Nanocoder v1.27.0 - Skills, a daemon, and custom tools

[will-lamerton]

Nanocoder v1.27.0 is out. This release is the biggest extension surface the project has shipped: a unified extension primitive called Skills, a per-project background daemon for event-triggered runs, and a markdown-defined custom tools layer. Also: two OOM fixes, a startup config lint, and a reworked OpenRouter model picker.

Built by the Nano Collective - a community collective building AI tooling not for profit, but for the community.

Skills: a unified extension primitive

Nanocoder has had custom commands, subagents, and tools as separate extension types. v1.27.0 replaces that with Skills: a single surface that brings all three under one ergonomic model.

A skill member lives in .nanocoder/commands|agents|tools/ as a .md file - fully backwards-compatible with how those directories already work. The new behaviour is the bundle form: a directory under .nanocoder/skills/<name>/ with a skill.yaml manifest and optional commands/, agents/, tools/ subdirectories. Bundles ship and version together as a unit. A bundle's subagent automatically gets its sibling tools, scoped tools are hidden from the global tool list, and bundle commands auto-namespace (commands/status.md in bundle git invokes as /git:status).

New /skills slash command lists everything loaded. Bundled members fan out into the existing CustomCommandLoader, SubagentLoader, and ToolManager.registry - downstream consumers keep using their familiar registries.

A per-project daemon for event-triggered skill runs

Skill members can now declare subscribe: blocks in their frontmatter (or a bundle manifest) that wake them on file.changed and schedule.cron events. The daemon handles dispatch.

Start it with nanocoder daemon start (stop, status, and logs are also available). It runs as a long-lived background process with a lockfile, Unix-socket IPC, and ready-to-use installers for launchd (macOS) and systemd user units (Linux). The interactive TUI never starts event sources - file watching and cron only run when you opt in via the daemon.

Triggered runs execute in a new internal headless mode (no ask_user, no foreground confirmations). Per-subscription confirm: true opts into plan mode instead. The daemon backpressure-caps per-subscription concurrency and trailing-debounces file.changed by 500 ms so chatty save loops do not pile up.

The legacy scheduler is gone

The old source/schedule/ module (runner, storage, index, types) has been removed. Schedules are now expressed as schedule.cron subscriptions on a skill member. The old useSchedulerMode hook, scheduler-view component, and .nanocoder/schedules.json / .nanocoder/schedules/*.md storage layout are gone. The /schedule command has been trimmed and refocused.

Custom Tools: markdown-defined, model-callable tools

Custom Tools sit between custom commands (prompt injection only) and full MCP servers (external process, full protocol). Drop a .md into .nanocoder/tools/ (project) or ~/.config/nanocoder/tools/ (personal), declare parameters in YAML frontmatter with JSON Schema-style types, and write a shell command body using template placeholders. All substitutions are shell-quoted to prevent injection.

Here is the full shape of a Custom Tool definition:

---
name: git_review
description: "Run git diff and collect changed files for review"
parameters:
  - name: since
    type: string
    required: true
    description: "Show changes more recent than this ref"
approval: never
read_only: true
---
git diff --name-only {{since}}..HEAD

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.

LLM-based context compaction

The default /compact strategy is now LLM-based. 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. 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).

OpenRouter request configuration

A new openrouter block on the provider config forwards OpenRouter-specific request body fields on every request: provider routing rules, reasoning, plugins, models fallback list, service_tier, route, user, and so on. Always-on transport and 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.

Config lint at startup

Startup now surfaces common misconfigurations as warnings before they cause silent failures: an openrouter block on a provider that is not OpenRouter, unknown fields, type mismatches. Includes a full test suite.

OpenRouter model selection reworked

The provider wizard's model selection list has been replaced with a paginated, searchable ModelSelectionList (12 visible items, page navigation, multi-select with running counter, select-all, error state). Browsing OpenRouter's hundreds of models is now usable from the wizard.

Unified Session Service

A new Session Service consolidates the various ad-hoc keying strategies scattered across commands, handlers, and tool result display. Touched 50 files, removing duplicated key-derivation logic and shrinking useAppState by ~38 lines.

Nanocoder Battlemap

A new long-form competitive comparison doc covers 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, and telemetry. Honest where Nanocoder leads, honest where it does not.

Fixes

• Streaming OOM on long responses. StreamingMessage was calling wrapWithTrimmedContinuations() on the entire growing assistant message on every ~150 ms flush. A ~37k-token response would overwhelm the GC and crash the 4 GB Node heap. Now slices to a bounded tail before wrapping so per-render work is constant. Thanks to @abhishekDeshmukh74.

• Performance entry buffer OOM in long sessions. React 19's react-reconciler (dev build) and Node's built-in fetch (undici) both write to the global performance buffer, which nanocoder was never draining. Over long subagent-heavy sessions the buffer accumulated millions of entries until V8 thrashed in mark-compact GC. Fix installs an unref'd 30s interval that clears marks and measures in the Ink render path. Thanks to @ragini-pandey.

• Auto-compact now respects the per-provider contextWindow override instead of computing the threshold against the model's default context window.

• Errors from AI SDK providers now surface meaningful messages (status codes, body excerpts, nested fields) instead of [object Object].

• Removed dead MCP server discovery hosts from the wizard templates (remote.mcpservers.org etc.). Companion guard in mcp-client prevents the wizard from surfacing entries it cannot reach.

• Markdown parser regressions fixed: indented fenced code blocks, fenced blocks inside blockquotes, and indented list continuations all extract correctly now. Indented-code extraction removed from the shared parser - only fenced blocks split out for copyable rendering.

• Code blocks are now copyable without surrounding ASCII box borders that previously got copied into the clipboard. Thanks to @aravindinduri.

• System prompt tightened to reduce over-eager task list creation on small jobs and to avoid markdown tables in model output (they render poorly in the terminal).

• Nix packaging updated for pnpm 11 via pkgs.pnpm_11. Drops version-override scaffolding and postInstall hacks. Re-enables the update-nix.yml workflow.

Full changelog at https://github.com/Nano-Collective/nanocoder/blob/main/CHANGELOG.md.