Make the pre-PR gate answerable; open-pr now pushes

Cargo.toml

@@ -6,7 +6,7 @@ members = [
 
 [package]
 name = "retch-cli"
-version = "0.6.20"
+version = "0.6.21"
 edition = "2021"
 authors = ["Ken Tobias"]
 description = "A fast, feature-rich system information fetcher written in Rust (similar to fastfetch or neofetch)"

Justfile

@@ -382,8 +382,32 @@ pr:
     echo "  [ ] GitHub wiki cloned and updated (Configuration-and-Theming.md, Workspace-Architecture.md)"
     echo "  [ ] Upstream tldr page updated / docs/retch.md synced (if CLI flags changed)"
     echo ""
-    echo -n "All manual items confirmed? [y/N] "
-    read -r CONFIRM
+    # A bare `read` makes this gate unanswerable by anything that is not a human at a
+    # terminal: a script, CI job or agent either blocks on a stdin that will never answer or
+    # dies without saying why -- and that failure reads as the gate refusing the change rather
+    # than asking a question nobody could hear. Three sources of an answer, in order:
+    #
+    #   1. PR_CONFIRM in the environment -- the explicit answer for a non-interactive caller.
+    #      It is NOT a bypass: setting it is the same act of confirmation as typing y, just
+    #      recorded where a script can supply it. Answer it AFTER checking each item.
+    #   2. An interactive stdin -- a human, prompted exactly as before.
+    #   3. Neither, so read whatever was piped in, bounded by a timeout. `echo y | just pr`
+    #      keeps working, and a stdin that never answers costs ten seconds rather than hanging.
+    #
+    # The failure names PR_CONFIRM, because a gate that cannot be satisfied from the context it
+    # failed in is a wall rather than a gate.
+    if [ -n "${PR_CONFIRM:-}" ]; then
+        CONFIRM="$PR_CONFIRM"
+        echo "All manual items confirmed? [y/N] $CONFIRM   (answered by PR_CONFIRM)"
+    elif [ -t 0 ]; then
+        echo -n "All manual items confirmed? [y/N] "
+        read -r CONFIRM
+    else
+        echo -n "All manual items confirmed? [y/N] "
+        read -r -t 10 CONFIRM || CONFIRM=""
+        echo "$CONFIRM"
+        [ -n "$CONFIRM" ] || { echo -e "${RED}Aborted.${NC} No terminal to confirm the checklist on, and nothing on stdin. Re-run with PR_CONFIRM=y once each item above is actually checked."; exit 1; }
+    fi
     [ "$CONFIRM" = "y" ] || [ "$CONFIRM" = "Y" ] \
         || { echo -e "${RED}Aborted.${NC} Complete the checklist first."; exit 1; }
 
@@ -395,6 +419,22 @@ open-pr *ARGS:
     #!/usr/bin/env bash
     set -euo pipefail
     just pr
+
+    # Push the branch if it has no upstream yet. Without this, on a never-pushed branch
+    # `gh pr create` has no remote branch to open a PR from and fails non-interactively --
+    # AFTER the gate has printed "Gate passed", which reads as the gate refusing a change it
+    # had just approved.
+    #
+    # Deliberately ONLY when there is no upstream. Pushing unconditionally would make this
+    # recipe silently publish existing commits on a branch that already has one -- a different
+    # and more surprising act than "put this branch where gh can see it".
+    if ! git rev-parse --abbrev-ref --symbolic-full-name '@{upstream}' >/dev/null 2>&1; then
+        BRANCH="$(git rev-parse --abbrev-ref HEAD)"
+        [ "$BRANCH" != HEAD ] || { echo "detached HEAD -- check out a branch first" >&2; exit 1; }
+        echo "no upstream for $BRANCH -- pushing it so gh has a remote branch to open from"
+        # pre-push runs `just check`, so this cannot publish a branch the gate would refuse.
+        git push -u origin "$BRANCH"
+    fi
     gh pr create "$@"
 
 # Generate a flamegraph for execution profiling (requires perf on Linux or dtrace on macOS)

NOTES.md

@@ -96,7 +96,34 @@ The `retch-sysinfo` crate can be used independently as a library for cross-platf
 
 ---
 
-## Current State (v0.6.20)
+## Current State (v0.6.21)
+- **v0.6.21 — two gates that could not be satisfied from the situation they failed in**
+  (tooling only; no runtime change, `retch-sysinfo` unchanged at `0.1.53`). Both were hit by hand
+  while landing v0.6.20, and both are fixes `rusticprofile` already had.
+  - **`just pr` ended in a bare `read`, so nothing but a human at a terminal could answer it.** A
+    script, CI job or agent either blocked on a stdin that would never answer or died without
+    saying why — and that failure reads as *the gate refusing the change*, not as a question
+    nobody could hear. It now takes its answer from `PR_CONFIRM`, from an interactive stdin, or
+    from piped input under a ten-second bound, and the failure message **names `PR_CONFIRM`**.
+    It is not a bypass: all four paths still require an explicit `y`, so this widens *who can
+    answer*, not *what counts as an answer* — and the checklist must still be answered after
+    each item is actually checked.
+  - **`just open-pr` did not push, so on a never-pushed branch it printed "Gate passed" and then
+    failed.** `gh pr create` had no remote branch to open from. The observable result was a
+    command that announced the gate passing and then exited non-zero, which reads as the gate
+    rejecting work it had just approved. It now pushes **only when there is no upstream** —
+    pushing unconditionally would silently publish existing commits on a branch that already has
+    one, a different and more surprising act. `pre-push` still runs `just check`, so this cannot
+    publish a branch the gate would refuse; the push is inside the gate, not around it.
+  - **Both were verified on this PR itself**, which is the only honest test for the second one:
+    the branch had no upstream when `open-pr` ran, and that condition cannot be reproduced after
+    the fact.
+  - *Why these existed at all:* they are `rusticprofile`'s `0.0.21` and `0.2.12`, which retch
+    never received — the same cross-repo staleness that left the nushell completion path wrong
+    here for months and that v0.6.20's `standard-check` now guards for the install family. The
+    `pr`/`open-pr` triad is **not** yet covered by that standard; `templates/justfile-common.just`
+    records it as out of scope, because these recipes legitimately differ per repo.
+  - `retch-cli` → 0.6.21. Patch bump.
 - **v0.6.20 — nushell completions went where Windows nushell never looks; the install helpers
   become a checked cross-repo standard** (tooling only; no runtime behavior change,
   `retch-sysinfo` unchanged at `0.1.53`).

docs/retch.1

@@ -1,4 +1,4 @@
-.TH "RETCH" "1" "August 2026" "retch 0.6.20" "System Information Fetcher"
+.TH "RETCH" "1" "August 2026" "retch 0.6.21" "System Information Fetcher"
 
 .SH "NAME"
 .PP