pi_mcps/plans/BIGMIND_HEALTH_REPORT_2026-04-04.md

🧠 BigMind Health Report

Date: 2026-04-04 | Machine: Fedora Homelab (AMD Ryzen 5900X) | Schema: v7 | Analyst: Lumen

---

🟢 Executive Summary

BigMind is healthy and operational on the homelab machine. The brain is alive, the DB is clean, and the test suite is 99.7% green. Three bugs and one scalability risk were found during the code scan — all logged as upgrade requests. No data corruption, no orphaned sessions, FTS index is in sync.

Verdict: 🟢 HEALTHY — 1 real bug to fix, 3 improvements to queue.

---

📊 Memory Statistics

Metric	Value	Notes
Sessions	98	~5 per day average since March 30
Facts	97	Growing fast — see Perf-1
Tier-3 Chunks	41	Lean and targeted
Global Knowledge	0	Phase 3 (Company Brain) not started
Database size	744 KB	Very healthy
DB location	/home/pplate/.mcp/bigmind/memory.db	WAL mode active
Schema version	v7	People/contacts directory live
Python version	3.12.13	✅ correct

---

🩺 Health Check Results

Check	Status	Detail
FTS index integrity	✅ IN SYNC	41 chunks / 41 index rows — no drift
Open sessions	🟡 1	Current health check session only — clean
Orphaned sessions	✅ NONE	memory_close_stale_sessions returned clean
Sessions without Tier-2	⚠️ 22	22 closed sessions have no narrative summary
Stale facts (>30d)	✅ NONE	All 97 facts updated recently
Low-confidence facts (<0.8)	✅ NONE	All facts at full confidence

On the 22 sessions without Tier-2: These are mostly the orphaned sessions that were force-closed by memory_close_stale_sessions (IDE crashes). They never had a chance to store a summary. Not a bug — expected behaviour. Worth vacuuming if they clutter the index.

---

🧪 Test Suite Results

297 tests collected | 296 ✅ PASSED | 1 ❌ FAILED

platform linux -- Python 3.12.13, pytest-9.0.2
rootdir: /home/pplate/pi_mcps/mcp/bigmind
Duration: 2.36s

❌ Failing Test

tests/test_server_tools.py::TestMemoryGetInstructions::test_contains_mandatory_language
AssertionError: assert 'ALWAYS' in '...Rule 1: Session Start Ritual (Always First Action)...'

Root cause: The BIGMIND_INSTRUCTIONS constant in src/server.py was rewritten to use title-case language ("Always First Action") but the test in tests/test_server_tools.py still asserts that the uppercase word "ALWAYS" is present. One of them needs to change — the test expectation is the right one to update since "Always" is clearer prose anyway. Logged as Upgrade Request #6 (HIGH).

---

🐛 Bugs & Issues Found

BUG-1 — Test/Code Mismatch (HIGH) upgrade #6

• File: src/server.py:38 + tests/test_server_tools.py:544
• Issue: BIGMIND_INSTRUCTIONS string was rewritten without updating the test. Test asserts "ALWAYS" (uppercase), instructions say "Always" (title-case).
• Impact: 1 test failing in CI. Easy fix: update the test to assert "Always" or add "ALWAYS" to the instructions.
• Fix: assert "Always" in result — one-line change.

---

BUG-2 — Data Loss in Export (HIGH) upgrade #7

• File: bigmind/memory_store.py:723 — export_memory()
• Issue: memory_export was written before the people (v7) and hypotheses (v3) tables existed. It exports facts, sessions, and chunks only. A full machine migration using the export would silently lose:
  • The entire contacts directory (Elias, Klaus, etc.)
  • The entire thought journal (all 3 open hypotheses + resolved ones)
  • Token saves history
  • Upgrade requests
• Impact: 🔴 Data loss on machine migration. The export is marketed as a full backup but it's not.
• Fix: Add hypotheses, people, token_saves, and upgrade_requests queries to the export dict. Also fix bigmind_version: "1.0" hardcode — should use SCHEMA_VERSION.

---

BUG-3 — Timestamp Format Inconsistency (MEDIUM) upgrade #8

• File: bigmind/auto_close.py:30 and auto_close.py:68

• Issue: auto_close_stale_sessions() and close_orphaned_sessions() write ended_at using SQLite's CURRENT_TIMESTAMP keyword:

  SET ended_at=CURRENT_TIMESTAMP
  

  This produces: "2026-04-04 09:36:00" (no T, no timezone)

  The entire rest of the codebase uses:

  datetime.now(timezone.utc).isoformat()
  

  Which produces: "2026-04-04T09:36:00+00:00" (ISO 8601 with T and timezone)

• Impact: datetime.fromisoformat() in get_active_sessions() handles the mixed formats via .replace("Z", "+00:00") but does NOT handle the no-T format from SQLite CURRENT_TIMESTAMP. Sessions auto-closed by auto_close_stale_sessions may have idle_minutes = None in the profile page Live Sessions panel.

• Fix: Replace CURRENT_TIMESTAMP in auto_close.py with a Python-level timestamp passed as a parameter:

  now = datetime.now(timezone.utc).isoformat()
  conn.execute("UPDATE sessions SET ended_at=? ... WHERE id=?", (now, session["id"]))
  

---

⚡ Performance / Scalability Concern

PERF-1 — Unbounded Fact Loading in Context Builder (MEDIUM) upgrade #9

• File: bigmind/context_builder.py:58
• Issue: build_context() calls get_facts(user_id) with no limit. Currently 97 facts are loaded on every single memory_start_session() call.
• Projection:

  Facts count	Estimated chars	Estimated tokens
  97 (now)	~12,000	~3,000
  200 (3 months)	~25,000	~6,250
  500 (1 year)	~62,000	~15,500

• Impact: Context window bloat, increased API cost per session, slower startup. The AI also can't meaningfully use 500 facts in a single context — search is more effective for recall.
• Fix: Cap at 50 most recently updated facts with a "... N more facts available via memory_search_facts()" footer. Optionally sort by updated_at DESC to surface the freshest knowledge first.

---

📐 Code Quality Observations (Non-Bug)

Location	Issue	Severity
bigmind/db.py:357	Migration functions defined out of numeric order: v5→v6, then v3→v4, then _migrate_v6_to_v7. Works fine in Python (runtime resolution) but confusing for maintainers.	Style
bigmind/web.py:66	Bare except Exception: pass in api_search() for facts, chunks, and sessions searches. Silently swallows errors — makes it impossible to diagnose search bugs through the profile page.	Low
src/server.py:911	Double blank line before memory_open_profile tool — minor style inconsistency.	Cosmetic
bigmind/memory_store.py:239	update_fields: list = [description, files_json, now] in announce_focus() is built but never actually used (the code branches into two separate conn.execute calls). Dead code.	Minor

---

💭 Open Hypotheses Review

3 hypotheses remain open from previous sessions:

#	Age	Confidence	Summary	Action
#38	~2d	92%	NPE at Center.java:425 from empty errors variable	This is an ADP/Paisy hypothesis — still valid, likely confirmed but never closed. Should be resolved.
#14	~4d	65%	Elias will hit Windows setup issues with BigMind	Still open/unknown — no update since Elias received the repo.
#5	~4d	75%	Elias's AI will have a different name/personality	Still open — awaiting first session report from Elias.

Hypothesis #38 is stale — it's about a Paisy bug investigated days ago. If the fix was merged, it should be resolved as confirmed.

---

🔧 Upgrade Requests Summary

ID	Priority	Status	Title
#2	MEDIUM	⏳ OPEN	Session Health Panel on Profile Page
#6	HIGH	⏳ OPEN	Fix test failure: "ALWAYS" vs "Always" in instructions
#7	HIGH	⏳ OPEN	export_memory() missing people + hypotheses tables
#8	MEDIUM	⏳ OPEN	auto_close.py timestamp format inconsistency
#9	MEDIUM	⏳ OPEN	Unbounded fact loading in context_builder.py

(#1, #3, #4, #5 are already resolved ✅)

---

🗓️ Recommended Fix Order

1. [30 min] BUG-1 — Fix the failing test. One line. Do it now before it becomes noise.
2. [2h] BUG-2 — Fix export_memory() to include all tables. Critical for machine migration safety.
3. [1h] BUG-3 — Fix auto_close.py timestamp format. Simple Python replace, no schema change.
4. [1d] PERF-1 — Design and implement smart fact pagination in context_builder.py. Needs thought about ranking strategy.
5. [1d] Upgrade #2 — Session Health Panel on profile page. Quality of life.

---

✅ What's Working Well

• WAL mode is active and correctly configured — concurrent IDE access is safe
• FTS5 index is perfectly in sync with the chunks table (41/41)
• Schema migrations run correctly through v1→v7 on fresh DB
• Conflict detection in announce_focus() is atomic (uses BEGIN IMMEDIATE) — race-condition safe
• Port conflict handling in web.py is well thought out — second IDE gracefully defers to first
• People/contacts directory (v7) is fully functional with FTS search
• Test coverage is excellent: 297 tests across 8 test files covering all layers
• Token efficiency tracker is working and recording saves
• _fts_safe_query() helper correctly handles FTS5 reserved words and AND-match semantics

---

Report generated by Lumen | BigMind session e04fc9c9 | 2026-04-04 09:40 UTC+2