CHANGELOG.md

Changelog

0.8.2 mistral large 4 known-good, mistral-vibe-cli model aliases, mistral system prompt mined from the vibe cli

• Mistral Large 4 known-good. Mistral's new frontier open-weight model ("le Chonk", 1.05T total / 49B active MoE, multimodal, public preview since Oct 6; weights promised end of month) is served as mistral-large-4 on api.mistral.ai with a 524,288-token window per /v1/models (the docs page advertises 1M; the API figure is the safe one). It carries the same reasoning_effort ladder as Medium 3.5 — exactly none/high, 400 on anything else — and streams thinking as chunked thinking content, the shape the Mistral-native parser already folds, so no wire changes were needed. Known-good row: reasoning high, temperature 0.2, 8192 output, tbNone (the platform's strict validator rejects reasoning_content on replayed assistant messages).

• The Vibe CLI routing aliases are curated under mistral. mistral-vibe-cli-latest (routes Medium 3.5, thinking high) and mistral-vibe-cli-fast (routes Small 4, thinking off) — the ids Mistral's own mistral-vibe CLI sends by default — are known-good rows on the plain mistral provider (one API key covers both, so the parked mistralvibe twin stays parked). Both ride the none/high ladder with temperature 1.0, what the Vibe CLI itself sends.

• The Mistral system prompt is now Mistral's own, 3codified. Mined from the open-source Vibe CLI (mistral-vibe 2.26.0, vibe/core/prompts/cli.md): the structure and content are Vibe's — instruction hierarchy (critical vs overridable, AGENTS.md precedence, external data as data), blast-radius rules (one-time approval doesn't generalize; state action and radius in one line), ambiguity handling (one question, no strategy menus, no silent partial completion), stop-when-stuck tripwires (same error twice, no-op edit, three edits to one file → re-read fresh, change strategy after two failures), and the voice rules (concise is not curt, no filler words, phase-transition signaling) — with 3code's tool surface and rules substituted and the phrasing tightened for density. Per-model variation is two splices on the shared body (header, reasoning section) instead of replace chains against GLM wording; Large 4, Large 3/Medium 3.5, and the vibe-cli aliases all render from it.

0.8.1 prompt text selection and transcript search, provider refresh commands and CLI, kolibri-1 and tesseracted providers, macOS universal tarball and Windows console keys fixed

• :provider update <name>. Re-runs just the models step of the provider wizard: the offered list refreshes (curated registry in regular mode, the provider's /models endpoint under --experimental), Enter keeps the current selection, and the key, url, and name pass through untouched. :provider edit remains the full wizard.

• :provider new / :provider pull. new lists the models a provider serves that its config doesn't list yet (curated registry growth in regular mode, the /models endpoint under --experimental), across every configured provider or just one by name. pull brings those models into the stored list, scoped the same way or down to named models (pull nvidia glm-5.4); nothing is removed, and under --experimental each addition is verified with a one-token call first.

• The macOS universal tarball was arm64-only. The OSX workflow lipos the arm64 and x86_64 slices into ./3code, then the test suite runs, and its first step rebuilds ./3code natively, silently replacing the fat binary before packaging. Every macos-universal tarball since the workflow appeared, including the 0.8.0 release, was thin arm64: fine on Apple Silicon, bad CPU type in executable on Intel Macs. The fat binary now lives outside the workspace where the test rebuild cannot touch it, and the build fails loudly if either slice is missing. Intel users of the 0.8.0 release asset stay broken until the next release; a fresh install from main works once CI has rebuilt.

• Tesseracted: Kolibri-1 without the GPUs. Tesseracted Labs (Munich) is hosting Kolibri-1 behind an OpenAI-compatible gateway at api.tesseracted.com/v1 (free during launch week, key via tesseracted.com/kolibri-1-chat/developers); :provider add tesseracted prefills the url. The known-good row reuses the kolibri family (same prompts, same Hermes tool surface, tbCurrentTurn, 262144 window per the gateway's own input+output cap), with two gateway differences: applyKolibriReasoning sends top-level reasoning_effort (true none for off) instead of chat_template_kwargs, and allowPrivate is off (third party, usage metadata retained ~31 days) so private mode stays opt-in. Output is capped at 16,384 tokens by the gateway (not the model). tl_live_ joins the key-prefix catalog, so pasting the key names the provider.

• The provider wizard masks api keys in its multipurpose field. The first field (provider name, url, or key) echoes as typed so names and urls stay readable, but the moment the buffer reads as a key (a recognized key prefix from the first character, an unrecognized secret once it outgrows any provider name) the line repaints masked, so at most the pre-detection prefix ever reaches the screen. The dedicated api-key fields remain fully hidden reads.

• Kolibri-1 known-good: first harness with curated support for Aleph Alpha's open-weight model. Kolibri-1 (78B MoE, 3.46B active, German/English, Apache 2.0) is served self-hosted on vLLM with the official aleph-alpha-inference plugin; :provider add kolibri prefills http://localhost:8000/v1 and the known-good rows cover both weight ids (FP8 and BF16) under the kolibri family (alias alephalpha). Reasoning is the model's own graded chat-template knob: :reasoning offers off/low/medium/high mapped to chat_template_kwargs.reasoning_effort / enable_thinking: false, matching what the kolibri1 reasoning parser reads server-side. Thinking replay is tbCurrentTurn, exactly the chat template's own contract (a turn's tool loop keeps its thinking, older turns drop it). Sampling follows the model card (temperature 1.0; top_p/top_k come from the repo's generation_config server-side), context window curated at the recommended 262144, allowPrivate on (self-hosted). The XML tool-call fallback now also parses Hermes-style JSON bodies (<tool_call>{"name": ..., "arguments": ...}</tool_call>), Kolibri's native emission, alongside GLM's arg_key markup. Dedicated Kolibri system prompt, tightened to the output-contract style and grounded in the model card; the model was RL-trained across randomized harnesses and prompts, so it takes the house style without a special recipe.

• Stealth models left the known-good registry. The curated table no longer carries anonymous/stealth mounts: stealth/space-bunny-alpha (OpenRouter), space-bunny-free (Zen), and omen-alpha (OpenCode Go) join 0xalpha and union-alpha off the table; a different mechanism will cover stealth models. They stay runnable via --experimental with a family override, and their context-window / output-cap substring fallbacks and the OpenRouter include_reasoning streaming fix remain, so an experimental run behaves as before.

• Markdown links are OSC 8 hyperlinks. [text](url) now renders as a clickable span instead of raw syntax. The target stays visible in the text (bare url when the text is the url, text - url otherwise), so terminals without OSC 8 support still show, and regex-detect, the address; width/wrap math skips OSC sequences like it already skipped CSI, and the library's plain-text surface strips them.

• Caret flicker in the spinner, finished. The drawn-caret rework left one ?25h at every turn end: nothing ever re-hid the cursor, so from the second turn on the terminal's own blinking caret sat parked on the bar row, in and around the braille spinner. The physical cursor now stays hidden for the whole session as designed (one hide at startup, one show at exit). Two masked bugs surfaced once it really was hidden: after a queued autosend the idle prompt drew no caret at all (a leftover pendingCaret suppressed the drawn cell; the visible cursor had been standing in for it), and the frame-artifact renderer drew the caret from the physical cursor, so it now draws the reverse-video cell instead.

• Space Bunny stealth model known-good on OpenRouter and OpenCode Zen. stealth/space-bunny-alpha (OpenRouter) and space-bunny-free (Zen) are a free anonymous preview (1M context, 524k output cap, text+image+video input) whose tokenizer and reasoning wire shape match the MiniMax M-series, so it rides the minimax family: enable_thinking + reasoning_split are accepted on both routes and thinking comes back split from content. Reasoning is mandatory: upstream offers an effort ladder minimal..max but rejects none ("Reasoning is mandatory for this endpoint and cannot be disabled") and ignores enable_thinking=false, so :reasoning offers no knob (same rule as kimi-k2.7-code). The OpenRouter stealth route also omits reasoning deltas from the stream unless include_reasoning is sent; 3code sends it for that mount only. OpenRouter banners the model as going away October 5, so expect that row to vanish then; Zen's free mount has no listed end date.

• Session listing scope, transcript search, and paging. -l still lists this directory's 20 newest sessions; -a/--all now widens it (and the new search) to every directory, printing each session's cwd. New -f/--find TERMS... searches saved .3log transcripts: case-insensitive, whitespace-agnostic phrases (a quoted argument is a phrase that matches across line breaks), OR across terms, ranked by match frequency with newest-first ties, one snippet per hit with the 3log formatting stripped. The scan reads each file once and matches through libc memmem (the grep -F engine), so ~500 MB of history searches in about a second; only displayed pages pay for snippet extraction. --page N pages both listings and search results, and :sessions all now lists everything instead of refusing.

• Windows: console keys are read as records and translated to the VT grammar in-process. ReadFile under ENABLE_VIRTUAL_TERMINAL_INPUT wedges forever when the record-queue head is a charless event (a key-up): later keys pile up behind it and never wake the read, and the bytes it has already translated sit in a kernel-side buffer invisible to every readiness probe - Ctrl-C went dead mid-turn and the prompt could not be quit. The editor now consumes INPUT_RECORDs with ReadConsoleInputW (which handles every event type) and emits the same CSI/Alt-chord grammar itself; the ESC-tail probe and the startup drain work on the same byte view, and the drain only drops structural bytes inside a sequence that began with ESC (it used to eat typed hex letters a-f that landed in its window). Selection and modifier keys still arrive as full VT sequences.

• Keyboard text selection and standard edit ops in the prompt (#44). The input buffer is now a real text widget: Shift+Arrows / Shift+Home / Shift+End extend a reverse-video selection (Ctrl+Shift+Arrows by word, Ctrl+Shift+Home/End to the buffer ends), typing or any delete replaces it, plain motion drops it, Ctrl+X cuts, Alt+W copies (OSC 52 to the system clipboard where the terminal allows), Ctrl+V pastes from the system clipboard (pbpaste / wl-paste / xclip / xsel / PowerShell, kill-ring fallback), Ctrl+Y yanks the last cut. New edit commands: Ctrl+K delete-to-eol, Alt+D / Ctrl+Delete delete-word-right (ctrl+delete used to delete the word on the LEFT), Ctrl+T transpose chars, Ctrl+Left/Right word jumps are now first-class configurable bindings. Every command is rebindable in [shortcuts], with modified-key names (ShiftLeft, CtrlShiftLeft, CtrlDelete, ...). Windows enables ENABLE_VIRTUAL_TERMINAL_INPUT so the console speaks the same VT key grammar as every other platform (legacy conhost keeps the _getch pair codes); the modified-key CSI decoding and selection rendering are shared verbatim. Keys are read via ReadFile under VT input: _getch decodes the console's key records and on live consoles still emits the legacy 224/0 pair encoding with the flag set, silently stripping modifiers.

• Byte-envelope 400 recovery (issue #48). Gateways reject oversized request bodies with an opaque 400 (inference_failed on OpenCode Zen) while the token count is still far under the advertised context window, so the token-based summarize threshold never fired: every resend carried the same history and 400'd again, and :summarize could not rescue either because its meta-call replayed the same oversized payload (upstream anomalyco/opencode#35013 documents the same envelope tripping OpenCode's own TUI). A turn that dies on a 400 with enough history now collapses it once and retries: a real recap when the summarizer call succeeds, otherwise the middle is dropped locally so the session still recovers instead of dead-ending forever. The summarizer payload is also clipped to ~768KB of message text, so the rescue call itself fits the envelope that killed the main request. On top of the reactive recovery, a proactive guard measures the exact serialized request body after every call and, for the Zen-family providers only (opencode, opencodego), collapses the history once it crosses 1MB — under the tightest route envelope observed — so the oversized request is never sent at all; token counts cannot predict a byte limit, but the body size 3code just sent can. Cached transport connections that the server closed (Connection: close) are no longer reused: a non-blocking peek drops the dead socket before send can wedge in an unobservable spin, which the recovery retry could hit when the gateway closed the rejected connection.

• Flail guard: streak signal no longer flags verification retries. The stuck-streak signal read a long same-tool run of near-identical calls as a doom loop even when the run was legitimate output-capture debugging: retrying one pytest command with cycling capture tails (| tail -2, | grep -E 'passed|failed' | tail -2, redirect-to-file) tokenizes to the same subject set, because the varying tails fall under the 6-char token minimum, so a converging retry sequence and a cosmetic-variant doom loop were shape-identical (20260930 astropy session; escalations landed mid-diagnosis and the step-2 message forbade the retry the work needed). The ring behind the signal now restarts on a subject change: a call sharing no distinctive token with the ring's consensus (the tokens a majority of ring members carry, so prose # comments cannot bridge a subject change) empties it. An uninterrupted run of near-identical calls still fills the ring and still escalates; interleaved genuinely different work no longer accumulates into one. Replayed against 400 recorded sessions: the mergepdf doom loop, ssh-boilerplate loop regions, and spaced failed repeats all still fire.

• 3code provider subcommand. Non-interactive provider management for scripts and keyless first setup: provider add <name|url|api-key> [--key KEY] [--models "a b c"], provider models <name> <model...> (replace the list), provider rm <name>, and provider list. Same first-field resolution as the wizard (catalog name, URL under --experimental, API key, subscription login via browser OAuth); without --models the provider's full known-good list is stored. No verification ping; a bad key surfaces on the first turn.

• Claude Sonnet 5.5, GPT-6 Sol / Luna, GPT-6.1 Sol. claude-sonnet-5-5 is known-good for anthropic (API key) and claudecode (Pro/Max subscription): 1M context, 128k output, adaptive thinking with effort low..max. Its off knob maps to Anthropic's new between_tools setting (Sonnet 5.5 rejects disabled), and its thinking blocks ride the drop_block fallback like Opus 5.5's. gpt-6-sol, gpt-6-luna, and gpt-6.1-sol are known-good for openai and chatgpt plus the openrouter / opencode / nanogpt / venice gateways: 1.05M context, 128k output cap; Sol/Luna keep none on the effort ladder, 6.1 Sol always thinks (low..max like Astra). Live-verified: gpt-6-sol, gpt-6-luna, and gpt-6.1-sol on openrouter; gpt-6-luna on the chatgpt subscription (the Codex backend serves a per-account model list and may not offer sol yet); opencode and venice reached upstream (402 balance); anthropic and nanogpt keys here lack credit/a valid session, so those rows stay table-verified.

• Test build: browser-backed web_fetch behind browserfetch = on. When the plain fetch survives tag-stripping as almost nothing (a JS shell, e.g. a Reddit profile), the fetch is redone in a shared headless Chrome and the rendered page text is returned instead. One browser per machine (own throwaway profile, 127.0.0.1:9223, launched by the first 3code that needs it and reused by every later one), one tab per fetch. Default off; ws is a new dependency. Anti-bot walls still apply: a fresh headless profile gets Reddit's "Prove your humanity" challenge.

• The occasional "submit deletes the line above the prompt" is fixed. A multi-row draft in the buffered mid-turn editor (history recall, shift+enter, typing during the stream) made the answer-start erase leave the cursor on the erased block top with the chrome row model wiped; the repaint after it then re-anchored up to editor rows - 1 rows too high and overwrote the wrapped tail of the just-committed prompt echo. Single-row drafts canceled the arithmetic out, which is why it only showed sometimes. The content-start transition now parks the cursor back on the editor caret row and keeps the painted-chrome count. The multiline golden fixture is regenerated: it had recorded the bug (the first echo's second line row missing).

• Wrapped :commands no longer strand stale rows above their echo. Both command submit paths cleared the editor's row model before the commit walked up (the idle path via its beforeRepaint hook, the mid-turn path in the submit handler itself), so a command wrapping the editor to three or more rows erased from inside its own block: the resting bar and the editor's top rows survived as junk between the previous content and the committed echo. The model reset now happens inside the commit, after the walk-up has consumed the pre-submit geometry (clearEditor). Single-row commands were never affected.

0.8.0 private mode, per-model params, cache-hot resume, commandcode and platte providers, PortableGit in the Windows installer

• anthropic provider: Claude on the native Messages wire (Opus 5.5, Sonnet 5, Fable 5.1, Haiku 4.5) with an sk-ant- key — system prompt as a top-level field, tool_use/tool_result content blocks, and the :reasoning knob on the model's thinking surface (adaptive effort low/medium/high/xhigh/max on 4.6+/5.x, the legacy token budget on 4.5; Opus 5.5+ and Fable cannot disable thinking). Thinking blocks replay across tool loops with their signatures; the prefix-binding models (Opus 5.5+, Fable) request drop_block so an interrupted or compacted history degrades instead of 400ing. No sampling params (the 4.6+ generations reject them). claudecode is the Claude Pro/Max subscription twin: browser OAuth on the Claude Code public client, Bearer + oauth beta against api.anthropic.com, the Claude Code line leading the system prompt
• xiaomi provider; MiMo 2.6 Pro/Flash on xiaomi, OpenRouter, OpenCode Zen/Go
• editor buffer hard-capped at 512KB: oversized inserts and pastes cut at a rune boundary and ring the bell instead of growing forever
• release-build crash typing large wrapped prompts fixed (caret past a wrap gap built a negative slice); the wrap walk can no longer spin on terminals too narrow for the prompt plus a rune, nor slice past the buffer when a raw paste ends mid-rune
• thinking replay (think-back) is configurable per provider in [params]
• dead oauth refresh token ends the turn with a re-login hint, not a crash
• harness autosends (steer, flail prods) show exact text in scrollback
• Grok 4.7 known-good: xai, OpenRouter, OpenCode Zen (same shape as 4.6)
• MiMo 2.6 known-good on DeepInfra and nano-gpt; Grok 4.7 on nano-gpt and Venice (flattened grok-4-7 id)
• retry-backoff notice boxed by blank rows, absorbed on reconnect
• -c/--config switch reads an alternate config file (docs always promised it)
• [settings] max_timeout raises the bash timeout ceiling; THREECODE_MAX_TIMEOUT wins per run
• reference: env-var section (XDG roots, diagnostics); -c and max_timeout documented
• shortModel canonicalizes after slash-strip; zai-glm-5-3 shows glm-5.3
• prompt caret carries the palette's white tone, not the terminal default
• token bar survives usage-less turn ends (repainted from last label)
• catalog sweep: crof dropped, dead combos gone, stragglers added
• session reservations 0600 again (decimal 0644 parsed as 0o1204)
• Windows installer bundles PortableGit; resolveBash prefers it
• Windows bash detection: bash_path, Git for Windows, MSYS2, legacy
• no bash found: tool dropped with a warning instead of a hard fail
• sandbox = off silences the Windows host-rules warning at launch too
• Windows net-fence state resolved at startup; warns when it is gone
• new commandcode provider: one key for the open-model lineup
• SSE parser strips the optional space; Hetzner data:{...} streams
• new platte provider: GLM 5.3 on EU infrastructure
• union-alpha stealth preview on OpenRouter; generic other family
• GLM-5.3 on Mistral: 1M context, thinking-only low/high/max
• session files store the skills catalog, not the verbatim prompt
• OpenCode zen gateways take top-level reasoning_effort
• startup error line when the OS sandbox cannot confine bash
• JSON error bodies: detail convention read, longest string leaf picked
• pre-TLS connect failures say "host is not responding", not TLS
• Windows sandboxed bash no longer wedges (env-passed cmd, bounded reap)
• caret parks at the right margin instead of wrapping to the next row
• patch edits with empty search/replace rejected, file untouched
• caret steps over typed trailing spaces
• drawn caret, hidden physical cursor: no flicker, fewer bytes
• payload-fresh frame-model copies end the Termux write-after-free
• retry waits show an hourglass; braille spinner is in-flight only
• long-session memory leaks fixed (SSL ctx cache, thread-local frees)
• GLM-5.2 on Mistral: reasoning_effort ladder, chunked thinking folded
• Windows sandbox works when setup ran as another account (sandwall)
• resolveBash falls back to Git for Windows; installer bundle optional
• Esc clears the draft; only an empty prompt cancels the turn
• Hy4 and Mistral join the known-good registry
• DeepSeek V4.1 Flash known-good on all five providers serving it
• typing no longer flickers the turn clock to 0s
• private mode: -p/--private, allow-private providers, magenta bar
• [params] per-(provider, model) overrides: temperature, max-tokens...
• retry notices live on one row instead of stacking in scrollback
• resume re-sends byte-identical history (prompt-cache friendly)
• GPT-6 Astra known-good on openai and chatgpt

0.7.0 key rebinding, editor integration, smarter stuck-loop guard

• [shortcuts] config rebinding for every key command
• Alt+E edits the draft in $EDITOR; :! runs a shell command yourself
• flail guard: nudges a spinning model twice, then aborts the turn
• faster rendering: frame skip, diff-painted keystrokes, ghostty fix
• Kimi K3 first-party reasoning knobs; GLM-5.3 effort ladder, 32k cap
• User-Agent on every request; opencode gateways get session header
• catalog: Qwen 3.5-3.8, DashScope, Omen-alpha, Aki, GreenPT, Lyceum
• resume replays through shared transcript formatters
• Termux release tarball with autoupdate
• transcript: checkpoint markers stripped, empty replies skipped

0.6.3 catalog sweep

• catalog: GLM-5.3, Qwen 3.8, DeepSeek V4 on 15 more providers

0.6.2 GLM-5.3-Flash

• catalog: GLM-5.3-Flash on zai/zaicode; GLM-5.3 on regular z.ai API

0.6.1 --no-sandbox, paste-aware input, catalog refresh

• --no-sandbox flag; pasted newlines kept as one draft
• wizard experimental mode lists the full /models output
• catalog: 0xalpha stealth preview, current free tiers
• terminal: late OSC 11 replies, Windows startup keys, macOS openpty

0.6.0 filesystem sandbox, subscription logins, network wall

• filesystem sandbox: one policy, deny/readonly/allow per line
• kernel enforcement: Landlock, Seatbelt, Windows restricted tokens
• network wall: host rules proxy bash egress (netns, Seatbelt, WFP)
• ChatGPT and SuperGrok subscription logins via browser OAuth
• OpenAI requests move to the Responses API; :reasoning lists levels
• patient retry: 429/5xx/network backoff, up to ~36 hours
• provider wizard: one field, parallel checks, known-good registry
• Windows: in-process bash capture; macOS: Seatbelt, own workflow
• AgentSession blocking API; example/webserve.nim web frontend
• Termux arm64 release archive
• terminal/input fixes: shared geometry, resize, ctrl-c/esc/ctrl-d

0.5.2 search engine overhaul: Exa, Brave, and Parallel backends

• search: Exa MCP default, Brave and Parallel REST engines
• per-engine keys (exa-key, brave-key); Exa runs keyless

0.5.1 new providers, GLM-5.2 reasoning fixes, Qwen family

• new providers: OpenCode Zen/Go, Kimi, nano-gpt, zaicode
• GLM-5.2 reasoning fixed on Together/OpenRouter; variant bug fixed
• Qwen3.x first-class family (vLLM enable_thinking)
• provider config: shared keys/URLs allowed, clear dup-name error
• internals: KnownGoodCombos named fields replace tuple indices

0.5.0 Windows support, new providers, big stability push

• Windows support: MSYS2 bash, session locking, color palette
• new providers: Ollama, Eurouter, Lyceum, Regolo, TensorX, Novita
• streaming: network-quiet timeout, truncated SSE retry, ctrl-c
• resume: O(1) per-cwd index, full replay, stale lock reclaim
• rendering: resize redraw, single GUI thread, one byte-path renderer
• bash tool: native timeout and computeDiff, no GNU tools needed
• robustness: UTF-8 sanitize, binary diff guard, graceful exits
• packaging: --version provenance, nightly branch+commit strings

0.4.0 error icons for failed tools, bottom-pinned bar, no raw JSON

0.3.5 $/r/w tool bullets, bright cyan receipts, bar ticks

0.3.4 initial docs site, -i/--interactive flag, streaming ping test

0.3.3 icon-based tool banners, history fixes, display polish

0.3.2 native read command, binary guard for bash, deepseek known-good

0.3.1 update_plan tool, gpt-oss reasoning tuning

0.3.0 --good subcommand, ctrl-c cancel during stream, linux-arm64 builds

0.2.7 inline receipts, gpt-oss grounding prompts

0.2.5 token bar with cache indicator, skill autoloader, web-research

0.2.1 per-model tool dispatch, multiline input, markdown table fit

0.2.0 initial public release