Repository navigation
Description
Product or interface
CLI - interactive TUI
Version
0.7.2 (source tree at commit 86f7772), Node.js v24.22.1
Platform
macOS
OS version and architecture
macOS 26 (darwin 25.6.0) arm64
Issue area
UI / UX
Steps to reproduce
- Start the interactive TUI in a workspace that has at least one enabled Skill. An enabled Skill named
reviewis exposed as/review, and/skills [filter]lists the roster. - On an empty composer, type
/. The slash menu opens and lists built-in commands together with the Skills — this works as expected. - Clear the composer, type
please review this fileand then type/. - Observe that no completion menu appears, and the
/is inserted as plain text. - Repeat with the cursor on the second line of a multi-line prompt and type
/. Again nothing appears.
Expected and actual behavior
Expected: a / typed at a token boundary anywhere in the composer opens the completion menu at the cursor, filtered by the token currently being typed, and selecting an Agent Skill inserts its /name token at the cursor. The user should be able to pick a Skill while a prompt is already half written, not only before the first character.
Actual: the menu only opens when the text before the cursor, on the first line, is empty or exactly /. A / typed after other text, or on any line other than the first, produces no completion at all. Because every Skill is only reachable through its /name token, this also means a half-written prompt cannot invoke any Skill unless the user abandons the text, retypes it after a leading /name, or restarts the message.
Source-level observations
The behaviour is gated in three places, so fixing the composer alone is not sufficient:
packages/tui/src/tui/engine/components/editor.ts:1337— the auto-trigger ischar === '/' && this.isAtStartOfMessage().packages/tui/src/tui/engine/components/editor.ts:2357—isAtStartOfMessage()requiresisSlashMenuAllowed()(line 0 only,:2352) andbeforeCursor.trim() === '' || beforeCursor.trim() === '/'.packages/tui/src/tui/engine/components/editor.ts:2364—isInSlashCommandContext()requirestextBeforeCursor.trimStart().startsWith('/'), so typing more letters after a mid-prompt/also never re-triggers the menu.packages/tui/src/tui/commands/catalog.ts:727—matchTuiCommandInput()splits the trimmed input and only treats the first whitespace-separated token as a command token. Even if the menu were shown and inserted/reviewmid-prompt, Enter would send it to the model as ordinary prompt text instead of dispatching a Skill invocation.
The last point needs a product decision as part of the fix: either move the chosen token to the leading position when the user selects a Skill, or define inline skill invocation explicitly so a mid-prompt token still resolves to the Skill and appends the rest of the message as its instructions.
Desired behavior
- Triggering: typing
/at any token boundary in the composer opens the Skill/command menu, including after existing text and on any line of a multi-line prompt. - Filtering: the menu filters on the token before the cursor, and
Escdismisses it without altering the prompt. - Insertion: the selected Skill's token is inserted at the cursor position, preserving the surrounding text.
- Consistency: the same set of Agent Skills is offered wherever the menu opens, so
/skillsand the/nametokens stay in sync with what/skillsreports as available.
Additional context
I searched the existing issues (open and closed) for this behaviour before filing; the closest ones are distinct:
- [Feature]: Support custom Skills via slash commands in ACP mode #230 — Skills as slash commands in ACP mode (closed); about protocol exposure, not composer completion.
- [Feature]: support disable-model-invocation in SKILL.md frontmatter #555 —
disable-model-invocationsupport. - [Feature]: Recursively discover nested SKILL.md directories #498 — recursive
SKILL.mddiscovery. - [Feature]: On-demand Skill and MCP discovery to reduce context overhead #263 — on-demand Skill discovery to reduce context overhead.
None covers where the completion menu may be triggered, or how a Skill token is resolved when it is not the leading token of the message.
Activity