (from v2.1.186), and `claude mcp logout <name>` clears its stored credentials. `--no-browser` prints the authorization URL instead of opening a browser — useful over SSH, where you open the URL locally and paste the redirect URL back at the prompt. Added to the mcp module, reference cheatsheet, and feature catalog.

The terminal simulator now responds to `claude mcp login` (with the --no-browser flag and an OAuth success output) and `claude mcp logout`, and the `/mcp` panel output lists `claude mcp login` under its manage-servers hint.

The Arabic mcp module, reference cheatsheet, terminal simulator, and feature catalog mirror the `claude mcp login` / `claude mcp logout` documentation, including the --no-browser / SSH note.

Path-scoped rules in .claude/rules/ trigger when Claude reads files matching the pattern, not on every tool use; rules without a paths field load at launch with the same priority as .claude/CLAUDE.md; circular symlinks are detected and handled gracefully. A new 'Write effective instructions' section covers the context-window visualization and the under-200-lines-per-CLAUDE.md target. Added to the memory module.

Channels lead with their research-preview status and v2.1.80 minimum (mcp); a single-skill plugin can place SKILL.md at the plugin root with the frontmatter name as the invocation name (plugins); the built-in Explore and Plan agents skip CLAUDE.md and git status while every other subagent loads both (subagents); and dynamic-workflow resume only works within the same Claude Code session, restarting fresh after you exit (workflows).

Auto-generated PR #315 replaced the English memory module's `<!-- terminal: guided -->` marker with its new paragraph, which would have broken the guided terminal step and left EN out of parity with AR (which kept the marker). The marker is restored and the new paragraph now sits before it, matching the Arabic structure.

The Arabic memory, mcp, plugins, subagents, and workflows modules mirror every English change — path-scoped rule behavior and the effective-instructions section, the Channels rewording, the single-skill SKILL.md note, the Explore/Plan skip note, and the workflow resume clarification.

MCP_TIMEOUT sets the server startup/connection timeout at launch. Tool execution has a separate cap: a `timeout` field in milliseconds on a server's `.mcp.json` entry (e.g. "timeout": 600000 for ten minutes) limits how long any single tool call may run, overrides MCP_TOOL_TIMEOUT for that server only, is a hard wall-clock limit that progress notifications don't extend, and ignores values below 1000. Added to the mcp module.

Audit PR #312 wanted to relabel MCP_TIMEOUT as 'a per-server startup timeout, not a general per-server timeout'. The docs call it the 'MCP server startup timeout' — a launch-time environment variable, not a per-server field — so the awkward clause was dropped in favor of the accurate startup-vs-tool-execution split.

The Arabic mcp module mirrors the startup-timeout (MCP_TIMEOUT) versus tool-execution-timeout (per-server `timeout` field / MCP_TOOL_TIMEOUT) split, matching the English change.

claude plugin prune (new in v2.1.121, aliased autoremove) removes auto-installed plugin dependencies that no other installed plugin still requires — directly-installed plugins are never touched, and 'uninstall <plugin> --prune' does both in one step. claude plugin details now lists the component inventory grouped as Skills, Agents, Hooks, MCP servers, and LSP servers with a projected per-session token cost. Added to the plugins module, the reference cheatsheet, and the feature catalog, EN and AR.

An audit PR removed the auto:N variant of ENABLE_TOOL_SEARCH. The MCP reference documents it verbatim — 'auto:N: Threshold mode with a custom percentage, where N is 0-100' (e.g. auto:5 for 5%) — so the variant stays. The mcp module and quiz were left unchanged.

The Arabic plugins module mirrors the corrected claude plugin prune description and the expanded claude plugin details inventory, matching the English changes.

The '(new in v2.1.147)' tag implied /code-review was new; the changelog records /simplify being renamed to /code-review in v2.1.147. The module and quiz now describe /code-review as the bundled skill formerly called /simplify, in EN and AR.

An audit PR dropped 'Console' from the supported account list. The install doc states 'Claude Code requires a Pro, Max, Team, Enterprise, or Console account' — Console stays. We adopted the docs' clearer 'requires … account' phrasing while keeping all five account types.

An audit PR removed two documented commands and re-dated the startup-retry behavior to v2.1.187. The MCP reference documents both 'claude mcp add-from-claude-desktop' and 'claude mcp serve', and dates the up-to-three startup retries on transient errors to v2.1.121 (v2.1.187 is the unrelated 5-minute remote-tool idle timeout). The original mcp content was correct and was kept unchanged.

The Arabic slash-commands module and quiz mirror the /code-review-formerly-/simplify clarification, and the Arabic getting-started module keeps Console in the account list with the 'requires' phrasing.

v1.24.5 renamed the command the wrong way round. The official name — per the v2.1.105–v2.1.113 release notes and the live /fewer-permission-prompts skill — is /fewer-permission-prompts. Reverted across the slash-commands and skills modules, the reference cheatsheet, and the feature catalog (entry id and name), so search and every surface now show the correct command.

The subagent definition priority no longer lists 'built-in' as a priority level. Per the features-overview docs, built-in subagents are always available and are not part of the name-override priority system, which runs managed > CLI flag > project > user > plugin.

An audit proposed moving the rename from v2.1.160 to v2.1.154, citing the GitHub CHANGELOG.md. The official workflows docs (code.claude.com/docs/en/workflows) state 'Before v2.1.160 the literal trigger keyword was workflow' and separately note workflows require v2.1.154+ — so v2.1.160 is the documented rename version and the existing content was kept unchanged.

The Arabic slash-commands, skills, and subagents modules, plus the Arabic feature catalog and reference cheatsheet, mirror the /fewer-permission-prompts revert and the subagent-priority correction.

The MCP module now spells out the override values from the docs: true forces tool search always on, false loads every definition upfront on every turn, and auto (or auto:N) activates tool search only when tool definitions exceed 10% (or N%) of the context window.

An audit PR claimed tool search only works after setting ENABLE_TOOL_SEARCH=auto and that the default loads every MCP definition upfront. The official tool search reference states tool search is enabled by default (definitions deferred and discovered on demand), off by default only on Vertex AI or non-first-party ANTHROPIC_BASE_URL proxies. The original content was correct; PR #298 was a regression and is closed.

Quiz q6 previously framed the 10% threshold as automatic default behavior. Per the docs, the 10% threshold is specifically the ENABLE_TOOL_SEARCH=auto mode — under the default (unset), tool search is already on regardless of size. The question now asks what auto does when definitions cross 10%, with explanations corrected to match.

The Arabic MCP module and quiz get the same correction: tool search on by default, the full ENABLE_TOOL_SEARCH value table, and q6 reframed around auto mode.

A reader said the memory hierarchy was confusing without context on why it matters. The Memory Hierarchy section now has an 'In practice' callout: use project memory for what a teammate needs to understand the codebase (setup, testing, architecture), user memory for how you personally like to work, and CLAUDE.local.md for entries that only matter to you.

The plugins module attributed claude plugin tag to v2.1.121. The official changelog shows v2.1.118 added claude plugin tag (release git tags with version validation) and v2.1.121 added the separate claude plugin prune command, so the annotation now reads v2.1.118.

An audit PR removed max from the subagents effort values, claiming v2.1.72 dropped it. The official sub-agents reference lists the effort options verbatim as low, medium, high, xhigh, max, so the existing content is correct and max was kept.

The Arabic memory module gets the same 'In practice' callout, the Arabic plugins module gets the v2.1.118 fix for claude plugin tag, and the Arabic subagents module keeps max in its effort values.

The command is now /less-permission-prompts across the slash-commands module (PR #291), the feature catalog, and the reference cheatsheet. The catalog and cheatsheet had been missed by PR #291 and still showed the old /fewer-permission-prompts name, so the command search index now matches the official spelling. Verified against the changelog (added in v2.1.112).

PR #290 repeated the already-rejected #288 proposal to change the skill-description budget from 1% to 2%. The official /en/skills docs state 'The budget scales at 1% of the model's context window' and give 0.02 = 2% only as the example for raising it via skillListingBudgetFraction, so the existing 1% content is correct and was kept.

The Arabic feature catalog and reference cheatsheet now use /less-permission-prompts, matching the English content and the slash-commands module.

The existing Agent Teams paragraph now notes that v2.1.178 removed the TeamCreate and TeamDelete tools — with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 set, every session already has one implicit team, so teammates are spawned directly through the Agent tool's name parameter with no setup step and cleanup happens automatically on session exit. Verified verbatim against the changelog.

The skills module and quiz keep '1% of the context window' as the default skill-description budget. The official skills docs say 1% verbatim; the 0.02 = 2% value the audit cited is the example for raising the budget with skillListingBudgetFraction, not the default, and the claimed changelog line does not exist.

The dynamic-workflow paragraph keeps its accurate wording: the trigger keyword was renamed from workflow to ultracode in v2.1.160. The changelog says 'Renamed the dynamic-workflow trigger keyword from workflow to ultracode', so the audit's 'two separate changes in v2.1.178, not a rename' reframing was incorrect and was not applied.

The Arabic subagents module mirrors the v2.1.178 Agent Teams change (TeamCreate/TeamDelete removed, implicit team per session, spawn via the Agent tool's name parameter, automatic cleanup on session exit).

The hooks module said Claude Code 'supports 30+ hook events'; the official hooks docs enumerate exactly 30, so the prose now states '30 hook events' for precision.

Added the Auto Mode fallback thresholds (3 consecutive blocks or 20 total blocks pause auto mode and resume prompting; not configurable) to the classifier-precedence paragraph, matching the permission-modes docs and making the existing quiz q2 answerable from the module.

The Arabic hooks and advanced-features modules mirror both changes — '30 hook event' and the Auto Mode block-threshold sentence.

Quiz q3 asks 'Which subscription plans include Claude Code access?', so the correct answer stays the four subscription plans (Pro, Max, Team, Enterprise) and the explanation now frames the Anthropic Console accurately as a non-subscription option — API access billed from pre-paid credits, with the same browser sign-in — instead of PR #281's inaccurate 'Console uses API key authentication rather than a subscription' framing. Cloud providers (Bedrock, Vertex AI) are noted too.

The Arabic getting-started quiz q3 mirrors the corrected wording: the four subscription plans remain the answer, and the Console is described as a non-subscription option (pre-paid API access, same browser sign-in).

Quiz q6 described `${CLAUDE_EFFORT}` as resolving to 'low, medium, high, xhigh, max, or ultra — the stored value for ultracode'. The docs do not define an 'ultra' effort value for this variable, so the explanation now lists the verified levels low/medium/high/xhigh/max.

The `CLAUDE_EFFORT` environment-variable entry no longer claims it can resolve to 'ultra for ultracode', matching the feature catalog's accurate list of low/medium/high/xhigh/max.

The Arabic skills quiz q6 mirrors the corrected `${CLAUDE_EFFORT}` level list (low/medium/high/xhigh/max).

The terminal playground simulates `/cd <directory>` (moving the session without breaking the prompt cache) and lists it in `/help`; the reference cheat sheet and feature catalog both gained a `/cd` entry.

The cheat sheet documents the `--safe-mode` flag and `CLAUDE_CODE_SAFE_MODE` env var (start with all customizations disabled for troubleshooting), and the feature catalog adds a `--safe-mode` flag entry.

Added `CLAUDE_CODE_DISABLE_BUNDLED_SKILLS` / `disableBundledSkills`, plus the `enforceAvailableModels` (v2.1.175) and `footerLinksRegexes` (v2.1.176) managed settings, to the reference cheat sheet's env-vars and config-options sections.

The Arabic terminal commands, cheat sheet (slash commands + CLI flags), and feature catalog mirror the `/cd` and `--safe-mode` additions.

The MCP module states that HTTP or SSE servers reconnect mid-session with exponential backoff (up to five attempts), and that as of v2.1.121 the initial connection at startup is retried up to three times on transient errors such as a 5xx response, connection refused, or timeout.

The Channels section now states that the Telegram, Discord, and iMessage channel plugins require Claude Code v2.1.80 or later, matching the official channels documentation.

The MCP quiz had two near-identical questions both testing the `alwaysLoad` config option; the redundant one is removed and the remaining questions renumbered to q1–q9.

The catalog described channels as 'Telegram, Slack, and other platforms', and the `channelsEnabled` managed setting likewise listed 'Telegram, Slack'; both now correctly list the Telegram, Discord, and iMessage channel plugins, and the feature entry notes the research-preview / v2.1.80 requirement. Slack is a separate integration, not a channel.

The Arabic MCP module mirrors the SSE/startup-retry and v2.1.80 channel updates, the Arabic MCP quiz is renumbered to a matching q1–q9, and the Arabic feature catalog's Channels entry is corrected.

The Subagents module no longer says `isolation: worktree` agents can use `ExitWorktree` to leave their worktree early — the tools reference states `ExitWorktree` is not available to subagents that run in their own working directory. The passage now covers worktree locking during the run and the `worktree.baseRef` base-branch behavior instead.

The module now lists `fresh` (branch from the remote default for a clean tree) as the default for `worktree.baseRef`, with `head` as the opt-in for branching from local HEAD — matching the official docs, the module quiz, and the reference cheatsheet.

The Arabic Subagents and Advanced Features pages mirror the ExitWorktree and baseRef corrections.

The Plugins module now lists only the managed-policy settings that exist in the official docs — `enabledPlugins`, `extraKnownMarketplaces`, `strictKnownMarketplaces`, and `blockedMarketplaces` — instead of the non-existent `deniedPlugins`.

The team-disable quiz question no longer presents `deniedPlugins` as a real setting; the distractor and its explanation now turn on the real `blockedMarketplaces` managed-policy blocklist.

The Arabic Plugins module and quiz mirror the English correction, dropping `deniedPlugins` for the real managed-policy settings.

The Project Setup module now states plainly that project settings are overridden by local settings, and that only managed settings and command-line arguments override local — matching the official precedence order (managed > command-line > local > project > user).

The Arabic Project Setup page mirrors the precedence clarification and now uses 'الإعدادات المُدارة' consistently for the managed layer instead of switching to 'سياسة المنظمة'.

The Workflows module no longer claims the 3 Pro/Max free /ultrareview runs expire on May 5, 2026. Per the official ultrareview docs, the three runs are a one-time allotment per account that does not refresh — there is no published expiry date. Corrected in EN and AR.

Quiz q7 keeps the doc-accurate 'renamed from /simplify in v2.1.147' framing (matching the commands quiz, reference cheatsheet, and templates), and its closing note now matches the docs: from v2.1.154 /simplify is a separate cleanup-only review that applies fixes without hunting for bugs, and /code-review --fix is the bug-finding-with-fixes path.

The Arabic Workflows module and slash-commands quiz mirror the /ultrareview free-runs correction and the /code-review wording, with natural phrasing in place of an earlier mixed-script glitch.

Skills module and quiz corrected from ~2% to ~1%, the documented default. The 2% figure is only an example of a raised budget, not the default. Raise it with the skillListingBudgetFraction setting or the SLASH_COMMAND_TOOL_CHAR_BUDGET environment variable. Verified against the official skills docs.

The Skills module and quiz now point to /doctor to see whether the skill-description budget is overflowing and which skills are dropped from the listing, matching the official docs.

The Arabic Skills module and quiz mirror the 1%/budget corrections, and a tanween spelling fix (مثلاً → مثلًا) is applied in the budget example.

Getting Started now explains that `/terminal-setup` writes the Shift+Enter keybinding for VS Code, Cursor, Devin Desktop, Alacritty, and Zed — not notification support for Kitty/Ghostty/Alacritty. Desktop notifications work by default in Ghostty, Kitty, and iTerm2; other terminals use `preferredNotifChannel: "terminal_bell"` or a Notification hook. Verified against the terminal-config docs page.

In VS Code, Cursor, and Devin Desktop, `/terminal-setup` turns off the integrated terminal's GPU acceleration (to prevent garbled text) and adjusts its scroll sensitivity for smoother fullscreen scrolling. This accurate detail was preserved and made precise rather than dropped. The feature-catalog description was sharpened to match.

The Arabic Getting Started module and feature catalog mirror the English corrections with consistent phrasing for the Shift+Enter keybinding scope, the notification defaults, and the GPU-acceleration/scroll-sensitivity behavior.

A new search island lets you open a command palette with ⌘K / Ctrl+K and find modules, commands, and reference material from any page, with keyboard navigation through results.

Commands module, slash-commands module, quizzes, reference cheatsheet, and terminal simulation now list the full documented effort range (low/medium/high/xhigh/max/ultra). `/code-review ultra` runs a deeper multi-agent review in the cloud — the same one previously surfaced as `/ultrareview`.

From v2.1.154 /simplify runs a separate cleanup-only review that applies fixes without hunting for bugs. The commands module, slash-commands module, both quizzes, reference cheatsheet, templates, terminal-steps, and terminal simulation all drop the 'still works as an alias' framing and now explain the v2.1.154 split, with /code-review as the bug-finding tool.

The workflows module, workflows quiz (Q9), reference cheatsheet /workflows entry, and terminal simulation now describe the trigger as `ultracode` instead of `workflow`. The rename landed in v2.1.160; asking for a workflow in your own words still works. Verified against the workflows docs page.

Advanced-features module clarifies that `MAX_THINKING_TOKENS=0` only disables thinking on Opus 4.6 and Sonnet 4.6. On Opus 4.7 and later, adaptive reasoning is always used and the variable has no effect. Verified against the model-config docs page.

The workflows module no longer states that `@claude` mentions run in a sandboxed environment — the github-actions docs say `runs-on: ubuntu-latest` and 'Your code stays on GitHub's runners' but do not use the 'sandboxed environment' framing.

Arabic versions of the commands, slash-commands, workflows, and advanced-features modules, plus the Arabic quizzes, reference cheatsheet, and terminal simulation all mirror the EN updates with consistent phrasing for the effort-range extension, the /simplify split, the ultracode trigger rename, and the MAX_THINKING_TOKENS scoping.

When `status` is `waiting`, the new `waitingFor` field (v2.1.162) says exactly what the session is blocked on — `permission prompt` or `input needed`. The subagents module, templates, terminal simulation, and reference cheatsheet now include a jq example that wakes only permission-prompt sessions with `claude respawn`, so a boot script can route those two cases to different handlers.

Advanced-features module now documents `OTEL_RESOURCE_ATTRIBUTES` as the supported way to attach team, department, or cost-center labels to every metric datapoint and event. Formatting rules (comma-separated, no spaces, percent-encode special characters) and the `OTEL_METRICS_INCLUDE_RESOURCE_ATTRIBUTES=false` cardinality escape hatch are called out. Built-in attributes like `user.id` and `session.id` always win on key collision.

Subagents module, terminal simulation, templates, quiz Q9 explanation, reference cheatsheet, and feature catalog now list the documented `pid`, `cwd`, `kind`, `startedAt`, `sessionId`, `name`, and `status` fields — replacing the older `id`/`state`/`project`/`last_activity` names. All jq examples are updated from `.state` to `.status` and from `.id` to `.sessionId` so they actually parse the live output.

Arabic subagents module, advanced-features module, terminal simulation, templates, quiz, reference cheatsheet, and feature catalog all mirror the field-name correction, the `waitingFor` addition, and the `OTEL_RESOURCE_ATTRIBUTES` env var, with consistent Arabic phrasing for the OTLP resource block and the cardinality escape hatch.

The official VS Code extension docs page now states the extension installs in 'other VS Code forks like Devin Desktop or Kiro'. The getting-started module and Q5 quiz explanation in the getting-started quiz both add Kiro alongside Cursor and Devin Desktop so learners see the full set of VS Code forks where the Claude Code extension is supported.

Getting-started module no longer presents Windsurf as a current VS Code fork name — it now describes Cursor, Devin Desktop (renamed from Windsurf in v2.1.162), and Kiro as the VS Code forks that support the Claude Code VS Code extension and the integrated terminal. The paragraph also notes that `/ide` and `/scroll-speed` use the new Devin Desktop label, on top of the existing `/terminal-setup` auto-configuration.

The Arabic getting-started module and Arabic getting-started quiz mirror the rename — Cursor وDevin Desktop (الاسم الجديد لـ Windsurf من v2.1.162) وKiro — with consistent phrasing for the `/ide`, `/terminal-setup`, and `/scroll-speed` label change.

Advanced-features module now documents Auto Mode platform availability: on the Anthropic API it's on by default once your account meets the model and admin requirements; on Bedrock, Vertex AI, and Foundry (v2.1.158+) it's gated behind `CLAUDE_CODE_ENABLE_AUTO_MODE=1` and only Opus 4.7 and Opus 4.8 are supported, while Sonnet 4.5, Opus 4.5, Haiku, and Claude 3 are not. The `permissions.disableAutoMode: "disable"` admin override is called out as the way to lock the mode off in managed settings.

The /ultrareview paragraph used to caveat that workflows generally work on Bedrock, Vertex AI, and Foundry while `/ultrareview` specifically might be restricted. The official workflows page now states plainly that dynamic workflows are available on all paid plans on Bedrock, Vertex AI, and Foundry with Claude Code v2.1.154+, so the hedge is removed in favor of the confirmed availability.

Both changes are mirrored in the Arabic advanced-features and workflows modules with consistent phrasing for platform availability, the `CLAUDE_CODE_ENABLE_AUTO_MODE` opt-in, and the model restriction to Opus 4.7/4.8 on third-party providers.

Advanced-features and commands modules now document the v2.1.160 rename of the trigger keyword from `workflow` to `ultracode`. The literal `workflow` keyword no longer triggers a run, but a natural-language request still works, and `/effort ultracode` activates dynamic workflows.

Removed all instructions to set this env var: from the commands module fast-mode section, the advanced-features environment-variable section and templates, and the feature catalog. Quiz Q23 in advanced-features now teaches that the variable was removed in v2.1.160 with no replacement.

Workflows module now reflects the official CHANGELOG — /ultrareview has been available since Claude Code v2.1.111, instead of being marked only as new in v2.1.112. (This entry originally read v2.1.86; corrected in v1.24.31.)

Every change is mirrored in the Arabic modules (advanced-features, commands, workflows), Arabic templates, Arabic feature catalog, and Arabic quizzes, with Arabic phrasing for the `workflow`/`ultracode` distinction clarified after CodeRabbit review.

Commands module no longer pins fast mode to Opus 4.6 alone. Opus 4.8 is the fast mode default in v2.1.154+, and the CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1 env var is documented for pinning back to 4.6 while it remains supported. The Opus 4.6 fast mode deprecation timeline is called out.

Commands module now lists all platforms where fast mode is unavailable — Bedrock, Vertex AI, Foundry, Claude Platform on AWS — and notes it is not supported in the VS Code extension. The rate-limit pool is shared across Opus 4.8, 4.7, and 4.6.

Project-setup quiz Q4 option B explanation was inverted: settings.local.json (not project settings) overrides project. The corrected explanation now reads: project settings override user settings, but they are overridden by local settings, managed policy, and command-line arguments — matching the official settings hierarchy.

Every change is mirrored in the Arabic commands module and the Arabic project-setup quiz, with consistent terminology for deprecation and platform exclusions.

Plugins module now shows monitors declared under experimental.monitors in plugin.json, with a note that top-level still works but claude plugin validate warns. Feature catalog entry updated to match.

Project-setup module now tells learners that monitors and themes are experimental plugin manifest keys and should be nested under experimental: {}.

Mirrored in the Arabic plugins and project-setup modules plus the AR feature catalog, with cleaner Arabic phrasing for the new project-setup paragraph.

Getting-started module now describes Cursor and Windsurf as VS Code forks that reuse the VS Code extension and integrated terminal, with /terminal-setup auto-configuring GPU acceleration and scroll sensitivity. Quiz updated to match.

Memory module now notes that subagents can maintain their own auto memory, with a link back to the subagents module for configuration details.

Project-setup module now distinguishes DISABLE_UPDATES=1 (blocks all update paths, including manual claude update) from DISABLE_AUTOUPDATER (suppresses package-manager update notifications only).

Skills module and quiz updated from the older ~1% figure to the current ~2% budget, with SLASH_COMMAND_TOOL_CHAR_BUDGET callable to raise the cap. All q3 option explanations now agree.

Slash-commands module clarifies that /code-review is a bundled skill that reports findings, /code-review --fix applies them to the working tree, and /simplify now invokes /code-review (report only) — the old cleanup-and-fix /simplify behavior was removed, but /simplify still works as an alias.

Subagents module and quiz now state that forked subagents work in interactive sessions, non-interactive mode, the Agent SDK, and on external builds — dropping the inaccurate 'interactive mode only' framing.

Workflows module no longer claims /ultrareview is unavailable on Bedrock, Vertex AI, or Foundry. Learners are pointed to the official docs to confirm, since the workflows feature itself works on those platforms.

Every fix is mirrored in the Arabic module pages and quizzes, including the cross-module link in memory.md and the consistent 2% phrasing across all skills q3 explanations.

MCP module now says the Telegram, Discord, and iMessage channel plugins are 'included in' the research preview, matching the official docs phrasing.

Workflows module now covers dynamic workflows (v2.1.154+) — orchestrating dozens to hundreds of subagents from a script Claude writes, the bundled /deep-research workflow, /effort ultracode, saving a run as a /command, and disabling via /config or CLAUDE_CODE_DISABLE_WORKFLOWS=1. Mirrored in the feature catalog, reference cheatsheet, and terminal simulation (/workflows, /deep-research).

Skills module now lists the full set of bundled skills available in every session — /code-review, /batch, /debug, /loop, and /claude-api — alongside the /run, /verify, and /run-skill-generator trio (v2.1.145+). Added /claude-api to the terminal simulation.

Added a workflows quiz question on how to start a dynamic workflow (the 'workflow' keyword or /effort ultracode, not /workflows) and a skills quiz question on bundled skills. Corrected the skills effort-variable answer to include the ultra value.

Translated the dynamic workflows section, bundled-skills table, new quiz questions, feature-catalog entries, and cheatsheet rows into Arabic.

Hooks and subagents modules now list the full effort set — low, medium, high, xhigh, max, auto — matching official docs. xhigh was added for Opus 4.7, sitting between high and max.

Skills module now notes that enabling the PowerShell tool with CLAUDE_CODE_USE_POWERSHELL_TOOL=1 on Linux and macOS also requires pwsh on your PATH.

Commands module (Arabic) now shows Shift+Tab cycling default → acceptEdits → plan, matching the English module and official docs.

Slash commands module, feature catalog, reference cheatsheet, and terminal simulation now cover /code-review ultra (alias /ultrareview) — a deep, multi-agent code review run in a cloud sandbox.

Arabic feature catalog now uses Arabic category labels for the new entries, and fenced code blocks carry language tags.

Commands module now covers automatic session recaps when returning to the terminal, /recap on-demand summaries, and Ctrl+R reverse history search with Ctrl+S scope cycling.

Workflows module now documents push notifications via Remote Control — Claude sends phone alerts when tasks finish or need input.

Advanced features module now documents hard_deny rules (unconditional blocks), four-tier precedence, and claude auto-mode defaults/config/critique CLI subcommands.

Slash commands module now includes /recap, /fewer-permission-prompts, /tui, /focus, and /undo (alias for /rewind).

Hooks module now documents continueOnBlock (feed rejection back to Claude instead of ending the turn) and terminalSequence (emit notifications without a terminal).

All new content mirrored in Arabic modules, quizzes, feature catalog, and reference cheatsheet.

Commands module now covers built-in output styles (Default, Proactive, Explanatory, Learning) and custom output style creation with Markdown files and keep-coding-instructions frontmatter.

MCP module now documents channels — MCP servers that push events (Telegram, Discord, iMessage) into running sessions. Covers --channels flag, pairing, allowlists, and enterprise controls.

Advanced features module now covers the computer-use MCP server for GUI control on macOS. Includes enabling, permissions, per-app approval, and safety features.

Workflows module now fully documents routines: API triggers, GitHub triggers, web UI at claude.ai/code/routines, connectors, one-off scheduled runs, and /schedule list/update/run CLI management.

All new content mirrored in Arabic modules, quizzes, feature catalog, and reference cheatsheet.

Project-setup module now correctly documents that Homebrew/WinGet can opt in to auto-updates via CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE=1, adds autoUpdatesChannel setting, and DISABLE_AUTOUPDATER env var. Previous PR #195 incorrectly claimed the env var was unsupported.

Getting-started module, quiz, and installation tabs no longer describe npm as deprecated — the official docs list it as a supported alternative installation method.

Documents the three bundled skills for launching, verifying, and recording app build recipes. Added to both English and Arabic skills module.

Documents the skillOverrides setting for controlling skill visibility from settings.json without editing SKILL.md frontmatter. Includes the /skills interactive menu workflow.

Documents ${CLAUDE_SESSION_ID}, ${CLAUDE_EFFORT}, ${CLAUDE_SKILL_DIR}, and named arguments frontmatter in the skills module.

Getting-started module now documents apt, dnf, and apk installation methods and the claude doctor verification command.

All module text, quiz corrections, and tab content mirrored in Arabic.

English corrected from 'interactive mode only' to 'interactive sessions, non-interactive mode, and the Agent SDK'. Arabic completely rewritten to match (was incorrectly described as external-builds-only).

#192

Version corrected from v2.1.112 to v2.1.111 (confirmed via CHANGELOG) in both English and Arabic skills module.

#191

PostToolUse hooks can replace tool output entirely via hookSpecificOutput.updatedToolOutput (works for all tools since v2.1.121). Quiz q8 tested it, but the module content was missing.

#106

--add-dir grants Read/Edit access to additional directories; --mcp-config loads MCP servers from external JSON for the current session. Both documented in subagents module and terminal simulator.

#171

The English hooks module documented the exec form (args array for shell-free process spawning) but Arabic was missing it entirely.

Reviews the current diff for correctness bugs and reports findings without editing files. Accepts effort levels (low/medium/high/xhigh/max) and --comment to post inline GitHub PR comments. Renames /simplify, which still works as an alias.

#184

Mirrored across Arabic module content, quizzes, terminal steps, terminal-commands.json, templates, the feature catalog, and the reference cheatsheet.

#184

claude agents --json prints the session roster as a JSON array — id, state, name, project, last activity — for tmux-resurrect-style boot scripts, status bars, and session pickers.

#179

Stop and SubagentStop hook inputs now include background_tasks (running background bash/subagents) and session_crons (queued scheduled tasks). A completion-gate hook can block until both arrays drain.

#179

The /plugin Discover and Browse screens now show a plugin's commands, agents, skills, hooks, and MCP/LSP servers inline before installation.

#179

When a whole-file read exceeds the token limit, Read now returns a truncated first page with a PARTIAL view notice instead of a hard error. Claude paginates with offset/limit or narrows with grep.

#179

v2.1.144 renamed 'extra usage' to 'usage credits' across the CLI; /extra-usage still works as an alias. Catalog, cheatsheet, slash-commands module, fast-mode copy, and terminal sim updated.

#174

All five features mirrored across Arabic module content, quizzes, terminal steps, templates, terminal-commands.json, feature catalog, and reference cheatsheet.

#174, #179

Fact-check fixes in memory, project-setup, hooks, and plugins modules from automated audit pipeline.

#176, #177, #178, #180, #181

Toggle installed plugins on or off without uninstalling. Accepts --scope (user/project/local) to control where the setting is stored.

#171

Plugin Discover/Browse screens now show projected token cost per session for each plugin.

#171

claude plugin details <name> shows component inventory and token cost; claude plugin tag creates release git tags with version validation.

#171

Changelog data moved from GitHub API scraping to a structured YAML file (changelog.yaml) with per-release entries, categories, and issue references.

#170

claude agents --cwd <path> filters the Agent view to only show agents running in that directory.

#168

Documents --permission-mode <mode> flag with the complete list of accepted permission modes.

#168

Documents ANTHROPIC_WORKSPACE_ID for Workload Identity Federation on cloud platforms.

#168

CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1 ensures plugins are cloned over HTTPS in CI environments.

#168

Higher-contrast text in expanded state with full text-primary color and thicker chevron stroke when collapsed.

#162

Corrected the pre-existing backwards description for CLAUDE_CODE_NATIVE_CURSOR.

#164

CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE, CLAUDE_CODE_RESUME_PROMPT, and CLAUDE_CODE_ENABLE_FEEDBACK_SURVEY_FOR_OTEL across modules, feature catalog, quizzes, and templates.

#164

Added a security note for --plugin-url in both English and Arabic plugins documentation.

#160

Gateway model discovery, alternate screen toggle, image pasting hint, --plugin-url flag, package manager auto-update, and SESSION_ID env var.

#160

New quiz questions covering the 6 verified features added this release, with correct numbering throughout.

#160

15 missing Arabic quiz questions added (getting-started q7-q8, hooks q9-q10, project-setup q6, advanced-features q13-q21), plus 4 terminal steps and 4 templates.

#160

Module content, quizzes, templates, and feature catalog entries for 7 newly verified Claude Code features.

#159

Module content, quizzes, and templates for 5+ newly verified Claude Code features.

#142

Shift+Tab now correctly cycles through permission modes in the right order instead of skipping states.

#135

Added a CI code review workflow template that runs ultrareview with --timeout for environments where long reviews are disruptive.

#121

Module content, quizzes, and templates for 7 newly verified Claude Code features.

#119

Full content, quiz questions, and templates for 10 verified features across commands, plugins, MCP, and advanced-features.

#114, #96

The Getting Started module now includes a step-by-step beginner guide and a glossary of terms for those new to Claude Code.

#95

Full quiz question and terminal step coverage for the Advisor Tool in both English and Arabic.

#94

Full Module 0 coverage for install, authentication, terminal, and IDE setup in English and Arabic.

#86

Advisor Tool description is now accurate. Undocumented env vars have been removed from the feature catalog.

#84

Learners can read explanations before deciding to continue. A 'Next Question' button replaces the automatic advance.

#20

/allowed-tools was listed as a built-in command but does not exist in Claude Code. Removed it from the slash-commands module.

#33

Arabic quiz questions in memory, skills, and slash-commands modules now match their English counterparts.

#21

The effort maximum is session-wide, not Opus-only. Documentation corrected across 4 modules.

#74

Documents sandbox.network.deniedDomains in advanced-features with quiz questions and templates.

#71

Module content updates for commands, hooks, plugins, skills, and subagents covering newly verified Claude Code features.

#58

Skill description budget corrected from 2% / 16,000 to 1% / 8,000 characters.

#28

Added documentation for model and shell frontmatter fields in the skills module.

#28

The /clear command in the terminal simulator now clears visible history instead of printing text.

#25

When html2canvas fails, a localized error message appears below the button instead of silently failing.

#25

Code template blocks push content below the copy button to prevent text rendering underneath it.

#25

Automated skill for checking quiz data integrity, structural validity, and EN/AR consistency.

#23

Automated skill for cross-referencing module content against terminal-commands.json and terminal-steps.yaml files.

#23

Terminal simulator now supports /allowed-tools for managing permitted tools with add/remove and wildcard support.

#22

Replaced physical CSS properties with logical equivalents across 8 components for proper RTL support.

#18

Added RTL-aware arrow-key navigation to BuildTabs, QuizBlock, and InteractiveDiagram.

#18

Self-assessment scores now correctly cap at 20. Expert users are no longer misclassified as Beginner.

#18

Free-text fields use dir='auto' for mixed Arabic/English input. Email fields use dir='ltr'.

#18

Fixed command discovery, model aliases, effort options, diagnostics, and Agent Teams documentation.

#15

Automated audit workflow with OpenCode agents, runner scripts, report templates, and tracker sync.

#15

Score ranges are Beginner 0–6, Intermediate 7–14, Advanced 15–20. Expert users are no longer classified as Beginner.

#11

New /feedback page showing active feedback organized by status with item counts and a submission form.

#7

New /changelog page showing what has been shipped and fixed, linked from the footer.

#7

Each module page has a link that opens the feedback form with the module name pre-filled.

#7

Build → Config Builder, Quick Reference → Cheat Sheet, Feature Catalog → Feature Index. Footer and sidebar reorganized.

#7