Files
everything-claude-code/skills/continuous-learning-v2/agents/observer-loop.sh
zucchini a3d8d8ab92 fix(observer): auto-scale max_turns by analysis batch size (#2062)
* fix(observer): auto-scale max_turns by analysis batch size (#2035)

The hardcoded default of MAX_TURNS=20 is insufficient when
MAX_ANALYSIS_LINES=500 (also the default). Claude exhausts its turn
budget before it can write all discovered instinct files, producing:

  Error: Reached max turns (20)

Fix: when ECC_OBSERVER_MAX_TURNS is not explicitly set, compute
max_turns proportionally to the actual analysis batch size:
  max_turns = clamp(analysis_count / 10, 20, 100)

This gives:
  - 20–199 lines → 20 turns  (existing floor, unchanged)
  - 500 lines    → 50 turns  (resolves the reported failure)
  - 1000 lines   → 100 turns (cap)

Explicitly setting ECC_OBSERVER_MAX_TURNS still overrides the
auto-scaled value, preserving the existing escape hatch.

* test(observer): update max_turns test for auto-scaling; document validation

The max-turns budget test in tests/hooks/hooks.test.js still asserted the removed literal max_turns="${ECC_OBSERVER_MAX_TURNS:-20}", which would fail against the new auto-scaling logic. Assert the auto-scale formula and the 20/100 clamp bounds instead.

Also add the explanatory comment CodeRabbit requested above the max_turns sanitization block, clarifying it guards the explicit ECC_OBSERVER_MAX_TURNS override path.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-07 13:25:29 +08:00

336 lines
11 KiB
Bash
Executable File

#!/usr/bin/env bash
# Continuous Learning v2 - Observer background loop
#
# Fix for #521: Added re-entrancy guard, cooldown throttle, and
# tail-based sampling to prevent memory explosion from runaway
# parallel Claude analysis processes.
set +e
unset CLAUDECODE
SLEEP_PID=""
USR1_FIRED=0
PENDING_ANALYSIS=0
ANALYZING=0
LAST_ANALYSIS_EPOCH=0
# Minimum seconds between analyses (prevents rapid re-triggering)
ANALYSIS_COOLDOWN="${ECC_OBSERVER_ANALYSIS_COOLDOWN:-60}"
IDLE_TIMEOUT_SECONDS="${ECC_OBSERVER_IDLE_TIMEOUT_SECONDS:-1800}"
SESSION_LEASE_DIR="${PROJECT_DIR}/.observer-sessions"
ACTIVITY_FILE="${PROJECT_DIR}/.observer-last-activity"
cleanup() {
[ -n "$SLEEP_PID" ] && kill "$SLEEP_PID" 2>/dev/null
if [ -f "$PID_FILE" ] && [ "$(cat "$PID_FILE" 2>/dev/null)" = "$$" ]; then
rm -f "$PID_FILE"
fi
exit 0
}
trap cleanup TERM INT
file_mtime_epoch() {
local file="$1"
if [ ! -f "$file" ]; then
printf '0\n'
return
fi
if stat -c %Y "$file" >/dev/null 2>&1; then
stat -c %Y "$file" 2>/dev/null || printf '0\n'
return
fi
if stat -f %m "$file" >/dev/null 2>&1; then
stat -f %m "$file" 2>/dev/null || printf '0\n'
return
fi
printf '0\n'
}
has_active_session_leases() {
if [ ! -d "$SESSION_LEASE_DIR" ]; then
return 1
fi
find "$SESSION_LEASE_DIR" -type f -name '*.json' -print -quit 2>/dev/null | grep -q .
}
latest_activity_epoch() {
local observations_epoch activity_epoch
observations_epoch="$(file_mtime_epoch "$OBSERVATIONS_FILE")"
activity_epoch="$(file_mtime_epoch "$ACTIVITY_FILE")"
if [ "$activity_epoch" -gt "$observations_epoch" ] 2>/dev/null; then
printf '%s\n' "$activity_epoch"
else
printf '%s\n' "$observations_epoch"
fi
}
exit_if_idle_without_sessions() {
if has_active_session_leases; then
return
fi
local last_activity now_epoch idle_for
last_activity="$(latest_activity_epoch)"
now_epoch="$(date +%s)"
idle_for=$(( now_epoch - last_activity ))
if [ "$last_activity" -eq 0 ] || [ "$idle_for" -ge "$IDLE_TIMEOUT_SECONDS" ]; then
echo "[$(date)] Observer idle without active session leases for ${idle_for}s; exiting" >> "$LOG_FILE"
cleanup
fi
}
wait_for_claude_analysis() {
local child_pid="$1"
local wait_status=0
while true; do
wait "$child_pid"
wait_status=$?
if [ "$wait_status" -eq 0 ]; then
return 0
fi
# SIGUSR1 can interrupt wait while the Claude child is still running.
# Re-wait in that case so a signal is not logged as a false child failure.
if kill -0 "$child_pid" 2>/dev/null; then
continue
fi
return "$wait_status"
done
}
analyze_observations() {
if [ ! -f "$OBSERVATIONS_FILE" ]; then
return
fi
obs_count=$(wc -l < "$OBSERVATIONS_FILE" 2>/dev/null || echo 0)
if [ "$obs_count" -lt "$MIN_OBSERVATIONS" ]; then
return
fi
echo "[$(date)] Analyzing $obs_count observations for project ${PROJECT_NAME}..." >> "$LOG_FILE"
if [ "${CLV2_IS_WINDOWS:-false}" = "true" ] && [ "${ECC_OBSERVER_ALLOW_WINDOWS:-false}" != "true" ]; then
echo "[$(date)] Skipping claude analysis on Windows due to known non-interactive hang issue (#295). Set ECC_OBSERVER_ALLOW_WINDOWS=true to override." >> "$LOG_FILE"
return
fi
if ! command -v claude >/dev/null 2>&1; then
echo "[$(date)] claude CLI not found, skipping analysis" >> "$LOG_FILE"
return
fi
# session-guardian: gate observer cycle (active hours, cooldown, idle detection)
if ! bash "$(dirname "$0")/session-guardian.sh"; then
echo "[$(date)] Observer cycle skipped by session-guardian" >> "$LOG_FILE"
return
fi
# Sample recent observations instead of loading the entire file (#521).
# This prevents multi-MB payloads from being passed to the LLM.
MAX_ANALYSIS_LINES="${ECC_OBSERVER_MAX_ANALYSIS_LINES:-500}"
observer_tmp_dir="${PROJECT_DIR}/.observer-tmp"
mkdir -p "$observer_tmp_dir"
analysis_file="$(mktemp "${observer_tmp_dir}/ecc-observer-analysis.XXXXXX.jsonl")"
tail -n "$MAX_ANALYSIS_LINES" "$OBSERVATIONS_FILE" > "$analysis_file"
analysis_count=$(wc -l < "$analysis_file" 2>/dev/null || echo 0)
echo "[$(date)] Using last $analysis_count of $obs_count observations for analysis" >> "$LOG_FILE"
# Use relative path from PROJECT_DIR for cross-platform compatibility (#842).
# On Windows (Git Bash/MSYS2), absolute paths from mktemp may use MSYS-style
# prefixes (e.g. /c/Users/...) that the Claude subprocess cannot resolve.
analysis_relpath=".observer-tmp/$(basename "$analysis_file")"
prompt_file="$(mktemp "${observer_tmp_dir}/ecc-observer-prompt.XXXXXX")"
cat > "$prompt_file" <<PROMPT
IMPORTANT: You are running in non-interactive --print mode. You MUST use the Write tool directly to create files. Do NOT ask for permission, do NOT ask for confirmation, do NOT output summaries instead of writing. Just read, analyze, and write.
Read ${analysis_relpath} and identify patterns for the project ${PROJECT_NAME} (user corrections, error resolutions, repeated workflows, tool preferences).
If you find 3+ occurrences of the same pattern, you MUST write an instinct file directly to ${INSTINCTS_DIR}/<id>.md using the Write tool.
Do NOT ask for permission to write files, do NOT describe what you would write, and do NOT stop at analysis when a qualifying pattern exists.
CRITICAL: Every instinct file MUST use this exact format:
---
id: kebab-case-name
trigger: when <specific condition>
confidence: <0.3-0.85 based on frequency: 3-5 times=0.5, 6-10=0.7, 11+=0.85>
domain: <one of: code-style, testing, git, debugging, workflow, file-patterns>
source: session-observation
scope: project
project_id: ${PROJECT_ID}
project_name: ${PROJECT_NAME}
---
# Title
## Action
<what to do, one clear sentence>
## Evidence
- Observed N times in session <id>
- Pattern: <description>
- Last observed: <date>
Rules:
- Be conservative, only clear patterns with 3+ observations
- Use narrow, specific triggers
- Never include actual code snippets, only describe patterns
- When a qualifying pattern exists, write or update the instinct file in this run instead of asking for confirmation
- If a similar instinct already exists in ${INSTINCTS_DIR}/, update it instead of creating a duplicate
- The YAML frontmatter (between --- markers) with id field is MANDATORY
- If a pattern seems universal (not project-specific), set scope to global instead of project
- Examples of global patterns: always validate user input, prefer explicit error handling
- Examples of project patterns: use React functional components, follow Django REST framework conventions
PROMPT
# Read the prompt into memory before the Claude subprocess is spawned.
# On Windows/MSYS2, the mktemp path can differ from the shell's later path
# resolution, so relying on cat "$prompt_file" inside the claude invocation
# can fail even though the file was created successfully.
prompt_content="$(cat "$prompt_file" 2>/dev/null || true)"
rm -f "$prompt_file"
if [ -z "$prompt_content" ]; then
echo "[$(date)] Failed to load observer prompt content, skipping analysis" >> "$LOG_FILE"
rm -f "$analysis_file"
return
fi
timeout_seconds="${ECC_OBSERVER_TIMEOUT_SECONDS:-120}"
# Auto-scale max_turns proportional to analysis batch size when not explicitly set.
# The old hardcoded default of 20 is insufficient for the 500-line MAX_ANALYSIS_LINES
# default: Claude hits --max-turns before it can write all discovered instinct files.
# Formula: 1 turn per 10 analysis lines, floor 20, cap 100. (#2035)
if [ -n "${ECC_OBSERVER_MAX_TURNS:-}" ]; then
max_turns="${ECC_OBSERVER_MAX_TURNS}"
else
max_turns=$(( analysis_count / 10 ))
if [ "$max_turns" -lt 20 ]; then max_turns=20; fi
if [ "$max_turns" -gt 100 ]; then max_turns=100; fi
fi
exit_code=0
# Sanitize max_turns. The auto-scaled path above always yields a valid value >=20,
# but an explicit ECC_OBSERVER_MAX_TURNS override may be non-numeric, empty, or too
# small, so guard here and fall back to the safe default of 20.
case "$max_turns" in
''|*[!0-9]*)
max_turns=20
;;
esac
if [ "$max_turns" -lt 4 ]; then
max_turns=20
fi
# Ensure CWD is PROJECT_DIR so the relative analysis_relpath resolves correctly
# on all platforms, not just when the observer happens to be launched from the project root.
cd "$PROJECT_DIR" || { echo "[$(date)] Failed to cd to PROJECT_DIR ($PROJECT_DIR), skipping analysis" >> "$LOG_FILE"; rm -f "$analysis_file"; return; }
# Prevent observe.sh from recording this automated Haiku session as observations.
# Pass prompt via -p flag instead of stdin redirect for Windows compatibility (#842).
# prompt_content is already loaded in-memory so this no longer depends on the
# mktemp absolute path continuing to resolve after cwd changes (#1296).
ECC_SKIP_OBSERVE=1 ECC_HOOK_PROFILE=minimal claude --model haiku --max-turns "$max_turns" --print \
--allowedTools "Read,Write" \
-p "$prompt_content" >> "$LOG_FILE" 2>&1 &
claude_pid=$!
(
sleep "$timeout_seconds"
if kill -0 "$claude_pid" 2>/dev/null; then
echo "[$(date)] Claude analysis timed out after ${timeout_seconds}s; terminating process" >> "$LOG_FILE"
kill "$claude_pid" 2>/dev/null || true
fi
) &
watchdog_pid=$!
wait_for_claude_analysis "$claude_pid"
exit_code=$?
kill "$watchdog_pid" 2>/dev/null || true
rm -f "$analysis_file"
if [ "$exit_code" -ne 0 ]; then
echo "[$(date)] Claude analysis failed (exit $exit_code)" >> "$LOG_FILE"
fi
if [ -f "$OBSERVATIONS_FILE" ]; then
archive_dir="${PROJECT_DIR}/observations.archive"
mkdir -p "$archive_dir"
mv "$OBSERVATIONS_FILE" "$archive_dir/processed-$(date +%Y%m%d-%H%M%S)-$$.jsonl" 2>/dev/null || true
fi
}
on_usr1() {
[ -n "$SLEEP_PID" ] && kill "$SLEEP_PID" 2>/dev/null
SLEEP_PID=""
# Re-entrancy guard: defer the nudge so the main loop runs a follow-up
# analysis immediately after the current analysis finishes.
if [ "$ANALYZING" -eq 1 ]; then
PENDING_ANALYSIS=1
echo "[$(date)] Analysis already in progress, deferring signal" >> "$LOG_FILE"
return
fi
USR1_FIRED=1
# Cooldown: skip if last analysis was too recent (#521)
now_epoch=$(date +%s)
elapsed=$(( now_epoch - LAST_ANALYSIS_EPOCH ))
if [ "$elapsed" -lt "$ANALYSIS_COOLDOWN" ]; then
echo "[$(date)] Analysis cooldown active (${elapsed}s < ${ANALYSIS_COOLDOWN}s), skipping" >> "$LOG_FILE"
return
fi
ANALYZING=1
analyze_observations
LAST_ANALYSIS_EPOCH=$(date +%s)
ANALYZING=0
}
trap on_usr1 USR1
echo "$$" > "$PID_FILE"
echo "[$(date)] Observer started for ${PROJECT_NAME} (PID: $$)" >> "$LOG_FILE"
# Prune expired pending instincts before analysis
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
"${CLV2_PYTHON_CMD:-python3}" "${SCRIPT_DIR}/../scripts/instinct-cli.py" prune --quiet >> "$LOG_FILE" 2>&1 || echo "[$(date)] Warning: instinct prune failed (non-fatal)" >> "$LOG_FILE"
while true; do
exit_if_idle_without_sessions
if [ "$PENDING_ANALYSIS" -eq 1 ]; then
PENDING_ANALYSIS=0
USR1_FIRED=0
ANALYZING=1
analyze_observations
LAST_ANALYSIS_EPOCH=$(date +%s)
ANALYZING=0
continue
fi
sleep "$OBSERVER_INTERVAL_SECONDS" &
SLEEP_PID=$!
wait "$SLEEP_PID" 2>/dev/null
SLEEP_PID=""
exit_if_idle_without_sessions
if [ "$USR1_FIRED" -eq 1 ]; then
USR1_FIRED=0
else
ANALYZING=1
analyze_observations
LAST_ANALYSIS_EPOCH=$(date +%s)
ANALYZING=0
fi
done