Guard the mesa driver swap on its end state

mesa-va-drivers-freeworld Provides mesa-va-drivers, so once the swap has run
once it becomes the only thing providing the name we ask dnf to remove. The
command then resolves to removing and installing the same package in a single
transaction and exits with "conflicting requests", which reported
install-packages as a failed stage on every run after the first even though
the machine was already in the desired state.

The failure also produced misleading output, because dnf explored removing
mesa-dri-drivers and surfaced unrelated complaints from steam, lutris and
tigervnc-server-common.

Check for the freeworld package with rpm -q and skip the swap when it is
already installed.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01TFG55mJBiH2tV38tXcNzJ9

.gitignore

@@ -12,6 +12,10 @@
 /config/dot/espanso/match/identity.yml
 # Ignore backups of old config files
 /config/old
+# MCP bearer tokens. The server list in user/agents/mcp/servers is tracked
+# and names which variable carries each token; the tokens themselves are
+# machine-local, because this repo is public.
+/user/agents/mcp/env
 # Ignore Wireguard config of course!
 /config/wg/**
 # Ignore Neovim lazy-lock file

README.md

@@ -250,7 +250,7 @@ docs/           Settings reference, and the design specs behind the work
 
 ## Tests
 
-189 of them, under `tests/`. `tests/contracts.manifest` classifies every
+190 of them, under `tests/`. `tests/contracts.manifest` classifies every
 contract by the capabilities it needs. Run the hermetic set, or grant a
 specific external capability when automation needs it:
 

config/dot/hypr/monitors.lua

@@ -244,10 +244,20 @@ local shipped_mode = "4500x3000@60"
 local shipped_scale = 1.5
 local shipped_transform = 0
 
--- 10-bit output. 4500x3000@60 at 10bpc is ~24 Gbps, right at the edge of DP 1.4
+-- 8-bit output. 4500x3000@60 at 10bpc is ~24 Gbps, right at the edge of DP 1.4
--- HBR3, so this relies on DSC. If the display fails to light up or falls back to
+-- HBR3 and reliant on DSC, and this panel's link is marginal: every modeset
--- a lower mode, drop this to 8 first.
+-- retrains it and blanks the screen. 8bpc keeps headroom on the link.
-local shipped_bitdepth = 10
+--
+-- Related: directScanoutPolicy is 0 in Panama settings (2026-09-13). With
+-- scanout on, a fullscreen game whose buffer depth differs from the desktop
+-- (games ship both 8- and 10-bit swapchains) makes Hyprland change the output
+-- format, and on amdgpu a format change is a full modeset. Compositing always
+-- keeps the format fixed, so the link never retrains mid-game.
+--
+-- vrrPolicy is also 0 there. VRR on this panel loses sync and blacks out
+-- (seen on GNOME in July 2026 and again here); a 60Hz panel gains little
+-- from it anyway.
+local shipped_bitdepth = 8
 
 -- "auto" = sRGB at 8bpc, wide gamut at 10bpc. Not HDR; see header.
 local shipped_cm = "auto"

config/dot/hypr/rules.lua

@@ -125,6 +125,16 @@ hl.window_rule({
     no_dim    = true,
 })
 
+-- Steam games never get true fullscreen. Everything Hyprland does only for a
+-- real fullscreen window (direct scanout, VRR, tearing, auto HDR) makes this
+-- panel's marginal DP link retrain and blank (2026-09-13). "1 2" keeps the
+-- window maximized internally while the game believes it is fullscreen, which
+-- is what borderless windowed looks like from the game's side.
+hl.window_rule({
+    match = { class = "^steam_app_\\d+$" },
+    fullscreen_state = "1 2",
+})
+
 -- Steam itself is a normal window, but its transient popups are a mess.
 hl.window_rule({
     match = { class = "^steam$", title = "^(Friends List|Steam Settings|Special Offer.*)$" },
@@ -269,16 +279,28 @@ hl.layer_rule({
     ignore_alpha = 0.2,
 })
 
--- Overview, capture and local screen-reading UI dim the desktop behind them.
+-- Overview and local screen-reading UI dim the desktop behind them.
 hl.layer_rule({
     name  = "qs-overlay",
-    match = { namespace = "^qs-(overview|capture|screen-intelligence)$" },
+    match = { namespace = "^qs-(overview|screen-intelligence)$" },
     blur         = true,
     ignore_alpha = 0.4,
     dim_around   = true,
     no_screen_share = true,
 })
 
+-- The screenshot picker paints an opaque frozen frame edge to edge, so blur and
+-- dim would be compositor work on pixels nobody sees. no_anim is load-bearing:
+-- services/Capture.qml unmaps the picker and runs grim 90ms later, and the
+-- 200ms layersOut fade would otherwise put the half-faded picker, dimming and
+-- all, into every screenshot.
+hl.layer_rule({
+    name  = "qs-capture",
+    match = { namespace = "^qs-capture$" },
+    no_anim         = true,
+    no_screen_share = true,
+})
+
 -- Notification toasts. Blurred like every other shell surface -- without this
 -- the cards are a near-transparent fill sitting directly on the wallpaper and
 -- read as washed out rather than as glass.

config/dot/quickshell/modules/capture/CaptureOverlay.qml

@@ -27,9 +27,13 @@ PanelWindow {
         right: true
     }
 
-    // Fullscreen overlays must not reserve space, or every window on the
+    // Ignore, and nothing else: reserve no space (or every window on the
-    // workspace gets resized as the picker opens and closes.
+    // workspace resizes as the picker opens and closes) and respect nobody's
-    exclusiveZone: 0
+    // (or the window starts under the bar, the full-output freeze frame is
+    // squeezed into a shorter box, and every selection lands one bar-height
+    // off in the real capture). Never pair this with exclusiveZone; see
+    // tests/quickshell/exclusion-idiom-contract.
+    exclusionMode: ExclusionMode.Ignore
 
     WlrLayershell.namespace: "qs-capture" // matched by a layerrule in hypr/rules.lua
     WlrLayershell.layer: WlrLayer.Overlay

config/dot/quickshell/modules/cheatsheet/Cheatsheet.qml

@@ -57,7 +57,6 @@ PanelWindow {
 
     anchors { top: true; bottom: true; left: true; right: true }
     color: "transparent"
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
 
     // The `^qs-popover` prefix rule in hypr/rules.lua blurs what is behind

config/dot/quickshell/modules/dock/Dock.qml

@@ -55,7 +55,6 @@ PanelWindow {
     color: "transparent"
 
     // A dock that reserved space would not be intellihiding.
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
 
     // Matched by the `qs-dock` layer rule in hypr/rules.lua — do not rename.

config/dot/quickshell/modules/dock/DockPickOverlay.qml

@@ -40,7 +40,6 @@ PanelWindow {
 
     anchors { top: true; bottom: true; left: true; right: true }
     color: "transparent"
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
 
     // Blurred by the `^qs-popover` rule in hypr/rules.lua; the scrim is painted

config/dot/quickshell/modules/notifications/VisualBell.qml

@@ -44,7 +44,6 @@ PanelWindow {
     // Reserve nothing and respect nothing: the glow is drawn over the whole
     // output including under the bar and the dock, which is what makes it
     // visible from wherever the eyes happen to be.
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
     color: "transparent"
 

config/dot/quickshell/modules/osd/Osd.qml

@@ -29,7 +29,6 @@ PanelWindow {
     screen: root.modelData
     anchors.bottom: true
     margins.bottom: Theme.dockIconSize + Theme.dockPadding * 2 + Theme.barGap * 3
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
     implicitWidth: root.desiredWidth
     implicitHeight: 64

config/dot/quickshell/modules/overview/Overview.qml

@@ -22,7 +22,6 @@ PanelWindow {
     }
 
     color: "transparent"
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
 
     // Matched by the `qs-overlay` layer rule in hypr/rules.lua — do not rename.

config/dot/quickshell/modules/quicksettings/BrightnessControl.qml

@@ -34,11 +34,23 @@ Item {
     visible: root.rowCount > 0
     implicitHeight: rows.implicitHeight
 
-    // Probing I2C takes on the order of a second, so it waits until the panel
+    // Probing I2C takes on the order of a second, so it waits until the quick
-    // is actually on screen rather than running at shell startup. Monitors do
+    // settings panel is actually on screen rather than running at shell
-    // not come and go, so once is enough.
+    // startup. Monitors do not come and go, so once is enough.
-    onVisibleChanged: if (visible && !Brightness.scanned) Brightness.refresh()
+    //
-    Component.onCompleted: if (root.visible && !Brightness.scanned) Brightness.refresh()
+    // The trigger is the panel opening, not this item becoming visible: on a
+    // machine with no backlight, `visible` stays false until a scan has found
+    // a DDC display, so keying the scan off it would never scan at all.
+    Connections {
+        target: ShellState
+        function onQuickSettingsOpenChanged(): void { root.scanIfOpen(); }
+    }
+    Component.onCompleted: root.scanIfOpen()
+
+    function scanIfOpen(): void {
+        if (ShellState.quickSettingsOpen && !Brightness.scanned)
+            Brightness.refresh();
+    }
 
     // `-m` is the machine-readable form: name,class,current,percent,max
     Process {

config/dot/quickshell/modules/settings/DisplayIdentify.qml

@@ -22,7 +22,6 @@ Variants {
         implicitWidth: 260
         implicitHeight: 172
         color: "transparent"
-        exclusiveZone: 0
         exclusionMode: ExclusionMode.Ignore
         mask: Region {}
 

config/dot/quickshell/modules/settings/README.md

@@ -383,6 +383,13 @@ Lua-configured Hyprland, prints the refusal to stdout, and exits 0. `eval` exits
 0 on syntax and runtime errors too. The only trustworthy signal that a write
 landed is reading the value back.
 
+**`exclusionMode: ExclusionMode.Ignore` goes alone.** Quickshell's `exclusiveZone`
+setter flips `exclusionMode` back to `Normal` as a side effect, so a window that
+declares both is at the mercy of which property the QML engine applies last.
+The 2026-09-14 Qt update changed that order and every full-screen overlay slid
+under the bar. `tests/quickshell/exclusion-idiom-contract` fails any file that
+pairs them.
+
 **The Settings window is tiled.** `implicitWidth` is a hint; the layout decides,
 and it ranges from a half-screen split to the full display. `SliderRow` stacks
 its control under the label below 520px. Test narrow.

config/dot/quickshell/modules/welcome/Welcome.qml

@@ -78,7 +78,6 @@ PanelWindow {
 
     anchors { top: true; bottom: true; left: true; right: true }
     color: "transparent"
-    exclusiveZone: 0
     exclusionMode: ExclusionMode.Ignore
 
     WlrLayershell.namespace: "qs-popover-welcome"

config/dot/quickshell/services/Capture.qml

@@ -341,7 +341,10 @@ esac'
         property bool record: false
         property bool intelligence: false
         // A handful of frames at 60Hz, enough for the compositor to recomposite
-        // the output without the overlay on it.
+        // the output without the overlay on it. Only enough because the
+        // qs-capture layer rule in hypr/rules.lua sets no_anim: with the
+        // default 200ms layersOut fade the half-faded picker would still be
+        // on screen when grim reads it.
         interval: 90
         onTriggered: {
             const r = commitDelay.rect;

install

@@ -343,7 +343,7 @@ fi
 if [[ "$PANAMA_ROLE" == server ]]; then
   STAGES=(install-packages link-dotfiles link-user setup-server link-server setup-identity)
 else
-  STAGES=(install-packages link-dotfiles link-skills link-user change-settings link-vicinae-scripts setup-identity install-hardware)
+  STAGES=(install-packages link-dotfiles link-skills link-user link-mcp change-settings link-vicinae-scripts setup-identity install-hardware)
 fi
 
 # The two an upgrade drops. Both exist only to act on interview answers, and

setup/scripts/install-packages

@@ -1405,8 +1405,17 @@ soft "the multimedia group update" \
 sync
 log "Swapping ffmpeg-free for ffmpeg"
 soft "the ffmpeg swap" sudo dnf swap -y 'ffmpeg-free' 'ffmpeg' --allowerasing
-log "Swapping mesa-va-drivers for mesa-va-drivers-freeworld"
+# mesa-va-drivers-freeworld Provides mesa-va-drivers, so once the swap has been
-soft "the mesa driver swap" sudo dnf swap -y mesa-va-drivers mesa-va-drivers-freeworld
+# done it is the only thing providing the name we ask dnf to remove. `dnf swap`
+# then resolves to removing and installing the same package in one transaction
+# and fails with "conflicting requests", which reports install-packages as a
+# failed stage on every subsequent run. Guard on the end state instead.
+if rpm -q mesa-va-drivers-freeworld >/dev/null 2>&1; then
+	log "mesa-va-drivers-freeworld is already installed; swap not needed"
+else
+	log "Swapping mesa-va-drivers for mesa-va-drivers-freeworld"
+	soft "the mesa driver swap" sudo dnf swap -y mesa-va-drivers mesa-va-drivers-freeworld
+fi
 log "Upgrading Multimedia group with optional packages"
 soft "the optional Multimedia upgrade" sudo dnf4 group upgrade -y --with-optional Multimedia
 log "Installing GStreamer plugins (bad, good, base)"

setup/scripts/link-mcp

@@ -0,0 +1,99 @@
+#!/usr/bin/env bash
+
+# MCP servers, registered with the agent runtimes on this machine.
+#
+# An MCP server is a URL plus a bearer token, and the token is the whole reason
+# this is a stage rather than a manifest line. Panama is a public repository, so
+# the tokens cannot live in it, and neither runtime keeps its server list in a
+# file that could be symlinked anyway: Codex writes them into config.toml beside
+# dozens of unrelated settings, and Claude Code into ~/.claude.json. There is no
+# file to point at, so this registers them through the runtime's own CLI.
+#
+# What is tracked is user/agents/mcp/servers, which names each server and which
+# variable carries its token. What is not tracked is user/agents/mcp/env, which
+# holds the tokens. A new machine gets the servers by dropping its own env file
+# beside the tracked one and re-running ./install.
+#
+# This is personal content, so it obeys the same interview decision link-user
+# does. A machine that never said yes registers nothing.
+#
+# Only Claude Code is handled. Codex stores its servers inside config.toml, and
+# rewriting a section of somebody's live TOML is a worse failure mode than
+# leaving two lines for them to paste once.
+
+set -euo pipefail
+
+log() { echo -e "\033[1;34m[INFO]\033[0m $*"; }
+warn() { echo -e "\033[1;33m[WARN]\033[0m $*" >&2; }
+
+PANAMA_PATH="${PANAMA_PATH:-$HOME/.local/share/Panama}"
+MCP_DIR="$PANAMA_PATH/user/agents/mcp"
+SERVERS="$MCP_DIR/servers"
+ENV_FILE="$MCP_DIR/env"
+STATE_DIR="${XDG_STATE_HOME:-$HOME/.local/state}/panama"
+DECISION="$STATE_DIR/user-content"
+
+[[ -r "$SERVERS" ]] || { log "No MCP server list; nothing to register."; exit 0; }
+
+# The same gate link-user uses, read the same way, so one answer governs all
+# personal content rather than two stages disagreeing about it.
+decision="$([[ -r "$DECISION" ]] && cat "$DECISION" || printf 'no')"
+if [[ "$decision" != "yes" ]]; then
+    log "Personal content is not enabled on this machine; no MCP servers registered."
+    exit 0
+fi
+
+if ! command -v claude >/dev/null 2>&1; then
+    log "Claude Code is not installed; nothing to register."
+    exit 0
+fi
+
+# Tokens are optional. A machine without the env file still registers any server
+# that needs no header, and says which ones it skipped rather than failing.
+if [[ -r "$ENV_FILE" ]]; then
+    set -a
+    # shellcheck source=/dev/null
+    . "$ENV_FILE"
+    set +a
+else
+    warn "No $ENV_FILE; servers needing a token will be skipped."
+fi
+
+registered=0
+skipped=0
+
+while read -r name transport url token_var _rest; do
+    case "${name:-}" in ''|'#'*) continue ;; esac
+    if [[ -z "${transport:-}" || -z "${url:-}" ]]; then
+        warn "Ignoring malformed row for '$name'."
+        continue
+    fi
+
+    header=()
+    if [[ -n "${token_var:-}" ]]; then
+        token="${!token_var:-}"
+        if [[ -z "$token" ]]; then
+            warn "Skipping $name: $token_var is not set in $ENV_FILE."
+            skipped=$((skipped + 1))
+            continue
+        fi
+        header=(-H "Authorization: $token")
+    fi
+
+    # Re-registering is how this stays idempotent across upgrades, and how a
+    # rotated token reaches the runtime. Removing first avoids the CLI refusing
+    # a name it already knows. Neither call may print the token, so both are
+    # quiet unless they fail.
+    claude mcp remove "$name" -s user >/dev/null 2>&1 || true
+    if claude mcp add --transport "$transport" "$name" "$url" "${header[@]}" \
+        -s user >/dev/null 2>&1; then
+        log "Registered $name."
+        registered=$((registered + 1))
+    else
+        warn "Could not register $name."
+        skipped=$((skipped + 1))
+    fi
+done <"$SERVERS"
+
+log "MCP servers: $registered registered, $skipped skipped."
+log "Claude Code loads them at start, so restart a running session to pick them up."

tests/contracts.manifest

@@ -82,6 +82,7 @@ live-desktop tests/quickshell/dock-position-contract
 # Reads Hyprland option descriptions from the live compositor to verify enum
 # mappings.
 live-compositor tests/quickshell/enum-hypr-map-contract
+hermetic tests/quickshell/exclusion-idiom-contract
 hermetic tests/quickshell/fingerprint-contract
 # Reads the host firewall state through the production firewall helper.
 live-host tests/quickshell/firewall-contract

tests/quickshell/exclusion-idiom-contract

@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+
+# A window that wants the whole output must ask for ExclusionMode.Ignore and
+# nothing else. Quickshell's exclusiveZone setter forces exclusionMode back to
+# Normal as a side effect, so a window declaring both
+#
+#     exclusiveZone: 0
+#     exclusionMode: ExclusionMode.Ignore
+#
+# ends up with whichever property the QML engine applied last, and that order
+# is not ours to control: the 2026-09-14 Qt/Quickshell update flipped it, every
+# full-screen overlay slid down under the bar, and the screenshot picker's
+# frozen frame stopped lining up with the screen it was a picture of. This pins
+# the order-independent idiom: Ignore alone, never paired with exclusiveZone.
+
+set -uo pipefail
+
+repo_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+qs="$repo_dir/config/dot/quickshell"
+
+fail() { printf 'exclusion idiom contract: %s\n' "$1" >&2; exit 1; }
+[[ -d "$qs/modules" ]] || fail "missing $qs/modules"
+
+python3 - "$qs" <<'PY'
+import re, sys, pathlib
+qs = pathlib.Path(sys.argv[1])
+problems = []
+ignore = re.compile(r'\bexclusionMode\s*:\s*ExclusionMode\.Ignore\b')
+zone = re.compile(r'^\s*(?:WlrLayershell\.)?exclusiveZone\s*:', re.M)
+checked = 0
+for path in sorted(qs.rglob('*.qml')):
+    text = path.read_text()
+    if not ignore.search(text):
+        continue
+    checked += 1
+    for m in zone.finditer(text):
+        line = text.count('\n', 0, m.start()) + 1
+        problems.append(f"{path.relative_to(qs)}:{line}: exclusiveZone set in a file that uses "
+                        "ExclusionMode.Ignore; drop it, Ignore already means -1")
+if checked == 0:
+    problems.append("no window uses ExclusionMode.Ignore; the overlays this pins are gone")
+if problems:
+    print('\n'.join(problems), file=sys.stderr)
+    sys.exit(1)
+PY

user/agents/AGENTS.md

@@ -62,6 +62,40 @@ I wanted to share some of my preferences here so we can be more aligned as we wo
 
 ---
 
+# The prose bar
+
+Anything a person other than me will read gets a pattern pass before it ships. MR and PR
+descriptions, ticket fields, deliverable documents, reports, anything handed to a
+teammate or a client. Invoke the `unslop` skill and apply its pattern detection.
+
+- Skip its "Adding soul" section. Opinions and deliberate mess are right for an essay and
+  wrong for a risk table or a status field. These want plain, factual, and specific.
+- No em dashes, en dashes, semicolons, or arrow glyphs in prose. This is stricter than
+  unslop's own punctuation pattern, and mine wins. Grep the file before calling it done.
+- Passive voice is the usual offender in technical writing, not puffery. "X was verified"
+  wants to be "I verified X", which is shorter and says who did it.
+- Short reference cells still count. Three sentences wrapped around a filename is prose.
+- Never rewrite generated output. If a tool produced a block, an audit verdict or a test
+  report, regenerate it rather than editing its words.
+- Write it as me, in first person. Never use my name or a stand-in for it. Not "Gib
+  decided", not "the author", not "per the developer". If a sentence needs an actor, it
+  is "I".
+- Never flag one of my own decisions as a deviation. Don't write that something was out
+  of scope, that an AC or dev review said otherwise, that an earlier story should have
+  caught it, or that a direction was reversed. I write the tickets, so I am allowed to
+  change them, and a dev review is a proposal rather than a contract. Calling it out
+  reads as my own AI tattling on me. It waves a red flag at a nothing burger, costs the
+  PM time investigating a non-issue, and leaves them thinking I did something wrong.
+  State what the change is and why it belongs where it is, then stop.
+- If a decision feels big enough that you want to narrate it, ask me about it before
+  writing rather than annotating it in the document. Being aligned first is the point,
+  and the writing should read as though we always were.
+
+Files only I read are exempt: plans, scratch notes, working records. Those can name me
+and record who decided what, because that history is useful.
+
+---
+
 # Merge Requests
 
 - Make sure titles follow conventions from the repo.
@@ -84,5 +118,8 @@ I wanted to share some of my preferences here so we can be more aligned as we wo
   - Henry reviews all of my code & up until recently, would also write all the dev reviews for all the stories I completed as well. Nowadays, I write them myself & he reviews those too.
   - Henry likes work to be very considerate & he always prefers solutions that result in 0 downtime.
     - Despite the fact that Command Center does not have many users & the impact of it being down for a few minutes is small, Henry still leans on the side of solutions to problems that don't result in prod being down ever, even for just a few minutes during the build process. So our solutions should always keep that in mind. Any code that isn't considering everything & could result in a bug will probably be flagged by him, so its worthwhile to do right the first time!
+- Josh - Tech Lead for the APSCA project (apsca_next)
+  - On APSCA, Josh is my lead instead of Henry. He reviews my code and writes the dev reviews, spikes, and proposed schemas on the APSCA stories.
+  - His dev reviews are detailed and often include draft code, schema, and state machines. Treat them as the starting point for a story, not as a finished spec.
 - Hunter Southworth - Engineering Manager / Senior Developer
 - Kelson - Owner of Ksense

user/agents/mcp/servers

@@ -0,0 +1,21 @@
+# MCP servers registered with the local agent runtimes.
+#
+#   <name>  <transport>  <url>  <token variable>
+#
+# The token variable names an entry in user/agents/mcp/env, which is ignored by
+# git. This file carries no secret, so it is tracked like the rest of user/ and
+# a new machine gets the same servers by dropping its own env file beside it.
+#
+# A row with no token variable registers without an Authorization header.
+#
+# Lines beginning with # are comments and blank lines are ignored.
+
+# The NanoKVM boards. These give an agent full console control of each machine,
+# including the parts SSH cannot reach: firmware, a locked screen, and the GUI
+# prompts Xcode and the keychain raise during a signed iOS build.
+nanokvm-fedora  http  https://nano.gbrown.org/api/mcp     NANOKVM_FEDORA_TOKEN
+nanokvm-mac     http  https://macnano.gbrown.org/api/mcp  NANOKVM_MAC_TOKEN
+
+# Bill Tracker's hosted MCP server. No token: it is an OAuth server of its own
+# and Claude Code signs in through Gib's Auth on first use.
+bill-tracker    http  https://mcp.billtracker.gbrown.org/mcp

user/agents/skills/ticket/SKILL.md

@@ -108,6 +108,63 @@ rerun it, and verify the device log instead of accepting the fallback transcript
 - Don't check a checklist box (in `mr.md` or anywhere else) unless you actually verified
   it. Leave it unchecked and say why in the notes rather than guessing.
 
+## The prose bar (anything another person reads)
+
+Everything this skill produces for someone else gets a pattern pass before it ships:
+`mr.md`'s Summary and Additional Notes, every Jira field including the proof and
+mitigation cells, and every deliverable document. Call the Skill tool with "unslop" and
+apply its pattern detection.
+
+- **Skip unslop's "Adding soul" section.** It asks for opinions, varied rhythm, and some
+  mess. That is right for an essay and wrong for a mitigation cell Henry is auditing.
+  These want plain, factual, and specific.
+- **Never edit generated output.** The pre-MR handoff pasted into `mr.md` is the audit's
+  own words. If it reads badly, regenerate it, don't rewrite it.
+- **Proof cells are prose too.** A cell that wraps three sentences around a filename is
+  writing, not a citation, and it picks up passive voice and jargon like any other
+  paragraph. Real examples caught this way: "affordance" for button, "vacuously" for
+  "for the wrong reason", "the test was confirmed able to fail" for "I emptied the
+  legacy render and confirmed the test fails".
+- **Passive voice is the usual offender here,** not puffery. Technical writing drifts
+  into "X was verified" when "I verified X" is shorter and says who did it.
+
+### Write as Gabriel, and never flag his own decisions
+
+Everything that ships under his name is written **as him**, in first person. Never write
+his name or any stand-in for it. Not "Gib decided", not "the author", not "per the
+developer". When a sentence needs an actor, it is "I".
+
+**Never frame one of his decisions as a deviation.** Do not write that something "was not
+in this story's scope", that an AC or dev review said otherwise, that an earlier story
+"should have caught it", or that a direction was "reversed". State what the change is and
+why it belongs where it is, then stop.
+
+He writes the tickets himself, so he is entitled to change them, and a dev review is a
+proposal rather than a contract. Calling the change out reads as his own AI tattling on
+him. It waves a red flag at a nothing burger, costs the PM time investigating a
+non-issue, and leaves them assuming he did something wrong. The real story is normally
+mundane and good: the better component already existed and cost almost nothing to use.
+
+Compare, on KACP-23058:
+
+> Bad: "This was not in this story's scope. KACP-23059 shipped that page still on the
+> legacy tables, and Gib reversed the reuse-legacy direction on 2026-09-02."
+>
+> Good: "Both detail pages now share one set of grid components instead of two divergent
+> implementations, so the move ships with the components it shares."
+
+Same information, no flag attached. **If a decision feels big enough that you want to
+narrate it, that is the signal to ask him about it before writing, not to annotate it in
+the document.** Being aligned first is the point, and the writing should read as though
+you always were.
+
+This governs `mr.md`, `customfield_10260` and `10261`, every other Jira field, and PM
+deliverables. It does not govern files only he reads. `plan.md`, `bugs.md`, audits and
+working notes can name him and record who decided what, because that history is useful.
+
+The House style punctuation rule above is the stricter one where the two overlap. Keep
+it. Personal working files (`plan.md`, `ticket.md`) are exempt, nobody else reads them.
+
 ## Verification tools available
 
 The goal isn't "a plausible-sounding plan" or "code that compiles" — it's a plan and
@@ -226,15 +283,19 @@ method for producing accurate screenshots, not part of the deliverable's story.
 Process, per deliverable:
 
 1. Write `<slug>.md`.
-2. Write `<slug>.typ` from it, per the styling notes above.
+2. Run the prose bar over `<slug>.md` before building anything from it. A spike
-3. Compile: `typst compile <slug>.typ <slug>.pdf`. `typst` should already be on PATH;
+   deliverable is the most prose-heavy thing this skill produces and it goes straight to
+   Conrad, who is skeptical of AI. Fixing it here means fixing it once, rather than in
+   the markdown and again in the Typst.
+3. Write `<slug>.typ` from it, per the styling notes above.
+4. Compile: `typst compile <slug>.typ <slug>.pdf`. `typst` should already be on PATH;
    if it isn't, tell the user rather than silently skipping the PDF.
-4. Actually look at the compiled PDF using the Read tool (it reads PDFs directly, page
+5. Actually look at the compiled PDF using the Read tool (it reads PDFs directly, page
    by page for longer documents). Check every page for real layout problems: text or
    a table overflowing a page, an awkward page break splitting a table or diagram,
    cramped or excessive spacing, a diagram that rendered wrong. A successful compile
    only means valid Typst, not that it looks right — actually look.
-5. If anything looks wrong, fix `<slug>.typ` and go back to step 3. Repeat until the
+6. If anything looks wrong, fix `<slug>.typ` and go back to step 4. Repeat until the
    PDF genuinely looks right, not just until it compiles without erroring.
 
 If `.claude/docs/` is tracked by git in this repo (check with
@@ -690,6 +751,35 @@ Entered only when the user has confirmed (per Phase 0) that the plan is approved
         write `mr.md` before that verdict is reached.
    Findings that need no code change (mentions, disclosures) get addressed in
    `mr.md`'s Additional Notes or the Jira fields as usual.
+
+   **The verdict is the gate and it is not negotiable.** Ready to Open MR is what
+   goes on record with the Lead, so never reason your way past it because the
+   remaining items look harmless. That judgment is not yours to make.
+
+   **Don't confuse the verdict with the handoff's recommendation.** They are
+   different fields in different files for different readers. The audit's
+   `Verdict` says whether the branch is ready to become an MR, and that is the
+   gate. The handoff's `Recommendation` (`Merge`, `Merge after conditions`) is
+   advice to the maintainer about merging it, and `Merge after conditions` is
+   normal and expected at a Ready verdict. Never treat the recommendation as a
+   blocker or report it as one.
+
+   **Reruns converge by giving the audit more, not by repeating the command.**
+   When a rerun is needed, the user is running it, so every rerun costs them.
+   Make each one count:
+   - **Feed it context on the first run.** It reviews the diff, and left alone it
+     re-derives intent from the branch name and re-raises decisions that are
+     already settled. Ask the user to point it at the ticket's
+     `resources/ticket.md` and `plan.md`. This is the cheapest thing that reduces
+     repeat findings.
+   - **Answer its decisions explicitly in the rerun.** When it parks a finding
+     under "Decide before opening", write the decision and its reasoning into the
+     prompt for the next run, so it can reclassify rather than re-ask. A silent
+     rerun on an unchanged diff re-asks the same question.
+   - **Say so when it finds new things in unchanged code.** That means the earlier
+     run was an incomplete pass, not that the branch got worse. Tell the user
+     plainly, because it is the audit's problem to fix and it changes how much
+     weight the next run deserves.
 8. Read `~/.agents/skills/ticket/templates/mr.md` — this is the org's MR template,
    copied into this skill so it still works even though the original
    `.claude/docs/mr/template.md` no longer exists in the command-center repo. `mr.md`
@@ -714,7 +804,9 @@ Entered only when the user has confirmed (per Phase 0) that the plan is approved
      verified" if you actually ran a build and it passed).
    - Replace `<!-- Paste the Pre-MR handoff here -->` with the exact content of the
      latest `pre-mr-review` handoff file (the one with verdict Ready to Open MR) —
-     paste it verbatim, don't summarize it.
+     paste it verbatim, don't summarize it. Its `Recommendation` line is the
+     audit's advice to the maintainer and is left exactly as written, including
+     `Merge after conditions`.
 
      The handoff section of `mr.md` is GENERATED OUTPUT, not something you author.
      It is the block that states whether the branch is ready to merge, so it must be
@@ -756,13 +848,10 @@ Entered only when the user has confirmed (per Phase 0) that the plan is approved
    content. A reviewer should be able to read the whole file in about a minute; if it
    has grown past roughly 120 lines including the handoff, it has drifted.
 
-   Before saving, call the Skill tool with "unslop" and apply its pattern detection to
+   Before saving, run the prose bar (see **The prose bar** above) over the Summary and
-   the Summary and Additional Notes ONLY: puffery, superficial -ing phrases, "not just
+   Additional Notes ONLY. The pasted handoff is generated output and is never edited.
-   X but Y", vague attributions, rule-of-three padding. The pasted handoff is generated
+   Henry reads every one of these, and prose that reads as machine-generated costs the
-   output and is never edited, per the rule above. Skip unslop's "Adding soul" section
+   MR its credibility before anyone looks at the diff.
-   too — an MR description wants plain and factual, not voice. Henry reads every one of
-   these, and prose that reads as machine-generated costs the MR its credibility before
-   anyone looks at the diff.
 9. Fill the Jira ticket fields directly, by issue type. Rich text fields are ADF:
    render markdown with `python3 ~/.agents/skills/review-ticket/scripts/review2adf.py
    render <file.md>` and PUT via `{"fields": {...}}` to `/rest/api/3/issue/<KEY>`.
@@ -771,11 +860,9 @@ Entered only when the user has confirmed (per Phase 0) that the plan is approved
    field-to-type map below is verified against the project's edit screens, don't PUT
    a field to a type that doesn't carry it.
 
-   **Same prose bar as `mr.md`.** Every field authored here is read by the Lead and the
+   **Run the prose bar** (see **The prose bar** above) over every field before rendering
-   PM. Run the `unslop` skill's pattern detection over the markdown before rendering it
+   it to ADF, proof and mitigation cells included. Everything here is read by the Lead
-   to ADF, skipping its "Adding soul" section — Jira fields want plain, factual, and
+   and the PM.
-   specific. This does not apply to proof cells, which are references to artifacts and
-   test names rather than prose.
 
    **Proof first.** Before filling any proof column, capture working feature proof
    yourself wherever possible: run the app (`run` skill) and screenshot the real
@@ -835,13 +922,12 @@ Entered only when the user has confirmed (per Phase 0) that the plan is approved
      handled and what covers it. Deliver the number asked for, not more, and pick the
      ones a reviewer most benefits from. This field, not `mr.md`.
    - `customfield_10261` User Story Patch Notes: the same change in user-facing words.
-   - `customfield_10142` Working Feature Proof: ALWAYS fill this on a Story, never
+   - `customfield_10142` Working Feature Proof: leave it alone. Write nothing here.
-     leave it empty. It renders as its own panel in the ticket's testing section, so
+     This field is where Gib drops the proof attachments, and dragging files into it
-     an empty field reads as unfilled testing even when every test-table cell is
+     replaces whatever text is sitting there, so a written list is destroyed by the
-     complete (this happened on KACP-23143). At minimum it lists the files staged in
+     upload it was describing. Naming each artifact is already the job of the Test
-     `proof/` for Gib to attach, one line each saying what the artifact shows, plus a
+     Cases proof cells, which reference it by filename. Step 10 tells Gib which files
-     pointer that per-row proof lives in the Test Cases table. A headline artifact (a
+     to upload. Do not PUT this field on a Story.
-     demo capture, a before/after pair) leads the list when one exists.
 
    **Spike** (deliverables, not code): no test or proof fields to fill. The output is
    the deliverables directory, and step 10 tells Gib which files to upload where (the