Write me about what you're working on. I read every message.
Private task system
Life OS
A calmer mind builds a freer life.
One control plane for life and business, with six separate boards. Life opens on one NOW card, then sends, decisions, next and waiting. A task can't be marked done without evidence. The database enforces it.
Done today: Pay the water billevidence: receipt saved
Example · not real tasks
6walled boards, one engine
8narrow tools in the Hermes task bridge
0delete commands in the agent CLI and task bridge
1,046tests passed (1 Nov 2026)
185commits, 5 Oct to 10 Nov 2026
01The problem
Nine to-do files, and no one place to look.
Tasks lived in plain text: seven division task files, an attention file and a separate command board. Earlier tries at the same idea (an AI agent, a separate status page, a cockpit) added more places, not fewer.
A checkbox in a text file can be ticked by anyone, at any time, with no proof. And as more AI sessions helped with the work, each one needed a safe way in that kept it to its own part.
"Unifying and harmonizing and symbiosis … it won't break things, it won't leave things behind."
02What I built
One control plane, six boards, narrow doors for agents.
Life Control, the one page: one NOW card, then sends, decisions, next and waiting. A "Done today" panel shows the evidence for each task closed.
Walled boards. Life is the top level; business, studio, tech, trading and education work each get their own board, so work never shows up in life's Now, Next or Today.
Done needs evidence. The API rejects blank or very short evidence, and the database rejects blank evidence. The habit is a commit, a link or a receipt.
Narrow doors for AI. Sessions use a small CLI; Hermes and the Slack bot use narrow tool bridges. The CLI and the eight-tool bridge have no delete or archive command.
Plans inside the page. Each project's plan opens inside Life Control, view-only, with no scripts allowed to run.
Slack briefs. A plain-text morning brief, and a ping when a new decision is waiting. No AI model in the loop.
It replaced the text files: on 5 Oct 2026 every open line was imported (583 rows) and eight Drive files were retired, with the originals kept.
Six boards, and who works where
LifeThe top level. Calm. Personal life only.HermesHousehold botLife session
BusinessClients, offers, money inBusiness session
StudioGames and designStudio session
TechMachines and upkeepTech session
TradingTo-dos only, no market dataTrading session
EduLearning new toolsEdu session
I see every board. The CLI works on the board named in each call, and the server refuses a change to a task on another board (409). Hermes's tools target the Life board; the household credential is held to household routes on the server.
lo, the CLI for AI sessionsExample · not real tasks
$ lo --board studio add "Fix the jump sound" --next "Find the clipping frame"
--done-when (the done condition) is required and cannot be empty.$ lo --board studio done 41
--evidence (proof it is done) is required and cannot be empty.$ lo --board studio done 41 --evidence "commit 3f2a1c, test 12 passes"
Closed #41 with evidence.$ lo --board studio done 88 --evidence "done"
Refused: task #88 is not on the studio board.# lo has no delete or archive command.
03How it works
The life of a task: asked, worked, closed with evidence, kept.
Animated diagram. Each box is a step in the code. No real tasks are shown anywhere on this page.
1 · Ask
Me, on the page: quick add or the task panel
An AI session, through the CLI
Hermes, through its eight tools
The household bot: its own short list of routes
2 · Doors
The CLI names a board on every change; a mismatch is refused
The CLI refuses a task with no next action or no done-when
No SQL, no shell: predefined routes only
3 · Work
FastAPI on SQLite, plus a separate AI runner
One NOW, then sends, decisions, next, waiting
Normal task actions are recorded as events
4 · Close
Done needs evidence, or the server answers 400
A database CHECK backs it up
Defer needs a date and a reason; blocked needs a reason
5 · Keep
No delete or archive in the CLI or the task bridge
A task with a recorded send is never deleted
Reopening clears the evidence; nightly backup keeps 14
How Hermes plugs in
Hermes Agent is an open-source AI agent (by Nous Research) that I host myself. Today it runs as the household Slack bot on the application server, using Claude Haiku. The first owner profile ran on the GPU node with a local model; that part is history.
Hermes AgentSelf-hosted, with its own memory and tools.
The Life OS task bridge: 8 tools
Read the board
Add a task (needs a next action)
Mark done, with evidence
Block, with a reason
Answer a decision
Preview a send
Send, with a confirm token
Record a send made by hand
Life OSThe tools target the Life board; the server checks the board on each change.
What the bridge can't do
No SQL, no shell, no routes of the caller's choosing: the tools use predefined Life OS routes. Who it acts as comes from its settings, not from the model. A real email needs a matching preview token, valid for 15 minutes.
Chat, the other way
Built 5 Oct 2026: chat with Hermes inside Life Control, picking the model per message (Claude, Grok, GPT or local Qwen). The original chat service is retired and its replacement is not set up, so the panel stays hidden.
Hidden by default since 29 Oct
Same agent, other walls
The Slack household bot is Hermes in its own profile, held to household routes by Life OS. A private profile's credential is limited to 13 allowlist patterns; everything else answers 403, and its decisions need a one-time confirm token.
Private profile: manual start
208→54 WATTS · 3 GPUS
Hermes once ran the briefs on a schedule. I paused its scheduled jobs on 16 Oct 2026; the recorded GPU draw fell from about 208 W to about 54 W. The briefs came back on 25 Oct as plain scripts, with no model.
04Tools
Small and plain.
Python + FastAPIweb server, plus a separate AI runner
Hermes Agentself-hosted; household bot on Claude Haiku
Ollama + Qwenlocal model; also Claude, Codex, Grok
Slacka bot and plain-text briefs
systemd + Caddyservices, timers, TLS
pytestfull suite required before deploys
Claude Code, CodexAI writes the code; I specify and verify
05Proof
How it grew, from the plans and the git log.
Life OS has no tagged releases. Its database schema versions are the closest thing, so each step below shows the one it brought.
The first planAreas, projects and task states; a calm Today view of three focus slots.
Build 1 Schema v4The Life Control backend: the NOW picker and the prepared-sends queue.
Build 2 Schema v5HermesDecisions, and owner and household roles. Hermes gets its eight narrow tools.
Build 3, and the move Schema v6Moved to its production server. The Life Control page moved in behind Life OS's login. The text files were imported and retired. AI buttons (Plan it, Break it down) write drafts only; nothing is saved until I approve. The git history starts here.
Build 4 HermesChat with Hermes inside Life Control.
The power fix HermesScheduled AI jobs paused: about 208 W down to about 54 W on the GPUs.
Briefs as Slack messagesMorning brief and decision ping, plain text, no model. The household bot can file requests.
Walled boards Schema v7 → v8Life, Business, Trading, Tech, then Studio. Plans inside the page, the lo CLI, one template with board themes. 95 commits that day; six rounds of review by Codex before the merge.
Edu board, private profile Schema v9HermesThe sixth board. Hermes's private profile limited to 13 allowlist patterns. An honest Today (overdue, today, this week) and the Done-today panel with evidence.
Business board layoutNOW plus the offer, money tracks, and a "parked, not lost" line.
The rebuilds that add a board keep every row and every id, in one transaction. The real board stays private: it holds real life.
06How it's tested and run
The rules are enforced, not hoped for.
Evidence, enforced twiceThe API rejects blank or very short evidence, and a CHECK in the database rejects blank evidence, even from a stray UPDATE.
Walls checked inside the writeWhen a caller sends the board it expects, closing a task checks it in the same transaction and a mismatch is refused. The CLI and the Hermes tools send it.
The full suite, by ruleProject rule: the full suite before commits and deploys. 1,046 tests passed on 1 Nov 2026, across 30 test modules and one integration script: the API, the evidence rule, the walls, security, the CLI, the briefs, the Hermes and household bridges, migrations and the page.
A second AI reviewsCodex is the independent reviewer. The walled boards went through six Codex rounds, ending in SHIP.
Secure by defaultIt won't start on a network address without a login. bcrypt passwords, cross-site writes refused, a strict content policy, nothing cached.
Integrations, kept narrowIt connects to AI providers, Hermes, and calendar, Drive and mail helpers. Each bridge exposes only its own short list of operations.
Backed upA nightly backup at 3:30 AM keeps the last 14. Project rule: a manual backup before each deploy.
07What I'd do nextProposed · not built
Real walls, all the way down.
One data file and one login per board, so a wall is a separate database, not only a check.
Agents may only suggest tasks: at most 10 waiting per board, gone after 14 days unanswered.
08Skill tags
AI agentsAutomationPythonFastAPISQLiteMCPHermes AgentSlackSecuritysystemdpytestClaude CodeCodex