@staff0rd/assist 0.745.0 → 0.747.0

package/README.md

@@ -82,6 +82,7 @@ After installation, the `assist` command will be available globally. You can als
 - `/test-review` - Review existing tests for quality, coverage gaps, and conventions
 - `/inspect` - Run .NET code inspections on changed files
 - `/screenshot` - Capture a screenshot of a running application window
+- `/ask [what to ask about]` - Compose a plan, summary or proposed decision and put it in the web preview pane via `assist ask`, run as a background task so a command timeout can't kill the pending preview, then proceed on approval (folding in inline comments) or revise and re-ask on rejection
 - `/show [what to show]` - Open long URLs, paths and snippets in the web preview pane via `assist show`, where the terminal's hard wrapping can't break them. With no arguments it gathers them from the previous reply
 - `/raven` - Query and manage RavenDB connections and collections
 - `/miro [board url | extract name]` - Dump a Miro frame's raw `board_list_items` pages and extract its boxes as an ordered YAML list via `assist miro extract`
@@ -347,6 +348,7 @@ The Config tab of the sessions web dashboard never receives secret values: `GET
 - `assist mermaid export [file.md]` - Render each fenced mermaid block to `<stem>-<index>.svg` via [Kroki](https://kroki.io) (`--out`, `--index`, `mermaid.krokiUrl`)
 - `assist prompts` - Show top 10 denied tool calls by frequency with count and repo breakdown
 - `assist chart [--title <title>]` - Draw a terminal line chart of a `label value` series piped in on stdin, one pair per line, separated by a comma, tab or whitespace. Points are charted in the order given — nothing is sorted or aggregated — and the chart closes on q, Esc or Ctrl-C. The y axis fits the data range with 20% padding rather than starting at zero, so a series that only moves in its third decimal still reads as a shape; a flat series is padded so the line does not sit on the axis. Blank lines are skipped, fewer than two points prints `Not enough data points to chart.`, and a non-numeric value exits 1 naming the line
+- `assist ask --title <title> --body <markdown|->` - Put a plan, summary or proposed decision in the session's web preview pane under a Review chip and block until the user approves or rejects it; `--body -` reads it from stdin. Inline comments (quote + note) attach to either decision. Approving exits 0 and prints each comment with its quoted excerpt; rejecting exits non-zero, printing the reason and every comment with advice to revise and re-run. An empty body is an error. Outside a web session it prints the markdown to stdout and exits 0. The `ask` advice fragment steers agents to use it instead of asking for sign-off in chat; turn it off with `assist config set advice.sections.ask false`
 - `assist show --body <markdown|-> [--title <title>]` - Render markdown in the session's web preview pane so long links, paths and code can be clicked and copied unbroken; `--body -` reads it from stdin. It returns straight away and the pane stays open, with links opening in a new tab, until the user clicks Close. Inline and fenced code each get a copy-to-clipboard button. A later `show` replaces it, but a pending approval preview (PR, backlog item, Slack post and so on) is left in place and `show` exits non-zero saying so. An empty body is an error. Outside a web session it prints the markdown to stdout. The `show-links` advice fragment steers agents to use it; turn it off with `assist config set advice.sections.show-links false`
 
 ### Project setup

package/allowed.cli-reads

@@ -64,6 +64,7 @@ assist sessions nodes
 assist sessions nodes doctor
 assist sessions nodes logs
 assist sessions output
+assist ask
 assist show
 assist sql auth list
 assist sql columns

package/claude/advice/ask.md

@@ -0,0 +1,6 @@
+---
+title: Asking for sign-off
+when: always
+---
+
+When a plan, summary or proposed decision needs the user's sign-off before you act on it, don't ask in chat. Write it as markdown to a scratch file and run `assist ask --title '<title>' --body - < <file>` as a background task, doing no other work until it returns. It opens in the session's web preview pane with approve, reject and inline comments, and blocks until the user decides. Approval exits 0 and prints any comments to fold in; rejection exits non-zero with the reason and comments — revise and re-run. Outside a web session it just prints the markdown, so ask in chat.

package/claude/commands/ask.md

@@ -0,0 +1,29 @@
+---
+description: Put an update in the web preview pane and wait for the user's sign-off
+allowed_args: "[what to ask about]"
+---
+
+`assist ask` renders markdown in the session's web preview pane with approve and reject controls and inline commenting, and blocks until the user decides. Use it for a plan, summary or proposed decision that needs sign-off before you act on it.
+
+## Step 1: compose the update
+
+If `$ARGUMENTS` is non-empty, compose the update it describes. Otherwise, compose the plan, summary or decision you were about to ask the user to confirm in chat.
+
+Keep it self-contained: the user reviews it in the pane, not alongside the conversation. Lead with what you intend to do, then the facts behind it. Use headings and lists so individual points can be commented on.
+
+Write the markdown to a scratch file with the Write tool. It will contain quotes, backticks and newlines, so don't inline it in a shell command.
+
+## Step 2: ask
+
+```
+assist ask --title '<short title>' --body - < <scratch file>
+```
+
+Always run it **as a background task**, and do no other work until it returns. In a web session it blocks until the user decides, which can take far longer than the default command timeout, and the pending preview dies with the process.
+
+Outside a web session it prints the markdown and exits 0 without a decision: ask the user in chat instead.
+
+## Step 3: act on the decision
+
+- **Approved (exit 0)** — any inline comments are printed, each with its quoted excerpt. Proceed with the update, folding in every comment.
+- **Rejected (non-zero exit)** — the reason and every inline comment are printed. Do not proceed. Address each one, revise the scratch file, and re-run Step 2.