Patchline

A Cursor-shaped coding harness you can read.
Local Electron desk. Python LangGraph worker. Human Accept on every skill and every diff.

---

Why this exists

Most “coding agents” hide the loop. Patchline puts the loop on screen, the way Cursor does, and keeps the model frozen.

You see	The harness does
Agent / Plan / Chat	Mode prompt + tool set, no silent writes in Chat or Plan
Todos, tool rows, streaming text	JSONL from harness_worker
Accept / Undo on a diff	Keep the patch, or restore and record a trace
Accept / Reject on a skill	Skill Box only after you say yes
Fullscreen graph	Import, call, and co-edit edges the retriever walks
“This model does not accept images.”	Text models never pretend they saw a screenshot

The desk is Patchline. The worker is python -m adaptive_agent.harness_worker run. Together they are adaptive-coding-ai-agent.

---

Architecture

Patchline (Electron)
  apps/web          React 19 · Vite 6 · Monaco · vis-network
  apps/desktop      IPC · ConPTY · spawn python -B -m adaptive_agent.harness_worker
        │  stdin: start_run JSON    stdout: JSONL events
        ▼
adaptive_agent/harness_worker.py
  router → adapter patch
  code graph retrieve (anchors + PageRank)
  LangGraph tools (read, grep, patch, todos, plan, skills)
  plan_gate.json  ·  skill review poll  ·  Memory SQLite
        │
        ▼
OpenCode Go   https://opencode.ai/zen/go/v1
  default model: deepseek-flash
  vision only if OPENCODE_VISION_MODEL is set

---

Product

Crops are from the running Electron window.

The desk

Modes

Mode	Writes files?	What you get
Agent	Yes	Todos, tools, patches
Plan	No, until Build	Questions, research, editable markdown
Chat	No	Answers from the workspace

Shift+Tab cycles modes. Send sits at the right edge and becomes Stop while a turn is running.

Chat the harness actually fills

Clarifying chips, Continue planning, sticky Build, skill Reject / Accept, todos, and read_file in one column. Chats are tabs.

Text models do not see images

Paste is refused unless OPENCODE_MODEL or OPENCODE_VISION_MODEL is a vision id (deepseek-v4-flash-vision-exp).

README Preview and Source

Diffs you keep or undo

Muted green and red. Accept (Ctrl+Shift+Y) keeps the file. Undo (Ctrl+N) restores it and writes a trace.

Graph, fullscreen

Title-bar Graph. Legend: files, imports, calls, co-edits. Retrieval uses this walk so the next prompt names real symbols.

---

Questions the desk asks, and answers you give

Plan mode does not invent a todo list and call it a plan. A real exchange:

You

Add skip for muted clips. I am not sure if it belongs in the ranker or the player.

Patchline

Where should skip happen? ranker · player

You pick ranker, then Continue planning.

Patchline writes a markdown plan you can edit:

# Skip muted clips
## Files
- feed/ranker.py
Filter clips with audio_gain 0 before scoring.

You press Build. Only then does Agent mode patch the file. The diff waits for Accept.

If the same failure shows up again, the harness drafts a skill. It does not install it:

skip_muted_clips v3 — two Undo traces, muted clips still ranked first.
Reject skill · Accept skill

Accept puts it in the Skill Box. The next system prompt can retrieve it. Reject throws the draft away.

---

Tech stack

Layer	Choice	Why
Desk	Electron 29	One window, real OS terminal, file IPC
UI	React 19, Vite 6, Tailwind 4	Fast desk, Monaco theme, no extra app server
Editor	Monaco	Diff review, TS path hints, Preview / Source
Graph view	vis-network	Fullscreen import / call / co-edit
Terminal	ConPTY (node-pty)	A shell, not a log
Harness	LangChain + LangGraph	Tools, streaming messages, mode prompts
Model	OpenCode Go, DeepSeek Flash	Frozen base; adapters are instruction patches
Memory	SQLite (data/agent.db)	Episodes, pending skills, Skill Box
Graph store	SQLite + JSON	Import, call, contains, git co-edit, PageRank
Tests	pytest, node:test, Playwright	Worker, path safety, Electron smoke

---

Quick start

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env
npm install
npm run dev

Put OPENCODE_API_KEY in .env. Keys come from opencode.ai/auth after a Go subscription.

Vite listens on 127.0.0.1:5173. Electron opens Patchline. Point it at a folder or set WORKSPACE_ROOT.

Installer: npm run dist → apps/desktop/release.

---

Config

Variable	Default
OPENCODE_API_KEY	—	Required
OPENCODE_BASE_URL	https://opencode.ai/zen/go/v1	Chat completions
OPENCODE_MODEL	deepseek-flash	Coding default
OPENCODE_VISION_MODEL	unset	Set to allow image paste
WORKSPACE_ROOT	repo root	Folder the desk opens

---

Layout

adaptive_agent/     harness, graph, memory, skills, router, plan gate
apps/web/           Patchline UI
apps/desktop/       Electron main, preload, PTY
lessons/            L2 skills, L3 graph, L4 retrieval
tests/              pytest, no live model required
docs/screenshots/   Electron crops
docs/architecture/  system, turn, skill loop

---

CLI

python -m adaptive_agent.router "implement the hash function" --run
python -m adaptive_agent.code_graph --repo C:\path\to\project --out data\graphs\example.json
python -m adaptive_agent.graph_viz --query "where do we verify a token?"
python lessons/l2_skill_induction.py --decision reject --reason "not ready"
python -m pytest -q
npm run test:desktop

---

What we will not do

• Auto-activate a skill.
• Send pixels to a text model and hope.
• Let Plan mode write files before Build.
• Commit .env, data/agent.db, or data/eval/.

Windows JSONL is UTF-8. The worker replaces lone surrogates so a pasted character cannot crash the pipe.