nanocoder v1.28.0

What's Changed

• Added Agent Client Protocol (ACP) support. Nanocoder can now run as an ACP agent (nanocoder --acp), exposing its conversation, tool-calling, and permission flows over the protocol so it can be driven by ACP-compatible editors and clients (Zed and others). New source/acp/ module covers the agent, server, session, conversation loop, content conversion, capability negotiation, and permission handling. Thanks to @Avtrkrb. Closes #529. Follow-up work added the missing ACP docs, kept the parallel tool-execution loop in sync with the plain shell, and made NANOCODER_MAX_TURNS configurable (with a raised default) so long ACP and plain-mode runs are not cut short.

• Consolidated the built-in tool surface from 33 tools down to 19 and added automatic tool profiling. Tasks collapse from four tools into a single TodoWrite-style write_tasks (replace-whole-list, no IDs for the model to juggle); file ops merge delete/move/copy/create_directory into one file_op; git shrinks from eleven tools to six (status/diff/log/add/commit/pr) with rarer operations going through execute_bash. A new auto tune profile - now the default - infers full/minimal/nano from the model's parameter count, so small local models automatically get the slim tool set and prompt while large/cloud models are unchanged. The resolved profile re-resolves live on model switch and is surfaced in the input indicator (e.g. tune: nano (auto)). /usage and the context indicator now rebuild from the live tune/mode/model with the profile-filtered tool count, so usage and context reporting are trustworthy. Roughly 6.3k lines removed.

• Added structured tool results and single-source validation. Tool arguments are now type-checked against each tool's JSON schema at the execution boundary (schema-validate.ts), so malformed model output returns a clear field-level error the model can self-correct from, instead of being coerced at the render layer. Validation runs through the single withValidation seam shared by both the native and XML-fallback execution paths, ahead of any per-tool validator. Approval policy is likewise centralised and the old global mode context removed, so the current development mode (including a mid-run switch) is honoured consistently across the main loop and subagents.

• Added session resume with full history replay. Resuming a saved session now replays the message and tool-call history into the chat view via a new session-history-renderer, so you pick up exactly where you left off with the conversation visible rather than an empty screen. Companion fixes resolved an autosave race, history-truncation, and resume edge cases, with regression coverage. Thanks to @akramcodez. Closes #545.

• Added the /copy command to copy the most recent assistant response to the system clipboard. Cross-platform via clipboardy (pbcopy / xclip / clip.exe), with a clear warning when there is no assistant message and a surfaced error when the clipboard write fails.

• Added read-before-edit guards and tool-call loop detection for local models. string_replace and write_file now require the file to have been read first, and a tool-signature tracker detects a model repeating the same tool call in a tight loop. Both make smaller local models materially safer to run unattended.

• Added a skill linter (/skills check <name> and a check_skill tool) that validates a bundle from disk with the same parsers the loader uses, so a PASS means it will load. It is wired into /skills create so the model verifies and self-corrects what it generates, and it lints member template bodies (unsupported mustache tags, unbalanced or unclosed sections, placeholders referencing undeclared parameters), not just frontmatter. Also added optional command arguments: inline defaults (name=default) and conditional sections ({{# name }}, {{^ name }}) in command bodies, plus inverted-section support in custom-tool templates via a shared section engine. Fixed subagents ignoring the development mode (the executor captured normal once at startup and never updated it).

• Added Atlas Cloud as a first-class provider and sponsor, with a wizard onboarding template and a dedicated provider docs page.

• Added the Requesty provider (https://requesty.ai) as a first-class OpenAI-compatible provider, mirroring the OpenRouter integration with a name matcher, app-attribution headers, a wizard onboarding template, and docs. Thanks to @Thibaultjaigu. Closes #589.

• Added API-reported context usage with an estimate fallback. The ctx: NN% indicator now reflects provider-accurate token counts when the model reports them (captured from the final step's usage, not the tool-loop-summed totalUsage), falling back to client-side estimation otherwise. Estimated figures are marked with a leading ~ (ctx: ~42%); API-reported figures render bare. Thanks to @JimStenstrom. Refs #381. Follow-up work anchored the API figure across turns, counted real tool definitions in the auto-compact gate, and added a tiktoken-based generic fallback tokenizer for more accurate counts on models without a native tokenizer.

• Added git branch display in the boot summary and /status panel, with a narrower-terminal-friendly rendering. Thanks to @ragini-pandey. Closes #539.

• Reworked provider model discovery UX in the wizard, surfacing real discovery errors instead of silently falling back, and showing model names in the selection list with long labels truncated by ellipsis. Thanks to @akramcodez and @Dhirenderchoudhary. Closes #554.

• Added arrow-key navigation and highlighting for slash-command completion. The completion menu now navigates with the arrow keys, with the double-submit-on-select bug fixed (TextInput ignores Enter; UserInput handles select-then-submit behind a guard flag) and customCompletion restored. Thanks to @rakshith1928. Closes #578.

• Added a force-compact-and-retry fallback. When the LLM server rejects a request for exceeding its context window, nanocoder now compacts and nudges the conversation to continue instead of failing outright. Thanks to @lordoski. Closes #546.

• Added timeout, output limits, and abort support to the bash executor, with regression tests covering the executor lifecycle. Thanks to @akramcodez. Closes #547.

• Custom display: failed tool results condense to a one-liner in compact mode. Generic Error: / Validation failed: results now render as a short red line (e.g. write_file failed.) in compact display mode, while the model still receives the full error in history and non-compact mode still shows it in full.

• Improved ask_user prompt rendering: dropped the leading ? from the header, fixed spacing and the hanging indent, added arrow-key navigation, and normalised object-shaped options to readable strings (preferring a label over a value id).

• Improved @-file-mention handling: mention placeholders are kept in the chat instead of being expanded inline, and large-file inlining is capped to avoid blowing up the prompt.

• Fix: Copilot GPT-5 reasoning streams are now handled correctly via an @ai-sdk/openai patch and chat-handler changes, so reasoning content streams properly instead of breaking the response. Thanks to @EntropyParadigm. Closes #577.

• Fix: propagate AbortSignal to streaming bash execution, so cancelling a turn actually stops a long-running streamed command. Thanks to @Dhirenderchoudhary. Closes #550.

• Fix: prevent silent overwrite of corrupted config files. A corrupted agents.config.json no longer gets silently clobbered; the wizard now honours the projectDir prop for config-path resolution and the infinite render loop on the existingProviders default param is resolved, all with regression coverage. Thanks to @Dhirenderchoudhary. Closes #551, #552.

• Fix: read_file returns an empty marker instead of throwing on empty files, including range and metadata-only reads, with the EMPTY_FILE_MARKER constant extracted (later renamed EMPTY_CONTENT_MARKER) and tests for files containing only a newline. Thanks to @ragini-pandey. Closes #530.

• Fix: prevent orphaned tool results from breaking LLM compaction. A tool-aware boundary in the converter and summariser stops a recent-tail slice from splitting a tool-call group, which previously produced empty model output after compaction.

• Fix: / key no longer scrolls the chat view to the bottom. Thanks to @itsishant. Closes #590.

• Fix: close the chokidar watcher if daemon startup throws, and move the daemon stop-handler setup before resources start so a failed lockfile write calls stop() and cleans up. Thanks to @OMEE-Y and @rakshith1928. Closes #553, #557.

• Fix: replaced a stray console.log in message-queue.tsx with the structured logger. Thanks to @A-S-Manoj. Closes #588.

• Fix: added subagent context window overrides so delegated agents can run with a different context limit than the main session, and resolve provider names case-insensitively with custom CA bundle support for provider TLS. Thanks to @zerone0x.

• Added a tmp >=0.2.6 override to resolve a path-traversal advisory in a transitive dependency. Thanks to @ragini-pandey.

• Large refactor and dead-code sweep: unified tool-call ID generation, extracted a shared useWizardForm hook, consolidated session-override managers, added message-factory helpers, deduped config loaders / git exec / command dispatch / conversation-loop flush, extracted StyledSelectInput and makeSimpleToolFormatter, and deleted several dead modules and orphaned exports (including fetch-local-models orphaned by an earlier removal).

• Updated the Nanocoder Battlemap competitive comparison and refreshed the README and docs.

• Dependency updates: ai 6.0.174 -> 6.0.193, @ai-sdk/openai, @ai-sdk/openai-compatible, @ai-sdk/anthropic, @agentclientprotocol/sdk 0.22.1 -> 0...

nanocoder v1.27.0

What's Changed

• Added Skills, a unified extension primitive that brings commands, subagents, and tools under a single ergonomic surface. Two forms over one primitive: single-file (a .md in .nanocoder/commands|agents|tools/, fully backwards-compatible with the existing flat dirs) and bundle (a directory under .nanocoder/skills/<name>/ with skill.yaml plus optional commands/, agents/, tools/ subdirs). Bundles are for multi-piece features that need to ship and version together: a bundle's subagent automatically gets its sibling tools, scoped tools are hidden from the global tool list (default for bundles, opt-in for single-file), and bundle commands auto-namespace (commands/status.md in bundle git invokes as /git:status). New /skills slash command lists everything loaded, and bundled members fan out into the existing CustomCommandLoader, SubagentLoader, and ToolManager.registry so downstream consumers keep using their familiar registries. Closes #515.

• Added event-triggered skill runs via a new per-project daemon. Skill members can declare subscribe: blocks in frontmatter (or a bundle manifest) and the daemon wakes them on file.changed and schedule.cron events. The daemon (nanocoder daemon <start|stop|status|logs>) runs as a long-lived background process with a lockfile, Unix-socket IPC, and installers for launchd (macOS) and systemd user units (Linux). The interactive TUI never starts event sources, so file watching and cron only run when you opt in. Triggered runs execute in a new internal headless mode (no ask_user, no foreground confirmations) which supersedes the legacy scheduler mode. Per-subscription confirm: true opts into plan mode instead. Backpressure caps per-subscription concurrency and trailing-debounces file.changed by 500ms so chatty save loops don't pile up.

• Removed the legacy source/schedule/ module (runner, storage, index, types). Schedules are now expressed as schedule.cron subscriptions on a skill member and dispatched by the daemon. The old useSchedulerMode hook, scheduler-view component, and the .nanocoder/schedules.json / .nanocoder/schedules/*.md storage layout are gone. The /schedule command has been trimmed and refocused around the new model. Closes #524.

• Added Custom Tools: markdown-defined, model-callable tools that sit between custom commands (prompt injection only) and MCP servers (full external process). Drop a .md into .nanocoder/tools/ (project) or ~/.config/nanocoder/tools/ (personal), declare parameters in YAML frontmatter (with JSON Schema-style type, required, pattern, maxLength, etc.), and write a shell command body using {{ name }} and {{# section }} template placeholders. All substitutions are shell-quoted to prevent injection. Includes approval (always/never) and read_only flags that participate in mode policy: plan mode requires approval=never && read_only=true, scheduler/headless requires approval=never. New /tools create <name> scaffolds a template and asks the model to help fill it in. Project tools shadow personal ones by name, and they register into the same ToolManager registry as built-ins and MCP tools so /tools, subagents, and mode filtering see them through a single unified registry. Closes #520.

• Added LLM-based context compaction as the default /compact strategy. The active model writes a structured markdown summary of the compressible segment (Context / Decisions / Files modified / Tools used / Open questions) using a dedicated summariser prompt, replacing the older messages with one synthetic summary while recent messages are kept verbatim. Higher fidelity than mechanical truncation at the cost of one extra round-trip. New strategy field in autoCompact config (llm default, mechanical legacy), new CLI flags (--llm, --mechanical, --strategy <name>), and automatic fallback to mechanical compression on LLM failure (network error, empty response, summary larger than original). Auto-compact uses the same strategy.

• Added OpenRouter request configuration via a dedicated openrouter block on the provider config. Forwards OpenRouter-specific request body fields (provider routing rules, reasoning, plugins, models fallback list, service_tier, route, user, etc.) on every request - always-on transport/routing concerns, not gated by /tune. Provider is detected by name (case-insensitive). tune.modelParameters.reasoningEffort populates openrouter.reasoning.effort when unset; explicit values on the provider config always win. Closes #519.

• Added config lint at startup. Surfaces common misconfigurations as warnings before they cause silent failures - e.g. an openrouter block on a provider that is not OpenRouter, unknown fields, type mismatches. Includes a full test suite. Closes #523.

• Reworked OpenRouter model selection in the provider wizard. Replaced the previous list with a paginated, searchable ModelSelectionList (12 visible items, page navigation, multi-select with running counter, select-all, error state). Makes browsing OpenRouter's hundreds of models actually usable from the wizard. Closes #516.

• Added unified Session Service for key generation and session infrastructure. Consolidates the various ad-hoc keying strategies scattered across commands, handlers, and tool result display into one place. Touched 50 files across handlers, commands, hooks, and the chat handler, removing duplicated key-derivation logic and shrinking useAppState by ~38 lines. Fully closes #229.

• Added the Nanocoder Battlemap - a long-form competitive comparison doc covering Nanocoder against Claude Code, OpenAI Codex CLI, Gemini CLI, Aider, OpenCode, Crush, and Pi across twelve axes (license, pricing, local-model support, MCP, extensibility, tool-calling, subagents, surface, plain mode, stars, contributors, telemetry). Honest where Nanocoder leads, honest where it does not. Closes #423.

• Made code blocks copyable without ASCII artifacts. Reworked the markdown parser and AssistantMessage rendering so fenced code blocks render as plain selectable text without surrounding box borders that previously got copied into the clipboard. Thanks to @aravindinduri.

• Fix: streaming OOM on long responses. StreamingMessage was calling wrapWithTrimmedContinuations() on the entire growing assistant message on every ~150ms flush, allocating per-line arrays proportional to the full message size. A ~37k-token (~150 KB) response would overwhelm GC and crash the 4 GB Node heap. Now slices the message to a bounded tail (MAX_LINES * textWidth * 4 chars, snapped to a newline boundary) before wrapping, so per-render work is constant in the visible window rather than linear in total streamed content. Thanks to @abhishekDeshmukh74. Closes #526.

• Fix: performance entry buffer OOM in long sessions. React 19's react-reconciler (dev build) calls performance.measure/performance.mark on every render, and Node's built-in fetch (undici) pushes a resource-timing entry per HTTP request. Both write to the same global buffer, which nanocoder never drained. Over long subagent-heavy sessions the buffer accumulated millions of entries until V8 thrashed in mark-compact GC and crashed with Ineffective mark-compacts near heap limit. Fix installs an unref'd 30s interval that calls performance.clearMarks() / performance.clearMeasures() in the Ink render path. Thanks to @ragini-pandey. Closes #521.

• Fix: auto-compact ignored the per-provider contextWindow override. The auto-compact threshold was computed against the model's default context window even when the provider config narrowed it, causing compaction to fire late (or not at all) on providers configured with a smaller usable window. Closes #525.

• Fix: surface real provider errors instead of [object Object]. The error parser now extracts meaningful messages from AI SDK errors (status codes, body excerpts, nested responseBody fields) before they reach the UI, so model and API failures show up as actionable text instead of an opaque stringified object.

• Fix: removed dead MCP server hosts from the wizard templates. remote.mcpservers.org and a handful of other defunct discovery endpoints were still listed in mcp-templates.ts and would silently fail when selected. Companion dead-URL guard in mcp-client prevents the wizard from surfacing entries it cannot reach.

• Fix: markdown parser regressions in fenced code extraction. Indented fenced code blocks (e.g. inside list items) now extract correctly, fenced blocks nested inside blockquotes are no longer split out (they belong to the quote), and indented list continuations are preserved as list text instead of being misparsed as indented code. Removed indented-code extraction from the shared markdown parser entirely - only fenced blocks split out for copyable rendering.

• Fix: tightened the task management section of the system prompt to reduce over-eager task list creation on small jobs.

• Minor: nudged the system prompt to avoid markdown tables in output. Tables render poorly in the terminal even though the markdown converter supports them, so the model is now steered toward bullet lists or short prose.

• Fix: Nix packaging - reproducible pnpm 11 builds via pkgs.pnpm_11 now that nixpkgs PR #505103 has landed. Drops the version-override scaffolding, the chmod +x pnpm.cjs postInstall hack, and the manage-package-manager-versions replaceStrings patch from the previous release. Also documents and works around an upstream fetch-pnpm-deps shell bug (export VAR value without =) that caused non-deterministic v11/index.db writes when side-effects-cache fell back to true. Re-enables the update-nix.yml workflow. Closes #511.

• Audit/security: updated pnpm-workspace.yaml auditConfig and minimumReleaseAgeExclude entries to align with the versions resolved ...

nanocoder v1.26.1

What's Changed

• Bumped @nanocollective/get-md to ^1.4.0, which makes node-llama-cpp an optional peer dependency. Drops the entire node-llama-cpp transitive — including ~500 MB of platform-specific native binaries (CUDA, Vulkan, Metal, ARM variants) that pnpm fetched eagerly regardless of host — from the install graph and the Nix closure. The fetch_url tool (the only consumer of get-md) uses the standard HTML→Markdown path which doesn't need the LLM converter, so there's no functional change.

• Patched flake.nix for pnpm 11 compat. Overrides nixpkgs's bundled pnpm (currently 10.x) to pnpm 11.0.9 to match packageManager, then handles two pnpm-11-specific quirks: chmod +x on bin/pnpm.cjs (a Corepack-compat shim shipped without the execute bit by upstream) and a replaceStrings patch on fetchPnpmDeps's installPhase to skip the now-rejected pnpm config set manage-package-manager-versions false global write. Unblocks the update-nix.yml workflow that's been failing on ERR_PNPM_LOCKFILE_CONFIG_MISMATCH.

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.26.0...v1.26.1

nanocoder v1.26.0

What's Changed

• BREAKING: Removed redundant nanocoderTools.alwaysAllow setting. It duplicated the top-level alwaysAllow with no extra behaviour. Move any entries from nanocoderTools.alwaysAllow to the top-level alwaysAllow array in agents.config.json. The deprecated agents.config.example.json has also been removed.

• Added nano mode — a third tool profile under /tune designed for the smallest open-weights models or low-end hardware running larger models locally. Strictly more aggressive than minimal: drops find_files, list_directory, and agent; trims CORE PRINCIPLES and CODING PRACTICES; uses ≤4-line TASK APPROACH, FILE OPERATIONS, and CONSTRAINTS sections; replaces the verbose SYSTEM INFORMATION block with a single-line ## SYSTEM line; and omits AGENTS.md from the prompt by default. Brings the system prompt from ~500–700 tokens (minimal) down to ~150–250 tokens. Includes a new "Nano (low-end hardware)" preset and an Include AGENTS.md toggle.

• Added reasoning trace support. Reasoning content from models (Codex GPT-5, DeepSeek-R1-style, Anthropic extended thinking, etc.) now streams in real time, renders as a collapsible Thought block above the response, persists across history, and is included in logs. Toggle expansion with Ctrl+R, configure default expansion via the new Display Settings panel, and pin per-session via the tune expandedReasoning option. <think> tags are now stripped on the native tool-calling path so reasoning never leaks into rendered output. Thanks to @Daniel5055. Closes #457.

• Added refined non-interactive mode. A new --plain flag streams output suitable for CI pipelines, scripts, and pipes — no Ink rendering, no interactive prompts, deterministic exit codes, and proper handling of stdin/stdout. Includes a dedicated non-interactive-shell component and full test coverage for the plain transport.

• Reworked VS Code extension integration. Removed the "Ask Nanocoder" command in favour of a more natural flow: highlighted text and the currently focused file are automatically pulled into Nanocoder as context. Backed by a rewritten extension protocol, a simplified extension entrypoint, a new useVSCodeServer hook, and tighter UI hooks for the development mode indicator and chat input.

• Added /rename command for renaming the current chat session. Accessible from the chat input and reflected in the development mode indicator. Thanks to @lordoski.

• Added defaultMode config option for interactive sessions. Set a starting mode (normal, auto-accept, yolo, plan) in agents.config.json instead of always launching in normal mode. Thanks to @lordoski.

• Added custom system prompt support via agents.config.json. Define a project-level system prompt that replaces or augments the built-in prompt sections, with full validation and prompt-builder integration. Closes #487.

• Added per-provider and per-model context window overrides in agents.config.json via contextWindow and contextWindows[model]. New resolution order: session override (/context-max, --context-max) → provider model override → provider default → NANOCODER_CONTEXT_LIMIT → models.dev / Ollama fallback. /usage, status bar, auto-compact, and context reporting all use the active provider config consistently. Closes #455.

• Added subagent context window overrides so delegated agents can run with a different context limit than the main session. Thanks to @zerone0x.

• Added disabledTools config option in agents.config.json to disable specific built-in tools project-wide. Honoured by both the main agent and subagents.

• Added --trust-directory CLI flag to bypass the first-run directory trust prompt for automation. Yolo mode now also skips file path validators in line with its "auto-accept everything" semantics.

• Added JSON tool fallback mode for non-tool-calling open-weights models (Qwen, Kimi, GLM). The tune menu now cycles through Native ON / OFF (XML) / OFF (JSON), with formatToolsForJSONPrompt() embedding literal JSON Schema and a JSON branch in the tool-call parser. Native path also gains JSON/XML hallucination recovery: parseToolCalls() is run against text content when the SDK reports no tool_calls, and stripEmbeddedToolCallText() removes echoed tool-call text so Ghost Echo no longer leaks into the UI or history. Reference #500.

• Added <function=...> format to the tool-call parser for models that emit OpenAI-style function tags.

• Added Display Settings panel under /settings ("Tool Results and Thinking") with two new toggles: Show Thinking by default (Ctrl+R) and Expand Tool Results by default (Ctrl+O), persisted to nanocoder-preferences.json. Thanks to @cleyesode. Closes #499.

• Added 12+ new themes to the bundled theme set.

• Removed the in-repo release content generator workflow. Release content is now generated from another repo via ContentForest.

• Bumped CI and devcontainer toolchain to Node.js 22 + pnpm 11 across pr-checks.yml, release.yml, update-badges.yml, and .devcontainer/Dockerfile. engines.node raised to >=22, with a packageManager field (pnpm@11.x.x) added to package.json so Corepack pins the pnpm version automatically for both contributors and CI — no more drift, no per-workflow version: lines. CONTRIBUTING.md / docs/getting-started/installation.md / .devcontainer/README.md updated to match. Fixes the node:sqlite ERR_UNKNOWN_BUILTIN_MODULE crash hit by pnpm 11 on Node 20 runners.

• Migrated pnpm config from package.json to pnpm-workspace.yaml for pnpm 11 compatibility: overrides, patchedDependencies, and the new allowBuilds block now live in the workspace file. While migrating, audited and removed all 16 existing version overrides — every one was redundant under current dependency resolution (each transitive resolved to a version already exceeding the override's floor), and removing them left pnpm audit output unchanged.

• Major spring-clean refactor across the codebase. Extracted shared oauth-login command from the duplicated codex-login and copilot-login flows; introduced shared item-selector component used by model-selector and provider-selector; centralised tool-needs-approval logic; moved AI SDK error-handling specs into a dedicated subdirectory; tightened logging method factory; and trimmed redundant code paths in conversation-loop, subagent-executor, and the plain shell.

• Refactored App.tsx and added missing test coverage for the app container and chat input components.

• Refactored settings theme selector UX. Replaced the scrollable SelectInput with arrow-key navigation, real-time theme preview, and a "Theme Name [n/N]" position indicator. Preferences are only persisted on Enter.

• Updated the system prompt to encourage subagent use more proactively for complex multi-step tasks.

• Strip leading and trailing whitespace from assistant messages so blank lines and stray newlines from the model no longer push content around in the UI.

• Fix: Empty model responses now trigger a capped auto-continue instead of crashing or burning tokens forever. Empty-response retry notices are coalesced into a single live counter so the chat history stays readable.

• Fix: Surface context-size failures explicitly instead of silently retrying empty responses. Thanks to @zerone0x. Closes #501.

• Fix: Persist compressed messages across recursive conversation turns. Auto-compact now also resets the streaming token count after compression, with regression coverage to keep it stable. Thanks to @lordoski. Closes #480.

• Fix: Throw stream errors immediately during streaming instead of swallowing them and ending the response cleanly. Thanks to @alexbrjo.

• Fix: Cap malformed-XML retries to prevent OOM when a model gets stuck emitting broken tool-call XML in a tight loop. Reference #500.

• Fix: Stale lastProvider after running the config wizard mid-session — the next /provider invocation picked up the old value. Includes regression coverage. Thanks to @electricwolfemarshmallowhypertext. Closes #477.

• Fix: Inaccurate tok/s reporting in the streaming status line.

• Fix: File search, content search, and @mention autocomplete now work cross-platform. Replaced Unix-only find/grep/which shell-outs with pure Node.js implementations honouring .gitignore and DEFAULT_IGNORE_DIRS, restored the 30s search timeout via AbortSignal, and added Windows path normalisation. Thanks for the cross-platform report (#469).

• Fix: Read file tool description corrected.

• Fix: Type MCPTool.inputSchema as JSONSchema7 instead of any, with isPlainObject guard at the MCP protocol boundary so malformed server schemas resolve to undefined rather than flowing through the system as a trusted type. Thanks to @ragini-pandey. Closes #467.

• Fix: Apply caCertPath TLS configuration to Anthropic and Google providers (previously only OpenAI-compatible providers honoured it), with cert path validation. Companion fix for custom CA bundle support across all providers. Thanks to @zerone0x. Closes #491 (#380).

• Fix: Reject ambiguous provider names with case-insensitive duplicates (e.g. both Ollama and ollama defined). Provider names are now resolved case-insensitively throughout. Thanks to @zerone0x. Closes #488.

• Fix: Quoted custom command arguments (double, single, and backtick) are now parsed as a single parameter so multi-word values survive tokenisation. Closes #478.

• Fix: Missing margin on the live sub-agent progress component.

• Fix: git_pr tool still required approval in yolo mode.

• Fix: bash mode no longer passes the currently selected file when not relevant.

• Fix: Incorrect passing of system prompt in some chat-handler paths.

• Fix: Default missing subagent_type and description parameters on the agent tool to safe values to prevent a UI crash when the model omits them.

• Fix: Mo...

nanocoder v1.25.2

What's Changed

• Fixed Nix package: copy themes.json and prompt section files into the Nix store so nix run no longer crashes at startup
• Fixed Nix package: corrected wrapper script heredoc indentation that broke the shebang line

If there are any problems, feedback or thoughts please drop an issue or message us through Discord! Thank you for using Nanocoder.

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.25.1...v1.25.2

nanocoder v1.25.1

What's Changed

• Removed rogue document from docs/

If there are any problems, feedback or thoughts please drop an issue or message us through Discord! Thank you for using Nanocoder.

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.25.0...v1.25.1

nanocoder v1.25.0

What's Changed

• Added yolo mode — a new development mode that auto-accepts every tool without exception, including bash commands and destructive git operations (hard reset, force delete, stash drop/clear). Cycles between normal → auto-accept → yolo → plan via Shift+Tab. The status bar turns red when yolo mode is active.

• Added subagents — isolated child conversations that the LLM can delegate work to. Ships with two built-in agents: Explore (read-only codebase investigation) and Reviewer (code review with actionable feedback). Subagents are defined as markdown files with YAML frontmatter specifying name, description, model, and allowed tools. User-defined subagents can be placed in .nanocoder/agents/. Managed via the /agents command (show, create). Thanks to @BRIJESHKR for the initial subagent implementation. Closes #414.

• Added concurrent subagent execution. The LLM can launch multiple subagents in parallel, each with independent tool sets and live in-place progress rendering via the AgentProgress component.

• Added subagent tool approval matching main agent behaviour. Subagent write tools prompt for approval and bash always prompts unless in alwaysAllow. The agent tool itself no longer requires approval since internal tools have their own gates.

• Redesigned the system prompt into a modular, composable architecture. The monolithic main-prompt.md has been replaced with individual section files under source/app/prompts/sections/ (identity, core principles, coding practices, file editing, tool rules, diagnostics, task management, etc.). The new prompt-builder assembles the prompt dynamically based on the current mode — normal, auto-accept, plan, and scheduler each get a tailored prompt with only the sections and tools relevant to that mode. Includes a generate-system-prompts script for offline token counting and prompt inspection.

• Made plan mode useful. Plan mode now enforces read-only tools at the policy level — all mutation tools (write, bash, git commit/push, task management) are blocked, leaving only exploration tools (read, search, find, list, git log/diff/status). The dedicated plan mode system prompt instructs the LLM to investigate thoroughly and produce a structured plan with summary, files to modify, step-by-step approach, dependencies/risks, and open questions — instead of trying to execute changes.

• Added /tune command for per-session prompt and tool customization. Includes a tune selector UI (Ctrl+T) with tool profiles (full, minimal), a disableNativeTools toggle for forcing XML fallback, and aggressive compact mode. Tune state persists for the session and is reflected in the mode indicator.

• Centralized tool policy into ToolManager so prompt-time tool filtering and runtime approval use the same source of truth. Extracted tool-registry.ts for cleaner separation of tool definitions from policy logic.

• Added ChatGPT Codex as a provider with OAuth device flow authentication (/codex-login), streaming response support via a dedicated StreamingMessage component, and Codex-specific credential management. Includes provider template and setup wizard integration.

• Migrated web_search tool from Brave Search scraping to the official Brave Search API. Now requires a webSearch.apiKey in agents.config.json under nanocoderTools. Removes the cheerio scraping dependency.

• Added /setup-config command that lists all config files (project and global agents.config.json, .mcp.json, nanocoder-preferences.json) with their paths and opens the selected one in your editor.

• Added configurable paste threshold for single-line paste handling, with tests for the configurable placeholder threshold. The threshold is a user preference in nanocoder-preferences.json. Changed the default config file for paste settings from agents.config.json to nanocoder-preferences.json. Thanks to @grenkoca.

• Added live in-place task list display. Task progress now updates in place instead of appending repeated static lists to the conversation.

• Improved tool output truncation across all tools. Every tool formatter now respects terminal width for cleaner output, including execute_bash, file ops, git tools, search_file_contents, and web_search.

• Redesigned the provider setup wizard with a unified model fetcher that auto-detects API compatibility (OpenAI-compatible, Ollama, Anthropic, Google) and fetches available models from the provider's endpoint. Simplified the provider step UI and added MiniMax and Kimi provider templates.

• Fix: alwaysAllow config not being respected for execute_bash. Three interconnected bugs prevented it from working: top-level alwaysAllow was never loaded from agents.config.json, isNanocoderToolAlwaysAllowed only checked nanocoderTools.alwaysAllow not the top-level list, and nonInteractiveAlwaysAllow set needsApproval on AI SDK tools but the conversation loop evaluated it from the original registry entries. Closes #431.

• Fix: dimColor making text inaccessible to reading on some screens. Closes #440.

• Fix: Tasks now clear when running /clear command.

• Fix: Prevent edit flow from resolving MiniMax/Kimi to Anthropic template.

• Fix: Set correct default model for MiniMax provider.

• Fix: /usage and context percentage showing stale system prompt length. Added setLastBuiltPrompt and fixed useContextPercentage overwriting cache.

• Added debug logging to 15 silent catch blocks in git utilities that were swallowing errors invisibly. Operational catches now log at debug level for easier diagnosis while boolean probes (e.g. isGitAvailable, branchExists) remain intentionally silent. Thanks to @ragini-pandey. Closes #452.

• Security: Address semgrep security finding in console.error. Thanks to @BRIJESHKR.

• Security: Fixed vulnerable packages. Thanks to @BRIJESHKR.

• Security: Replaced exec() with execFile() in VS Code extension installer to prevent command injection. Removes shell interpretation from CLI discovery and extension status checks. Thanks to @ragini-pandey.

• Added /credits command showing project contributors (auto-generated from git history via pnpm run build:credits) and dependency versions.

• Added desktop notifications for tool confirmations, question prompts, and generation completions. Supports macOS (terminal-notifier with osascript fallback), Linux (notify-send), and Windows (PowerShell). Configurable per-event in nanocoder-preferences.json and in /settings with custom messages and optional sound. Includes a "Notifications" settings menu for preference management.

• Added CLI quality metrics framework (benchmarks/measure.ts, benchmarks/report.ts) tracking correctness, performance (module count, boot time, bundle size), stability (tool/command counts, help text hash), and health (test counts, vulnerabilities).

• Reduced startup time by parallelizing and deferring non-critical initialization. LLM client and subagent init run in parallel on the critical path; update checks, MCP servers, and LSP servers now initialize in the background without blocking chat. Replaced the full Status box with a lightweight one-line BootSummary component.

• Fix: Scheduler mode memory leak — chat messages and components now clear before each scheduled job execution, preventing memory accumulation across repeated runs.

• Fix: Updated agent documentation (AGENTS.md, CLAUDE.md) with current project structure (766 files across 97 directories), added missing directory mappings, removed deprecated references, and documented the lazy-loading command system.

• Fix: Improved Ollama timeout handling for slow local models. Timeout errors are now detected case-insensitively, and IPv6 loopback addresses are correctly resolved. Thanks to @ragini-pandey.

• Dependency updates

If there are any problems, feedback or thoughts please drop an issue or message us through Discord! Thank you for using Nanocoder.

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.24.1...v1.25.0

nanocoder v1.24.1

What's Changed

• Added --context-max CLI flag for setting the context limit from the command line, complementing the existing /context-max command and NANOCODER_CONTEXT_LIMIT env variable.

• Removed time from the system prompt to keep the KV cache more stable across requests. Thanks to @initialxy. Closes #415.

• Task tool results are no longer displayed as compacted results during ensuring task progress remains visible in the conversation.

• User input now uses the same text wrapping as assistant messages for a more consistent chat appearance.

• Improved search_file_contents tool robustness.

If there are any problems, feedback or thoughts please drop an issue or message us through Discord! Thank you for using Nanocoder.

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.24.0...v1.24.1

nanocoder v1.24.0

What's Changed

• BREAKING: Removed legacy ~/.agents.config.json config file support. Nanocoder no longer checks the home directory for a dot-prefixed config file. If you are still using this path, move your config to the platform-specific directory: ~/Library/Preferences/nanocoder/agents.config.json (macOS), ~/.config/nanocoder/agents.config.json (Linux), or %APPDATA%\nanocoder\agents.config.json (Windows).

• BREAKING: Removed legacy ~/.nanocoder-preferences.json preferences file support. Preferences are now only loaded from the platform-specific directory (e.g. ~/Library/Preferences/nanocoder/nanocoder-preferences.json on macOS) or a project-level nanocoder-preferences.json in your working directory. To migrate, move your existing file: mv ~/.nanocoder-preferences.json ~/Library/Preferences/nanocoder/nanocoder-preferences.json

• BREAKING: Removed deprecated array format for MCP server configuration. Only the object format is now supported: { "mcpServers": { "serverName": { ... } } }. If you are using the array format in .mcp.json, convert each array entry to an object key using the server name.

• BREAKING: Removed agents.config.json fallback for MCP server loading. Global MCP servers must now be configured in ~/.config/nanocoder/.mcp.json (Linux), ~/Library/Preferences/nanocoder/.mcp.json (macOS), or %APPDATA%\nanocoder\.mcp.json (Windows). Provider configuration still uses agents.config.json.

• BREAKING: Removed auth and reconnect fields from MCP server configuration. The auth field was never functional (both HTTP and WebSocket transports logged warnings that it was unsupported). The reconnect field was never implemented. Use headers for HTTP authentication instead (e.g. "headers": { "Authorization": "Bearer $TOKEN" }).

• Added /resume command for restoring previous chat sessions. Sessions are automatically saved and can be resumed from an interactive selector. Sessions are filtered by the current project directory by default, with an --all flag to show all sessions. Thanks to @yashksaini-coder.

• Added --provider and --model CLI flags for non-interactive provider and model specification, allowing CI/CD scripts and automation to skip the setup wizard. Closes #394. Thanks to @james2doyle.

• Added NANOCODER_PROVIDERS environment variable support for configuring providers without config files, useful for Docker containers and CI environments. Closes #307. Thanks to @kaustubha07.

• Added GitHub Copilot as a provider template with OAuth device flow authentication and /copilot-login command. Thanks to @yashksaini-coder.

• Added MLX Server provider template for local Apple Silicon inference. Closes #318.

• Added parallel tool execution allowing the model to run multiple independent tool calls concurrently for faster task completion.

• Added compact mode toggle via Ctrl+L in chat input to collapse the conversation view.

• Added VS Code fork support for IDE integration (Cursor, Windsurf, VSCodium, etc.). Thanks to @kapsner.

• Added Aurora Borealis theme.

• Added notice when the model falls back to XML tool calls, informing users they can switch to a model with native tool calling support.

• Adopted AI SDK human-in-the-loop pattern for tool approval. Tool confirmation now uses the SDK's built-in tool-approval-request/tool-approval-response flow instead of manual tool-call splitting, improving reliability and reducing code complexity.

• Simplified tool processing by removing double XML parsing and the JSON tool call parser. Tool call parsing now happens in a single place and only on the XML fallback path for non-tool-calling models.

• Restructured documentation into a Nextra-compatible docs/ folder structure with nested sections for getting-started, configuration, and features. The README is now a concise landing page linking to the full docs.

• Refactored app-utils into focused handler files, extracted shared utilities, unified mode state, and stubbed commands for cleaner architecture.

• Fix: alwaysAllow field in MCP server configuration was silently dropped during config loading due to a missing field mapping. MCP tools configured with alwaysAllow now correctly skip confirmation prompts as documented.

• Fix: Provider timeouts are now respected in non-interactive mode. Thanks to @kaustubha07. Closes #402.

• Fix: Non-interactive mode no longer exits prematurely when the prompt or response contains the word "error".

• Fix: Invalid CLI arguments no longer trigger the setup wizard. Thanks to @james2doyle.

• Fix: Installation detector no longer falsely reports Homebrew on macOS when HOMEBREW_PREFIX is set but Nanocoder was installed via npm. Closes #392.

• Fix: Preserve draft message when navigating through history with arrow keys.

• Fix: fetch_url display now truncates to fit terminal width.

• Fix: Validation failures no longer incorrectly prompt for tool confirmation.

• Fix: Various error message and execute_bash formatter improvements.

• Fix: Safe process metrics refactored into shared module to prevent duplicate declarations. Thanks to @cleyesode.

• Security: Package audit failures resolved.

• Dependency updates: ai 6.0.116, @ai-sdk/anthropic 3.0.58, @ai-sdk/google 3.0.33, @ai-sdk/openai-compatible 2.0.30, @modelcontextprotocol/sdk 1.27.1, undici 7.24.0, cheerio 1.2.0, dotenv 17.3.1, wrap-ansi 10.0.0.

If there are any problems, feedback or thoughts please drop an issue or message us through Discord! Thank you for using Nanocoder.

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.23.0...v1.24.0

nanocoder v1.23.0

What's Changed

• Added ask_user tool for interactive question prompts. The LLM can now present the user with a question and selectable options during a conversation, returning their answer to guide the next step. Uses a global question-queue to bridge the tool's suspended Promise with the Ink UI component.

• Added per-project cron scheduler for running AI tasks on a schedule. Schedule files live in .nanocoder/schedules/ as markdown prompts with YAML frontmatter, managed via the /schedule command (create, add, remove, list, logs, start). Includes cron expression parsing, sequential job queue with deduplication, dedicated scheduler mode with auto-accept, and run history logging.

• Added centralized graceful shutdown system. A ShutdownManager now coordinates cleanup of all services (VS Code server, MCP client, LSP manager, health monitor, logger) on exit, preventing orphaned child processes and dangling connections. Configurable via NANOCODER_DEFAULT_SHUTDOWN_TIMEOUT env variable. Closes #239.

• Added file operation tools: delete_file, move_file, create_directory, and copy_file. Reorganized existing file tools into a file-ops/ directory group.

• Added readline keybind support to text input. Replaces ink-text-input with a custom TextInput component supporting Ctrl+W (delete word), Ctrl+U (kill to start), Ctrl+K (kill to end), Ctrl+A/E (jump to start/end), and Ctrl+B/F (move char). Closes #354.

• Added /context-max command and NANOCODER_CONTEXT_LIMIT env variable for manual context length override on models not listed on models.dev. Resolution order: session override > env variable > models.dev > null. Closes #379.

• Added /ide command matching the --vscode flag for toggling VS Code integration from within a session.

• Added persistent context percentage display in the mode indicator, replacing the previous context checker component.

• Added include and path parameters to search_file_contents tool for scoping searches to specific file patterns and directories.

• Added Kanagawa theme.

• Refactored the skills system into custom commands, eliminating redundant parsers, loaders, and test suites. Commands gain optional skill-like fields (tags, triggers, estimated-tokens, resources) for auto-injection and relevance scoring. The /skills command is removed and its functionality absorbed into /commands with new subcommands (show, refresh). Thanks to @yashksaini-coder for the initial skills implementation in PR #370.

• V2 type-safe tool system overhaul with defensive parsing. Implements a three-tiered defense system for handling chaotic LLM outputs, preventing crashes from non-string responses and enabling robust self-correction. Includes universal type safety with ensureString(), response normalization, confidence system inversion, ghost echo deduplication, and AI SDK contract fixes. Local LLM experience is now significantly more stable. Thanks to @cleyesode. Closes #362.

• Fix: XML parser now uses optimistic matching for consistency with the JSON parser. Thanks to @cleyesode.

• Fix: Bash tool now emits progress immediately on stdout/stderr data instead of waiting for the 500ms timer, so fast-completing commands show streaming output.

• Fix: Recognize 127.0.0.1 as a local server URL and tighten error classification. Ollama users configuring 127.0.0.1 instead of localhost no longer experience misleading connection errors. Replaced broad connect substring match with specific error codes to prevent misclassifying "disconnect"/"reconnect". Closes #366.

• Fix: Skip loading git tools when not inside a git repository.

• Fix: Strip ANSI escape codes before running regex matching in tool formatters.

• Fix: Gap in layout during auto-compact.

• Fix: Hardened write_file validation and MCP client type safety.

• Fix: Use local TextInput component instead of the missing ink-text-input package.

• Fix(mcp): Use Python-based mcp-server-fetch instead of non-existent npm package.

• Security: Semgrep and audit fixes.

• Dependency updates: ai 6.0.95, @ai-sdk/anthropic 3.0.46, @ai-sdk/google 3.0.30, undici 7.22.0, sonic-boom 4.2.1.

If there are any problems, feedback or thoughts please drop an issue or message us through Discord! Thank you for using Nanocoder. 🙌

Installation

npm install -g @nanocollective/nanocoder

Usage

nanocoder

Full Changelog: v1.22.5...v1.23.0