diff --git a/DOCS-reload.md b/DOCS-reload.md
new file mode 100644
index 0000000..811f1a7
--- /dev/null
+++ b/DOCS-reload.md
@@ -0,0 +1,59 @@
+# Picking up a change without a restart
+
+Two kinds of change reach the running bot, and they are handled differently.
+
+## Data: `POST /_reload`
+
+Aliases, an album's `people.json`/`tags.json`, and every setting a page in the
+management UI writes are *data*. The command table is built from that data by
+one function (`signal_bot/wiring.py`), and `RegistryHolder` can run it again:
+the new table is built whole, then swapped in with a single assignment. Nothing
+restarts, the websocket stays up, and no frame is lost or replayed.
+
+```sh
+scripts/reload.sh                 # rebuild the table, report how many commands are live
+scripts/reload.sh '!nowaosoba'    # ...and then type that at the bot in testbot
+```
+
+The answer is what the bot is serving **now**:
+
+| status     | code | meaning                                                            |
+|------------|------|--------------------------------------------------------------------|
+| `reloaded` | 200  | the new table is live; `commands` is its size                      |
+| `kept`     | 409  | the data no longer builds; the previous table is still answering, and `reason` says what was wrong |
+| `connecting` | 503 | the bot has not reached Signal yet; there is no table to rebuild   |
+
+A rejected reload is deliberate: the builder reads files people edit by hand
+and pages write mid-session, and the only safe answer to a file that does not
+load is the table that was already answering. Only bad *data* is survived
+(`ConfigError`, `DuplicateTriggerError`); a programming error still surfaces
+as a 500 and a traceback in the journal, because a bug is not a file.
+
+Startup is the one time a bad table is fatal: there is nothing to fall back
+on, and `Restart=always` looping on a `ConfigError` is a clearer signal than a
+bot that answers nothing.
+
+## Code: `scripts/reload.sh --code`
+
+A Python change needs the process replaced. The restart is about a second and
+lossless (signal-cli replays what arrived meanwhile), but it is a deploy step,
+so it is the flag rather than the default:
+
+```sh
+scripts/reload.sh --code '!pomoc'
+```
+
+restarts the unit, waits for `/healthz` to report `ok` again, prints the
+version it came back as, and -- with a command given -- types it in the test
+group so "restarted" means "answered". There is no hot-reload of Python on
+purpose: `importlib.reload` leaves every object the old module already
+created bound to the old classes, which is a bug that looks like a feature
+until the first `isinstance` check.
+
+## Proven where
+
+`scripts/e2e.sh` brackets a `scripts/reload.sh` with the two scenarios in
+`fake_signal/proofs/`: `!delta` is unknown, a fifth person is tagged into the
+synthetic pack, the script runs, and the same process posts delta's photo.
+The route itself is covered in `tests/test_app.py`, the holder in
+`tests/test_registry_holder.py`, and the album case in `tests/test_bot.py`.
diff --git a/README.md b/README.md
index f45dcbb..1e2f9dd 100644
--- a/README.md
+++ b/README.md
@@ -227,9 +227,9 @@ things: `!staty` `!wczoraj` `!tydzień` `!miesiąc` `!rok` `!rekordy` `!korona`
 
 ## Managing it
 
-`./web.sh` serves one page on `127.0.0.1:8140`: health, the configuration
-(editable) and the album's tagger. Started by hand and stopped again, because
-it can rewrite `.env`: [DOCS-web-ui.md](DOCS-web-ui.md).
+`./web.sh` serves one page on `127.0.0.1:8140`: health, the configuration (editable) and the
+album's tagger. Started by hand and stopped again, because it can rewrite `.env`: [DOCS-web-ui.md](DOCS-web-ui.md).
+A tagged person goes live with `scripts/reload.sh`, no restart; code with `--code`: [DOCS-reload.md](DOCS-reload.md).
 
 ## Develop
 
diff --git a/fake_signal/proofs/reload_after.yaml b/fake_signal/proofs/reload_after.yaml
new file mode 100644
index 0000000..4bb858a
--- /dev/null
+++ b/fake_signal/proofs/reload_after.yaml
@@ -0,0 +1,7 @@
+# The other half of the bracket: the same process, one POST /_reload later,
+# and the person tagged into the album meanwhile answers with a photo.
+name: "!delta posts a photo after the reload"
+inject:
+  - { text: "!delta", uuid: uuid-reload-after, name: ReloadAfter }
+expect:
+  - { kind: image, to: reply }
diff --git a/fake_signal/proofs/reload_before.yaml b/fake_signal/proofs/reload_before.yaml
new file mode 100644
index 0000000..263dc2e
--- /dev/null
+++ b/fake_signal/proofs/reload_before.yaml
@@ -0,0 +1,8 @@
+# Not in scenarios/: that glob replays against the pack as first written,
+# and this pair brackets the POST /_reload that scripts/e2e_proofs.sh makes
+# in between. Before it, !delta is a word the bot has never heard.
+name: "!delta is unknown before the reload"
+inject:
+  - { text: "!delta", uuid: uuid-reload-before, name: ReloadBefore }
+expect:
+  - { kind: text, to: reply, contains: "Nie znam !delta" }
diff --git a/scripts/e2e.sh b/scripts/e2e.sh
index c5d5897..2b02a15 100755
--- a/scripts/e2e.sh
+++ b/scripts/e2e.sh
@@ -95,6 +95,24 @@ wait_for_subscription() {
     done
 }
 
+# Drive the fake through these scenario files, against the bot booted above.
+# One place for the group map, so a proof that injects after the main replay
+# resolves the same symbolic names the scenarios do.
+replay() {
+    "$PYTHON" -m fake_signal.cli "$@" \
+        --settle "${E2E_SETTLE:-0.15}" \
+        --base-url "$FAKE_URL" \
+        --group "watched=${WATCHED_GROUP_INTERNAL}" \
+        --group "reply=${WATCHED_GROUP}" \
+        --group "second=${SECOND_GROUP_INTERNAL}" \
+        --group "second_reply=${SECOND_GROUP}" \
+        --group "third=${THIRD_GROUP_INTERNAL}" \
+        --group "third_reply=${THIRD_GROUP}" \
+        --group "fourth=${FOURTH_GROUP_INTERNAL}" \
+        --group "fourth_reply=${FOURTH_GROUP}" \
+        --group "other=${OTHER_GROUP}"
+}
+
 validate_requirements() {
     if [[ ! -x "$PYTHON" ]]; then
         echo "Error: no interpreter at $PYTHON. Run: uv venv .venv && uv pip install -e '.[dev]'" >&2
@@ -159,23 +177,12 @@ main() {
     wait_for_subscription
 
     echo "== replaying scenarios"
-    "$PYTHON" -m fake_signal.cli \
-        "$REPO_ROOT"/fake_signal/scenarios/*.yaml \
-        --settle "${E2E_SETTLE:-0.15}" \
-        --base-url "$FAKE_URL" \
-        --group "watched=${WATCHED_GROUP_INTERNAL}" \
-        --group "reply=${WATCHED_GROUP}" \
-        --group "second=${SECOND_GROUP_INTERNAL}" \
-        --group "second_reply=${SECOND_GROUP}" \
-        --group "third=${THIRD_GROUP_INTERNAL}" \
-        --group "third_reply=${THIRD_GROUP}" \
-        --group "fourth=${FOURTH_GROUP_INTERNAL}" \
-        --group "fourth_reply=${FOURTH_GROUP}" \
-        --group "other=${OTHER_GROUP}"
+    replay "$REPO_ROOT"/fake_signal/scenarios/*.yaml
 
     # !dodaj is unwired until photos land in an approval queue instead of
     # the album (2026-09-04). Its scenarios, the media generation step and
     # prove_the_add_photo_invariants left with it; git has them at 64852ae.
+    prove_the_table_reloads_without_a_restart
     prove_the_deafness_gate
     enforce_the_time_budget "$STARTED_AT"
 }
diff --git a/scripts/e2e_proofs.sh b/scripts/e2e_proofs.sh
index 86e40c4..4f12661 100755
--- a/scripts/e2e_proofs.sh
+++ b/scripts/e2e_proofs.sh
@@ -43,6 +43,28 @@ prove_the_deafness_gate() {
     echo "== /healthz recovered on its own after ${waited}s"
 }
 
+# A person tagged while the bot is up becomes a command after POST /_reload,
+# with no restart: the same process that just answered "Nie znam !delta"
+# posts delta's photo a moment later. The scenarios live next to the main
+# set but outside its glob, because the second one is only true after the
+# reload that this proof performs between them.
+prove_the_table_reloads_without_a_restart() {
+    local proofs="${REPO_ROOT}/fake_signal/proofs"
+    echo "== before the reload, !delta is nobody"
+    replay "${proofs}/reload_before.yaml"
+
+    echo "== tagging a fifth person into the pack, then scripts/reload.sh"
+    "$PYTHON" "$REPO_ROOT/tool/e2e_pack.py" \
+        --dir "${TEMP_DIR}/pack" --group "$WATCHED_GROUP" --extra delta >/dev/null
+    # The operator's script, not a bare curl: this is the one place it runs
+    # against a real process in CI, so its exit code is the proof.
+    BIND_PORT="$BOT_PORT" "$REPO_ROOT/scripts/reload.sh" ||
+        die "scripts/reload.sh did not get the new table live"
+
+    echo "== after the reload, !delta posts a photo"
+    replay "${proofs}/reload_after.yaml"
+}
+
 # A budget, not a benchmark. This run was four minutes until the liveness
 # tick became configurable, and nothing announced the cost -- slower is the
 # one regression a green suite never reports. Exceeding it fails the run;
diff --git a/scripts/reload.sh b/scripts/reload.sh
new file mode 100755
index 0000000..8ebeed4
--- /dev/null
+++ b/scripts/reload.sh
@@ -0,0 +1,121 @@
+#!/bin/bash
+
+# ============================================================================
+# Make the running bot pick up a change, and prove that it did.
+#
+# Two kinds of change, two mechanisms:
+#
+#   scripts/reload.sh [cmd]          data changed (aliases, an album, a
+#                                    command switched off): POST /_reload,
+#                                    the table is rebuilt in place, nothing
+#                                    restarts and no frame is lost.
+#   scripts/reload.sh --code [cmd]   code changed: restart the unit, wait for
+#                                    /healthz to go green again. Lossless too
+#                                    (signal-cli replays what arrived), but
+#                                    a deploy step -- so it is the flag, not
+#                                    the default.
+#
+# With [cmd], the bot then types that command at itself in the test group,
+# so "reloaded" means "answered", not "returned 200". E.g.
+#
+#   scripts/reload.sh '!nowaosoba'
+#   scripts/reload.sh --code '!pomoc'
+# ============================================================================
+
+set -euo pipefail
+
+SCRIPT_NAME="$(basename "$0")"
+readonly SCRIPT_NAME
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+readonly SCRIPT_DIR
+readonly BOT="http://127.0.0.1:${BIND_PORT:-8137}"
+readonly UNIT="signal-bot"
+readonly TEST_GROUP="${RELOAD_GROUP:-testbot}"
+
+RESTART=0
+PROBE=""
+
+usage() {
+    echo "Usage: $SCRIPT_NAME [--code] [command-to-selftest]"
+    echo "  --code   restart the unit (code changed) instead of rebuilding the table"
+    exit 0
+}
+
+validate_requirements() {
+    if ! command -v jq >/dev/null; then
+        echo "Error: jq is required" >&2
+        exit 1
+    fi
+}
+
+wait_healthy() {
+    local attempt=0
+    # 150 x 0.4 s: a restart is ~1 s, but signal-cli can take a while to
+    # accept the websocket again after its own reconnect.
+    until [[ "$(curl -s "${BOT}/healthz" | jq -r '.status // "down"')" == "ok" ]]; do
+        attempt=$((attempt + 1))
+        if [[ $attempt -ge 150 ]]; then
+            echo "Error: ${BOT}/healthz never reported ok" >&2
+            curl -s "${BOT}/healthz" >&2 || true
+            exit 1
+        fi
+        sleep 0.4
+    done
+}
+
+reload_data() {
+    local reply status
+    reply="$(curl -s -X POST "${BOT}/_reload")"
+    status="$(jq -r '.status // "?"' <<<"$reply")"
+    case "$status" in
+    reloaded)
+        echo "reloaded: $(jq -r '.commands' <<<"$reply") commands live"
+        ;;
+    kept)
+        # The old table is still answering; the new data is what is broken.
+        echo "Error: reload refused, previous table kept: $(jq -r '.reason' <<<"$reply")" >&2
+        exit 1
+        ;;
+    *)
+        echo "Error: unexpected answer from ${BOT}/_reload: ${reply}" >&2
+        exit 1
+        ;;
+    esac
+}
+
+restart_code() {
+    systemctl --user restart "$UNIT"
+    wait_healthy
+    echo "restarted: $(curl -s "${BOT}/healthz" | jq -r '"\(.version) \(.pope)"')"
+}
+
+main() {
+    validate_requirements
+    if [[ $RESTART -eq 1 ]]; then
+        restart_code
+    else
+        wait_healthy
+        reload_data
+    fi
+    if [[ -n "$PROBE" ]]; then
+        "${SCRIPT_DIR}/selftest.sh" "$TEST_GROUP" "$PROBE"
+    fi
+}
+
+while [[ $# -gt 0 ]]; do
+    case $1 in
+    --code)
+        RESTART=1
+        shift
+        ;;
+    -h | --help)
+        usage
+        ;;
+    *)
+        PROBE="$1"
+        shift
+        ;;
+    esac
+done
+
+main
diff --git a/signal_bot/app.py b/signal_bot/app.py
index 4424c88..79cb2d6 100644
--- a/signal_bot/app.py
+++ b/signal_bot/app.py
@@ -143,3 +143,25 @@ async def selftest(request: SelfTestRequest, response: Response) -> dict[str, st
         response.status_code = status.HTTP_403_FORBIDDEN
         return {"status": "refused", "reason": str(refusal)}
     return {"status": "delivered", "group": group_id, "text": request.text}
+
+
+@app.post("/_reload")
+async def reload_commands(response: Response) -> dict[str, str | int]:
+    """Rebuild the command table from the data on disk, without a restart.
+
+    The answer is what the bot is serving *now*: after a rejected reload
+    that is still the previous table, and ``reason`` says what was wrong
+    with the new one. 409 rather than 500, because the data was bad and the
+    bot is fine. Loopback only, like everything else on this port.
+    """
+    starter: Starter = app.state.starter
+    runtime = starter.runtime
+    if runtime is None:
+        response.status_code = status.HTTP_503_SERVICE_UNAVAILABLE
+        return {"status": "connecting", "reason": starter.reason}
+    bot = runtime.bot
+    if bot.reload_commands():
+        return {"status": "reloaded", "commands": bot.command_count}
+    response.status_code = status.HTTP_409_CONFLICT
+    holder = bot.dispatcher.holder
+    return {"status": "kept", "commands": bot.command_count, "reason": holder.last_error}
diff --git a/signal_bot/bot.py b/signal_bot/bot.py
index 73d9249..d2bac13 100644
--- a/signal_bot/bot.py
+++ b/signal_bot/bot.py
@@ -14,38 +14,22 @@ from collections.abc import Callable
 from dataclasses import dataclass
 from typing import TYPE_CHECKING
 
-from signal_bot.analytics_command import register_analytics_command
-from signal_bot.animals import load_aliases, register_animal_commands
 from signal_bot.announce import announce_all
-from signal_bot.art_command import register_art_command
 from signal_bot.changelog import load as load_changelog
-from signal_bot.decorators import cooldown
 from signal_bot.dispatcher import Dispatcher
-from signal_bot.help_command import register_help_command
 from signal_bot.http_client import ImageFetcher
-from signal_bot.pack_command import register_pack_commands
-from signal_bot.pack_holder import PackHolder
-from signal_bot.pack_load import load as load_pack
-from signal_bot.pack_random import register_random_command
-from signal_bot.poem_command import register_poem_command
-from signal_bot.popes_command import register_pope_command
 from signal_bot.portraits import PortraitCache
-from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from signal_bot.reporting import compose
 from signal_bot.stats import periods
-from signal_bot.stats.commands import register_stats_commands
 from signal_bot.stats.nightly import days_to_report, mark_reported
-from signal_bot.suggest_command import register_suggest_command
-from signal_bot.traps_command import register_trap_command
 from signal_bot.versioning import POPE
+from signal_bot.wiring import COOLDOWN_MESSAGE, build_registry
 
 if TYPE_CHECKING:
-    from pathlib import Path
-
     import httpx
 
     from signal_bot.config import Config
-    from signal_bot.decorators import Decorator
     from signal_bot.discovery import GroupDirectory
     from signal_bot.liveness import Liveness
     from signal_bot.store import Store
@@ -55,7 +39,9 @@ logger = logging.getLogger(__name__)
 
 Clock = Callable[[], dt.datetime]
 
-COOLDOWN_MESSAGE = "BEEP BOOP POCZEKAJ {seconds} {unit}."
+# Re-exported: the message lives with the wiring now, and this is where
+# every test and the docs have always imported it from.
+__all__ = ["COOLDOWN_MESSAGE", "Bot", "Clock", "build_bot", "now_local"]
 
 
 def now_local() -> dt.datetime:
@@ -164,18 +150,14 @@ class Bot:
             self.liveness.note_envelope()
             await self.dispatcher.handle(frame)
 
+    def reload_commands(self) -> bool:
+        """Rebuild the command table from the data on disk, keeping the old one on failure."""
+        return self.dispatcher.holder.reload()
 
-def _register_album(registry: CommandRegistry, directory: Path, rate_limit: Decorator) -> None:
-    """Wire the album's commands, all of them reading one mutable holder.
-
-    !random takes the same cooldown as every other photo command; !add_photo
-    takes no decorator at all, per "NO CAP, NO LIMITS" -- no cooldown, no
-    size cap, no daily limit.
-    """
-    aliases = load_aliases()
-    holder = PackHolder(load_pack(directory))
-    register_pack_commands(registry, holder, decorate=rate_limit)
-    register_random_command(registry, holder, tuple(aliases["random"]), decorate=rate_limit)
+    @property
+    def command_count(self) -> int:
+        """How many commands the live table holds -- what /_reload reports back."""
+        return len(self.dispatcher.registry.commands)
 
 
 def build_bot(
@@ -187,46 +169,20 @@ def build_bot(
     liveness: Liveness,
     clock: Clock = now_local,
 ) -> Bot:
-    """Assemble the object graph from its injected dependencies."""
-    registry = CommandRegistry()
+    """Assemble the object graph from its injected dependencies.
+
+    The registry is built through a holder so that the data behind it --
+    aliases, albums -- can be re-read without a restart: see registry_holder.
+    The fetcher and the portrait cache outlive any one table and are shared
+    with every rebuild.
+    """
     fetcher = ImageFetcher(client)
     portraits = PortraitCache(fetcher, config.commons_api_url)
-    rate_limit = cooldown(config.cooldown_seconds, COOLDOWN_MESSAGE)
-    register_animal_commands(
-        registry,
-        fetcher,
-        decorate=rate_limit,
-        endpoints=config.image_api_urls,
-    )
-    register_pope_command(registry, portraits, decorate=rate_limit)
-    register_trap_command(registry, fetcher, config.commons_api_url, decorate=rate_limit)
-    register_art_command(registry, fetcher, config.commons_api_url, decorate=rate_limit)
-    register_poem_command(registry, fetcher, config.poem_text_url, decorate=rate_limit)
-    register_stats_commands(registry, decorate=rate_limit)
-    # Reads the *command* history rather than the message history, and
-    # reads the registry when it runs -- so the album's commands, which
-    # are registered below this line, are still in the near-miss set.
-    register_analytics_command(registry, decorate=rate_limit)
-    # Rate-limited like every other command, and unlike !add_photo, whose
-    # "NO CAP, NO LIMITS" is a deliberate exception rather than a default.
-    # The cost is real and accepted: somebody told "napisz !zaproponuj ..."
-    # who then files two ideas in a row is asked to wait for the second. The
-    # alternative is one person being able to fill a table nothing prunes.
-    register_suggest_command(registry, decorate=rate_limit)
-    # Absent unless this installation has an album, and scoped to that
-    # album's own groups when it does -- so on every other machine, and in
-    # every other group, these commands simply do not exist.
-    # "album", not "directory": build_bot's own `directory` is the group
-    # directory, and shadowing it here silently passed the wrong object.
-    for album in config.photo_pack_dirs:
-        _register_album(registry, album, rate_limit)
-    # Last, so !pomoc trails the list it prints; it reads the registry
-    # when it runs, so the order is presentation only.
-    register_help_command(registry, decorate=rate_limit)
+    holder = RegistryHolder(lambda: build_registry(config, fetcher, portraits))
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=holder,
         store=store,
         transport=transport,
         clock=clock,
diff --git a/signal_bot/dispatcher.py b/signal_bot/dispatcher.py
index 8e172db..eb0eb93 100644
--- a/signal_bot/dispatcher.py
+++ b/signal_bot/dispatcher.py
@@ -22,6 +22,7 @@ if TYPE_CHECKING:
     from signal_bot.discovery import GroupDirectory
     from signal_bot.models import IncomingMessage
     from signal_bot.registry import CommandRegistry
+    from signal_bot.registry_holder import RegistryHolder
     from signal_bot.store import Store
     from signal_bot.transport_base import Transport
 
@@ -36,11 +37,19 @@ class Dispatcher:
 
     config: Config
     directory: GroupDirectory
-    registry: CommandRegistry
+    # The box, not the table: the table is rebuilt and swapped when the
+    # data behind it changes, and a dispatcher holding the table itself
+    # would keep answering from the one it was built with.
+    holder: RegistryHolder
     store: Store
     transport: Transport
     clock: Clock
 
+    @property
+    def registry(self) -> CommandRegistry:
+        """The command table as it stands right now."""
+        return self.holder.registry
+
     async def _answer_unknown(self, message: IncomingMessage, group: str) -> None:
         """Answer a "!" word nothing claims, and remember it.
 
diff --git a/signal_bot/pack_load.py b/signal_bot/pack_load.py
index b3f1c42..fdea3df 100644
--- a/signal_bot/pack_load.py
+++ b/signal_bot/pack_load.py
@@ -4,34 +4,25 @@
 Every check here fails startup rather than a command, for the reason
 config_env.py gives: a manifest that names a person who does not exist, a
 trigger the folder can never match, or a file that has since been renamed is
-an answer, not an outage, and no amount of waiting fixes it. The one thing
-that is *not* an error is a pack that has not been tagged yet -- that is an
-album waiting for the tagger, and it says so in the log.
+an answer, not an outage, and no amount of waiting fixes it. What is *not*
+an error lives in pack_tags.py: a pack that has not been tagged yet, and a
+tag whose file has since gone.
 """
 
 from __future__ import annotations
 
-import json
-import logging
 from typing import TYPE_CHECKING, Any, Final
 
 from signal_bot.folding import fold
 from signal_bot.groups import ConfigError
-from signal_bot.pack import (
-    MEDIA_DIR,
-    MEDIA_TYPES,
-    PEOPLE_FILE,
-    TAGS_FILE,
-    Entry,
-    Pack,
-)
+from signal_bot.pack import PEOPLE_FILE, Entry, Pack
 from signal_bot.pack_custom import load as load_custom
+from signal_bot.pack_tags import document as _document
+from signal_bot.pack_tags import tags as _tags
 
 if TYPE_CHECKING:
     from pathlib import Path
 
-logger = logging.getLogger(__name__)
-
 # "!" plus at least one character to fold.
 MIN_TRIGGER_LENGTH: Final = 2
 
@@ -66,24 +57,6 @@ def load(directory: Path) -> Pack:
     )
 
 
-def _document(path: Path) -> dict[str, Any]:
-    """One JSON manifest, read strictly."""
-    try:
-        text = path.read_text(encoding="utf-8")
-    except OSError as error:
-        msg = f"photo pack manifest {path} cannot be read: {error}"
-        raise ConfigError(msg) from error
-    try:
-        document = json.loads(text)
-    except json.JSONDecodeError as error:
-        msg = f"photo pack manifest {path} is not valid JSON: {error}"
-        raise ConfigError(msg) from error
-    if not isinstance(document, dict):
-        msg = f"photo pack manifest {path} must be an object"
-        raise ConfigError(msg)
-    return document
-
-
 def _groups(document: dict[str, Any]) -> frozenset[str]:
     """The groups the album may be posted into, which may not be empty.
 
@@ -213,38 +186,3 @@ def _unique_triggers(entries: list[Entry]) -> None:
             if owner != entry.key:
                 msg = f"{PEOPLE_FILE} trigger {trigger!r} is claimed by {owner!r} and {entry.key!r}"
                 raise ConfigError(msg)
-
-
-def _tags(directory: Path, people: set[str]) -> dict[str, frozenset[str]]:
-    """Which file shows whom, checked against the album and the manifest.
-
-    A missing tags.json is not an error: that is a pack the tagger has not
-    been run on yet, and the commands answer "no such photo" until it has.
-    A tag naming a file that is not there *is* an error -- a renamed file
-    would otherwise drop out of the album silently.
-    """
-    path = directory / TAGS_FILE
-    if not path.exists():
-        logger.warning("photo pack %s has no %s yet; run ./web.sh", directory, TAGS_FILE)
-        return {}
-    document = _document(path)
-    tags: dict[str, frozenset[str]] = {}
-    for filename, keys in document.items():
-        _media_file(directory, str(filename))
-        if not isinstance(keys, list) or any(str(key) not in people for key in keys):
-            msg = f"{TAGS_FILE} entry {filename!r} names somebody who is not in {PEOPLE_FILE}"
-            raise ConfigError(msg)
-        if keys:
-            tags[str(filename)] = frozenset(str(key) for key in keys)
-    return tags
-
-
-def _media_file(directory: Path, filename: str) -> None:
-    """The tagged file has to exist, and be something Signal can render."""
-    path = directory / MEDIA_DIR / filename
-    if path.name != filename or not path.is_file():
-        msg = f"{TAGS_FILE} names {filename!r}, which is not a file in {MEDIA_DIR}/"
-        raise ConfigError(msg)
-    if path.suffix.lower() not in MEDIA_TYPES:
-        msg = f"{TAGS_FILE} names {filename!r}, whose type is not one Signal is told about"
-        raise ConfigError(msg)
diff --git a/signal_bot/pack_tags.py b/signal_bot/pack_tags.py
new file mode 100644
index 0000000..cede643
--- /dev/null
+++ b/signal_bot/pack_tags.py
@@ -0,0 +1,96 @@
+# Copyright (c) 2026 Krzysztof Rudnicki. MIT licence.
+"""Reading tags.json: which file shows whom, checked against the album.
+
+Split out of pack_load.py at the 250-line cap, and the half whose checks are
+*not* all fatal. The manifest half refuses anything it cannot honour, for
+the reason given there; this half has two conditions it survives, because
+both are states a healthy album passes through: no tags.json yet, and a tag
+whose file is no longer there.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from typing import TYPE_CHECKING, Any
+
+from signal_bot.groups import ConfigError
+from signal_bot.pack import MEDIA_DIR, MEDIA_TYPES, PEOPLE_FILE, TAGS_FILE
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+def document(path: Path) -> dict[str, Any]:
+    """One JSON manifest, read strictly."""
+    try:
+        text = path.read_text(encoding="utf-8")
+    except OSError as error:
+        msg = f"photo pack manifest {path} cannot be read: {error}"
+        raise ConfigError(msg) from error
+    try:
+        document = json.loads(text)
+    except json.JSONDecodeError as error:
+        msg = f"photo pack manifest {path} is not valid JSON: {error}"
+        raise ConfigError(msg) from error
+    if not isinstance(document, dict):
+        msg = f"photo pack manifest {path} must be an object"
+        raise ConfigError(msg)
+    return document
+
+
+def tags(directory: Path, people: set[str]) -> dict[str, frozenset[str]]:
+    """Which file shows whom, checked against the album and the manifest.
+
+    A missing tags.json is not an error: that is a pack the tagger has not
+    been run on yet, and the commands answer "no such photo" until it has.
+    Nor is a tag naming a file that is no longer there: that entry is
+    skipped, with a warning naming it. It used to be fatal, so a renamed
+    file could not drop out of the album silently -- and on 2026-09-01 two
+    such entries appeared in the live pack, which meant the next restart
+    would have looped on ConfigError. One stale row is drift in a file a
+    page edits mid-session, not a broken album, and now that the table is
+    rebuilt from that file while the bot runs, the only safe reading of
+    drift is "everything else still works".
+    """
+    path = directory / TAGS_FILE
+    if not path.exists():
+        logger.warning("photo pack %s has no %s yet; run ./web.sh", directory, TAGS_FILE)
+        return {}
+    tags: dict[str, frozenset[str]] = {}
+    for filename, keys in document(path).items():
+        if not _media_file(directory, str(filename)):
+            logger.warning(
+                "%s names %r, which is not a file in %s/ -- skipped", TAGS_FILE, filename, MEDIA_DIR
+            )
+            continue
+        if not isinstance(keys, list) or any(str(key) not in people for key in keys):
+            msg = f"{TAGS_FILE} entry {filename!r} names somebody who is not in {PEOPLE_FILE}"
+            raise ConfigError(msg)
+        if keys:
+            tags[str(filename)] = frozenset(str(key) for key in keys)
+    return tags
+
+
+def _media_file(directory: Path, filename: str) -> bool:
+    """Whether the tagged file is there; if it is, it must be something Signal can render.
+
+    A name that reaches outside the album is refused outright rather than
+    skipped: that is not a file that went missing, it is a tag that should
+    never have been written.
+
+    Raises:
+        ConfigError: The name is a path, or the type is one Signal cannot show.
+    """
+    path = directory / MEDIA_DIR / filename
+    if path.name != filename:
+        msg = f"{TAGS_FILE} names {filename!r}, which is not a file in {MEDIA_DIR}/"
+        raise ConfigError(msg)
+    if not path.is_file():
+        return False
+    if path.suffix.lower() not in MEDIA_TYPES:
+        msg = f"{TAGS_FILE} names {filename!r}, whose type is not one Signal is told about"
+        raise ConfigError(msg)
+    return True
diff --git a/signal_bot/registry_holder.py b/signal_bot/registry_holder.py
new file mode 100644
index 0000000..c82af17
--- /dev/null
+++ b/signal_bot/registry_holder.py
@@ -0,0 +1,77 @@
+# Copyright (c) 2026 Krzysztof Rudnicki. MIT licence.
+"""The one mutable box the command table lives in, so data changes need no restart.
+
+``PackHolder`` made a *photo* visible the moment it was stored, and said out
+loud what it would never do: add a command. A new person in the tagger, a
+new alias, a command switched off from the management page -- every one of
+those changes the registry itself, and until now every one of them meant
+restarting the unit. Restarting is cheap and lossless (signal-cli replays
+what arrived meanwhile), but it is a deploy step somebody has to remember,
+and a 21:37 report can land in the middle of it.
+
+So the registry is rebuilt, whole, from the same builder that made it at
+startup, and swapped in one assignment. Never patched in place: a table
+that is edited command by command can be observed half-edited, and a
+duplicate trigger found on the third command of five leaves the first two
+already live. Building a fresh table first means ``DuplicateTriggerError``
+and ``ConfigError`` are raised before anything the dispatcher reads has
+changed.
+
+``reload`` never raises on bad *data*. The builder reads files people edit
+by hand and pages write mid-session, and the only safe answer to a file
+that does not load is the table that was already answering. A programming
+error is not data and is not swallowed: it reaches the caller, which for
+``POST /_reload`` means a 500 and a traceback in the journal.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+from signal_bot.groups import ConfigError
+from signal_bot.registry import DuplicateTriggerError
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+    from signal_bot.registry import CommandRegistry
+
+logger = logging.getLogger(__name__)
+
+# A `type` alias is evaluated lazily, so the TYPE_CHECKING-only names are fine.
+type Builder = Callable[[], CommandRegistry]
+
+
+class RegistryHolder:
+    """A mutable reference to the command table the dispatcher reads through."""
+
+    def __init__(self, build: Builder) -> None:
+        """Build the first table now, loudly.
+
+        Startup is the one time a bad table must *not* be survived: there is
+        no previous one to fall back on, and ``Restart=always`` looping on a
+        ConfigError is a clearer signal than a bot that answers nothing.
+        """
+        self._build = build
+        self.registry = build()
+        self.last_error = ""
+
+    def reload(self) -> bool:
+        """Rebuild the table from whatever is on disk now and publish it.
+
+        Returns:
+            True when the new table is live, False when the old one is still
+            being served because the data no longer builds -- and then
+            ``last_error`` says why, for whoever asked.
+        """
+        try:
+            fresh = self._build()
+        except (ConfigError, DuplicateTriggerError) as error:
+            self.last_error = str(error)
+            logger.exception("command registry no longer builds; keeping the previous one")
+            return False
+        self.registry = fresh
+        self.last_error = ""
+        logger.info("command registry rebuilt: %d commands", len(fresh.commands))
+        return True
diff --git a/signal_bot/wiring.py b/signal_bot/wiring.py
new file mode 100644
index 0000000..03bb109
--- /dev/null
+++ b/signal_bot/wiring.py
@@ -0,0 +1,95 @@
+# Copyright (c) 2026 Krzysztof Rudnicki. MIT licence.
+"""Which commands exist: the one function that fills a CommandRegistry.
+
+Split out of bot.py when the registry stopped being built once. ``build_bot``
+still assembles the object graph; this is the part of it that reads data --
+aliases.json, every album's manifests -- and so the part ``RegistryHolder``
+runs again when that data changes. Everything with a lifetime longer than
+one table (the HTTP client, the portrait cache, the store) is passed in, so
+a rebuild reuses them rather than opening a second of each.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from signal_bot.analytics_command import register_analytics_command
+from signal_bot.animals import load_aliases, register_animal_commands
+from signal_bot.art_command import register_art_command
+from signal_bot.decorators import cooldown
+from signal_bot.help_command import register_help_command
+from signal_bot.pack_command import register_pack_commands
+from signal_bot.pack_holder import PackHolder
+from signal_bot.pack_load import load as load_pack
+from signal_bot.pack_random import register_random_command
+from signal_bot.poem_command import register_poem_command
+from signal_bot.popes_command import register_pope_command
+from signal_bot.registry import CommandRegistry
+from signal_bot.stats.commands import register_stats_commands
+from signal_bot.suggest_command import register_suggest_command
+from signal_bot.traps_command import register_trap_command
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+    from signal_bot.config import Config
+    from signal_bot.decorators import Decorator
+    from signal_bot.http_client import ImageFetcher
+    from signal_bot.portraits import PortraitCache
+
+COOLDOWN_MESSAGE = "BEEP BOOP POCZEKAJ {seconds} {unit}."
+
+
+def _register_album(registry: CommandRegistry, directory: Path, rate_limit: Decorator) -> None:
+    """Wire the album's commands, all of them reading one mutable holder.
+
+    !random takes the same cooldown as every other photo command. !add_photo
+    is not wired: a photo posted to a group must not reach the album until
+    somebody has approved it, and no approval queue exists yet.
+    """
+    aliases = load_aliases()
+    holder = PackHolder(load_pack(directory))
+    register_pack_commands(registry, holder, decorate=rate_limit)
+    register_random_command(registry, holder, tuple(aliases["random"]), decorate=rate_limit)
+
+
+def build_registry(
+    config: Config, fetcher: ImageFetcher, portraits: PortraitCache
+) -> CommandRegistry:
+    """Read the data and register every command on a fresh table.
+
+    Raises:
+        ConfigError: An album no longer loads.
+        DuplicateTriggerError: Two commands live in one group claim a word.
+    """
+    registry = CommandRegistry()
+    rate_limit = cooldown(config.cooldown_seconds, COOLDOWN_MESSAGE)
+    register_animal_commands(
+        registry,
+        fetcher,
+        decorate=rate_limit,
+        endpoints=config.image_api_urls,
+    )
+    register_pope_command(registry, portraits, decorate=rate_limit)
+    register_trap_command(registry, fetcher, config.commons_api_url, decorate=rate_limit)
+    register_art_command(registry, fetcher, config.commons_api_url, decorate=rate_limit)
+    register_poem_command(registry, fetcher, config.poem_text_url, decorate=rate_limit)
+    register_stats_commands(registry, decorate=rate_limit)
+    # Reads the *command* history rather than the message history, and
+    # reads the registry when it runs -- so the album's commands, which
+    # are registered below this line, are still in the near-miss set.
+    register_analytics_command(registry, decorate=rate_limit)
+    # Rate-limited like every other command. The cost is real and accepted:
+    # somebody told "napisz !zaproponuj ..." who then files two ideas in a
+    # row is asked to wait for the second. The alternative is one person
+    # being able to fill a table nothing prunes.
+    register_suggest_command(registry, decorate=rate_limit)
+    # Absent unless this installation has an album, and scoped to that
+    # album's own groups when it does -- so on every other machine, and in
+    # every other group, these commands simply do not exist.
+    for album in config.photo_pack_dirs:
+        _register_album(registry, album, rate_limit)
+    # Last, so !pomoc trails the list it prints; it reads the registry
+    # when it runs, so the order is presentation only.
+    register_help_command(registry, decorate=rate_limit)
+    return registry
diff --git a/tests/test_analytics_command.py b/tests/test_analytics_command.py
index 67ebb35..17f5630 100644
--- a/tests/test_analytics_command.py
+++ b/tests/test_analytics_command.py
@@ -15,6 +15,7 @@ from signal_bot.analytics.usage import CommandUse, PersonUse
 from signal_bot.analytics_command import register_analytics_command, triggers_for
 from signal_bot.dispatcher import Dispatcher
 from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from signal_bot.stats.events import MessageEvent
 from tests.analytics_seed import ALA, BASIA, Shape, ran
 from tests.conftest import WATCHED_GROUP, envelope
@@ -220,7 +221,7 @@ async def test_the_command_answers_in_the_group_that_asked(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=a_registry(),
+        holder=RegistryHolder(a_registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
diff --git a/tests/test_app.py b/tests/test_app.py
index 01708ca..a17feae 100644
--- a/tests/test_app.py
+++ b/tests/test_app.py
@@ -10,6 +10,7 @@ from fastapi.testclient import TestClient
 
 from signal_bot import app as app_module
 from signal_bot.app import lifespan
+from signal_bot.groups import ConfigError
 from signal_bot.versioning import __version__
 from tests.conftest import WATCHED_GROUP, mock_client, signal_client
 
@@ -79,3 +80,53 @@ async def test_lifespan_is_an_async_context_manager(monkeypatch: pytest.MonkeyPa
         # Present immediately, before anything has reached Signal: /healthz
         # reads it, and an absent attribute would be a 500 rather than a 503.
         assert app_module.app.state.starter is not None
+
+
+def test_reload_is_refused_while_still_connecting(monkeypatch: pytest.MonkeyPatch) -> None:
+    with _serving(monkeypatch, mock_client(_offline)) as client:
+        response = client.post("/_reload")
+        assert response.status_code == httpx.codes.SERVICE_UNAVAILABLE
+        assert response.json()["status"] == "connecting"
+
+
+def _connected(client: TestClient) -> None:
+    """Spin the loop until the connecting task has produced a runtime."""
+    for _ in range(50):
+        if client.get("/healthz").json()["status"] == "ok":
+            return
+    msg = "the bot never connected"
+    raise AssertionError(msg)
+
+
+def test_reload_rebuilds_the_command_table(monkeypatch: pytest.MonkeyPatch) -> None:
+    with _serving(monkeypatch, signal_client(WATCHED_GROUP)) as client:
+        _connected(client)
+        response = client.post("/_reload")
+        assert response.status_code == httpx.codes.OK
+        body = response.json()
+        assert body["status"] == "reloaded"
+        assert body["commands"] > 0
+
+
+def test_reload_keeps_the_old_table_when_the_data_is_bad(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """409, not 500: the bot is fine, the file somebody edited is not."""
+    with _serving(monkeypatch, signal_client(WATCHED_GROUP)) as client:
+        _connected(client)
+        before = client.post("/_reload").json()["commands"]
+
+        complaint = ConfigError("people.json: !alfa claimed twice")
+
+        def broken(*args: object) -> object:
+            raise complaint
+
+        monkeypatch.setattr("signal_bot.bot.build_registry", broken)
+        response = client.post("/_reload")
+        assert response.status_code == httpx.codes.CONFLICT
+        body = response.json()
+        assert body["status"] == "kept"
+        assert body["commands"] == before
+        assert "claimed twice" in body["reason"]
+        # And the old table still answers.
+        assert client.get("/healthz").json()["status"] == "ok"
diff --git a/tests/test_bot.py b/tests/test_bot.py
index cea67e8..fd69f1d 100644
--- a/tests/test_bot.py
+++ b/tests/test_bot.py
@@ -8,8 +8,10 @@ from dataclasses import replace
 from typing import TYPE_CHECKING
 
 import httpx
+import pytest
 
 from signal_bot.bot import Bot, build_bot, now_local
+from signal_bot.groups import ConfigError
 from signal_bot.liveness import Liveness
 from signal_bot.stats.events import MessageEvent
 from signal_bot.stats.periods import WARSAW
@@ -125,3 +127,44 @@ def test_an_album_is_wired_up_only_where_there_is_one(
     assert command is not None
     assert command.name == "pack:p1"
     assert command.groups == frozenset({WATCHED_GROUP})
+
+
+def test_reload_commands_picks_up_a_person_added_to_the_album(
+    tmp_path: Path,
+    config: Config,
+    store: Store,
+    transport: FakeTransport,
+    directory: GroupDirectory,
+) -> None:
+    """A person tagged after startup is a command after a reload, not a restart."""
+    root = pack_fixture.build(tmp_path / "pack")
+    bot = _bot(replace(config, photo_pack_dirs=(root,)), store, transport, directory)
+    before = bot.command_count
+    assert bot.dispatcher.registry.resolve("!delta") is None
+
+    people = [*pack_fixture.PEOPLE, {"key": "p4", "triggers": ["!delta"]}]
+    pack_fixture.build(root, people=people, tags={**pack_fixture.TAGS, "delta.png": ["p4"]})
+
+    assert bot.reload_commands() is True
+    assert bot.dispatcher.registry.resolve("!delta") is not None
+    assert bot.command_count == before + 1
+
+
+def test_reload_commands_keeps_serving_when_the_album_breaks(
+    tmp_path: Path,
+    config: Config,
+    store: Store,
+    transport: FakeTransport,
+    directory: GroupDirectory,
+) -> None:
+    """A half-edited manifest must not take the working commands down with it."""
+    root = pack_fixture.build(tmp_path / "pack")
+    bot = _bot(replace(config, photo_pack_dirs=(root,)), store, transport, directory)
+    (root / "people.json").write_text("{not json", encoding="utf-8")
+
+    assert bot.reload_commands() is False
+    assert bot.dispatcher.registry.resolve("!alfa") is not None
+    assert bot.dispatcher.holder.last_error
+    # And the real loader is what said no -- the same error a restart would die on.
+    with pytest.raises(ConfigError):
+        _bot(replace(config, photo_pack_dirs=(root,)), store, transport, directory)
diff --git a/tests/test_dispatcher.py b/tests/test_dispatcher.py
index 89b90fe..6f43a7c 100644
--- a/tests/test_dispatcher.py
+++ b/tests/test_dispatcher.py
@@ -10,6 +10,7 @@ import pytest
 
 from signal_bot.dispatcher import Dispatcher
 from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from tests.conftest import OTHER_GROUP, SECOND_GROUP, WATCHED_GROUP, envelope
 
 if TYPE_CHECKING:
@@ -56,7 +57,7 @@ def dispatcher(
     return Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
@@ -156,7 +157,7 @@ async def test_a_failing_command_does_not_kill_the_loop(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
@@ -184,7 +185,7 @@ async def test_a_reply_goes_back_to_the_group_that_asked(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
diff --git a/tests/test_dispatcher_gates.py b/tests/test_dispatcher_gates.py
index a3063b0..bb2187d 100644
--- a/tests/test_dispatcher_gates.py
+++ b/tests/test_dispatcher_gates.py
@@ -9,6 +9,7 @@ from typing import TYPE_CHECKING
 from signal_bot.decorators import cooldown
 from signal_bot.dispatcher import Dispatcher
 from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from signal_bot.stats.queries_day import day_rows
 from tests.conftest import SECOND_GROUP, WATCHED_GROUP, envelope
 
@@ -40,7 +41,7 @@ async def test_a_refused_command_records_nothing(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
@@ -72,7 +73,7 @@ async def test_waiting_the_stated_time_lets_the_command_run(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: clock[0],
@@ -110,7 +111,7 @@ async def test_a_scoped_command_in_another_group_is_not_a_command_at_all(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
@@ -141,7 +142,7 @@ async def test_the_words_after_a_trigger_reach_the_handler(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
diff --git a/tests/test_dispatcher_recording.py b/tests/test_dispatcher_recording.py
index b3d0946..89bf718 100644
--- a/tests/test_dispatcher_recording.py
+++ b/tests/test_dispatcher_recording.py
@@ -19,6 +19,7 @@ import pytest
 from signal_bot.analytics import probe
 from signal_bot.dispatcher import Dispatcher
 from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from tests.conftest import WATCHED_GROUP, envelope
 
 if TYPE_CHECKING:
@@ -61,7 +62,7 @@ def build(
     return Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
diff --git a/tests/test_dispatcher_unknown.py b/tests/test_dispatcher_unknown.py
index 273bdb0..457a103 100644
--- a/tests/test_dispatcher_unknown.py
+++ b/tests/test_dispatcher_unknown.py
@@ -17,6 +17,7 @@ from signal_bot.context import CommandContext
 from signal_bot.discovery import GroupDirectory
 from signal_bot.dispatcher import Dispatcher
 from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from tests.conftest import WATCHED_GROUP, envelope
 
 if TYPE_CHECKING:
@@ -53,7 +54,7 @@ def dispatcher(
     return Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: NOW,
diff --git a/tests/test_pack_add_invariants.py b/tests/test_pack_add_invariants.py
index 95ee5a7..178c952 100644
--- a/tests/test_pack_add_invariants.py
+++ b/tests/test_pack_add_invariants.py
@@ -27,6 +27,7 @@ from signal_bot.pack_load import load
 from signal_bot.pack_replies import FAILED
 from signal_bot.pack_store import ADDS_FILE
 from signal_bot.registry import CommandRegistry
+from signal_bot.registry_holder import RegistryHolder
 from tests import pack_fixture
 from tests.conftest import WATCHED_GROUP, envelope
 from tests.pack_add_fixture import (
@@ -71,7 +72,7 @@ async def test_a_typed_label_reaches_no_log_line_no_row_and_no_reply(
     dispatcher = Dispatcher(
         config=config,
         directory=directory,
-        registry=registry,
+        holder=RegistryHolder(lambda: registry),
         store=store,
         transport=transport,
         clock=lambda: dt.datetime(2026, 8, 31, 12, 0, tzinfo=dt.UTC),
diff --git a/tests/test_pack_holder.py b/tests/test_pack_holder.py
index 9ac945a..7b49dcf 100644
--- a/tests/test_pack_holder.py
+++ b/tests/test_pack_holder.py
@@ -38,9 +38,10 @@ def test_a_pack_that_no_longer_loads_keeps_the_old_one_and_does_not_raise(
     holder = PackHolder(load(root))
     before = holder.pack
 
-    # A tag naming a file that is not there: exactly what a half-rolled-back
-    # write leaves behind.
-    (root / "tags.json").write_text(json.dumps({"vanished.png": ["p1"]}), encoding="utf-8")
+    # Half a write: exactly what a page interrupted mid-save leaves behind.
+    # (A tag naming a vanished file used to be the example here; since
+    # 2026-09-04 that is skipped with a warning rather than refused.)
+    (root / "tags.json").write_text('{"alfa.png": ["p1"', encoding="utf-8")
 
     assert holder.reload() is False
     assert holder.pack is before
diff --git a/tests/test_pack_load.py b/tests/test_pack_load.py
index 1513286..6903d32 100644
--- a/tests/test_pack_load.py
+++ b/tests/test_pack_load.py
@@ -158,10 +158,21 @@ def test_an_untagged_pack_loads_and_warns(tmp_path: Path, caplog: pytest.LogCapt
     assert "./web.sh" in caplog.text
 
 
-def test_a_tag_naming_a_file_that_is_gone_is_refused(tmp_path: Path) -> None:
-    """A renamed file would otherwise drop out of the album in silence."""
-    with pytest.raises(ConfigError, match="not a file"):
-        _load(tmp_path, tags={"alfa.png": ["p1"]}, media={})
+def test_a_tag_naming_a_file_that_is_gone_is_skipped_out_loud(
+    tmp_path: Path, caplog: pytest.LogCaptureFixture
+) -> None:
+    """Drift in one row must not take the album down; it is named, then ignored.
+
+    Fatal until 2026-09-04, when two such rows in the live pack meant the
+    next restart would have looped on ConfigError.
+    """
+    pack = _load(
+        tmp_path,
+        tags={"alfa.png": ["p1"], "gone.jpg": ["p2"]},
+        media={"alfa.png": pack_fixture.PIXEL},
+    )
+    assert set(pack.tags) == {"alfa.png"}
+    assert "'gone.jpg', which is not a file in media/ -- skipped" in caplog.text
 
 
 def test_a_tag_may_not_reach_outside_the_album(tmp_path: Path) -> None:
diff --git a/tests/test_registry_holder.py b/tests/test_registry_holder.py
new file mode 100644
index 0000000..b924887
--- /dev/null
+++ b/tests/test_registry_holder.py
@@ -0,0 +1,109 @@
+# Copyright (c) 2026 Krzysztof Rudnicki. MIT licence.
+"""The registry holder: rebuilt whole, swapped once, never left half-built."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import pytest
+
+from signal_bot.groups import ConfigError
+from signal_bot.registry import CommandRegistry, DuplicateTriggerError
+from signal_bot.registry_holder import RegistryHolder
+
+if TYPE_CHECKING:
+    from signal_bot.context import CommandContext
+
+
+async def _noop(ctx: CommandContext) -> None:
+    """A handler that does nothing; only the table's shape matters here."""
+
+
+def _table(*names: str) -> CommandRegistry:
+    registry = CommandRegistry()
+    for name in names:
+        registry.register(name, [f"!{name}"], _noop, f"{name} description")
+    return registry
+
+
+class _Builder:
+    """A builder whose answer can be changed between calls, like data on disk."""
+
+    def __init__(self, *names: str) -> None:
+        self.names = names
+        self.error: Exception | None = None
+        self.calls = 0
+
+    def __call__(self) -> CommandRegistry:
+        self.calls += 1
+        if self.error is not None:
+            raise self.error
+        return _table(*self.names)
+
+
+def test_the_first_table_is_built_at_construction() -> None:
+    builder = _Builder("cat")
+    holder = RegistryHolder(builder)
+    assert builder.calls == 1
+    assert holder.registry.resolve("!cat") is not None
+    assert holder.last_error == ""
+
+
+def test_a_bad_table_at_startup_is_fatal() -> None:
+    """No previous table to keep, so the unit must fail rather than answer nothing."""
+    builder = _Builder("cat")
+    builder.error = ConfigError("tags.json names a file that is not there")
+    with pytest.raises(ConfigError):
+        RegistryHolder(builder)
+
+
+def test_reload_publishes_the_new_table() -> None:
+    builder = _Builder("cat")
+    holder = RegistryHolder(builder)
+    builder.names = ("cat", "dog")
+
+    assert holder.reload() is True
+    assert holder.registry.resolve("!dog") is not None
+    assert {command.name for command in holder.registry.commands} == {"cat", "dog"}
+
+
+@pytest.mark.parametrize(
+    "error",
+    [
+        ConfigError("people.json is not valid JSON"),
+        DuplicateTriggerError("!cat is claimed twice"),
+    ],
+)
+def test_bad_data_keeps_the_previous_table(
+    error: Exception, caplog: pytest.LogCaptureFixture
+) -> None:
+    """The table that was answering keeps answering; the caller learns why."""
+    builder = _Builder("cat")
+    holder = RegistryHolder(builder)
+    before = holder.registry
+    builder.error = error
+
+    assert holder.reload() is False
+    assert holder.registry is before
+    assert holder.last_error == str(error)
+    assert "keeping the previous one" in caplog.text
+
+
+def test_a_later_good_reload_clears_the_error() -> None:
+    builder = _Builder("cat")
+    holder = RegistryHolder(builder)
+    builder.error = ConfigError("broken")
+    holder.reload()
+    builder.error = None
+
+    assert holder.reload() is True
+    assert holder.last_error == ""
+
+
+def test_a_programming_error_is_not_swallowed() -> None:
+    """Only bad *data* is survivable; a bug must surface, not become a 409."""
+    builder = _Builder("cat")
+    holder = RegistryHolder(builder)
+    builder.error = TypeError("register() got an unexpected keyword")
+    with pytest.raises(TypeError):
+        holder.reload()
diff --git a/tool/e2e_pack.py b/tool/e2e_pack.py
index a886e0e..e9c8cee 100755
--- a/tool/e2e_pack.py
+++ b/tool/e2e_pack.py
@@ -42,15 +42,25 @@ TAGS: Final = {
 }
 
 
-def write(directory: Path, group: str) -> None:
-    """Write the manifest, the tags and the files themselves."""
+def write(directory: Path, group: str, extra: str | None = None) -> None:
+    """Write the manifest, the tags and the files themselves.
+
+    ``extra`` adds one more person, ``!<extra>`` with one photo, on top of
+    the fixed four -- run a second time over the same directory, that is
+    what "somebody tagged a new person while the bot was up" looks like,
+    and what POST /_reload has to pick up without a restart.
+    """
+    people, tags = list(PEOPLE), dict(TAGS)
+    if extra is not None:
+        people.append({"key": "p5", "triggers": [f"!{extra}"]})
+        tags[f"{extra}.png"] = ["p5"]
     album = directory / "media"
     album.mkdir(parents=True, exist_ok=True)
-    for name in TAGS:
+    for name in tags:
         (album / name).write_bytes(CLIP if name.endswith(".mp4") else PIXEL)
-    manifest = {"groups": [group], "people": PEOPLE, "combos": COMBOS}
+    manifest = {"groups": [group], "people": people, "combos": COMBOS}
     (directory / "people.json").write_text(json.dumps(manifest, indent=2), encoding="utf-8")
-    (directory / "tags.json").write_text(json.dumps(TAGS, indent=2), encoding="utf-8")
+    (directory / "tags.json").write_text(json.dumps(tags, indent=2), encoding="utf-8")
 
 
 def main() -> int:
@@ -58,8 +68,9 @@ def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--dir", required=True, type=Path, help="where to write the pack")
     parser.add_argument("--group", required=True, help="the group id the pack belongs to")
+    parser.add_argument("--extra", help="one more person to add, by trigger word")
     arguments = parser.parse_args()
-    write(arguments.dir, arguments.group)
+    write(arguments.dir, arguments.group, arguments.extra)
     sys.stdout.write(f"== wrote a synthetic pack to {arguments.dir}\n")
     return 0