majiayu000 Parzival Bootstrap — Cross-Session Memory
Load cross-session context from previous Parzival sessions stored in Qdrant. This replaces the automatic startup injection with an on-demand skill invocation.
Steps
- Run the following Python script to retrieve cross-session context:
import sys
import os
import time
from datetime import datetime, timezone
# Set up import path for ai-memory source
_install_dir = os.path.expanduser("~/.ai-memory")
sys.path.insert(0, os.path.join(_install_dir, "src"))
start_ms = time.perf_counter()
_trace_start = datetime.now(tz=timezone.utc)
try:
from memory.config import MemoryConfig
from memory.search import MemorySearch
from memory.injection import (
retrieve_bootstrap_context,
select_results_greedy,
format_injection_output,
init_session_state,
log_injection_event,
)
from memory.project import detect_project
from memory.qdrant_client import QdrantUnavailable
except ImportError as e:
print(f"## Cross-Session Memory (Parzival Bootstrap)\n")
print(f"**Unavailable**: AI Memory module not installed ({e})")
print(f"\nBootstrap: import error | Qdrant: unknown")
sys.exit(0)
# Optional: Prometheus metrics (best-effort, never blocks)
try:
from memory.metrics_push import push_skill_metrics_async
except ImportError:
push_skill_metrics_async = None
# Optional: Langfuse trace events (best-effort, never blocks)
# LANGFUSE: V3 ONLY. See LANGFUSE-INTEGRATION-SPEC.md
try:
from memory.trace_buffer import emit_trace_event
except ImportError:
emit_trace_event = None
TRACE_CONTENT_MAX = 10000
try:
config = MemoryConfig()
except Exception as e:
print(f"## Cross-Session Memory (Parzival Bootstrap)\n")
print(f"**Unavailable**: Failed to load configuration ({e})")
print(f"\nBootstrap: config error | Qdrant: unknown")
sys.exit(0)
if not config.parzival_enabled:
print("## Cross-Session Memory (Parzival Bootstrap)\n")
print("Parzival is not enabled. Set `PARZIVAL_ENABLED=true` in .env to activate.")
sys.exit(0)
try:
project_name = detect_project(os.getcwd())
search_client = MemorySearch(config)
session_id = os.environ.get("CLAUDE_SESSION_ID", "unknown")
# Retrieve bootstrap context from Qdrant
results = retrieve_bootstrap_context(search_client, project_name, config)
# Greedy-fill within token budget
selected, tokens_used = select_results_greedy(results, config.bootstrap_token_budget)
# Format as markdown with attribution
formatted = format_injection_output(selected, tier=1)
elapsed_ms = int((time.perf_counter() - start_ms) * 1000)
duration_seconds = time.perf_counter() - start_ms
# Initialize session state for Tier 2 deduplication (HIGH)
injected_ids = [str(r.get("id", "")) for r in selected if r.get("id")]
init_session_state(session_id, injected_ids)
# Audit log (HIGH)
from pathlib import Path
audit_dir = Path(os.getcwd()) / ".audit"
log_injection_event(
tier=1,
trigger="skill:aim-parzival-bootstrap",
project=project_name,
session_id=session_id,
results_considered=len(results),
results_selected=len(selected),
tokens_used=tokens_used,
budget=config.bootstrap_token_budget,
audit_dir=audit_dir,
)
# Build output
print("## Cross-Session Memory (Parzival Bootstrap)\n")
if not selected:
print("No cross-session memories found for this project.\n")
else:
# Group results by type for organized display
handoffs = [r for r in selected if r.get("type") == "agent_handoff"]
decisions = [r for r in selected if r.get("type") in ("decision", "agent_memory")]
insights = [r for r in selected if r.get("type") == "agent_insight"]
github = [r for r in selected if r.get("type", "").startswith("github_")]
other = [r for r in selected if r not in handoffs + decisions + insights + github]
if handoffs:
print("### Last Handoff\n")
for h in handoffs:
print(h.get("content", "").strip())
print()
if decisions:
print("### Recent Decisions\n")
for d in decisions:
score_pct = int(d.get("score", 0) * 100)
print(f"- **[{score_pct}%]** {d.get('content', '').strip()[:200]}")
print()
if insights:
print("### Insights\n")
for i in insights:
score_pct = int(i.get("score", 0) * 100)
print(f"- **[{score_pct}%]** {i.get('content', '').strip()[:200]}")
print()
if github:
print("### GitHub Activity (since last session)\n")
for g in github:
score_pct = int(g.get("score", 0) * 100)
print(f"- **[{g.get('type', 'github')}|{score_pct}%]** {g.get('content', '').strip()[:200]}")
print()
if other:
print("### Other Context\n")
for o in other:
score_pct = int(o.get("score", 0) * 100)
print(f"- **[{o.get('type', 'unknown')}|{score_pct}%]** {o.get('content', '').strip()[:200]}")
print()
# Include raw formatted output for full context
print("<details><summary>Raw retrieved context</summary>\n")
print(formatted)
print("\n</details>\n")
print("---")
print(f"Bootstrap: {len(selected)} results | {tokens_used} tokens | {elapsed_ms}ms | Qdrant: available")
# Prometheus metrics (CRITICAL — best-effort, never blocks)
if push_skill_metrics_async:
try:
push_skill_metrics_async(
"aim-parzival-bootstrap",
"success" if selected else "empty",
duration_seconds,
)
except Exception:
pass
# Top-level Langfuse trace (MEDIUM — best-effort, never blocks)
# LANGFUSE: V3 trace buffer pattern. See LANGFUSE-INTEGRATION-SPEC.md §3.1
if emit_trace_event:
try:
emit_trace_event(
event_type="skill_bootstrap",
data={
"input": f"Parzival bootstrap skill for project: {project_name}",
"output": f"Selected {len(selected)} results, {tokens_used} tokens, {elapsed_ms}ms"[:TRACE_CONTENT_MAX],
"metadata": {
"skill_name": "aim-parzival-bootstrap",
"project_name": project_name,
"results_considered": len(results),
"results_selected": len(selected),
"tokens_used": tokens_used,
"elapsed_ms": elapsed_ms,
"agent_name": os.environ.get("CLAUDE_AGENT_NAME", "main"),
"agent_role": os.environ.get("CLAUDE_AGENT_ROLE", "user"),
},
},
project_id=project_name,
session_id=session_id,
start_time=_trace_start,
end_time=datetime.now(tz=timezone.utc),
tags=["skill", "bootstrap"],
)
except Exception:
pass
except (QdrantUnavailable, ConnectionError, TimeoutError) as e:
elapsed_ms = int((time.perf_counter() - start_ms) * 1000)
print("## Cross-Session Memory (Parzival Bootstrap)\n")
print(f"**Qdrant unavailable**: {e}\n")
print("Continuing with file-based context only.\n")
print("---")
print(f"Bootstrap: 0 results | 0 tokens | {elapsed_ms}ms | Qdrant: unavailable")
if push_skill_metrics_async:
try:
push_skill_metrics_async("aim-parzival-bootstrap", "failed", time.perf_counter() - start_ms)
except Exception:
pass
except Exception as e:
elapsed_ms = int((time.perf_counter() - start_ms) * 1000)
error_type = type(e).__name__
print("## Cross-Session Memory (Parzival Bootstrap)\n")
print(f"**Error retrieving context**: {error_type}: {e}\n")
print("Continuing with file-based context only.\n")
print("---")
print(f"Bootstrap: 0 results | 0 tokens | {elapsed_ms}ms | Qdrant: error")
if push_skill_metrics_async:
try:
push_skill_metrics_async("aim-parzival-bootstrap", "failed", time.perf_counter() - start_ms)
except Exception:
pass
Include the script output in your current context as cross-session memory from previous Parzival sessions.
If the script reports Qdrant unavailable or an error, note this and continue with file-based context only (MEMORY.md, oversight/ files).