Changelog

ByteAsk now has one plan for Context, the Cloud Agent and the ByteAsk CLI. It is billed per workspace (a person or a GitHub organization), with no per-seat charge, and the plan fee is all usage credit.

• Plans: Free includes $5 of usage credit every month, with no card. Pro is $20 a month, Max $100 and Ultra $200, each with that much credit every month. Annual plans get 2 months free. Enterprise is custom.

• Models are billed at the provider's list price + 20%. Bring your own Anthropic or OpenAI key on any plan, with no markup.

• The cloud sandbox (4 CPU, 12 GB, for a Cloud Agent task's build and tests) is $0.40 an hour, billed per minute. Setup and environment builds are free, and Context lookups and indexing are included within your plan's limits.

• Existing paid CLI subscribers keep their current terms until their paid period ends.

Everything is on the pricing page.

One setup prompt

Every install button on byteask.ai now copies one short prompt. Paste it into Claude Code, Codex, Cursor or another MCP client, and the agent follows the setup guide: it asks which products you want, tells you the cost first, and shows a dry run of every change before it makes it. Claude Code and Codex get the full Context setup; Cursor and other MCP clients get the MCP server. The same steps by hand are in the docs.

ByteAsk's best-measured harness configuration is now the default. Nothing to set up: every setting below is on for new and updated installs, and each one can be changed in ~/.byteask/config.toml.

With ByteAsk Context connected (byteask context connect):

• ByteAsk follows a short set of working rules: reproduce the problem first, make the change complete across the related places (Context's find_related_changes and get_symbol list candidates), keep the repository's own helpers and names, and check the change with cpp_finish_check before finishing. Your own developer_instructions replace them.

• Context's code pack waits up to 2 minutes on large repositories, and a top-level CMakeLists.txt, meson.build or configure.ac marks a repository as C/C++.

• cpp_finish_check lists the places Context says may need the same change, right after its verdict.

• Headless runs (byteask exec) offer a smaller tool set; interactive sessions keep every tool.

C/C++ checks:

• cpp_finish_check builds and runs the nearest tests' programs first and names the other tests the change reaches.

• A change that no compiled file reaches is reported as not checked, never as passed.

• When a turn ends with a C/C++ change nobody checked, ByteAsk runs the check once and shows the result.

• C/C++ tools whose program is not installed are not offered.

Builds and long commands:

• ByteAsk's background builds run at idle priority, give way to your builds in the same tree, and stop first under memory pressure.

• The environment note states the memory limit and a job count that fits it.

• On a long-running command without a terminal, ByteAsk waits 30 s before the first check and up to 3 minutes between checks; every wait ends as soon as the command exits.

Platforms: Linux (x86_64, arm64) and Windows carry everything above. The macOS engine in this release is the 0.2.3 engine; the 0.2.4 macOS engine follows.

Your own provider key now unlocks that provider's Pro models. On the Free plan with your own Anthropic (or other provider) key saved and BYOK active, that provider's models are now selectable in /model. They used to show as "Available on ByteAsk Pro". Those turns go straight to the provider on your key.

ByteAsk Context: a one-line check for the change. With ByteAsk Context connected (byteask context connect) in a CMake project, ByteAsk now finds the tests that cover the place Context points to and starts building them in the background. It then gives the agent one command that builds and runs only those tests, instead of building and testing the whole project. This applies in full-access sessions.

0.2.1 and 0.2.2 were skipped so that every channel carries the same version.

The previous release was 0.1.16 on 9 October. This one rebuilds the engine on all four platforms.

Update with byteask --update, or pip install -U byteask, or npm i -g @byteask/cli.

---

Ultra plan

• Ultra ($200 for 30 days) is now on /upgrade next to Pro and Max: $20 of model use per 5 hours and $60 per week. A Max subscriber can renew Max or move up to Ultra from the same screen.

• Every plan row now shows its real limits (for example $4 per 5h, $8 per week on Pro) instead of a multiple of Free.

Leaner sessions

• Rarely used tools now load on demand, so each request carries a smaller fixed prompt. Interactive tools (questions in Plan mode, images, web search) are always loaded.

• The always-on C/C++ guidance is now a short set of rules: build and test with the project's own build system, keep your build tree and files, use sanitizers and debuggers when the symptoms call for them.

• A long wait on a running command now ends as soon as you send a message.

C/C++ tools

• New: cpp_finish_check. Before saying a change works, the agent can build the affected code and run the project's tests, within 90 seconds, and report PASS, FAIL or NOT RUN with the reason. It never deletes files and never sends the agent back on its own.

• New debugging tools: debug_core (diagnose a core dump), debug_reverse (step backwards through an rr recording, Linux), cpp_kdb (inspect a stopped Linux kernel over KGDB/KDB, Linux), plus cpp_impact (related places a change may need), cpp_test_select (run the tests a change affects) and cpp_repo_history.

• Safer by default: on Linux the C/C++ tools run in their own sandbox (where none can start, they run without one and say so), cpp_compile_db never writes into your project, and cpp_format only formats the lines you changed, only where the project has a .clang-format.

• More accurate cpp_check, cpp_tidy and cpp_sanitize (MemorySanitizer mode, allocation and free stacks in sanitizer reports, builds sized to the machine's memory).

Full access means full access

• With full permission (approval policy never and no sandbox), ByteAsk no longer stops to ask: the C/C++ tools run without consent prompts and the agent works around changes it did not make instead of pausing.

ByteAsk Context

• byteask context connect adds ByteAsk Context (the hosted, build-aware code index) as an MCP server and signs you in. byteask context disconnect removes it.

Other changes

• Locked models in /model show a lock and cannot be selected; /usage shows when extra credit expires.

• byteask policy shows or resets the local policy file (~/.byteask/policy.toml).

Read the full release notes

• /model now reliably shows what your plan includes. Models you can't use are marked 🔒 Pro, and retired models are hidden. If the list ever looked out of date for you, this fixes it.

• Bring-your-own-key users on an Anthropic key: Claude Fable 5.1 and Claude Haiku 5.5 now work, and Claude Haiku 4.5 no longer fails on effort settings.

The engine is unchanged from v0.1.15.

The previous release was 0.1.14 on 23 September. This one rebuilds the engine on all four platforms.

Update with byteask --update, or pip install -U byteask, or npm i -g @byteask/cli.

---

Debugging on a Mac

• debug_session now runs on lldb, so interactive debugging works on Apple Silicon Macs, where gdb does not run. lldb is the default on macOS, and a Mac is never told to install gdb.

• A crash dump from another machine opens with that machine's libraries: pass sysroot to debug or debug_session.

• After a crash, ByteAsk runs the post-mortem commands first and tells you what failed and where.

C/C++ tools you can trust

• Program arguments reach your program exactly as written, including empty and quoted ones.

• A tool no longer reports a run as clean when it never ran, was cut short or was only partly instrumented.

• No C/C++ tool can end your turn or hang it any more.

• A misnamed argument is refused with a clear message instead of being silently ignored.

• cpp_perf, cpp_bench, cpp_sanitize and cpp_layout give more precise results.

• Security: the debug tools refuse commands that could run programs on your machine without your approval, and a function name passed to cpp_decompile can no longer run commands.

Other changes

• /model is now three short steps: provider, then model, then reasoning effort.

• When your free messages run out, the sign-in prompt opens by itself. Type your email; you no longer need to type /login.

• The first-run setup menu lists Use ByteAsk managed (recommended) first, already selected.

• Settings you set with byteask features now survive signing in, signing out and switching between ByteAsk, your own keys and a subscription.

• Update prompt: an interactive launch asks Update now? [Y/n] again when a new version is out.

• The model list refreshes on every launch, including before you sign in.

• Conversation titles and catch-up summaries use a small, fast model.

Read the full release notes

The previous release was 0.1.13 on 20 September. This one rebuilds the engine on all four platforms.

Update with byteask --update, or pip install -U byteask, or npm i -g @byteask/cli.

---

Use your ChatGPT subscription

ByteAsk can run on a ChatGPT Plus (or higher) subscription instead of ByteAsk credits. Choose it from /login. This option is being switched on in stages over the coming days, so you may not see it yet.

• Sign-in happens in a pane inside ByteAsk. When it finishes or you press esc, the pane closes and leaves nothing behind in your terminal.

• ChatGPT Free and Go plans are refused at sign-in with a clear message, because OpenAI serves no coding models on them. The refusal happens before anything is saved, so an existing sign-in is kept. From the refusal screen you can try a different account or go back.

• To switch back, use /login and choose Use ByteAsk managed.

Fixes

• Questions about ByteAsk itself are now answered from ByteAsk's documentation.

• The Token usage: … line no longer prints when a session ends.

• The background app-server no longer tries to install or update another product.

• The browser pages at the end of a sign-in, and the device-code screen, now show ByteAsk branding.

Read the full release notes

The previous release was 0.1.12 on 8 September. This one rebuilds the engine on all four platforms.

Update with byteask --update, or pip install -U byteask, or npm i -g @byteask/cli.

---

Less noise in the model's context

Command output is now cleaned before the model reads it. Colour codes and progress-bar redraws are removed, repeated lines are folded into a count, and long build, install and test logs keep every warning, error and summary line while the routine lines are counted. The full original output is always saved on your machine, and the model can open it when it needs a line that was folded away. What you see in your own terminal does not change.

ByteAsk can also read several files, or several ranges of one file, in a single step, and search with context lines, instead of one shell command per file.

Long sessions hold up better after the history is compacted: the agent keeps a note of the files it had open and its most recent results, so it does less re-exploring.

C and C++

• cpp_sanitize accepts sanitizer: "tsan", so data races can be found. It can also build your project's own CMake or Make target with sanitizers on, so a multi-file program no longer needs every source file listed.

• On macOS a clean sanitizer run now says that leaks were not checked (LeakSanitizer does not exist on macOS). It used to read as a clean result.

• debug_record (rr) handles bugs that fail one run in twenty: it can record under rr's chaos scheduler until a run fails and then replay that run, including backwards. When rr cannot work on a machine it says why, and what to use instead.

• cpp_symbolize resolved caller frames to the wrong line. Fixed.

• cpp_layout reported nested members of unknown size as padding. Fixed.

• cpp_godbolt could pick the wrong compiler for a request like "clang 18". Fixed.

• cpp_binary can explain a link failure caused by a libstdc++ ABI mismatch, and its header view shows the NX and RELRO state.

• cpp_compile_db can generate compile_commands.json for Makefile projects.

• A Python project with a Makefile no longer gets the C/C++ guidance, and a repository whose C++ sits several folders down now does.

• debug_attach no longer offers a gdb session that has already exited.

Account and updates

• /upgrade: on a Pro plan the arrow keys did not move. Fixed; Pro can be renewed from the picker.

• /usage shows your plan in its title.

• The update check now reads ByteAsk's own feed, and it is on. It had been off since the first release.

• The referral programme no longer grants a free Pro month, and the footer and hints that still mentioned it are gone. Referral credit is unchanged.

• The plugin marketplace is off by default. features.plugins = true in ~/.byteask/config.toml turns it back on.

• /export suggested a file name that was not ours. It now suggests byteask-session.md.

• Old state from before the rename is cleared from ~/.byteask on launch. Your sign-in, settings and history are kept.

Read the full release notes

The previous release was 0.1.11 on 30 July. This one rebuilds the engine on all four platforms and carries six weeks of our own work plus a large upstream merge.

Update with byteask --update, or pip install -U byteask, or npm i -g @byteask/cli.

---

Your transcript stays yours

/usage, /referral-code, /upgrade and /voucher used to print into your terminal's real scrollback. Checking your quota wedged a dashboard permanently into the middle of your conversation, and nothing could take it back.

They are panes now, like /model. They open over the composer, and esc closes them and leaves the transcript exactly as it was.

enter never closes a pane. It only does what the footer says, so muscle memory trained by /usage cannot fire a payment link on /upgrade.

The operator announcement banner is gone for the same reason. It inserted itself at the top of every session and, being scrollback, could not be dismissed.

/usage tells you more, and tells the truth

The meters are twice the resolution and now show when each window resets.

5-hour  ████████████████░░░░  82%   resets in 47m
Weekly  ██████░░░░░░░░░░░░░░  31%   resets in 6d 23h
Extra   $12.00 left of $15.00 · covers spend past your limits

Plan pricing used to read "$20/month". That was wrong. We sell a one-time 30-day pass, and the pane now says so: one payment for 30 days, no auto-renew, your plan switches as soon as payment completes. Your current plan is marked in the list.

Our pricing page still describes a monthly subscription. The page is what is out of date, and we are fixing it.

You will stop being logged out

If you were ever stuck behind Reconnecting… 4/5 with no explanation, this is the release that fixes it.

Sessions lasted 30 days and nothing renewed them. When yours expired, the client kept sending the dead credential, and running /login could appear to succeed while the new credential silently failed to save. One account sat locked out for six weeks.

Your session now renews itself quietly at launch, well before it expires. Anyone already expired is healed on their next launch with no email round-trip. Renewal edits only the credential, so your trusted projects, reasoning effort and configured tools all survive it. And a credential that cannot be written now says so loudly instead of reporting success.

Errors say what happened, immediately

Every unexpected response used to be treated as temporary. A "please sign in" or a plan limit would spin Reconnecting… for about six seconds and then be replaced by a generic retry error, so you never saw the sentence that told you what to do.

Only genuine server errors and rate limits retry now. Everything else is shown at once, in our words. A failed payment link is one line telling you to contact us, rather than an internal cause you cannot act on.

Models

The picker shows 20 models, up from 18, with nine new ones. auto is now a chain across three models and two providers rather than a single fixed choice, so if one is having a bad day your turn quietly lands on the next. Its first choice also got five times cheaper.

Model descriptions were rewritten so that "fast", "balanced" and "deep" mean the same thing everywhere, instead of each vendor's marketing language.

Anthropic models are currently out of service and hidden.

Accessibility

A screen-reader mode renders panes as linear, announce-once text instead of in-place repaints, and substitutes a plain symbol set when your terminal is not UTF-8. Turn it on with --ax-screen-reader, the BYTEASK_AX_SCREEN_READER environment variable, or tui.screen_reader in your config.

---

From upstream

We merged roughly 2,000 commits of upstream work. The parts you will notice:

New commands. /recap summarizes the conversation on demand. /cd and /pwd change and show the working directory mid-conversation. /agents opens a command center grouping your agent sessions into needs-input, working, ready and finished. /copy can now target the last code block or quote, not just the whole response.

New subcommands. byteask queue --thread <session> --message "…" sends a message into a running session from another terminal. byteask agents browses sessions across the local daemon.

Sessions reconnect themselves when a connection drops, instead of dying and taking your draft with it.

/keymap can genuinely rebind keys now, including two-key chords like g g.

Links in model output are clickable where your terminal supports it, including local file paths that keep their line and column.

Vim mode gained undo and redo, dot-repeat, search motions, and character find.

The model can ask you structured questions - a small form with options rather than only free text - and /status now shows this thread's token usage.

Two changes worth reading before you update

An auto-reviewer can now approve some risky actions without asking you, and it is on by default. The permission menu offers an explicit choice between "Approve for me" and "Ask for approval", and anything auto-denied is labelled in your approval history. If you want to be asked about everything, choose "Ask for approval". Relatedly, what counts as auto-approvable under "ask only when needed" is now decided by explicit rules rather than a built-in guess about which commands are safe.

byteask exec --full-auto has been removed. It was deprecated with a warning in 0.1.11; it is now an error. Use --sandbox workspace-write. If you have scripts passing it, they will fail on this version.

byteask mcp-server now warns that it is deprecated. It still works.

---

Under the hood

Our C/C++ and reference tools are no longer exposed to the internal thread that summarizes your conversation. Upstream strips every tool from that thread because a summarization prompt can contain untrusted text from your transcript; ours had no way to be stripped, which meant tools that compile and run your code, and tools that make network calls, were reachable from it. They are now excluded before that thread reads anything.

/referral-code prints all three share commands complete, rather than one whole and one truncated. The installer tells you if a referral code was rejected instead of failing silently, and reports where it put things and whether your PATH needs a change.

---

Known issues

• Upgrading is currently failing. Our payment provider began requiring a contact number we do not collect. Everything else works normally. Email contact@byteask.ai and we will sort your plan out directly.

• The update prompt still blocks startup with a yes/no question. It should be a line you can ignore, and will be.

• The pricing page still describes a monthly subscription rather than the 30-day pass we actually sell.

Highlights:

• Operator alerts rendered at the top of the session

• /login rebuilt as a native bottom-pane settings surface

• byteask exec now renders extension tool calls (cpp_*/ref_*) - previously invisible

• cpp-tools: real sanitizer output, project builds, bounded execution, trust gate

• Hardware-aware performance verification

Read the full release notes

What's new

• Status/result messages (e.g. key-add errors) no longer leak onto unrelated screens after navigating away.

• install.sh / install.ps1 now say plainly when a best-effort ripgrep install didn't work, instead of failing silently.

All 4 platform binaries included below.

Read the full release notes

• Security & abuse-resistance hardening.

Engine unchanged from 0.1.8 - the binaries below are identical; 0.1.9 ships client + gateway improvements.

Install: curl -fsSL https://code.byteask.ai/install.sh | sh · pip install byteask · npm i -g @byteask/cli

Binaries are gzip-compressed; SHA256SUMS lists the checksums.

Read the full release notes

• See your tools working - every ByteAsk tool call (autosar_ref, mcu_ref, cpp_godbolt, debug, …) now shows a live cell in the TUI with the tool name, arguments, and result - the same visibility web search always had. Interrupted calls show cleanly; retryable tool replies no longer render as failures.

• Referral upgrades - you AND your friend now each get $10 of usage per referral, and 3 successful referrals earn your first month of Pro free. /referral-code shows your live progress.

• Reliability: turn-abort no longer leaves stuck in-progress items in thread history.

Install: curl -fsSL https://code.byteask.ai/install.sh | sh · pip install byteask · npm i -g @byteask/cli

Binaries are gzip-compressed; SHA256SUMS lists the checksums.

• Grounded reference tools - autosar_ref, mcu_ref, fpga_ref, hft_ref, grid_ref, scpi_ref, x86_ref: page-cited lookups over the ByteAsk corpus for embedded / firmware / hardware / safety-critical C++ (MISRA, AUTOSAR, Intel SDM, IEEE 1547, SCPI, Xilinx, ITCH, and more).

• Renamed the native-debugger tools - cpp_debug/cpp_debug_session/cpp_debug_attach/cpp_record → debug/debug_session/debug_attach/debug_record (they debug any native binary, including Rust).

• Removed the /app slash command.

• Gateway maintenance: auto model routing, Gemini/Anthropic reliability fixes.

Install: curl -fsSL https://code.byteask.ai/install.sh | sh · pip install byteask · npm i -g @byteask/cli

Binaries below are gzip-compressed; SHA256SUMS lists the checksums.

ByteAsk CLI v0.1.6 - fixes API-key paste hanging on macOS Terminal (wrapper terminal-reset). Engine binaries unchanged from v0.1.5.

Read the full release notes

ByteAsk CLI v0.1.4 - Gemini/Anthropic fixes (CRLF SSE, tool-schema, thought-signature), gateway model allowlist, Linux bubblewrap auto-install + sandbox fallback, referral across curl/pip/npm, catalog trim (Sonnet 4.5 removed).

ByteAsk CLI v0.1.3 - the AI coding agent for your terminal.

Engine binaries for all supported platforms are attached below; installers fetch the latest automatically.

Install

• macOS / Linux: curl -fsSL https://code.byteask.ai/install.sh | sh

• Windows (PowerShell): irm https://code.byteask.ai/install.ps1 | iex

• pip install byteask · npx @byteask/cli

Verify downloads against SHA256SUMS.