LLeMbas-CLI/CHANGELOG.md

Changelog

All notable changes to LLeMbas CLI. The format follows Keep a Changelog, and the project uses Semantic Versioning.

[2.0.0-beta.2] — 2026-11-07

Harness spec 3.1.0: what the model has already read stays exactly as it was, so its server can reuse it instead of reading the whole session again.

Changed

• The system prompt is made once per session — at the start, after a compaction, on a model switch — kept with the session and sent byte for byte on every request; a resumed session sends the one it kept. It holds no date any more.
• What changes during a session is said in a note after your words on your next message: the date (on the first message, and when it changes), a permission mode you chose, a plan you changed, memory or an instruction file you edited, a personality you picked, an MCP server's instructions that arrived later. The model's own changes (its plan, its edits, a plan approval) are not said again. Notes are not shown as your words.
• The tool list is the same for the whole session. plan_submit is offered in every mode and answers that it works only in plan mode; show is offered from the start where the session can have a viewer; knowledge search is decided as the session starts. The tools stay listed when a step limit or budget winds a prompt up (a call in that answer is not run), and a tool refused for good stays listed and answers that refusal. The first request waits up to 3 seconds for the configured MCP servers, so their tools are in the first list.
• Compaction asks for its summary as the next turn of the same conversation instead of a separate request with a pasted transcript, so it reuses what the server has read and no longer loses the middle of a long session. A tool call in the summary's reply is not run.

Added

• Cached input tokens are recorded for every request (the usage's own count, or llama.cpp's timings and progress), kept with the session, and shown by /context for the last request.

[2.0.0-beta.1] — 2026-11-03

LLeMbas CLI 2.0 works with LLeMbas 2.0 and nothing else: it uses the server it is logged in to, keeps tool results short without ever rewriting what the model has already read, and reads every instruction file on the way down to where it was started. Harness spec 3.0.0.

Changed

• The CLI talks only to the LLeMbas server it is logged in to. Custom connections are gone: the OpenAI-compatible, Responses, Anthropic, Gemini and Ollama connections in connections.yaml, with their dialect, key_cmd, auth, discover, unload_url and one_model_at_a_time, and a model's fallback. Log in with lembas login <address>; without a login the CLI says so and how. An old connections.yaml still starts: an entry that is not the login is named in a warning and ignored. One server at a time: logging in to another signs this machine out of the first.
• Models are named as the server serves them (deepseek/deepseek-flash), with no connection in front; a config.yaml naming one the old way is rewritten once. The context window, efforts and capacity are the server's word.
• Memory is the server's, shared with your chats there (when the login's token may reach it): the memory tool and the memory in the prompt. Notes, knowledge bases and skills stay on this machine.
• Logging in asks for the memory scope in place of library; a login made before keeps working with the one it has.
• Usage is money, not credits. The status bar shows what the month has cost so far, and /usage the month's tokens and cost and this device's tokens — each currency's amount on its own (€2.00 · $0.75), never added together. Plans, balances, rolling windows and the unlimited of an administrator are gone with the server's credits.
• Instruction files stack. Every AGENTS.md (or CLAUDE.md) from the project root down to the working directory goes to the model, outermost first — not only the nearest one.
• Tool results are kept short when they come in, and never changed afterwards, so the server's prompt cache holds. Every result has an id ([#0001]); a read says its lines and a hash of the file; reading an unchanged file again, or the same listing or search with the same result, answers with a pointer to the earlier result; output over 400 lines or 40,000 characters keeps its beginning and its end; edits answer with their diff. compaction.prune, which cut old outputs and made the server read the whole conversation again, is removed.
• The link speaks protocol 3 with a LLeMbas 2.0 server. A 1.x server is refused at login and by the link, saying to update the server (or use LLeMbas CLI 1.x); a newer one says to update the CLI.
• A chat started on a model runs on that model. When the web UI opens a session on this device with a model the device cannot run, it is refused with the reason instead of being answered on the device's own model.
• File tools speak the server's words. A missing file is No such file or directory: <path> from read, edit, multiedit, show, list, glob and grep alike; a file write or edit may not change yet says it has not been read, or changed since it was last read, with its full path. glob and grep give full paths. list sorts names exactly (capitals first) and says when it stopped at 400 lines.
• A new skill's name is lowercase letters, digits and -, up to 64 characters, as the Agent Skills format has it; _ and . are no longer accepted when skill_manage makes one. Skills that already have such a name still load.

Added

• recall: the model can see any earlier tool result again, in full or from a line on — the part of a long output left out, or a result from before the conversation was summarised. After a summary the model also gets a table of the files it had read.
• show, in a session the web UI or LLeMbas Desktop opened: the model can open a file it made — a page, a document — in your viewer beside the chat.
• lembas login --token-stdin --server <url>: sign in with a token a program already holds — LLeMbas Desktop, from the one approval that gives it the agent's token beside its own. The token is read from stdin only, never from an argument or the environment, and checked against the server first: it must speak protocol 3, accept the token, and say it carries link. The login is then stored exactly as a sign-in by code (an earlier one replaced and its token signed out), and --json prints the same done (or error) line.

[1.5.2] — 2026-10-25

Logging in again leaves one device, and a link that cannot connect says why.

Fixed

• Logging in again left an offline twin of this device on the instance. A second lembas login to the same instance overwrote the old token without signing it out, so the web UI showed the device twice -- the new one linked, the old one offline and holding its agent chats. The old token is now signed out, naming the new one, and a LLeMbas server 1.7.2 or later moves the old device's agent chats to it. A sign-out that fails is said, with where to revoke it.
• lembas service could not link when the login was given its CA as a relative path. login --ca kept the path as typed, and the service, started in another directory, did not find it and dialled with no CA at all. The path is now kept in full, and a link to an instance whose certificate is signed by a CA in the system's store gets through even without --ca, as the rest of the CLI already did.
• "Could not reach the instance" now says why: a CA file given at login that is gone, a certificate that is not trusted, or what else the connection met.

[1.5.1] — 2026-10-23

A screenshot no longer makes the session compact.

Fixed

• An image counted as half a million tokens. The context estimate measured an image's base64 as text, so a 1–3 MB screenshot read as "context at 150–280%" the moment view_image (or an @ attachment) added it, and the session was compacted -- after every image, since the summary drops pictures and the model looked again. The provider's own count, a few thousand tokens, was never in question. An image now counts by its dimensions (28-pixel patches, at most 16k tokens; a flat 1600 when the size cannot be read), everywhere the CLI estimates: the context meter, pruning, /context, and the estimate a provider that reports no usage falls back to.
• Screenshots are sent at 1600 px. Any image over 1600 px or 500 KB is shrunk to 1600 px on its longer edge when ImageMagick is there (only files over 3.5 MB were before): no more detail than the model is billed for, and a lighter request.

[1.5.0] — 2026-10-20

Harness spec 2.0.0: background work is one thing — a job — for commands and subagents alike, and the person sees it.

Added

• Subagents in the background. task takes background: true: the call returns at once with a job id and the model carries on; the subagent's report arrives as a message in the conversation when it ends — at the next step while the model works, as a turn of its own otherwise. A running subagent can be messaged (job_send: given to it at its next step) and a finished one started again with its context intact; a subagent can message the main agent before its final reply (report: a finding to act on now, or a question). Approval cards from a subagent say which one asks.
• A subagent has a session of its own, kept with the parent: its transcript survives a resume and can be read later; /sessions and session_search leave children out.
• job_list, job_output, job_stop, job_send replace bash_list, bash_output and bash_kill (which still work as former names, and a rule written for them still applies): one list of the session's background commands and subagents, one way to read, stop and message.
• Jobs belong to the session that started them, not to the process: another session's jobs are invisible, ids run per session, and a job's end note goes to its own session — kept for it when it is not open, delivered when it is resumed. Jobs are stored: a resumed session lists what finished while it was away, and a job the process died under is shown as lost with the output it had.
• Jobs in the terminal. The status bar counts the session's running jobs (2 jobs, with ! when one ended unseen or has output the model has not read). /jobs lists them — kind, id, description or command, state, duration — and from there Enter opens one and k stops it: a command opens as its live output; a subagent is entered, its transcript shown in the conversation's place with a header saying where it is, and anything typed goes to it (a finished one starts again with the message, context intact). Esc leaves. /jobs <id> and /jobs stop <id> do the same from the prompt. A background call's line shows its job under it (→ job7 (explore: description) running…), and an approval card from a subagent names it.
• Ctrl+B moves the command the model is running in the foreground to the background now; the call returns with what it had printed and the job id.
• Quitting (/quit, Ctrl+C twice, Ctrl+D on an empty box) or /reload with jobs running asks first: N jobs running — stop them and quit?
• Over the link (capability jobs, harness spec 2.0.0): a LLeMbas server lists a session's jobs, reads a job's output, stops one, messages a subagent and reads its transcript (_lembas/jobs, _lembas/job/output|stop|send|transcript); job and job_note events, the jobs in _lembas/session/status, the subagent on an approval request, and a job's note as a machine prompt. A session a terminal holds answers the same through the hub.
• limits.jobs (background commands a session may run at once, default 8) and limits.subagent_steps (a subagent's step backstop, default 50) in the configuration.
• lembas -p waits for the jobs the prompt started and gives the model their notes before it exits (--no-wait skips that); /agents lists and describes agents over the link too.

Changed

• A command's timeout is when it stops blocking, not when it dies. A foreground bash still running at its timeout (default limits.bash_timeout, 2 minutes) goes on in the background as a job, and the call returns what it had printed with the job id; Esc still kills it, so does job_stop. background: true returns at once instead of after 300 ms.
• A finished job wakes the model in every front end — the terminal, lembas -p and a session the web UI holds — where it used to reach only the terminal, and there as an unmarked message to whatever session was open. Every such message now carries its mark, and the terminal shows it as the harness's turn, not the person's.
• A stopped command reads "stopped (SIGTERM)" rather than "exited with null".
• The subagent's preamble and the built-in explore and general instructions are prompt texts of the harness spec (blocks/subagent.md, agents/*.md), shared with the server; a subagent is no longer told about plan_write and ask_user, which it does not have, nor about skill_view when it lacks it. explore gets the job tools; every subagent gets report.
• The capacity rule also counts the subagents already running.

Fixed

• A subagent in a worktree that was stopped or failed kept its branch silently; the branch is named in its report now. In a worktree, the edit-mode guard and the plan directory pointed at the main checkout's .agent.
• A subagent's error or notice was dropped: its report said ended: error with no message.
• /reload with background jobs running lost them as orphans; so did quitting with one that ignored SIGTERM (the kill timer never fired at exit). A command whose child kept stdout open stayed "running" after it had exited.
• An agent file with broken YAML frontmatter was silently taken as having none; it is reported at start.
• A !command in the terminal could not be stopped.
• Headless runs opened a second line-reader for a parallel subagent's approval.
• Ctrl+D on an empty prompt quit only in a comment; it does now.
• bun test ran with a 5-second timeout per test, whatever bunfig.toml said: Bun 1.4 does not read [test] timeout. The test script and scripts/ci.sh pass --timeout explicitly.

[1.4.0] — 2026-10-20

Harness spec 1.2.0: the library's tools are the same tools in LLeMbas.

Changed

• The harness spec's memory, notes and skills tools now fit both projects. The CLI's facts in their descriptions — memory files and targets, project notes, supporting files, project skills, session_search — are the spec's variables, filled with the CLI's own values, so what a model is sent here is word for word what it was; LLeMbas sends the same tools with its own facts (LLeMbas 1.5.0). A CLI linked to an instance (library: lembas) therefore sees memory, note_view, note_manage, skills_list, skill_view and skill_manage whether the library is its own or the instance's, where it used to see the instance's memory_add, notes_create and the rest.
• Note and knowledge-document ids are strings in the tools' parameters, as they are on an instance; the number is still accepted.
• Tool descriptions are filled by the prompt texts' rules, and the descriptions inside a tool's parameters are filled too.

Added

• narrow in the harness spec: what a project offers of a shared tool's parameters where it offers less than all. Three conformance areas, run by both projects: fill, narrow, and memory — one change or a batch, applied together, against the size limit, messages included.

[1.3.0] — 2026-10-18

Harness spec 1.1.0: planning as one structure shared with LLeMbas.

Added

• plan_write and plan_update, offered in every mode: the model writes its plan as a structure — title, summary, findings, objectives, phases of tasks — and keeps it current as it works (a task doing, done, or dropped with why; work added; a finding recorded). Ids (p1, t3, o2…) are given by the CLI, never by the model, and an unknown id is an error naming the known ones. The plan is the session's: it is stored with it, comes back on resume, and is gone on /new. The panel above the prompt shows the active phase, the task being done, what is next, the last two done and the count while anything is open; /plan show prints the whole plan with its ids, /plan clear drops it, and /plans says whether this session has one.
• A plan block kind: a plan change is not drawn as a tool line in the transcript; the panel and the card follow the plan itself.
• Over the link, an instance now hands the device what only it knows about a chat (_meta.lembas of session/new and session/prompt, capabilities context, library, prototypes): the chat's own instructions reach the prompt under ## Instructions from <instance>, the account's personalization is used for that session in place of the device's, the knowledge bases the chat may search are named to the model, and whether it may write a prototype. A later prompt's context replaces the earlier one.
• The plan travels as a structure (capability plan_state): the plan session update carries it as _meta.lembas.plan beside ACP's flat entries (null when cleared), _lembas/plan carries it with the card, _lembas/session/status answers it, and a session opened from the store says its plan at once.
• _lembas/session/rewind (capability rewind): the web UI editing or regenerating from a turn takes that turn and every later one out of the conversation and the history — as /undo does, through the same store marking — without touching the files; a session a terminal runs is rewound there, and its transcript redrawn. A rewound turn's id is forgotten, so the web UI re-sending it runs the turn anew rather than being answered what the reverted turn said.
• Prototypes in the terminal: a completed prototype-chat / prototype-page fence in a reply is saved as a file — the model's output as it is, under ~/.local/state/lembas/prototypes/<session>/<n>-<slug>.html, never in the project — and drawn as one line saying where; /prototype lists this session's, /prototype <n> prints one's path. The prompt offers them only when the instance says so for the chat, or with prototypes: true in the global config for a terminal session.
• The library's scope: every call to the instance's library made for a session the instance gave a library context carries _meta.lembas.scope (the session id), so the instance answers within that chat's bases and data group; knowledge_search's base is read against the bases in scope as LLeMbas reads it — by id, then by name whatever its case, an unknown one answered with the bases there are (resolveBase).
• Conformance areas plan (normalise, update, todos), toolcall_index, prototype and library_scope; an area the spec has and this implementation does not answer yet is skipped by name rather than failed.

Changed

• plan_submit presents the plan written with plan_write, not a file: the card shows the plan with its ids, and a copy still goes under .agent/plans/<date>-<slug>.md for the record. Its path argument is kept for callers that send it and ignored. With no plan written it says to call plan_write first.
• The system prompt is the spec's one shared text (system/default.md), with the CLI's facts filled into its variables — the opening line, the terminal as the surface, the file tools, the background-command and git sections, and the current plan with its ids on every step while one exists; the shared blocks/library.md and blocks/prototypes.md join it when their tools are offered. A bullet whose variable is empty is left out, and a variable left unfilled is an error rather than text sent to a model.
• The loop's numbers come from the spec's loop.json (the step backstop, how many calls run together, how often an empty reply is asked again), and so does the wrap-up message, which now names the limit reached: "The time budget for this prompt has been reached…".
• The default theme is moria, the LLeMbas web UI's own; lembas and lembas-light remain selectable. The working line is a plain braille spinner with the verb, without the faces, and the welcome box no longer carries the ☤ glyph.
• A tool-call fragment streamed without an index continues the call started last, not call 0, so a server that numbers only the first fragment of each call no longer merges two calls.
• A permission rule written for todo applies to plan (the key plan_write and plan_update are checked under), as question applies to ask_user.

Removed

• The todo tool. todo is now a former name of plan_write: a call under it — from a stored session, an agent file, or a model trained on it — becomes one phase called "Work" with its items translated (pending → todo, in_progress → doing, completed → done, cancelled → dropped).
• The ☤ glyph and the face frames of the spinner, from the TUI.

[1.2.0] — 2026-10-14

Added

• lembas remote list | add <dir>… | remove <dir>… | enable | disable | mode <m> | terminal <on|off> manages the remote: block without editing YAML, keeping the file's comments. add stores each directory as an absolute real path, skips one already covered by a root, refuses a missing one (nothing is written), and turns remote work on. --json prints one object, {"enabled", "roots", "max_mode", "terminal"}, with the defaults applied — for a program such as LLeMbas Desktop.
• lembas -l / lembas --local: local mode. Config, sessions, memory, skills, the library, snapshots, worktrees, caches, trust and history live in the project's git-ignored .agent/local/lembas/ (.lembas/ where .agent is another tool's), and nothing is written anywhere else — not even temporary files: the TUI starts with TMPDIR there, and the TUI library's cache goes there too (the commands the agent runs keep your TMPDIR). Your logins and connections.yaml are read, never written; your config.yaml is not read. A LLeMbas instance is used for its models, chat, speech in and out and the balance only: no link, /remote, service or hub, no deletes sent, no server library, no search, fetch or embeddings through it, no account personalization, and no self-update. The instance still records each request's usage, as any API does. -l does not trust the project by itself; the first start asks, and the answer is kept in the local directory. A local directory the repository carries is refused, hardline_disable is ignored, and tools treat the local directory as outside the project: no tool writes there (but in its worktrees), and reading it asks as anything outside the project does. The status bar says local, and lembas -l --help explains it.

Changed

• A running lembas service follows the remote: block without a restart: roots, mode, trust, terminals and enabled are read again whenever they are used (a stat of config.yaml; parsed only when it changed). When what the instance was told changes, the link is dialled again once nothing is working, so the web UI hears the new limits.
• lembas service links even before any folder is allowed; it then accepts no session until one is added. With remote work off or no roots it tells the instance roots: [] and refuses every session, directory and terminal; before, service run exited at once and the instance could not tell "remote off" from "not running".
• Harness spec 1.0.1: acp.md says so.

[1.1.0] — 2026-10-14

Added

• lembas login --json prints the code to approve and the outcome as one JSON object per line, for a program that drives the login (LLeMbas Desktop).

[1.0.0] — 2026-10-09

The first public release of LLeMbas CLI: a terminal coding agent and project manager for any LLM API, with permission modes, git snapshots, memory and skills, knowledge bases, MCP, voice, and a link to a LLeMbas instance whose web UI can work its sessions too. Signed Linux binaries for x64 and arm64.