fix: v0.9.1 robustness cleanups

- run_resume falls back to a new conversation only when the transcript file
  is gone; other failures end at the menu with the recorded id intact.
- The SessionStart hook resolves the session name under the registry lock;
  rename holds it across tmux rename-session and the registry move
  (dd_registry_locked, nestable).
- dd_load_env parses KEY=value like systemd instead of sourcing the files.
- devdesk-deploy: DPkg::Lock::Timeout, explicit FAILED lines on stdout and in
  the log; the hook answers synchronously and CI checks the HTTP status.
  logrotate for the deploy log.
- postinst drops duplicate terminal-features entries; hooks sweep stale
  temp files.

Closes #12

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

libexec/devdesk-claude-run

@@ -42,12 +42,21 @@ run_new() {
 }
 
 run_resume() {
-    local id args started rc fork=()
+    local id args fork=()
     id=$(dd_registry_get "$NAME" claude_session_id)
     if [ -z "$id" ]; then
         run_new
         return
     fi
+    # Start fresh only when the conversation's file is really gone. Any other
+    # failure (an early Ctrl+C, a network error) leaves the recorded id alone
+    # and ends at the menu, where ⏎ tries again (devdesk#12).
+    if [ -z "$(dd_convo_file "$id")" ]; then
+        printf "%sConversation %s is no longer in Claude's history — starting a new one.%s\n" "$MUTED" "$id" "$RESET"
+        sleep 1
+        run_new
+        return
+    fi
     mapfile -d '' args < <(common_args)
     # "Fork" from conversation search (devdesk#24): continue as a new
     # conversation, once; the SessionStart hook then records the new id.
@@ -56,17 +65,7 @@ run_resume() {
         # shellcheck disable=SC2016 # $n is a jq variable
         dd_registry_update 'del(.sessions[$n].fork)' --arg n "$NAME" 2>/dev/null || true
     fi
-    started=$SECONDS
     "$CLAUDE_BIN" --resume "$id" "${fork[@]}" "${args[@]}"
-    rc=$?
-    # A resume that fails straight away means the transcript is gone; start fresh.
-    if [ "$rc" -ne 0 ] && [ $((SECONDS - started)) -lt 10 ]; then
-        printf '%sCould not resume conversation %s — starting a new one.%s\n' "$MUTED" "$id" "$RESET"
-        sleep 1
-        run_new
-        return
-    fi
-    return "$rc"
 }
 
 menu() {

libexec/devdesk-deploy

@@ -3,6 +3,9 @@
 # listener (hook "devdesk-deploy") when CI publishes a v* tag; the tag's ref is
 # passed as $1. Safe for running sessions: package upgrades restart only
 # devdesk.service (ttyd), never devdesk-tmux.service where Claude lives.
+#
+# The hook runs this synchronously and returns its output, with HTTP 500 when
+# it fails, so CI's "Deploy to CT 126" job shows the real result (devdesk#12).
 set -euo pipefail
 
 LOG=${DEVDESK_DEPLOY_LOG:-/var/log/devdesk-deploy.log}
@@ -12,11 +15,28 @@ log() {
     printf '%s %s\n' "$(date '+%F %T')" "$*" >>"$LOG"
 }
 
-exec 9>/run/devdesk-deploy.lock
-flock -w 900 9 || {
-    log "another deploy is still running; giving up"
+# The outcome goes to the log and to stdout (the webhook response).
+say() {
+    log "$*"
+    printf '%s\n' "$*"
+}
+fail() {
+    say "deploy FAILED: $*"
     exit 1
 }
+on_error() {
+    local rc=$?
+    say "deploy FAILED: line $1 exited $rc (details in $LOG)"
+    exit "$rc"
+}
+trap 'on_error $LINENO' ERR
+
+# Wait for dpkg/apt rather than fail when unattended-upgrades or a manual apt
+# run holds the lock.
+APT_OPTS=(-o DPkg::Lock::Timeout=120 -o Dpkg::Use-Pty=0)
+
+exec 9>/run/devdesk-deploy.lock
+flock -w 900 9 || fail "another deploy is still running"
 
 want=${1:-}
 want=${want#refs/tags/}
@@ -25,7 +45,7 @@ before=$(dpkg-query -W -f='${Version}' devdesk 2>/dev/null || echo none)
 log "=== deploy: installed $before, release ${want:-latest}"
 
 refresh() {
-    apt-get update -q -o Dir::Etc::sourcelist="$SOURCES" -o Dir::Etc::sourceparts=- \
+    apt-get update -q "${APT_OPTS[@]}" -o Dir::Etc::sourcelist="$SOURCES" -o Dir::Etc::sourceparts=- \
         -o APT::Get::List-Cleanup=0 >>"$LOG" 2>&1
 }
 
@@ -41,14 +61,17 @@ for _ in $(seq 1 18); do
 done
 
 if [ -n "$want" ] && [ "$candidate" != "$want" ]; then
-    log "registry still offers $candidate, not $want; giving up"
-    exit 1
+    fail "the registry still offers ${candidate:-nothing}, not $want"
 fi
 if [ "$candidate" = "$before" ]; then
-    log "already at $before; nothing to do"
+    say "already at $before; nothing to do"
     exit 0
 fi
 
-DEBIAN_FRONTEND=noninteractive apt-get install -y -q --only-upgrade \
+DEBIAN_FRONTEND=noninteractive apt-get install -y -q --only-upgrade "${APT_OPTS[@]}" \
     -o Dpkg::Options::=--force-confold devdesk </dev/null >>"$LOG" 2>&1
-log "devdesk $before -> $(dpkg-query -W -f='${Version}' devdesk)"
+after=$(dpkg-query -W -f='${Version}' devdesk)
+if [ -n "$want" ] && [ "$after" != "$want" ]; then
+    fail "installed $after, not $want"
+fi
+say "devdesk $before -> $after"

libexec/devdesk-history.sh

@@ -2,15 +2,6 @@
 # Past conversations (devdesk#24) and client presence (devdesk#26), sourced by
 # bin/devdesk and reached through `devdesk attach --search|--resume-convo|--presence`.
 
-dd_projects_dir() {
-    printf '%s/projects\n' "${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
-}
-
-dd_convo_file() {
-    [[ $1 =~ ^[0-9a-f-]{36}$ ]] || return 1
-    compgen -G "$(dd_projects_dir)/*/$1.jsonl" | head -n1
-}
-
 # --search QUERY: full-text search over every Claude conversation on this host
 # (DevDesk and code-server alike). Case-insensitive, plain text. Newest first;
 # per conversation: title, folder, dates and up to 3 snippets (what you asked,

libexec/devdesk-inbox.sh

@@ -50,8 +50,7 @@ cmd_answer() {
 dd_transcript() {
     local id
     id=$(dd_registry_get "$1" claude_session_id)
-    [[ $id =~ ^[0-9a-f-]{36}$ ]] || return 1
-    compgen -G "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/projects/*/$id.jsonl" | head -n1
+    dd_convo_file "$id"
 }
 
 # --peek SESSION: what the session is doing and the last exchange, from the end

libexec/devdesk-lib.sh

@@ -8,13 +8,34 @@ DEVDESK_ENV=${DEVDESK_ENV:-/etc/devdesk/env}
 # Local overrides go in env.local: it is never shipped, so upgrades don't prompt about it.
 DEVDESK_ENV_LOCAL=${DEVDESK_ENV_LOCAL:-${DEVDESK_ENV}.local}
 
+# The env files are read by systemd (EnvironmentFile=) as well, so they are
+# parsed here the way systemd parses them, never sourced: KEY=value lines, the
+# value taken literally (no $ expansion, no command substitution), optionally
+# in matching quotes. Anything else is reported and skipped (devdesk#12).
 dd_load_env() {
-    local f
-    for f in "$DEVDESK_ENV" "$DEVDESK_ENV_LOCAL"; do
-        if [ -r "$f" ]; then
-            # shellcheck source=/dev/null
-            . "$f"
-        fi
+    local _dd_f _dd_line _dd_n _dd_key _dd_val
+    for _dd_f in "$DEVDESK_ENV" "$DEVDESK_ENV_LOCAL"; do
+        [ -r "$_dd_f" ] || continue
+        _dd_n=0
+        while IFS= read -r _dd_line || [ -n "$_dd_line" ]; do
+            _dd_n=$((_dd_n + 1))
+            _dd_line=${_dd_line%$'\r'}
+            [[ $_dd_line =~ ^[[:space:]]*([#\;].*)?$ ]] && continue
+            if ! [[ $_dd_line =~ ^[[:space:]]*([A-Z_][A-Z0-9_]*)=(.*)$ ]]; then
+                printf 'devdesk: %s:%s: not a KEY=value line, ignored\n' "$_dd_f" "$_dd_n" >&2
+                continue
+            fi
+            _dd_key=${BASH_REMATCH[1]}
+            _dd_val=${BASH_REMATCH[2]}
+            _dd_val=${_dd_val%"${_dd_val##*[![:space:]]}"} # trailing blanks, like systemd
+            if [[ $_dd_val =~ ^\"(.*)\"$ || $_dd_val =~ ^\'(.*)\'$ ]]; then
+                _dd_val=${BASH_REMATCH[1]}
+            elif [[ $_dd_val == [\"\']* ]]; then
+                printf 'devdesk: %s:%s: unmatched quote, ignored\n' "$_dd_f" "$_dd_n" >&2
+                continue
+            fi
+            printf -v "$_dd_key" '%s' "$_dd_val"
+        done <"$_dd_f"
     done
 }
 dd_load_env
@@ -101,26 +122,59 @@ dd_registry_get() {
     jq -Mr --arg n "$1" --arg f "$2" '.sessions[$n][$f] // empty' "$DEVDESK_REGISTRY" 2>/dev/null
 }
 
-# dd_registry_update JQ_FILTER [jq args...] — read-modify-write under flock.
-dd_registry_update() {
-    local filter=$1
-    shift
+# dd_registry_locked CMD [ARGS...] — run CMD holding the registry lock, so a
+# step that reads tmux and then writes the registry (the SessionStart hook,
+# rename) can't interleave with another. Nests: an update inside reuses it.
+dd_registry_locked() {
+    if [ -n "${DD_REGISTRY_LOCKED:-}" ]; then
+        "$@"
+        return
+    fi
     mkdir -p "$DEVDESK_STATE_DIR"
     (
         flock -w 5 9 || exit 1
-        [ -s "$DEVDESK_REGISTRY" ] || printf '{"sessions":{}}\n' >"$DEVDESK_REGISTRY"
-        local tmp
-        tmp=$(mktemp "$DEVDESK_REGISTRY.XXXXXX") || exit 1
-        if jq "$@" "$filter" "$DEVDESK_REGISTRY" >"$tmp"; then
-            chmod 600 "$tmp"
-            mv -f "$tmp" "$DEVDESK_REGISTRY"
-        else
-            rm -f "$tmp"
-            exit 1
-        fi
+        DD_REGISTRY_LOCKED=1
+        "$@"
     ) 9>"$DEVDESK_REGISTRY.lock"
 }
 
+# dd_registry_update JQ_FILTER [jq args...] — read-modify-write under flock.
+dd_registry_update() {
+    dd_registry_locked _dd_registry_apply "$@"
+}
+
+_dd_registry_apply() {
+    local filter=$1 tmp
+    shift
+    [ -s "$DEVDESK_REGISTRY" ] || printf '{"sessions":{}}\n' >"$DEVDESK_REGISTRY"
+    tmp=$(mktemp "$DEVDESK_REGISTRY.XXXXXX") || return 1
+    if jq "$@" "$filter" "$DEVDESK_REGISTRY" >"$tmp"; then
+        chmod 600 "$tmp"
+        mv -f "$tmp" "$DEVDESK_REGISTRY"
+    else
+        rm -f "$tmp"
+        return 1
+    fi
+}
+
+# Claude's conversation files: ~/.claude/projects/<folder>/<id>.jsonl.
+dd_projects_dir() {
+    printf '%s/projects\n' "${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
+}
+
+dd_convo_file() {
+    [[ $1 =~ ^[0-9a-f-]{36}$ ]] || return 1
+    compgen -G "$(dd_projects_dir)/*/$1.jsonl" | head -n1
+}
+
+# Temp files of writers that were killed mid-write (hooks run under Claude's
+# timeouts): .<session>.XXXXXX and registry.json.XXXXXX, older than 10 minutes.
+dd_sweep_tmp() {
+    find "$DEVDESK_CONTEXT_DIR" "$DEVDESK_STATUS_DIR" -maxdepth 1 -type f -name '.*.??????' -mmin +10 -delete 2>/dev/null
+    find "$DEVDESK_STATE_DIR" -maxdepth 1 -type f -name 'registry.json.??????' -mmin +10 -delete 2>/dev/null
+    return 0
+}
+
 dd_registry_set_dir() {
     dd_registry_update '.sessions[$n] = ((.sessions[$n] // {}) + {dir: $d, updated: $t})' \
         --arg n "$1" --arg d "$2" --arg t "$(date -Is)"

libexec/devdesk-session-hook

@@ -7,14 +7,20 @@
 # shellcheck source=libexec/devdesk-lib.sh
 . "${DEVDESK_LIBEXEC:-/usr/lib/devdesk}/devdesk-lib.sh" || exit 0
 
-session=$(dd_current_session)
-[ -n "$session" ] || exit 0
-dd_valid_name "$session" || exit 0
-
 input=$(cat)
 id=$(jq -r '.session_id // empty' <<<"$input" 2>/dev/null)
 source=$(jq -r '.source // empty' <<<"$input" 2>/dev/null)
 [ -n "$id" ] || exit 0
 
-dd_registry_set_claude "$session" "$id" "$source" 2>/dev/null
+# The session's name is read under the registry lock: a rename holds it too,
+# so the id can't land under the old name after the entry moved.
+# shellcheck disable=SC2317 # called through dd_registry_locked
+record() {
+    local session
+    session=$(dd_current_session)
+    if [ -z "$session" ] || ! dd_valid_name "$session"; then return 0; fi
+    dd_registry_set_claude "$session" "$id" "$source"
+}
+dd_registry_locked record 2>/dev/null
+dd_sweep_tmp
 exit 0