ChatGPT Is Expensive & Forgetful? Try Local Agent CLI on Markdown

Royan GagasOctober 25, 2026

developer experience

agentic coding

I love AI. But one thing frustrates me deeply as a System Analyst:

AI chatbots are forgetful.

3 Sins of Web Chatbots for Long FSDs

1 Expensive at Team Scale

Chatgpt Team / Claude Cowork: $20-30/seat/month. A team of 5 SAs = $100-150/month just for chat. Excluding API costs. For individuals or small teams in Indonesia, that's not an "investment", it's a burden.

We already have our own API keys, or could even use local models. Why pay for a closed workspace?

2 Context Drift - AI Forgets Mid-Way

This is the paintful one.

I pasted FSD chapters 1-3 (business definitions), then had a long discussion about edge cases in chapter 5. Halfway through, AI forgot:

The orders -> payments relationship we agreed on at the start
The business rule "refund only if status = PAID"
The table list we had already finalized

Once the chat hits 50+ pages or a long conversation, AI loses context. I have to re-paste the document, repeat the prompt, and hope it remembers.

That's not assitance. That's babysitting.

3. Siloed - Chat in Browser, Files Locally

Browser AI doesn't know my file is at D:/projects/client-xyz/input/fsd/. Every analysis requires re-uploading, copy-pasting, or explaining folder structure.

Developers work in a local repo. Why should SAs work in a separate cloud?

Breakthrough: AI That Works on Files - Via the Agent Terminal, Not a Chat Bubble

In Onesist I flipped the paradigm:

Don't bring files to the AI. Bring AI to the files and watch it work in a terminal.

Instead of a headless "Run" button, the highlight is the Agent Terminal inside the dashboard xterm.js + a real PTY. You open the terminal, type a natural language prompt, and the agent works in the actual project root.

How it works - Terminal focus (`src/components/agent/AgentTerminal.tsx + src/server/terminal/terminal-server.ts`):

Real terminal - PTY node-pty (macOS/Linux) / ConPTY (Windows) spawns opencode with cwd = projectRoot. Skills in .agents/skills/ (fsd-analyzer, markitdown, diagram-svg, query-writer) are immediately available.
Type a prompt - Use the fsd-analyzer skill. Generate ERD from FD1..FDn to output /erd/schema.dbml. the agents does readFile/writeFile directly.

Honest note: the headless Agent Runner (src/server/services/agent-runner.ts — —format json + JSONL parsing) still has many errors — that's why the Terminal is the primary control center, not the runner. The runner is for internal automation only.

Results because of the Terminal:

Zero workspace cost - use your own API key / local model.
Privacy - client data never leaves the machine (SQLite src/server/db/clients.ts).
No lost context - readFile(root, relPath) 200-page FSD read as files.
Transparent - every log visible, copyable, interruptible.

Anti-Hallucination Guardrail: The `fsd-analyzer` skill

AI without guardrails is like a smart but inventive junior frequent hallucinations.

Without constraints AI can:

Create a refunds table that doesn't exist in the FSD
Forget the critical POST /orders/refund endpoint
create tasks not traceable to requirements

Onesist's solution: Skill System - rules of engagement binding AI to SA methodology.

Two skills are bundled in vendor/skills/:

markitdown - PDF/DOCX/PPTX -> Markdown conversion (so FSD becomes a file the agent can read).
fsd-analyzer - completeness analysis. Not just prompts, but validation scripts:

bashCopy

vendor/skills/fsd-analyzer/scripts/
├── extract_entities.py    # extract business entities from FSD
├── validate_erd.py        # does ERD cover all entities?
├── validate_spec.py       # does Spec cover all use cases?
└── compare_artifacts.py   # compare FSD vs ERD vs Spec vs Task

On the frontend, the Completeness Checker (lib/fsd-completeness.ts):

Is User Journey complete?
Are Edge Cases covered?
Is Security & Permissions defined?
Is Data Model consistent?

wikiCopy

[GAP] FSD Chapter 3.2 mentions "refund" but no refunds table in ERD
[GAP] Endpoint POST /orders/refund in Spec but no task in output/task/
[OK]  All FSD entities have tables (12/12)

AI is no longer free to invent. It must pass the checklist before an artifact is considered final. That's what makes AI consistent not long prompts, but standardization.

Because Markdown is the only format that is:

Human-readable - SAs can read it directly.
Machine-readable - AI can generate & parse it.
Git-friendly - diffable and reviewable like code.
Renderable - from one file, Onesist renders an ERD canvas (ReactFlow + DBML), Spec cards, Task Kanban, and Mermaid diagrams.

sqlCopy

Table orders {
  id uuid [pk]
  user_id uuid [ref: > users.id]
  status varchar [note: "DRAFT | PAID | REFUNDED"]
  total decimal
}

Change 5 lines -> ERD updates. AI changes the file -> UI auto-refreshes.

Living Artifacts. Files are the truth.

Watch live & steer

- thinking & tool calls stream in xterm.js. Wrong direction?

don't touch MASTER_ERD.md

-> Enter same session continues. Unlike the headless runner where errors become toasts.

File Watcher - file-watcher.ts + SSE (/api/events) detects write -> UI re-renders. Files on disk are the state.

Attach/detach - close the panel -> reopen -> same session still running. Switch agents via dropdown without changing prompts.