Skill Creator v2.1.0: the input catches up, in-session tools respond

2026-09-16 jixoai v2.1.0 skill-creator

Skill Creator 2.1.0 bundles the 30 commits since 2.0.2. The composer input reaches the official web client semantics: Enter is IME-safe, attachments are checked as a batch, / carries both commands and skills, and a running session accepts queue or steer. @ references files and past sessions, $ references skill documents, with cross-workspace fuzzy search. The queue gains edit, remove, and steer. Roles move onto DSH native subagents. The in-session mcp__skill-creator__* tools that never connected are fixed, and the kernel's DSH data directory defaults into the app directory. No persisted-state breaking changes.

$ git log v2.0.2..v2.1.0 --oneline | wc -l
30

Skill Creator 2.1.0 is out (2026-09-16), bundling the 30 commits since 2.0.2: 13 feat, 5 fix, and 9 docs by commit prefix, with 3 more across chore and test. Three things stand out. The composer input reaches the semantics of the official web client (the web UI shipped with the upstream DSH kernel), and @ and $ references arrive with it. The in-session MCP tools that never connected are fixed. The kernel's DSH data directory moves into the app directory by default. No persisted-state breaking changes; upgrading is just the version bump.

The composer catches up with the official web client

The v2.0 input worked; this release aligns it with the official web client, in four steps.

Typing. Enter no longer submits half-typed IME text, and an Enter within 10ms after composition ends does not count as sending either. Pasted external text passes a sanitizer for invisible placeholder characters. Clearing the input after a successful send drops the native undo stack with it, so Cmd+Z cannot bring sent text back. The placeholder picks the most specific of four states: editing, disconnected, session unavailable, current mode.

Attachments. Images and files are checked as a batch; any count or size violation rejects the whole batch, at 4 images of 4MiB each and 2 files of 512KiB each. Enter holds while files are still being read, with a notice to retry shortly. The drop target covers the whole window, and a drop anywhere goes through the same checks.

Trigger. / carries both commands and skills, commands first; the + button expands the full directory with nothing typed. The command catalog currently holds /compact, /queue, and /steer, each executing on selection.

Submission. While a turn is running, Enter follows your preference to queue or steer, and Cmd/Ctrl+Enter inverts it. The preference sticks, editable from both the settings page and the /queue and /steer commands, one change applying to both. Drafts follow their sessions, survive switches, and clear only on successful send. The stop button cancels the current turn and the queue keeps running in order.

Mechanism: the input-taking command claim machine is ready, no such command ships yet

The claim state machine (the command name settles, the menu yields; backspacing the name exits) is in place. The first catalog has no input-taking commands, so that surface stays unrendered.

References bring files, sessions, and skills into the conversation

@ opens two reference groups. Files browse directories: directory rows drill deeper, .. goes up. The sessions group excludes the current session. A pick drops a @name reference into the input, drawn as a solid block of background color that backspace deletes whole. Submission carries only references still present in the text, at most 4 per message, and the server expands the content: files go through the same checks as attachments (absolute path, real path, the 512KiB ceiling, textual extensions), sessions become bounded transcript excerpts (last 30 text entries, 24k characters), injected as [reference: …] text blocks.

$ is the skill reference, split cleanly from /: /name runs a skill as a command, $name brings the skill document into the conversation as reference. The menu spans every workspace, grouped by Workspace / provider: 916 skills across 44 groups on the dev machine, disabled ones excluded. Search is fuzzy: case-insensitive, letters matching in order, contiguous runs and word starts ranking higher, description hits adding a small bonus. Same-named skills from different workspaces list side by side and pair by order of appearance.

A pick drops $name into the input with exactly the @ interaction. The server receives the workspace, provider, and skill identifiers, resolves them, reads SKILL.md, and expands it into a [reference: skill name · provider] block truncated at 200k characters. A skill that no longer exists rejects the whole message with an explicit error.

The $agents-sdk reference inside the input: a light green rounded background, with the toolbar row intact below

Queued messages can be edited, removed, and steered

The queue goes from a read-only list to something you can operate on. The pencil button edits text in place, and images and attachments survive the save. The x removes, addressed by the kernel's message identifier. Steer turns a running session's next queued message into a next-step intervention, and the button disables while the session is idle.

In-flight messages show as sending until the kernel confirms and the real queue takes over, with identical text deduplicated. /name entries matching the skills catalog pick up a background tint.

Roles ride DSH native subagents

The kernel's DSH family moves from 0.1.5-rc.2 to 0.1.6-alpha.1, with 21 dsh packages upgraded. The role system uses the official subagent capability directly instead of building its own bridge. A role is a subagent definition: a versioned persona, tools gated by session mode, and a unique role tool name; Reviewer, Researcher, and Writer ship built in.

A subagent's spawn, join, and cancel all show in the agent panel. The tool-narrowing rules apply to subagents too, so focused modes cannot be bypassed.

Behavior change: the kernel data directory defaults into the app directory

The daemon used to read and write ~/.dsh directly. Incompatible state in your DSH directory, say a corrupted profile, broke the whole agent kernel mount. The default is now <home>/.skill-creator/dsh-home: first boot creates the empty directory, the model-route settings.yaml and credentials write into the same place, and your ~/.dsh is left alone by default. To keep using the original directory, set one variable.

DSH_HOME=~/.dsh skill-creator start

In-session tools respond again

The in-session mcp__skill-creator__* tools never actually connected; this release fixes them. A regression test drives a real v2 client through the HTTP endpoint, negotiates, and calls a tool; the dev environment log shows zero protocol errors, with the kernel mounting 94 entries. The stdio form moves in the same pass.

Mechanism: root cause and the dual-protocol fix

The kernel's MCP client speaks the 2026-07-28 protocol version; the daemon endpoint (SDK 1.30) accepted at most 2025-11-25 and rejected the newer header with a 400. The endpoint moves to @modelcontextprotocol/server@2, where one per-request factory serves both lines: newer clients negotiate through server/discover, 2025-line clients keep the initialize handshake, and external clients are unaffected.

Other changes

Settings → Model scrolling: the right pane switches scroll ownership per section; inside Model, the tab-content area is the only vertical scroller, the provider gallery loses its own scrollbar, and the tab strip only captures the wheel while it can actually scroll. Verified container by container, desktop and narrow, light and dark (79e51ec).
Opening the agent panel before the connection was up froze the app: an infinite Svelte effect loop killed app reactivity. The lazy load now gates on connection status (same commit).
Menu selected rows get a darker tint, queue row actions move from bare 12px icons to 24px pads with larger click targets, and the steer button gains a visible disabled state.
CI restores the full gate: hermetic global skill roots in tests (299207f), and the web-mode browser smoke test runs in a dedicated ubuntu CI job (998a9ff).
The project site took three passes: typography, motion, and content closeout (caption truth, accessibility fixes, JSON-LD). The README becomes a bilingual pair, English canon plus a Chinese mirror, maintained in sync.

Upgrading

npm install -g skill-creator@2.1.0

No persisted migration. The one behavior change is the kernel data directory default; the earlier section carries the rollback switch.

Acknowledgments

The native subagents and the official web client's input semantics from the upstream DSH project are the sources behind this release's two main capabilities.