Hermes Local was the user-facing name for Life OS's local safeguards when the project still shipped a Layer 4 Python tools/ package. As of v1.8.1 Wave 2 (zero-python pivot), Layer 4 is gone; the entire skill is bash hooks + markdown prompts + agent definitions. The dangerous- command pattern guard (~40 patterns) lives inline in
scripts/hooks/pre-bash-approval.sh. Pattern provenance preserved: forked from NousResearch/hermes-agent under the MIT License.
Life OS installs into your AI terminal (Claude Code, Gemini CLI, or Codex CLI) and transforms it into a personal cabinet โ multiple independent AI agents that analyze your decisions from every angle, argue with each other, and hold both the plan and you accountable.
The decision engine is the same for everyone: plan, review, veto, execute, audit. What changes is the world it speaks.
When you first start a session, you pick a theme:
๐จ Choose your theme:
English:
a) ๐๏ธ Roman Republic โ Consul, Tribune, Senate
b) ๐บ๐ธ US Government โ Chief of Staff, Attorney General, GAO
c) ๐ข C-Suite โ CEO, General Counsel, CFO
ไธญๆ:
d) ๐๏ธ ไธ็ๅ
ญ้จ โ ไธ็ธใไธญไนฆ็ใ้จไธ็
e) ๐จ๐ณ ไธญๅฝๆฟๅบ โ ๅฝๅก้ขๆป็ใๅๆนๅงใไบบๅคงๅธธๅงไผ
f) ๐ข ๅ
ฌๅธ้จ้จ โ ๆป็ป็ใๆ็ฅ่งๅ้จใๆณๅกๅ่ง้จ
ๆฅๆฌ่ช:
g) ๐๏ธ ๆๆฒปๆฟๅบ โ ๅ
้ฃ็ท็ๅคง่ฃใๅ่ญฐใๆขๅฏ้ข
h) ๐๏ธ ้ใ้ข โ ๅ
้ฃๅฎๆฟ้ทๅฎใๅ
้ฃๆณๅถๅฑใ่ฒกๅ็
i) ๐ข ไผๆฅญ โ ็คพ้ทๅฎคใ็ตๅถไผ็ป้จใๆณๅ้จ
Type a-i
Here is the same decision โ "Should I leave my job?" โ through all three:
ไธ็ๅ
ญ้จ ้ใ้ข C-Suite
โโโโโโโโโ โโโโโโโโโ โโโโโโโโโ
๐ ไธญไนฆ็ ่ตท่ๆนๆก ๐ ๅ
้ฃๅบ ่ตทๆก ๐ VP Strategy drafts plan
๐ ้จไธ็ ๅฐ้ฉณ๏ผ ๐ ๅ
้ฃๆณๅถๅฑ ๅทฎใๆปใ๏ผ ๐ General Counsel vetoes:
"่ดขๅก่ท้ๆช่งฃๅณ" "่ฒกๅ็ไฝ่ฃใไธๆ" "Runway not addressed"
๐ฐ ๆท้จ 5/10 ๐ฐ ่ฒกๅ็ 5/10 ๐ฐ CFO 5/10
โ๏ธ ๅ
ต้จ 6/10 โ๏ธ ้ฒ่ก็ 6/10 โ๏ธ VP Ops 6/10
โ๏ธ ๅ้จ 4/10 โ๏ธ ๆณๅ็ 4/10 โ๏ธ CCO 4/10
๐ฑ ๅพกๅฒๅฐ ๅฎกๆ ธ ๐ฑ ไผ่จๆคๆป้ข ็ฃๆป ๐ฑ Internal Audit audits
๐ฌ ่ฐๅฎ ่ฟฝ้ฎไฝ ๐ฌ ๅ
้ฃๅไธใๅใ่ฟใ ๐ฌ Exec Coach challenges you
๐ ๅฅๆ: 5.8/10 ๐ ้ฃ่ญฐๆฑบๅฎๆธ: 5.8/10 ๐ Executive Brief: 5.8/10
Nine different worlds. Identical rigor underneath. Each language offers three governance styles โ historical, modern government, and corporate โ so every user finds a voice that fits. A Japanese user picks between Meiji-era ministers, Kasumigaseki bureaucrats, or corporate departments. An English user chooses Roman senators, US federal officials, or C-Suite executives. A Chinese user selects Tang Dynasty mandarins, modern State Council roles, or corporate divisions.
Theme determines language. After you pick a theme, every word of output โ every agent, every report, every response โ is in that theme's language. Chinese for ไธ็ๅ ญ้จ. Japanese for ้ใ้ข. English for C-Suite. No mixing, no exceptions.
Theme is per-session. Each terminal window can use a different theme independently. You can switch mid-session at any time by saying "switch theme" (or "ๅๆขไธป้ข" or "ใใผใๅใๆฟใ"). The engine never changes โ only the voice.
Auto-inference from trigger words. Say "ไธๆ" and the ไธ็ๅ ญ้จ theme loads automatically (ๅๆ-specific). Say "้ฃ่ญฐ้ๅง" and the ้ใ้ข theme loads (modern government-specific). Generic triggers like "ๅผๅง", "ใฏใใใ", or "start" show that language's three sub-choices โ because the word alone does not distinguish historical, government, or corporate.
Not role-playing. Each agent runs as a real, isolated subagent. They cannot see each other's reasoning. They score independently. They disagree.
Closing the privacy gap on the way out. v1.8.2 hardened what enters your vault (pre-write-scan.sh defends SOUL.md, wiki/, _meta/concepts/, user-patterns.md against secrets, prompt injection, invisible Unicode). But v1.8.2 said nothing about what leaves it. Decision/Task/Journal bodies โ full of raw user prose, third-party names, specific amounts โ were synced from _meta/outbox/<sid>/ straight to Notion at Step 10a with no privacy gate at all. v1.8.3 adds the missing outbound counterpart.
What can safely live in _meta/outbox/<sid>/decisions/ under your private git repo is not the same as what should travel to Notion:
- Notion workspaces may be shared (team Notion, accidental public link)
- Notion AI may index page content for org-wide assistants
- Mobile-device theft exposes the Notion app
- Notion has had data-breach incidents historically
The same sentence ("ๆ่ๅฉ่งๅพ X ๅ ฌๅธ็ป็ ยฅ850 ไธ โฆ") can be acceptable in your private journal and unacceptable on Notion's servers. v1.8.3 treats them differently.
Every Notion MCP write call is intercepted. The hook scans tool_input against references/outbound-pii-patterns.md and applies a three-tier action model:
| Group hit | Verdict | Action |
|---|---|---|
| A โ private key, AWS / GitHub / Slack token, full credit-card, SSN, JP MyNumber, โฅ40-char high-entropy | block |
exit 2, cancels the call, logs CLASS_F violation |
| B โ third-party name + sensitive event (ๅบ่ฝจ/็ ดไบง/่ขซ่ฃ/divorced/fired) | warn |
reminder; orchestrator MUST pause and ask (a) sanitize / (b) skip / (c) override |
| C โ company name + specific amount, bank-account-shaped numbers | warn |
same |
| D โ email / international phone / JP+CN mobile / JP postal address | warn |
same |
| E โ URL trackers, JWT shape | info |
quiet log only |
| no hit | pass |
proceeds normally |
Audit trail at _meta/runtime/<sid>/notion-pii-scan-<ts>.json records the matched pattern category IDs (never raw content), so AUDITOR Mode 3 patrol can track outbound risk frequency over time โ if you keep triggering Group B in 40 % of adjourns, that's a behavioral signal worth reviewing.
Group A patterns are unambiguous; hard-blocking is correct. Groups B/C/D have non-zero false-positive rates โ bash POSIX regex can't cleanly express CJK ranges, so B3 (Chinese name + sensitive predicate) relies on the rare predicate verbs alone; C2 (bank-account-shaped digits) is inherently ambiguous. Hard-blocking those would derail every adjourn. Advisory warns let you see the matched category and choose.
Claude Code PreToolUse hooks cannot rewrite tool_input. Sanitization happens one level up: orchestrator reads the warn reminder, generates a sanitized version, and re-issues the Notion call. The hook is the detector, not the rewriter. The orchestration contract for that handoff lives in pro/CLAUDE.md Step 10a.
cd ~/.claude/skills/life_OS && git pull
bash scripts/setup-hooks.sh # registers pre-notion-writeIdempotent โ running twice is a no-op. To uninstall: bash scripts/setup-hooks.sh --uninstall.
For full detail (5 pattern groups, audit schema, JSON parser tiers): see CHANGELOG.md.
Previously in v1.8.2 โ global Obsidian readability HARD RULE (#11), 4 specialized wiki templates (
kind:field),/wiki-obsidian-upgradebatch migrator, binary output redirect to~/Downloads/. See CHANGELOG for the v1.8.2 detail.
v1.8.1 is the largest reduction in Life OS history. Two waves shipped May 1-2, both under the same v1.8.1 tag:
/inbox-processโ drop any.mdinto_meta/inbox/to-process/, then say "ๅค็ inbox" or/inbox-process. ROUTER scans, proposes per-item disposition (acceptโwiki / updateโwiki / archive / reject / defer / merge), waits for confirmation, executes, logs. v1.8.1 Wave 2 added LLM-based duplicate detection (no SHA256, no FTS5) + delta tracking via_meta/inbox/manifest.json./research <topic>โ spawns 5 (or 8 with--depth deep) parallelgeneral-purposesubagents covering academic / practitioner / contrarian / origin / adjacent angles. Synthesizes a SCHEMA-compliant wiki draft with mandatoryCounterpointssection + automatic counter-bias check. v1.8.1 Wave 2 added decoupled CitationAgent (Phase 4, Anthropic pattern). Total wall time โค 9 min with citation verifier.wiki/log.mdactivity timeline convention โ every wiki Write/Edit/move appends one line with action enum (created/updated/promoted/deprecated/merged/renamed/rejected/bulk)./inbox-processand/researchwrite log entries automatically.- Zero-setup vault bootstrap โ the first time you open a Claude Code session in a vault on v1.8.1, the SessionStart hook automatically detects missing scaffolding and creates it silently. You see one
โจ Life OS v1.8.1 vault auto-bootstrap: wrote N filesline, then nothing..obsidian/graph.jsonis auto-patched (with backup) if present.
User feedback driving Wave 2: "ๆๆณๆ่ฟไบไธ่ฅฟๅ จ็ ๆใๆไธ็่งฃไธบไปไน้่ฆๅ้ๆฐๅญฆ๏ผๆ็็ณป็ป้ไนๆฒกๆ FTS5ใ" โ the entire Layer 4 Python tools/ package was carrying complexity (FTS5, vectors, mypy/ruff/pytest CI gates, 50+ optional deps) for capabilities the user never asked for or used. Wave 2 deletes every line of Python in the repo:
- 11 Python modules deleted:
tools/{approval,embed,export,reconcile,research,search,seed,sync_notion,skill_manager,stats,__init__}.pyplustools/lib/{__init__,config,llm,notion,second_brain}.pyand the package README. - 17 pytest test files deleted (matching the deleted modules).
- Packaging deleted:
pyproject.toml,uv.lock,.python-version,.github/workflows/integration.yml, 9evals/scenarios/tool-*.mdeval scenarios. - Security guard now bash-native: the only security-critical Python (~40 dangerous-command pattern guard from
tools/approval.py) was ported inline intoscripts/hooks/pre-bash-approval.shas a bash regex array. 100% pure bash; jq for JSON parsing (python3 fallback for stdin extraction only). Provenance preserved: forked from NousResearch/hermes-agent (MIT) commit59b56d4.... Smoke-tested with 17 fixtures (10 base + 7 edge cases). - CI workflow rewrite:
.github/workflows/test.ymlwas 6 jobs (pytest + mypy + ruff ร 3 OS ร 2 Python). Now 3 jobs (bash -n+ tests/hooks/ + spec drift, across ubuntu/macOS/windows). ~3ร faster, no Python toolchain to install. - Bash-that-wrapped-Python deleted:
scripts/wiki/wiki-link-audit.sh(311 lines of bash + awk frontmatter parsing) replaced byscripts/prompts/wiki-link-audit.md(LLM-driven Glob + Grep + Read).
| Wiki improvement | What changed |
|---|---|
aliases: [] field |
Obsidian uses for [[wikilink]] resolution |
source โ sources: [] |
Plural array; list every contributing URL/citation |
confidence is now a 5-bucket enum |
impossible | unlikely | possible | likely | certain (was 0.0โ1.0 float). Run /migrate-confidence to convert legacy entries; /wiki-decay auto-maps on the fly. |
last_tended field |
ISO date โ last time you actively reviewed (vs cosmetic edit) |
review_by field |
ISO date โ when wiki-decay should re-surface this entry |
| Provenance tags on every fact | ^[extracted] (paraphrased from a source), ^[inferred] (your synthesis), ^[ambiguous] (sources disagree). Required by /research agents and /inbox-process accept/update. |
LLM-based dedup in /inbox-process |
Replaces SHA256 hashing โ catches paraphrased near-duplicates, not just byte-identical. Pure LLM judgment using grep + Read. |
_meta/inbox/manifest.json delta tracking |
Each /inbox-process run marks proposal-table rows as ฮ-new vs carried-over. |
Decoupled CitationAgent in /research |
Anthropic-pattern Phase 4: WebFetch-verifies every ^[extracted] claim against sources[]. Auto-downgrades unverified claims. Confidence drops one bucket if 30%+ unverified. Opt-out with --no-citations. |
Across the 6 major LLM-Wiki / multi-agent research projects we benchmarked against (Anthropic's research-system blog post, LangChain langgraph agent supervisor, GPT-Researcher, CrewAI, QX-Labs, OpenAI Deep Research), no single one combines all of these:
| Axis | Life OS Wiki | Most others |
|---|---|---|
| Persistent backing store | Integrated wiki (markdown + frontmatter), survives sessions | Ephemeral chat output |
| Per-bullet provenance | ^[extracted]/^[inferred]/^[ambiguous] required |
None or run-level only |
| Citation verification | Decoupled CitationAgent runs AFTER synthesis (preserves analytical depth) | Interleaved (drives toward shallow easy-to-cite facts) |
| Confidence calibration | 5-bucket enum forces honest assessment | Float (illusory precision) or unstated |
| Freshness model | last_tended + review_by + /wiki-decay re-surfacing |
Append-only, no decay model |
| Inbox handoff | User triages with full proposal table, confirms before write | Autonomous append (you find out later) |
| Deployment posture | Zero Python (bash + markdown only) | Pip / npm / requirements.txt |
This is not a critique of those projects โ they solve different problems. But if you've tried "LLM wiki" or "multi-agent research" tools and found that (a) the output disappears, (b) you can't tell what's a real source vs a synthesis, or (c) the agent confidently cites URLs that don't say what it claims โ those are the gaps Life OS Wiki is built to close.
- macOS portability:
pre-bash-approval.shhad 5 barepython -cinvocations. macOS 12+ removed barepythonโ hook fail-CLOSED โ blocked every Bash command. R-1.8.0-020 commit title claimed this was fixed; it wasn't until Wave 1. - Scanner false positive:
pre-write-scan.shpattern #5 was blocking legitimate markdown inline code. Tightened to require shell metacharacters inside backticks. - session-start-inbox UX: 2 wrong task names; NEVER_RUN bucket compressed from 8+ lines to 1.
- Notion sync was hardcoded to 4 entities โ now config-driven; reads
_meta/config.md.
cd ~/.claude/skills/life_OS && git pull
bash scripts/setup-hooks.sh # installs new /inbox-process + /research + /migrate-confidence + /wiki-link-auditThen open any Claude Code session inside your second-brain vault. Vault scaffolding is auto-created. Existing wiki entries with legacy float confidence keep working โ /wiki-decay auto-maps. To permanently migrate to the enum, run /migrate-confidence from your vault (idempotent, with proposal preview).
For users who had python -m tools.<X> invocations: those Python entrypoints no longer exist. See CHANGELOG.md for the full replacement table.
v1.8.0 originally shipped with cron autonomy + always-on Cortex. After two days of production testing the cron architecture failed every reliability test (silent data loss, LLM-in-cron permission stalls, stale script paths, multiple bash compatibility bugs). v1.8.0 was pivoted in-place (same version tag) to a simpler design: user invokes everything, ROUTER does the work directly.
Core principle: cron required determinism, LLMs are non-deterministic โ that mismatch can't be patched. Replace cron with explicit user prompts. The user types "rebuild index" or "monthly review", ROUTER reads scripts/prompts/<job>.md and executes inline. No background processes. Nothing runs without you watching.
Two session modes:
- Mode 1 ยท Business session โ your standard Claude Code chat. Long-lived: a session can span days/weeks. ไธๆ/้ๆ are optional soft triggers. Cortex is now pull-based (ROUTER decides per-message whether to launch hippocampus / concept-lookup / soul-check / gwt-arbitrator) instead of always-on.
- Mode 2 ยท Monitor session (
/monitor) โ view-and-invoke operations console. Shows maintenance task timestamps + recent reports + action items. You say "่ท X" / "้ฝ่ท", monitor reads the matching prompt and executes. No cron, no background.
10 user-invoked maintenance jobs (each is a markdown prompt at scripts/prompts/<job>.md that ROUTER reads + executes via Read/Write/Bash):
reindexยทdaily-briefingยทbackupยทspec-complianceยทwiki-decay(the v1.7.x "python tool" jobs, now LLM-driven)archiver-recoveryยทauditor-mode-2ยทadvisor-monthlyยทeval-history-monthlyยทstrategic-consistency(the v1.8.0 "prompt cron" jobs, now user-invoked)
Hooks (only 1 fires automatically):
session-start-inboxโ at session start, scans the 10 maintenance jobs' last-run timestamps and shows a one-line "what's overdue" status. Does not execute anything; you decide what to invoke.pre-prompt-guardโ memory keyword auto-detect + ไธๆ/้ๆ soft trigger. Cortex always-on enforcement REMOVED.pre-bash-approval(kept) โ security gate against dangerous bash.post-task-audit-trail(weakened) โ only enforces R11 audit trail for archiver + knowledge-extractor (Cortex agents no longer required to write trails).
Removed in pivot:
- Cron infrastructure:
scripts/setup-cron.sh,scripts/run-cron-now.sh,scripts/commands/run-cron.md,tools/missed_cron_check.py,tools/cron_health_report.py, all launchd plists. - Python middleware:
tools/memory.py(now Write/Read directly to~/.claude/lifeos-memory/),tools/session_search.py(now Grep directly),tools/cli.py(no longer needed), 5 maintenance python tools (replaced by user-invoked prompts above). - Cortex artifacts:
pro/agents/narrator-validator.md(validator was tied to always-on flow). - Spec docs:
references/automation-spec.md,references/session-modes-spec.md,docs/architecture/hermes-local.md(cron-era specs).
Migration: re-pull the repo, then re-run bash ~/.claude/skills/life_OS/scripts/setup-hooks.sh (re-registers the simplified hook set). On macOS, launchctl unload ~/Library/LaunchAgents/com.lifeos.hermes-local.*.plist && rm ~/Library/LaunchAgents/com.lifeos.hermes-local.*.plist to remove the dead cron jobs. No second-brain data migration required. v1.7.x sessions / wiki / SOUL fully compatible.
v1.7.3 (Cortex always-on hook injection + narrator-validator + 4 slash commands wired) and earlier v1.7.x release notes have been consolidated into
CHANGELOG.md. v1.7.3's "always-on Cortex" behavior was REVERSED in v1.8.0 R-1.8.0-011 pivot โ current architecture is pull-based (see "What's New in v1.8.0" above). The README only describes the CURRENT architecture; previous designs live in CHANGELOG so this section doesn't mislead users.
v1.7.2.3 clarifies Mode 0 ownership for RETROSPECTIVE. ROUTER now owns the Bash-rendered briefing skeleton and pre-renders roughly 80% of the report, while the subagent contributes only the <!-- LLM_FILL: today_focus_and_pending_decisions --> block with a short Today's Focus and Pending Decisions section. ROUTER splices that block back into the skeleton, keeping the briefing grounded, compact, and easier to audit.
v1.7.2.1 is a subtraction hotfix: it restores the theme aesthetic while removing extra reporting ceremony. The user-visible report shape is reduced from 17 H2 blocks to 6, the compressed wrapper requirement is removed, and version markers stay in fixed positions for easier checking. The result is fewer rules, cleaner briefings, and the same auditability without the visual clutter.
v1.7.2 turns the local execution surface into a clearer user story: Hermes Local is now the public name for the safeguards and automation that make Life OS enforceable outside the prompt. It covers Layer 3 hooks and Layer 4 Python tools, while internal specs keep the stable execution layer, Layer 3, and Layer 4 labels. Cortex is now described as an always-on cognitive path for enabled workspaces: every user message can receive pre-router memory, concept, and SOUL signals, with graceful degradation when indexes or subagents are unavailable. The version-check hook now invalidates its daily cache against the remote SHA and supports --force, so same-day updates are no longer hidden by stale cache output. Hermes-derived prompt-cache and context-compression helpers improve speed and help large pasted transcripts stay manageable. Compression is only for local context management; subagent reports and audit evidence still stay literal where the workflow requires full fidelity.
v1.7.1 is a hardening release for transparency and evidence handling. Token use is now surfaced more explicitly so users can see why work was run, skipped, or escalated. ROUTER must paste subagent output without compressing it, preserving the full report trail for review. AUDITOR checks are more programmatic, and the release groups 27 fixes across hooks, i18n drift, Cortex emission, GWT clarity, DREAM output, force-push handling, marker disambiguation, and markdown frame resolution. R10 architectural shift: 11 of 18 retrospective steps moved from LLM to ROUTER Bash. LLM compliance gap closed by program substitution, not more spec rules. R11 adds runtime audit trail files so AUDITOR can verify across subagents directly, breaking the information-isolation bottleneck without exposing agent reasoning. R12: every 'ไธๆ' is fresh โ LLM cannot reuse previous briefing. Forbidden phrases + length collapse + missing fresh markers all trigger C-fresh-skip P0.
Patch update: final briefing contracts are now explicit, Mode 0 self-checks Claude Code hooks, and Cortex is OFF / opt-in via _meta/config.md. Hook auto-install closes the test-machine deployment gap.
Anti-confabulation hardening prevents fabricated failure explanations from reaching users.
Source-grounded briefings now include PRIMARY-SOURCE measured-count markers, STATUS.md staleness suppression, automatic 30d-โฅ3 Compliance Watch banners, and ROUTER Bash fact-checks for numeric/version/path claims before user display.
Cortex Cognitive Layer โ GA release
- 5 new Cortex capabilities: cross-session memory (hippocampus), signal arbitration (GWT), cited generation (narrator), concept graph, SOUL dimension detection
- 5 runtime hooks enforcing HARD RULEs (prevents COURT-START-001 class of violations)
- 10 Python tools + 3 libs (reindex / reconcile / stats / research / daily_briefing / export / sync_notion / seed / migrate / search / embed)
- 6 Cortex user-guides + v1.7-migration UX chapter
- cortex-spec + hippocampus-spec translated to Chinese and Japanese
Upgrade (v1.6 โ v1.7): see docs/guides/v1.7-migration.md. The previous uv run life-os-tool migrate command was removed in R-1.8.0-011 along with the life-os-tool dispatcher; current invocation uses scripts/prompts/migrate-from-v1.6.md (LLM-driven) per pro/CLAUDE.md ยง0.5.
See CHANGELOG for the full v1.7 commit chain and the COURT-START-001 v1.6.3 incident archive.
Trust guard โ five-layer defense against HARD RULE violations. In testing, the author said "ไธๆ" in the Life OS dev repo and the LLM bypassed the retrospective subagent, simulated 18 steps inline, and fabricated non-existent paths as authority. Documentation alone doesn't enforce anything โ every HARD RULE was descriptive, with zero real enforcement. v1.6.3 ships five independent layers so every trigger word actually launches a real subagent:
- UserPromptSubmit hook (
scripts/lifeos-pre-prompt-guard.sh) โ detects ไธๆ / start / ้ฃ่ญฐ้ๅง / ้ๆ etc. across all 9 themes, injects HARD RULE reminder into the model's context before any response - Pre-flight Compliance Check โ ROUTER must output
๐ Trigger: [word] โ Theme: [name] โ Action: Launch([agent]) [Mode]before any tool call; missing line = logged violation - Subagent self-check โ retrospective Mode 0 first line proves the subagent was actually launched (not main-context simulation)
- AUDITOR Compliance Patrol (Mode 3) โ 7-class violation taxonomy (A1 skip subagent, A2 skip directory check, A3 skip Pre-flight, B fabricate, C incomplete phase, D placeholder, E main-context phase) runs after every session start and archive
- Eval regression โ
evals/scenarios/start-session-compliance.mdcodifies all 6 COURT-START-001 failure modes
Dual-repo violations log (md + git, per user's storage constraint): violations persist to pro/compliance/violations.md (dev repo, public) and _meta/compliance/violations.md (user second-brain, private). Escalation ladder: โฅ3 same type in 30 days โ stricter hook reminder; โฅ5 โ briefing ๐จ Compliance Watch; โฅ10 in 90 days โ AUDITOR patrol every session.
Still current from v1.6.2: Bulletproof adjourn ยท Wiki auto-writes ยท SOUL auto-writes ยท DREAM 10 auto-triggers ยท SOUL trend arrows ยท REVIEWER 3-tier SOUL strategy ยท SOUL Health Report in briefing.
v1.6.3a hot patch (2026-04-21) โ closes the Layer 1 install gap.
scripts/setup-hooks.shnow auto-registers the UserPromptSubmit hook (run once:bash ~/.claude/skills/life_OS/scripts/setup-hooks.sh). Hook regex tightened (first-line + length checks) to reduce false positives on pasted content. New Class F (false positive) added to violation taxonomy.
See CHANGELOG for the full list and the original COURT-START-001 incident archive.
Life OS rests on five pillars. The decision engine is the core โ everything else grows from it.
The engine runs multiple agents organized around a principle that is 1,400 years old: no single voice goes unchecked. The theme gives those agents names from your culture. The logic is always the same.
| Agent | Function |
|---|---|
| ROUTER | Your primary entry point โ handles casual chat, routes complex matters to the engine |
| PLANNER | Breaks your situation into 3-6 dimensions and builds a structured plan |
| REVIEWER | Independent review with emotional audit, 10/10/10 regret test, SOUL consistency check, red-team challenge, and veto power (max 2 rounds) |
| DISPATCHER | Detects dependencies between domains, dispatches parallel or sequential execution |
| PEOPLE | Relationships, stakeholders, team dynamics |
| FINANCE | Money, assets, risk exposure |
| GROWTH | Learning, personal brand, communication |
| EXECUTION | Action plans, logistics, timelines |
| GOVERNANCE | Rules, risk, compliance, self-discipline |
| INFRA | Health, energy, living environment, digital infrastructure |
| AUDITOR | Audits the agents' work quality after every flow |
| ADVISOR | Audits you โ surfaces behavioral patterns across your decisions |
| COUNCIL | Cross-domain debate โ auto-triggers when domain scores differ by 3+ points |
| RETROSPECTIVE | Session start: sync, briefing, strategic overview |
| ARCHIVER | Session close: archive, knowledge extraction, DREAM, sync |
| STRATEGIST | Hall of Wisdom โ 93 thinkers across 18 domains, 3 dialogue modes |
Every major decision passes through three stages. No shortcuts.
Draft โ The PLANNER breaks your situation into dimensions and builds a plan.
Review โ The REVIEWER examines the plan independently. It runs an emotional audit: Is fear driving this? Will your family support it? Will you regret this in 10 minutes, 10 months, 10 years? It checks alignment with your SOUL (values archive), your wiki (established knowledge), and your strategic map (cross-project impact). It red-teams the weakest assumption. If it finds a blind spot, it sends a veto โ the plan goes back for revision. Maximum two veto rounds; the third pass must be approved.
Execute โ Six domain analysts each score the plan 1-10 from their domain, independently:
| Domain | What it covers | The question it asks |
|---|---|---|
| People | Relationships, stakeholders | "Are the right people involved?" |
| Finance | Money, assets, resources | "Can you afford this โ including the worst case?" |
| Growth | Learning, expression, culture | "What do you need to learn first?" |
| Execution | Action, logistics, timelines | "What's the concrete plan, week by week?" |
| Governance | Rules, risk, compliance | "What happens if everything goes wrong?" |
| Infrastructure | Health, energy, environment | "Can your body and environment sustain this?" |
Each domain has four specialized divisions. Finance, for example, has Income (salary, side income), Spending (budgets, habits), Assets (investments, real estate), and Reserves (emergency fund, insurance, retirement). The DISPATCHER routes to the right divisions automatically.
After all six report, the REVIEWER does a final review โ and if scores conflict by 3+ points, the COUNCIL convenes a structured three-round debate. Then the composite Summary Report is produced. The AUDITOR checks the agents' work ("The execution plan has no milestones past month 3 โ flag it"). The ADVISOR turns to you: "You've avoided addressing finances in your last four decisions. Why?"
Here is what the full flow looks like:
You: "I'm thinking about leaving my job to start something new."
Draft
๐ Planner โ Breaks it into 6 dimensions, builds the plan
Review
๐ Reviewer โ Emotional audit: running away or running toward?
10/10/10 test: will you regret this in 10 years?
SOUL check: aligned with your stated values?
Veto: "Financial runway not addressed. Revise."
Execute (after revision passes review)
๐ฅ People 7/10 "Co-founder chemistry is untested"
๐ฐ Finance 5/10 "18 months runway, but only if nothing goes wrong"
๐ Growth 8/10 "Strong domain expertise, credibility is real"
โ๏ธ Execution 6/10 "No milestone plan past month 3"
โ๏ธ Governance 4/10 "Non-compete clause needs legal review"
๐๏ธ Infra 7/10 "Health is good, but stress plan is vague"
Audit
๐ฑ Auditor โ "Execution plan is vague past month 3 โ request revision"
๐ฌ Advisor โ "You've been consuming startup content for weeks.
Confirmation bias is likely. When did you last
seriously consider staying?"
๐ Final Report โ Composite: 6.2/10 โ Proceed with conditions
Not every request needs a full court process. The ROUTER handles casual chat, quick questions, and emotional support directly. For questions that need domain expertise but are not decisions โ say, "what tax rules apply to freelancers?" โ an express path sends it to 1-3 relevant domain analysts directly, skipping the PLANNER, REVIEWER, AUDITOR, and ADVISOR. Quick answer, then: "This was an express analysis. Want the full deliberation?"
Some questions do not have a "correct answer." They need perspective.
The STRATEGIST gives you access to 93 of history's greatest thinkers across 18 domains โ from Socrates to Buffett, Laozi to Mandela, Dostoevsky to Feynman. Each one runs as a real subagent with their own voice, their own examples, their own way of pushing you.
Three modes:
- One-on-one โ Deep dialogue with a single thinker. Socrates will not let you off easy.
- Roundtable (2-4 thinkers) โ Multiple thinkers discuss your question, each from their worldview. Watch Seneca and Wang Yangming find unexpected common ground.
- Debate (2 thinkers) โ Two thinkers with opposing views argue directly over three structured rounds. You judge.
You: "I keep starting things and never finishing them."
Strategist: Recommended โ Seneca (on time) + Wang Yangming (on action)
Seneca: "You do not lack time. You waste it on things you have not
examined. Which of your current pursuits would you begin again,
knowing what you know now?"
Wang Yangming: "Knowledge and action are one. If you truly knew what you
wanted, you would already be doing it. The gap between
knowing and doing is the gap between wanting and truly
wanting."
โ Parting words from each thinker
โ Summary of your thinking journey saved to your knowledge base
18 domains: Philosophy, Eastern Thought, Science, Strategy, Business, Psychology, Systems, Human Nature, Civilization, Adversity, Aesthetics, Politics, Economics, Mathematics, Medicine, Exploration, Communication, Law.
You can name anyone not on the built-in list โ any historical figure โ and the STRATEGIST will honor the request with equal depth. If your SOUL archive exists, the STRATEGIST uses it to recommend thinkers who address your specific tensions and contradictions.
Every decision, insight, pattern, and action item is written to a persistent knowledge base โ structured markdown files that you own, in a storage system you choose.
second-brain/
โโโ SOUL.md # Who you are โ values, identity, aspirations
โโโ user-patterns.md # How you behave โ the advisor's observations
โโโ inbox/ # Quick captures from your phone
โโโ _meta/
โ โโโ STATUS.md # Global status dashboard
โ โโโ STRATEGIC-MAP.md # Relationships between projects
โ โโโ strategic-lines.md # Strategic line definitions
โ โโโ journal/ # Session reports, DREAM reports
โ โโโ outbox/ # Session staging area (one subdirectory per session)
โโโ projects/{name}/ # Active projects with tasks + decisions
โโโ areas/{name}/ # Ongoing life areas with goals
โโโ wiki/ # Reusable knowledge โ grows from DREAM
โโโ archive/ # Completed work
Three storage backends โ pick what fits your life:
| Backend | Best for | Tradeoff |
|---|---|---|
| GitHub | Version control, works with Obsidian | Requires basic Git knowledge |
| Google Drive | Zero setup, just works | Less structured |
| Notion | Mobile-friendly, database views | Best for phone capture |
You can use one, two, or all three. Multi-backend: writes go to all selected backends, reads come from the primary (auto-priority: GitHub > Google Drive > Notion). Conflict resolution follows last-write-wins with timestamp comparison.
Cross-device sync: Capture a thought on your phone (Notion inbox) at lunch. When you sit down at your computer and start a session, the system pulls it in, processes it, and syncs results back.
Parallel sessions: Work on project-alpha in one terminal window, project-beta in another. Each session writes to its own outbox directory. The next time you start a session, everything merges cleanly โ no conflicts, no locks.
First-run setup: On your very first session, the system detects that no second-brain exists and walks you through creating one โ choose your backend(s), and the full directory structure is created automatically.
You are good at thinking about individual projects. You are probably bad at seeing how they connect โ which ones feed into each other, which ones compete for your time, and what happens to the rest when one stalls.
Strategic Map adds the relationship layer.
Strategic Lines โ Group projects by the purpose they serve. Each line has a stated purpose and a driving force (what actually motivates you โ these can differ, and that tension is worth examining). Health signals define what to watch. Multiple projects can serve one line with different roles: critical-path (if this stalls, the line stalls), enabler (must complete first), accelerator (makes it faster), insurance (plan B).
Flow Graph โ Define what flows between projects: knowledge, deliverables, decisions, relationship capital. When a decision in one project invalidates another project's assumptions, the system flags it immediately. When knowledge flows are defined but no wiki entries actually carry the knowledge, that is a broken flow.
Health Archetypes โ No abstract numerical scores. The system matches each project to a pattern โ steady progress, controlled wait, momentum decay, uncontrolled stall, direction drift, or dormant โ and writes a narrative: what is happening, what it means, what to do.
Blind Spot Detection โ Actively looks for what is missing: unaffiliated projects (not assigned to any strategic line), broken flows (defined but not flowing), driving force neglect (behavior misaligned with what you say matters), dimension gaps (entire life areas absent from all strategic lines), and approaching deadlines with no preparation.
Your morning briefing becomes strategic:
๐บ๏ธ Strategic Overview
๐ฐ market-expansion ๐ก Controlled wait
project-alpha critical-path โธ on-hold (legal review)
project-beta enabler ๐ข active
The legal review creates a natural window.
โ Push project-beta prep work (2-3h) โ high leverage, low risk.
โก Today
๐ฅ Push project-beta prep โ exploit the waiting period
๐ข Safe to ignore: project-gamma (on track), side-project (non-critical)
โ Decide: project-delta is unaffiliated โ which strategic line does it serve?
Strategic Map integrates with SOUL (are your driving forces aligned with your values?), Wiki (do the knowledge flows actually carry real knowledge?), and DREAM (the sleep cycle uses the flow graph to discover cross-layer insights).
Strategic Map grows from zero. If you have not defined any strategic relationships, the system operates normally with a flat project list. After a few sessions with multiple projects, DREAM may propose: "You have N active projects but no strategic relationships defined. Would you like to map how they relate?"
SOUL records who you are โ not what you do, but what you value, what you believe, and who you aspire to be. Each entry has two sides: what IS (observed from your decisions) and what SHOULD BE (your stated aspiration). The gap between them is where growth happens.
Auto-writes, with you in control (v1.6.2). The system no longer asks you to confirm every entry. ADVISOR runs after every decision and increments evidence or challenges on existing SOUL dimensions. When a new value pattern accumulates 2+ pieces of evidence, SOUL auto-writes a new dimension at low confidence (0.3) โ with the "What SHOULD BE" field deliberately left empty for you to fill in when you're ready. You stay in charge: edit freely, delete dimensions that don't fit, or say "undo recent SOUL" to roll back.
Every session starts with a SOUL Health Report โ fixed position at the top of the briefing. Current profile with trend arrows, newly auto-detected dimensions awaiting your input, conflict warnings (dimensions your last 3 decisions all challenged), and dormant dimensions (30+ days without activation). You see the system's model of you every single time.
REVIEWER references SOUL in every decision. If a decision challenges a stated value, it surfaces the contradiction instead of rubber-stamping.
DREAM is the AI sleep cycle. After every session ends, the system "sleeps" โ inspired by human sleep architecture:
- Light sleep (N1-N2) โ Organize loose ends: classify unprocessed inbox items, flag expired tasks, detect orphan files
- Deep sleep (N3) โ Consolidate: extract Wiki knowledge and update SOUL dimensions from the last 3 days of activity
- REM (creative connections + 10 auto-triggers) โ Discovers cross-domain links you have not noticed, and automatically acts on 10 specific patterns: new project relationships, behavior diverging from stated values, wiki contradictions, dormant SOUL dimensions, unused cross-project cognition, decision fatigue, value drift, stale commitments ("30 days ago you said you would do X โ what happened?"), emotional-state decision clusters, and repeated identical decisions ("You're deciding X for the 4th time โ are you avoiding commitment?")
All triggered actions flow into the next session's briefing in a fixed "DREAM Auto-Triggers" block. Next morning: "Last session I noticed you're deciding the contract question again โ this is the 3rd time. What's actually blocking you?"
Wiki captures reusable knowledge about the world โ not about you. After every session, the ARCHIVER auto-writes wiki entries that pass all 6 strict criteria: cross-project reusable, about the world (not you), zero personal privacy (no names, amounts, IDs, or traceable details โ if stripping privacy makes the conclusion meaningless, it's discarded), factual/methodological, โฅ2 independent evidence points, no contradiction with existing entries. Personal material belongs in SOUL; reusable knowledge belongs in Wiki. The two never mix.
All three systems grow from zero. On day one, the system knows nothing about you. It learns only from your decisions and observations โ and it now learns continuously, not just when you ask.
You: Start session.
๐
Session Start:
Pick your theme: a-i (3 per language โ historical / government / corporate)
You: b
๐
ๅฎไพ้ฃ่ญฐ:
Syncing second-brain... 3 inbox items pulled from Notion.
๐ฅ "Look into certification programs" โ captured yesterday on phone
๐ฅ "project-alpha: supplier replied" โ forwarded from email
๐ฅ Quick note: "revisit budget assumptions"
๐บ๏ธ Strategic overview: [see Strategic Map above]
๐ค DREAM report: Last session noticed your wiki entry on negotiation
tactics could apply to the supplier conversation in project-alpha.
๐ Recommended: Process supplier reply first (time-sensitive).
Behind the scenes, the session boot sequence ran 18 steps: theme resolution, directory detection, data layer check, git health check, full sync pull, outbox merge, platform and version check, project binding, context loading (user-patterns, SOUL, STATUS, lint state, project context, global overview), Strategic Map compilation, DREAM report presentation, wiki health check, and the final briefing.
You: I'm considering switching from full-time to freelance.
โ 2-3 rounds of intent clarification (cannot be skipped โ HARD RULE)
โ Full engine flow: draft โ review (with possible veto) โ execute (6 domains
in parallel, each displayed as it completes) โ final review โ Summary Report
โ audit โ advisor
โ Report: 5.8/10 โ "Viable but timing is premature. Revenue runway
is 11 months, not the 18 you assumed. Recommendation: build 3 more
months of savings and one anchor client before transitioning."
You: I keep saying yes to things I don't care about.
โ ROUTER detects abstract thinking need, asks: "Would you like to activate
the Strategist?"
โ One-on-one with Marcus Aurelius on priorities and refusal
โ Parting words, journey summary, insights saved to second-brain
You: End session.
๐ Archiver:
Phase 1 โ Archive: decisions, tasks, journal โ outbox
Phase 2 โ Knowledge extraction (auto-write under strict criteria):
๐ฎ SOUL auto-written: "Values autonomy over stability" โ confidence 0.3
(evidence: 2 decisions this session; "What SHOULD BE" left empty for you)
๐ Wiki auto-written: "Freelance runway formula" โ wiki/career/
(passes 6 criteria + privacy filter: zero personal details)
โ 1 wiki candidate discarded โ contained personal amount, couldn't strip
Phase 3 โ DREAM:
๐ค N1-N2: 2 inbox items need classification
๐ค N3: new evidence for "deliberate decision-maker" dimension (+1)
๐ค REM: ๐จ Stale commitment detected โ 32 days ago you said you would
draft the freelance plan. Triggered for next session's briefing.
Phase 4 โ Sync: git push... Notion sync... done.
โ
Completion checklist verified. Session archived.
โฉ๏ธ To undo any auto-write: delete the file, or say "undo recent wiki/SOUL"
next session.
Life OS comes pre-configured for the decisions people actually face:
| # | Scenario | Domains involved | What the reviewer asks |
|---|---|---|---|
| 1 | Career transition | All Six | "Running away or pursuing something?" |
| 2 | Investment decisions | Finance, Execution, Governance, People | "FOMO or rational? Can you survive total loss?" |
| 3 | Relocation | All Six | "Do you really know the destination?" |
| 4 | Annual planning | All Six | "Too many goals? Measurable? Aligned with values?" |
| 5 | Startup decisions | All Six | "Solving a real pain point? Are you the right person?" |
| 6 | Major purchases | Finance, Execution, Governance | "Need or want? Would you still want it in a month?" |
| 7 | Relationships | People, Infra, Governance, Growth | "Are you evaluating the other person with bias?" |
| 8 | Periodic reviews | Retrospective | Daily, weekly, monthly, quarterly, yearly |
| 9 | Health management | Infra, Execution, Finance, Governance | "Sustainable, or another short burst?" |
| 10 | Learning plans | Growth, Execution, Finance, People | "Learning for growth, or avoiding real work?" |
| 11 | Time management | Execution, Finance, Governance, Infra | "Really no time, or avoiding something?" |
| 12 | Major family decisions | All Six | "Whose voice hasn't been heard?" |
Life OS installs in one command. It requires a Pro Mode terminal โ that means real subagents running in parallel with information isolation, not a chatbot.
| Platform | Command |
|---|---|
| Claude Code | /install-skill https://github.com/jasonhnd/life_OS |
| Gemini CLI / Antigravity | npx skills add jasonhnd/life_OS |
| OpenAI Codex CLI | npx skills add jasonhnd/life_OS |
On first start, you pick your theme. The system auto-detects your language and recommends a match, but the choice is always yours. You can switch at any time by saying "switch theme."
First run: The system detects that no second-brain exists and walks you through setup โ pick your storage backend(s), and the full directory structure is created automatically. On subsequent sessions, the system detects what kind of directory you are in: Life OS system repo (development), second-brain (normal use), or a project repo (connects to configured second-brain path).
Set up auto-updates (Claude Code):
bash ~/.claude/skills/life_OS/scripts/setup-hooks.shThis checks for updates once a day when you start a session.
After bash scripts/setup-hooks.sh, life_OS auto-registers its Task-spawnable agents under ~/.claude/agents/lifeos-*.md. Claude Code then recognizes calls such as Task(lifeos-retrospective) and Task(lifeos-archiver) as first-class targets instead of falling back to general-purpose.
The lifeos- prefix avoids collisions with other skills. Wrappers point at the canonical definitions under pro/agents/*.md in the skill, so updating the skill and rerunning setup refreshes agent behavior. There are multiple agents definition files; 21 are Task-spawnable wrappers, while narrator.md remains ROUTER-internal.
Uninstall: bash scripts/unregister-claude-agents.sh.
Manual update: Say "update" (or "ๆดๆฐ" or "ใขใใใใผใ") in any session.
Not supported: ChatGPT, Gemini Web, or any single-context chat interface. Life OS requires multiple independent subagents with true information isolation โ a single chat window cannot do this.
For detailed setup including storage backend configuration, see the full installation guide.
๐ You
โ
โโ ๐จ Theme Layer
โ 9 themes across 3 languages (3 per language: historical / government / corporate)
โ zh: ไธ็ๅ
ญ้จ ยท ไธญๅฝๆฟๅบ ยท ๅ
ฌๅธ้จ้จ
โ ja: ๆๆฒปๆฟๅบ ยท ้ใ้ข ยท ไผๆฅญ
โ en: Roman Republic ยท US Government ยท C-Suite
โ Maps the functional IDs โ display names, tone, trigger words
โ One file per theme (~60 lines). Adding a new theme = one new file.
โ
โโ โ๏ธ Decision Engine (multiple agents, culture-neutral)
โ โ
โ โโ ๐๏ธ ROUTER โ Daily entry point
โ โ Direct handling: casual chat, emotional support, quick questions
โ โ Express ๐: 1-3 domains for non-decision analysis
โ โ Full deliberation โ๏ธ: 2-3 rounds of intent clarification โ three stages
โ โ Detects confusion/values questions โ offers STRATEGIST
โ โ
โ โโ Three Stages โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ โ ๐ PLANNER (Draft)
โ โ Break into 3-6 dimensions, assign domains, set quality criteria
โ โ Reference SOUL for value dimensions user didn't mention
โ โ Check Strategic Map: does this affect downstream projects?
โ โ
โ โ ๐ REVIEWER (Review โ has veto power)
โ โ ๐ฐ Emotional audit: fear? impulse? avoidance?
โ โ ๐จโ๐ฉโ๐ง Relationship impact: how will family react?
โ โ ๐ฎ SOUL alignment: contradicts your stated values?
โ โ โฐ Regret test: 10 minutes / 10 months / 10 years?
โ โ ๐ฏ Red team: assume it fails โ weakest assumption?
โ โ ๐บ๏ธ Strategic propagation: invalidates downstream premises?
โ โ ๐ซ Veto โ back to PLANNER (max 2 rounds)
โ โ
โ โ ๐จ DISPATCHER (Dispatch)
โ โ Detect data dependencies โ parallel or sequential execution
โ โ Inject wiki known premises: "start from these conclusions"
โ โ
โ โ Six Domains (parallel execution, independent scoring 1-10)
โ โ ๐ฅ PEOPLE โ relationships, stakeholders, team dynamics
โ โ ๐ฐ FINANCE โ income, spending, assets, reserves
โ โ ๐ GROWTH โ learning, personal brand, expression, cross-cultural
โ โ โ๏ธ EXECUTION โ project mgmt, tools, research, energy
โ โ โ๏ธ GOVERNANCE โ legal, audit, discipline, info security
โ โ ๐๏ธ INFRA โ fitness, housing, digital infrastructure, routines
โ โ Each domain has 4 specialized divisions (24 total)
โ โ
โ โ ๐ REVIEWER (Final review) โ ๐ Summary Report
โ โ ๐ฑ AUDITOR (Audit agent work quality)
โ โ ๐ฌ ADVISOR (Audit YOUR behavioral patterns)
โ โ
โ โโ ๐๏ธ COUNCIL โ Cross-domain debate
โ โ Auto-triggers when domain scores differ by โฅ3 points
โ โ Manual trigger: "debate" / theme equivalent
โ โ 3 structured rounds: position โ rebuttal โ final statement
โ โ Moderator assessment + recommendation (not decision)
โ โ
โ โโ ๐
RETROSPECTIVE โ Session start (18 steps)
โ โ Step 1: ๐จ Theme selection (trigger word inference or a/b/c)
โ โ Step 2: ๐ Directory type detection (system repo / second-brain / project)
โ โ Step 3: ๐ฆ Data layer check (first-run โ create directory structure)
โ โ Step 4-7: ๐ Sync (config โ git health โ full pull โ outbox merge)
โ โ Step 8-9: ๐ Version check + project binding
โ โ Step 10-14: ๐ Context loading (patterns โ SOUL โ STATUS โ project โ overview)
โ โ Step 15: ๐บ๏ธ Strategic Map compilation (archetype matching + narrative + actions)
โ โ Step 16: ๐ค DREAM report (present last session's discoveries + candidates)
โ โ Step 17: ๐ Wiki health check (compile INDEX)
โ โ Step 18: ๐ Generate briefing (strategic overview + โก today's actions + metrics)
โ โ
โ โโ ๐ ARCHIVER โ Session close (4 phases)
โ โ Phase 1 ๐ฆ Archive: decisions / tasks / journal โ outbox
โ โ Phase 2 ๐ Knowledge extraction (core mission ยท v1.6.2 auto-write):
โ โ ๐ Wiki โ auto-written if passes 6 criteria + privacy filter
โ โ ๐ฎ SOUL โ auto-written at confidence 0.3 if โฅ2 evidence
โ โ ๐บ๏ธ Strategic relationship candidates โ user confirms on the spot
โ โ ๐ last_activity auto-update for touched projects
โ โ โฉ๏ธ User nudges post-hoc: delete file or "undo recent wiki/SOUL"
โ โ Phase 3 ๐ค DREAM (AI sleep cycle):
โ โ N1-N2 ๐ญ Light sleep: organize inbox, flag expired tasks
โ โ N3 ๐ง Deep sleep: consolidate Wiki knowledge + SOUL updates
โ โ REM ๐ Dreaming: creative connections + 10 auto-triggered actions
โ โ ยท Stale commitments, value drift, decision fatigue, repeated decisions...
โ โ ยท SOUL ร strategy: driving forces aligned with values?
โ โ ยท Wiki ร flows: knowledge actually transferring between projects?
โ โ ยท Patterns ร priorities: avoiding a critical-path project?
โ โ Phase 4 ๐ Sync: git push + Notion (4 operations)
โ โ โ
Completion checklist: every item must have a concrete value
โ โ
โ โโ ๐ STRATEGIST โ Hall of Human Wisdom
โ 93 thinkers across 18 domains
โ Socrates ยท Laozi ยท Buffett ยท Mandela ยท Feynman ยท Wang Yangming โฆ
โ ๐ฃ๏ธ One-on-one: deep dialogue with one thinker
โ ๐ช Roundtable (2-4): multi-perspective discussion
โ โ๏ธ Debate (2): opposing views, 3 rounds, you judge
โ Each thinker is an independent subagent with their own voice
โ Ending: parting words โ thinking journey saved to knowledge base
โ
โโ ๐พ Storage Layer
GitHub / Google Drive / Notion (pick 1-3)
โโโ SOUL.md ๐ฎ Personality archive (grows from zero)
โโโ user-patterns.md ๐ Behavioral patterns (ADVISOR observations)
โโโ _meta/
โ โโโ STATUS.md ๐ Global status dashboard
โ โโโ STRATEGIC-MAP.md ๐บ๏ธ Strategic relationship map
โ โโโ journal/ ๐ Reports + DREAM logs
โ โโโ outbox/ ๐ฎ Session staging
โโโ projects/ ๐ฏ Active projects with tasks + decisions
โโโ areas/ ๐ Ongoing life areas with goals
โโโ wiki/ ๐ Reusable knowledge (grows from DREAM)
โโโ archive/ ๐๏ธ Completed work
Each domain has four specialized divisions:
| Domain | Divisions |
|---|---|
| People | Talent (identifying people, evaluating partners), Evaluation (relationship health, social ROI), Relations (cultivation, reciprocity, important dates), Allocation (team building, delegation, family labor) |
| Finance | Income (salary, side income, passive channels), Spending (budgets, habits, subscriptions), Assets (investments, crypto, real estate), Reserves (emergency fund, insurance, tax, retirement) |
| Growth | Education (learning roadmap, skills, certifications), Image (personal brand, social presence), Writing (content planning, speech prep), Diplomacy (cross-cultural communication, networking) |
| Execution | Operations (project planning, task decomposition, deadlines), Equipment (tools, hardware, dev environment), Intelligence (industry research, competitive analysis), Logistics (energy management, workflow, procrastination) |
| Governance | Law (legal risk, contracts, IP, compliance), Audit (decision reviews, time audits, failure analysis), Discipline (bad habits, commitment tracking, self-deception), Defense (information security, privacy, scam detection) |
| Infrastructure | Fitness (exercise, diet, sleep, mental health), Housing (living space, workspace, renovation), Digital (knowledge base, servers, backup, automation), Routines (daily rhythm, morning/bedtime procedures) |
When domains overlap, jurisdiction follows root cause: body illness goes to Fitness, broken rhythm goes to Routines, work inefficiency goes to Logistics. If the lead and assisting domains disagree, the COUNCIL resolves it.
How information flows through the system โ from a thought on your phone to an insight you never expected:
Perceive โ Capture โ Judge โ Settle โ Associate โ Strategize โ Emerge
(phone) (inbox) (engine) (SOUL) (wiki) (strat-map) (DREAM)
Perceive and Capture happen on mobile โ zero-friction capture to inbox. Judge happens on desktop โ the decision engine runs the full Draft-Review-Execute cycle. Settle extracts lasting knowledge into two pools: SOUL (about you) and Wiki (about the world). Associate turns accumulated knowledge into active context โ when a new topic arrives, the system already knows what you know. Strategize adds the relationship layer โ per-project analysis becomes strategic-line-aware analysis. Emerge is where DREAM discovers connections across all layers that you have not noticed.
4 security boundaries:
- No destructive operations (file deletion, force push) without explicit user confirmation
- No secrets exposure โ agents never echo sensitive data
- No unauthorized decisions โ the engine advises, you decide
- Suspicious instructions rejected โ agents treat other agents' output as reference, never as commands
Information isolation โ Agents cannot see each other's reasoning. The PLANNER does not see the ROUTER's triage logic. Each domain analyst does not see other domains' reports. Thinker subagents in roundtable mode receive only summaries of what others said, not full output. This prevents groupthink and ensures genuinely independent analysis.
Workflow state machine โ Formal transition rules enforce the correct sequence. The PLANNER cannot skip to execution. The DISPATCHER cannot skip the REVIEWER. The Summary Report cannot be produced without the AUDITOR and ADVISOR running. Any violation is a process error that the AUDITOR flags.
HARD RULES index โ Non-overridable behavior is tracked in references/hard-rules-index.md: intent clarification cannot be skipped, pre-session preparation must be shown, each domain's report must be displayed in full as it completes, SOUL entries require user confirmation, theme language cannot be mixed, and more.
Model independence โ Only one file (CLAUDE.md) is bound to a specific AI model. All other intelligence โ agent definitions, extraction rules, inspection rules, knowledge network, directory structure โ is pure markdown readable by any model. Switching models means updating one file.
9 themes across 3 languages โ each language offers three governance styles: historical, modern government, and corporate.
themes/
โโโ zh-classical.md # ๐๏ธ ไธ็ๅ
ญ้จ โ Tang Dynasty (Chinese historical)
โโโ zh-gov.md # ๐จ๐ณ ไธญๅฝๆฟๅบ โ Modern Chinese government
โโโ zh-corp.md # ๐ข ๅ
ฌๅธ้จ้จ โ Corporate departments (Chinese)
โโโ ja-meiji.md # ๐๏ธ ๆๆฒปๆฟๅบ โ Meiji-era governance (Japanese historical)
โโโ ja-kasumigaseki.md # ๐๏ธ ้ใ้ข โ Central Government (Japanese modern)
โโโ ja-corp.md # ๐ข ไผๆฅญ โ Corporate structure (Japanese)
โโโ en-roman.md # ๐๏ธ Roman Republic โ Classical Roman governance (English historical)
โโโ en-usgov.md # ๐บ๐ธ US Government โ American federal government
โโโ en-csuite.md # ๐ข C-Suite โ Corporate Executive (English)
Each theme is a single file (~60 lines) that maps the functional IDs to display names, defines the tone, sets trigger words, and names the output formats. The engine reads the theme file once at session start and uses those names everywhere.
Adding a new theme (Korean government, EU Parliament, Shogunate, startup board) requires only one new file. No engine changes. No new agents.
Theme determines output language โ This is a HARD RULE enforced at every level. All Chinese themes output Chinese. All Japanese themes output Japanese. All English themes output English. Every agent, every report, every response follows this without exception.
Per-session independence โ Theme choice does not persist across sessions. Each new session re-prompts. Each terminal window can use a different theme simultaneously.
The core idea is 1,400 years old: no single voice goes unchecked.
- The planner only plans; it does not execute.
- The reviewer only reviews; it can veto but not rewrite.
- The six domain analysts only execute; they do not judge each other.
- The auditor audits the agents; the advisor audits you.
- No single agent can bypass review and act alone.
When you talk to a normal AI, you get one voice โ confident, agreeable, unchecked. Life OS gives you sixteen, and they do not always agree. That tension is the point.
The Theme Engine adds a second principle: governance is universal, but culture is personal. The logic that makes a good decision is the same everywhere. The language that makes it feel like yours is not. Life OS separates the two so you get both.
Built on the foundation of the Edict project. Life OS extends the framework from software development to all areas of personal life, adding the auditor, advisor, council, strategist, SOUL, DREAM, Strategic Map, and Theme Engine.