Regression: slash command autocomplete no longer triggers mid-input (worked per v2.1.0, fixed once in v2.1.136)

Experiencing the same issue here on 2.1.197 (Claude Code) for MacOS Tahoe 26.5

Confirming the same regression on macOS (existing reports here are Windows).

Environment

• Claude Desktop 1.37937.3 on macOS (Darwin 25.6.0)
• App bundle last updated Aug 26 2026 ~20:01 local; the behavior change appeared on the next relaunch (Aug 28), consistent with the update taking effect on restart.

Observed

Slash-command autocomplete only opens when / is the first character of the input. Typing / after any preceding text (e.g. check this repo /my-skill) shows no menu and no command highlighting, and the token renders as plain text. @ file mentions still trigger anywhere in the input; / does not.

Steps to reproduce

1. Focus the prompt input.
2. Type / as the first character — menu opens (works).
3. Clear, type some words first, then type / + a skill-name prefix.
4. No menu appears; the /… is not recognized as a command.

Same result in the stacked-skill position (/skill-a /skil… — only the leading command is highlighted). Reproduces consistently across app restarts. Matches the desktop report in #89771; adding macOS + the 1.37937.x line for platform coverage.

Similar issues were closed as "not planned" however please consider the case where we are in the middle of a prompt and can't recall the exact name of the skill (especially where there are many with similar names), for example:

"UI looks great, now we need to make the end point - you may find /ajax and /db-update-for-admin-user skills useful".

Checking the name via / at the start means abandoning a half-written prompt.

Confirming this on Linux with the native CLI — existing reports cover Windows Desktop and macOS Desktop, so this closes the "desktop-only" question. Also the highest version reported so far.

Environment

• Claude Code 2.1.252 (native binary), Linux 7.0.0-30-generic
• Terminal CLI, not the desktop app

One detail that may help narrow it down

Mid-input, the ghost text still works — it's only the selectable list that never opens. Typing ciao /pon renders ytail:ponytail inline in grey, and Tab accepts it. So mid-input:

• the command registry is populated ✅
• the / is parsed as a command token ✅
• the prefix match runs and returns the correct candidate ✅
• the suggestion list does not render ❌

Two code paths, one broken. This is consistent with v2.1.20 ("Fixed ghost text flickering when typing slash commands mid-input") still holding while the v2.1.0 menu path does not.

Repro on a clean profile

Ruled out local config, plugins and user skills entirely:

mkdir -p /tmp/cc-clean/.claude
cp ~/.claude/.credentials.json /tmp/cc-clean/.claude/
HOME=/tmp/cc-clean claude

Then type ciao /mo (built-in /model, no plugins loaded). Ghost text appears, no menu. Identical to the normal profile.

Note on prior closures

Re: the earlier reports closed as "not planned" — worth weighing that the changelog advertises this as a shipped feature in two separate entries (v2.1.0 introducing it, v2.1.136 fixing it), and none of the 99 releases between v2.1.136 and v2.1.252 announces its removal. From a user's side this reads as a silent regression, not a deprecation.

The workaround (/ on an empty input, select, then write the rest) inverts the natural order of composing a prompt, and Tab-accepting a single blind candidate doesn't help when several skills share a prefix.