lake

 ██╗      █████╗ ██╗  ██╗███████╗
 ██║     ██╔══██╗██║ ██╔╝██╔════╝
 ██║     ███████║█████╔╝ █████╗
 ██║     ██╔══██║██╔═██╗ ██╔══╝
 ███████╗██║  ██║██║  ██╗███████╗
 ╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝╚══════╝
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

A terminal LLM agent for Windows and Linux.

What it is

lake is a CLI coding agent that runs against your own LLM server — local llama.cpp / ollama / LM Studio / DeepSeek, or any OpenAI-compatible endpoint you control. It ships as a single AOT-published binary per RID and has a focused built-in tool family: filesystem, shell, terminal sessions, HTTP, web search, PDF, image, plus a multi-modal code knowledge graph (workspace.*) for cross-file understanding.

Every session persists locally, so /resume, /revert, /compact, and filesystem time-machine rollback all work out of the box. Subagents spawn nested task loops with their own context budget. Memory recalls and stores across sessions. Each tool can be approved, denied, denied-with-reason, reviewed inline, or compared as a diff before it runs.

Features

• Bring your own LLM Runs against any OpenAI-compatible endpoint — local llama.cpp, ollama, LM Studio, or remote DeepSeek / Gemini. No vendor lock-in.
• Single AOT binary One native executable per RID (win-x64, linux-x64). No .NET runtime install, no dependency hunt, cold-start in milliseconds.
• Built-in tool family fs, shell, terminal, http, web search, email, pdf, image — all native, all AOT-clean, all approval-gated.
• Rendering trio diagram_create (Mermaid), chart_create (Vega-Lite), dot_render (Graphviz). Descriptions cross-reference so the model picks the right one for the shape of the task.
• Skills library skills_search / skills_use fetch distilled how-to recipes from a central repo; /skills create and /skills refine capture and iterate on new ones from live sessions.
• Context handoff /handoff <task> distills the current session into a briefing and spawns a fresh successor (inherited config / role / overrides); /resume jumps back if the successor derails.
• Code knowledge graph workspace.* tools (find, tree, summary, search, outline) give the model cross-file understanding with background indexing.
• Time-machine /revert rolls back not just the conversation buffer but every fs.* file mutation since the chosen turn. ESC-ESC picker, single keystroke.
• Session persistence Every turn appended to events.jsonl on disk. /resume, /compact, full replay — sessions outlive the process, the terminal, and the machine.
• Tool approval gating Per-tool prompt with [a]pprove, [d]eny, deny-with-reason, [f]ormat, [c]ompare diff, and inline code.review cross-ref.
• Subagents agent.spawn launches a nested task loop with its own catalog, model, and event stream. One primary /revert undoes the whole spawn.
• Memory across sessions Auto-recall surfaces relevant prior context at turn start; auto-store captures new facts on the way out. Persists outside the conversation buffer.
• Headless / scripting lake -p "…" --json emits an NDJSON event stream on stdout — same code path as interactive, machine-readable for pipelines and CI.
• Overflow-safe context When the backend rejects a turn for exceeding its context window, oversized tool results are shed and the call is retried automatically — the loop keeps running instead of aborting.
• Buffer inspection /history renders the raw message buffer with per-message sizes — the debug affordance for questions /revert can't answer.

Recent releases

• 0.5.208 shell.exec guard: shell:true + non-empty args is rejected up front instead of silently dropping args (which used to run bare python and loop on a broken REPL until the 120 s timeout).
• 0.5.207 dot_render — Graphviz DOT graphs with every layout engine (dot / neato / fdp / sfdp / twopi / circo). Third of a rendering trio alongside diagram_create and chart_create; auto-registers when dot resolves on PATH.
• 0.5.206 Tool descriptions reference sibling tools by wire form (fs_read, workspace_stats, diagram_create, …) — matches what the sanitized backend catalog presents to the model so the sibling references actually resolve.
• 0.5.205 Unified tool-name wire form across Ollama + OpenAI-compat backends via a shared sanitizer; dotted config names (sys.info) present as underscored names (sys_info) to satisfy OpenAI's function-name regex on every backend.
• 0.5.204 chart_create — Vega-Lite statistical charts (bar, line, area, scatter, histogram, boxplot, heatmap, faceted small-multiples) via vl-convert. No Chromium, no Node — self-contained native binary.
• 0.5.203 diagram_create — Mermaid.js rendering via mmdc. Every Mermaid diagram type: flowchart, sequenceDiagram, classDiagram, stateDiagram-v2, erDiagram, gantt, pie, mindmap, timeline, gitgraph, quadrantChart, journey, sankey-beta.
• 0.5.200 /handoff v2 + agent.handoff: successor session inherits mode / role / per-model overrides; companion tool agent.handoff chains inside subagent trees with a per-branch depth cap (hidden from the child catalog).
• 0.5.196 /handoff <one-line task> — brief-and-branch context handoff. Drafts a briefing from the current session, spawns a fresh successor session (empty history, inherited config, workspace); the predecessor stays inspectable via /resume.
• 0.5.189 /optimize: iterative revise + undo across chat-shaped refiner history — the drafter sees the prior draft and your revision instruction, so successive rounds build on the same context.
• 0.5.184 image.create wires the local lake-image service (FastAPI wrapper for SDXL-Turbo, FLUX.1-schnell, and hidream-fast for text-in-image) into the tool belt.
• 0.5.175 Skills PR A: skills_search + skills_use — search and import distilled how-to recipes from a central skill repo, gated by LAKE_TOKEN.
• 0.5.176 /skills create — 3-pass sequential distillation of a live session into a new skill; /skills refine (0.5.178) for surgical iteration on existing ones. Distillation reads events.jsonl so a compacted session still yields the full skill.
• 0.5.171 email.auto_approve_recipients — whitelisted addresses skip the Auto-mode High-risk escalation on email.send, so long-running digest loops don't need YOLO.
• 0.5.170 code.review + approval reviewer response cap (64 KiB); runaway generations get snipped with a truncation notice instead of blowing up the next turn's context.
• 0.5.169 image.read hardening: reasoning forced off + 32 KiB response cap. Vision models that fail to emit EOS no longer fill the entire context window.
• 0.5.168 Context-overflow recovery + /history. When the backend returns "context exceeded", oversized tool results are shed and the call retried; /history shows the raw buffer with per-message sizes.
• 0.5.167 /optimize <text> — refine a rough prompt via purposes.optimize, review in a panel, accept / edit / discard.
• 0.5.166 Approval [r]eview: the reviewer's findings thread into the tool result the model sees on the next turn, not just the human-facing panel.
• 0.5.165 /context [<n>|reset] — per-session override for the effective context window; drives auto-compact and the status-row token gauge.
• 0.5.163 workspace.index tool — model-triggered rescan of the code knowledge graph. Complements background indexing.
• 0.5.162 Workspace parsers for C/C++, Java, Kotlin, Shell, and SQL; the graph now understands more repos out of the box.
• 0.5.155 Session settings pin — backend / model / role / per-model config survive resume across intervening config edits.
• 0.5.150 Approval [f]ormat action — per-tool formattable fields (e.g. email.send/reply text + html) rendered inline before you approve.
• 0.5.145 Roles: Markets category with 11 markets specialists.
• 0.5.144 Headless / scripting mode — lake -p "…" --json streams NDJSON events on stdout, same code path as interactive.
• 0.5.142 email.* family: eight verbs (send / reply / list / get / update / list_threads / get_thread / search).
• 0.5.141 agent.sleep + agent.wake_after — time-aware tools for scheduling future self-prompts with a live countdown.
• 0.5.140 Auto-compaction: tri-state mode + threshold-gated trigger, so long sessions no longer require manual /compact.
• 0.5.139 Subagent feature: agent.spawn + agent.completed — delegate bounded tasks to a nested agent loop with its own catalog, model, and event stream; one /revert at the primary level undoes the whole spawn.
• 0.5.136 lake config models retrieve [backend] and gemini config models retrieve — auto-discover the backend's model list and append to config.
• 0.5.130 code.review — model-callable single-file code review tool, with [c]ompare action and risk-eval cross-refs in approval panels.
• 0.5.127 /system + prompts.user — inspect the five-layer system prompt (base / role / user / mode / memory) and customise the user layer in config or per session.
• 0.5.124 DeepSeek backend (OpenAI-compatible) + LM Studio bootstrap config example.
• 0.5.122 /tools — interactive tool browser with per-session enable / disable.
• 0.5.120 /resume picker upgrade: turns / checkpoints / fs-ops counts, role, and ★ for compacted sessions.
• 0.5.118 workspace.* — multi-modal code knowledge graph (find, tree, summary, stats, search, outline) with background indexing.
• 0.5.117 Per-tool master switch in config.json — opt any tool family in or out at startup.
• 0.5.114 /compact — structured-summary context compaction. Preserves first user message + last assistant message; drops checkpoints and pre-compact filesystem snapshots.
• 0.5.113 Time-machine v2a — /revert also rolls back fs.* file mutations, not just the conversation buffer.
• 0.5.110 Time-machine v1: /revert + ESC-ESC picker for jumping back to a prior turn.
• 0.5.108 Diff colourisation (git-style red / green) + Rust syntax highlighting.
• 0.5.101 Memory adoption: prompt layer, surfaced auto-recall, auto-store from turns.

Currently shipping: 0.5.207. Installation, manifest, and per-RID artifacts are under a separate access-gated path on this host.