Skip to content

Commit dcebb04

Browse files
jiten-singh-shahijiten-singh-shahi
committed
feat: create learning-observe.sh hook for continuous learning
Adapted from ECC continuous-learning-v2/hooks/observe.sh. Captures tool use events (PreToolUse + PostToolUse) to project-scoped JSONL observation logs for pattern analysis by the learning-engine agent. Features: session guards (skip subagents/minimal profile), project detection via git hash, Python-based JSON parsing, secret scrubbing, auto-archive at 10MB, auto-purge after 30 days.
1 parent 7ef5cda commit dcebb04

2 files changed

Lines changed: 217 additions & 0 deletions

File tree

‎hooks/hooks.json‎

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -46,6 +46,11 @@
4646
],
4747
"description": "Check MCP server health before MCP tool execution and block unhealthy MCP calls"
4848
},
49+
{
50+
"matcher": "*",
51+
"hooks": [{ "type": "command", "command": "bash \"${CLAUDE_PLUGIN_ROOT}/scripts/hooks/run-with-flags-shell.sh\" \"pre:observe\" \"scripts/hooks/learning-observe.sh\" \"standard,strict\"", "async": true, "timeout": 10 }],
52+
"description": "Capture tool use observations for continuous learning"
53+
},
4954
{
5055
"matcher": "Bash",
5156
"hooks": [
@@ -132,6 +137,11 @@
132137
}
133138
],
134139
"PostToolUse": [
140+
{
141+
"matcher": "*",
142+
"hooks": [{ "type": "command", "command": "bash \"${CLAUDE_PLUGIN_ROOT}/scripts/hooks/run-with-flags-shell.sh\" \"post:observe\" \"scripts/hooks/learning-observe.sh\" \"standard,strict\"", "async": true, "timeout": 10 }],
143+
"description": "Capture tool use results for continuous learning"
144+
},
135145
{
136146
"matcher": "Write",
137147
"hooks": [

‎scripts/hooks/learning-observe.sh‎

Lines changed: 207 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,207 @@
1+
#!/usr/bin/env bash
2+
set -e
3+
4+
# Continuous Learning — Observation Hook (SCC)
5+
#
6+
# Captures tool use events for pattern analysis by the learning-engine agent.
7+
# Claude Code passes hook data via stdin as JSON.
8+
#
9+
# Adapted from ECC continuous-learning-v2/hooks/observe.sh for SCC.
10+
# Registered via hooks/hooks.json (PreToolUse + PostToolUse, standard+strict profiles).
11+
12+
HOOK_PHASE="${1:-post}"
13+
14+
# ─────────────────────────────────────────────
15+
# Read stdin (before any processing)
16+
# ─────────────────────────────────────────────
17+
18+
INPUT_JSON=$(cat)
19+
20+
if [ -z "$INPUT_JSON" ]; then
21+
exit 0
22+
fi
23+
24+
# ─────────────────────────────────────────────
25+
# Find a Python interpreter
26+
# ─────────────────────────────────────────────
27+
28+
resolve_python_cmd() {
29+
if command -v python3 >/dev/null 2>&1; then
30+
printf '%s\n' python3
31+
return 0
32+
fi
33+
if command -v python >/dev/null 2>&1; then
34+
printf '%s\n' python
35+
return 0
36+
fi
37+
return 1
38+
}
39+
40+
PYTHON_CMD="$(resolve_python_cmd 2>/dev/null || true)"
41+
if [ -z "$PYTHON_CMD" ]; then
42+
exit 0
43+
fi
44+
45+
# ─────────────────────────────────────────────
46+
# Session guards — skip automated/subagent sessions
47+
# ─────────────────────────────────────────────
48+
49+
# Only run for interactive CLI sessions
50+
case "${CLAUDE_CODE_ENTRYPOINT:-cli}" in
51+
cli|sdk-ts) ;;
52+
*) exit 0 ;;
53+
esac
54+
55+
# Minimal profile suppresses non-essential hooks
56+
[ "${SCC_HOOK_PROFILE:-standard}" = "minimal" ] && exit 0
57+
58+
# Cooperative skip for automated sessions
59+
[ "${SCC_SKIP_OBSERVE:-0}" = "1" ] && exit 0
60+
61+
# Skip subagent sessions
62+
_AGENT_ID=$(echo "$INPUT_JSON" | "$PYTHON_CMD" -c "import json,sys; print(json.load(sys.stdin).get('agent_id',''))" 2>/dev/null || true)
63+
[ -n "$_AGENT_ID" ] && exit 0
64+
65+
# ─────────────────────────────────────────────
66+
# Project detection
67+
# ─────────────────────────────────────────────
68+
69+
STDIN_CWD=$(echo "$INPUT_JSON" | "$PYTHON_CMD" -c '
70+
import json, sys
71+
try:
72+
data = json.load(sys.stdin)
73+
print(data.get("cwd", ""))
74+
except (KeyError, TypeError, ValueError):
75+
print("")
76+
' 2>/dev/null || echo "")
77+
78+
# Determine project ID from git or cwd
79+
if [ -n "$STDIN_CWD" ] && [ -d "$STDIN_CWD" ]; then
80+
PROJECT_ROOT="$STDIN_CWD"
81+
else
82+
PROJECT_ROOT="${CLAUDE_PROJECT_DIR:-$(pwd)}"
83+
fi
84+
85+
PROJECT_ID=$(cd "$PROJECT_ROOT" 2>/dev/null && git rev-parse --show-toplevel 2>/dev/null | shasum -a 256 | cut -c1-16 || echo "global")
86+
87+
# ─────────────────────────────────────────────
88+
# Configuration
89+
# ─────────────────────────────────────────────
90+
91+
CONFIG_DIR="${HOME}/.claude/homunculus"
92+
PROJECT_DIR="${CONFIG_DIR}/projects/${PROJECT_ID}"
93+
mkdir -p "$PROJECT_DIR"
94+
95+
OBSERVATIONS_FILE="${PROJECT_DIR}/observations.jsonl"
96+
MAX_FILE_SIZE_MB=10
97+
98+
# Skip if disabled
99+
if [ -f "${CONFIG_DIR}/disabled" ]; then
100+
exit 0
101+
fi
102+
103+
# Auto-purge observation files older than 30 days (runs once per day)
104+
PURGE_MARKER="${PROJECT_DIR}/.last-purge"
105+
if [ ! -f "$PURGE_MARKER" ] || [ "$(find "$PURGE_MARKER" -mtime +1 2>/dev/null)" ]; then
106+
find "${PROJECT_DIR}" -name "observations-*.jsonl" -mtime +30 -delete 2>/dev/null || true
107+
touch "$PURGE_MARKER" 2>/dev/null || true
108+
fi
109+
110+
# ─────────────────────────────────────────────
111+
# Parse tool event and write observation
112+
# ─────────────────────────────────────────────
113+
114+
PARSED=$(echo "$INPUT_JSON" | HOOK_PHASE="$HOOK_PHASE" "$PYTHON_CMD" -c '
115+
import json, sys, os
116+
117+
try:
118+
data = json.load(sys.stdin)
119+
hook_phase = os.environ.get("HOOK_PHASE", "post")
120+
event = "tool_start" if hook_phase == "pre" else "tool_complete"
121+
122+
tool_name = data.get("tool_name", data.get("tool", "unknown"))
123+
tool_input = data.get("tool_input", data.get("input", {}))
124+
tool_output = data.get("tool_response", data.get("tool_output", data.get("output", "")))
125+
session_id = data.get("session_id", "unknown")
126+
tool_use_id = data.get("tool_use_id", "")
127+
128+
# Truncate large values
129+
if isinstance(tool_input, dict):
130+
tool_input_str = json.dumps(tool_input)[:5000]
131+
else:
132+
tool_input_str = str(tool_input)[:5000]
133+
134+
if isinstance(tool_output, dict):
135+
tool_output_str = json.dumps(tool_output)[:5000]
136+
else:
137+
tool_output_str = str(tool_output)[:5000]
138+
139+
print(json.dumps({
140+
"parsed": True,
141+
"event": event,
142+
"tool": tool_name,
143+
"input": tool_input_str if event == "tool_start" else None,
144+
"output": tool_output_str if event == "tool_complete" else None,
145+
"session": session_id,
146+
"tool_use_id": tool_use_id
147+
}))
148+
except Exception as e:
149+
print(json.dumps({"parsed": False, "error": str(e)}))
150+
')
151+
152+
PARSED_OK=$(echo "$PARSED" | "$PYTHON_CMD" -c "import json,sys; print(json.load(sys.stdin).get('parsed', False))" 2>/dev/null || echo "False")
153+
154+
if [ "$PARSED_OK" != "True" ]; then
155+
exit 0
156+
fi
157+
158+
# Archive if file too large
159+
if [ -f "$OBSERVATIONS_FILE" ]; then
160+
file_size_mb=$(du -m "$OBSERVATIONS_FILE" 2>/dev/null | cut -f1)
161+
if [ "${file_size_mb:-0}" -ge "$MAX_FILE_SIZE_MB" ]; then
162+
archive_dir="${PROJECT_DIR}/observations.archive"
163+
mkdir -p "$archive_dir"
164+
mv "$OBSERVATIONS_FILE" "$archive_dir/observations-$(date +%Y%m%d-%H%M%S)-$$.jsonl" 2>/dev/null || true
165+
fi
166+
fi
167+
168+
# Write observation with secret scrubbing
169+
timestamp=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
170+
171+
export PROJECT_ID_ENV="$PROJECT_ID"
172+
export TIMESTAMP="$timestamp"
173+
174+
echo "$PARSED" | "$PYTHON_CMD" -c '
175+
import json, sys, os, re
176+
177+
parsed = json.load(sys.stdin)
178+
observation = {
179+
"timestamp": os.environ["TIMESTAMP"],
180+
"event": parsed["event"],
181+
"tool": parsed["tool"],
182+
"session": parsed["session"],
183+
"project_id": os.environ.get("PROJECT_ID_ENV", "global")
184+
}
185+
186+
# Scrub secrets
187+
_SECRET_RE = re.compile(
188+
r"(?i)(api[_-]?key|token|secret|password|authorization|credentials?|auth)"
189+
r"""([\"'"'"'"'"'"'\s:=]+)"""
190+
r"([A-Za-z]+\s+)?"
191+
r"([A-Za-z0-9_\-/.+=]{8,})"
192+
)
193+
194+
def scrub(val):
195+
if val is None:
196+
return None
197+
return _SECRET_RE.sub(lambda m: m.group(1) + m.group(2) + (m.group(3) or "") + "[REDACTED]", str(val))
198+
199+
if parsed["input"]:
200+
observation["input"] = scrub(parsed["input"])
201+
if parsed["output"] is not None:
202+
observation["output"] = scrub(parsed["output"])
203+
204+
print(json.dumps(observation))
205+
' >> "$OBSERVATIONS_FILE"
206+
207+
exit 0

0 commit comments

Comments
 (0)