Pull the template text transform out of SessionPrompt.command() into a
reusable expandCommandTemplate(cmd, args) helper: positional ($1..$n) +
$ARGUMENTS substitution, inline shell interpolation, trim. Behavior is
unchanged for command(); the helper deliberately excludes agent/model/
subtask selection, plugin hooks, and command events so the upcoming
inline-skill resolvePart branch can reuse only the text expansion.

No behavior change. Existing command tests pass.

Add SkillPartInput to PromptInput.parts and a skill branch to resolvePart,
mirroring the agent branch: a skill part resolves to the persisted structured
chip part plus a synthetic, model-visible text part carrying the expanded
template (argless — the surrounding prose is the turn body). This makes skill
invocation position-independent (activation reads the parts array, not a
leading token) without touching the session.command endpoint, command(), or
toModelMessages (synthetic text flows through; the structured part is skipped
like AgentPart). Unknown skill names keep the chip but inject nothing.

Tests: skill part after prose resolves to chip + synthetic template + intact
prose; unknown skill keeps chip, injects no synthetic text.

Regenerated src/v2/gen from the opencode server schema (bun run build):
adds SkillPart to the Part union and SkillPartInput to PromptInput.parts.
Consumer-side only; the checked-in packages/sdk/openapi.json snapshot is
untouched (route-inventory reads paths only).

Introduce SkillAttachmentPart in the prompt content union so an inline skill chip can live anywhere in the composer, mirroring the @-agent part. Thread the new member through clonePart, prompt equality, history cloning, and the composer preview helpers.

Round-trip a contenteditable=false skill pill (data-type=skill, data-name, data-source) through createPill / parseEditorToParts / renderPartsToEditor, accept it in isNormalizedEditor so the editor does not rebuild every keystroke, and count it as a non-editable boundary in the caret-placement predicates.

Replace the start-anchored slash regex with a cursor-relative, CJK-aware trigger so "/" opens the command picker mid-sentence without firing inside paths, URLs, or fractions. When the trigger fires mid-text, restrict the picker to inline-eligible entries (skills + simple custom commands), excluding action builtins and agent/arg-heavy commands which keep the leading path.

Selecting a skill or simple command from the picker now replaces the typed "/query" with a position-independent skill chip via addPart, instead of prepending a leading command mark. Arg/agent-heavy commands and action builtins keep their existing leading-pill / trigger paths.

Extract inline skill chips into SkillPartInput on send (mirroring agents), tagging the "/name" span via source so the bubble can render it. Add a toOptimisticPart skill case before the agent fallthrough so the optimistic bubble shows the chip and does not flicker when the server-persisted SkillPart arrives.

When a prompt or followup draft contains a structured skill part, its flattened text can start with "/name" (chip at offset 0). Guard both legacy session.command routing branches (submit + followup) and the canSubmitPrompt command_ready gate so such prompts flow through promptAsync, where the skill part resolves into synthetic model text. Path D (marked leading TextPart) is unaffected: a skill chip carries no command metadata.

Add a skill reference segment to HighlightedText keyed off the skill part's source span, and pass skill parts alongside inline files in the user-message prose branch (not the commandInvocation XOR branch, which suppresses the body). The "/name" token stays inline and is colored with the brand accent, matching the leading command mark.

Bridge the skill glyph from the chrome icon registry into the command-icon registry so both the composer pill and the sent bubble can resolve it. The composer skill pill now renders glyph + /name label (textContent stays /name for caret math); the bubble skill segment renders the glyph before the brand-accent /name token. The persisted SkillPart drops the source kind, so the glyph is uniform across surfaces. Adds a focused message-skill-inline snap target.

… test

Move SLASH_TRIGGER + matchSlashTrigger into slash-trigger.ts so it can be unit-tested without importing editor-input (which pulls @solidjs/router and throws in the test env). DRYs the two offset computations in editor-input. Adds a regex matrix covering start/space/newline/CJK triggers and path/URL/fraction/word-char negatives.

…ness

editor-serialize: skill pill attributes, '/name' textContent invariant, prose+skill+prose round-trip, position independence in isNormalizedEditor. build-request-parts: SkillPartInput emission + matching optimistic skill part (no agent fallthrough). session-action-readiness: inline skill bypasses the command-hydration gate but still respects submitReady.

Seeds a project-scoped skill (.agents/skills/summarize/SKILL.md), types prose then "/summarize" mid-sentence, asserts the picker opens with the skill and excludes action builtins, clicks it, and verifies a position-independent skill chip lands in the composer.

When a leading marked command owns the turn (Path D / session.command, which carries only image parts), a "/" in its args no longer opens the skill picker — an inserted skill chip would be silently dropped on submit. The slash stays literal command-argument text. Addresses a /codex review P2.

The skill glyph (✦) already signals "this is a /command"; showing the
slash as well produces "✦ /screenshot" instead of "✦ screenshot".
Strip the leading "/" when the segment type is "skill" so the chip reads
as glyph + name only.

The skill glyph already signals "this is a slash command", so the
leading "/" in the pill label is redundant visual noise. Strip it from
the label span; pushSkill rebuilds the canonical "/<name>" token from
data-name so flattened text and source offsets are unchanged.

editor-dom already computes logical pill length as 1 + data-name.length
(not textContent), so caret math is unaffected.

The skill chip now renders glyph + bare name (no leading "/") in both
the composer pill and the sent bubble. Update comments in
editor-serialize, message-part.css, highlighted-text.tsx, and the
message-skill-inline snap description to match the current behavior.

Three gaps after stripping the leading "/" from the pill label:

1. getNodeLength/getTextLength only checked data-cmd-mark, not
   data-type="skill". Skill chips reported textContent length (bare
   name) while the prompt model used "/name" length, causing cursor
   offsets to drift by one after any skill chip.

2. The copy handler (command-copy.ts) only intercepted data-cmd-mark
   pills. Copying a selection with a skill chip produced bare "name"
   on the clipboard instead of "/name", losing the slash on paste.

3. The e2e assertion in prompt-slash-open.spec.ts still expected
   "/summarize" in the chip's visible text.

Fix all three: extend the pill detection in editor-dom and
command-copy to include [data-type="skill"], update the e2e assertion.