Closing the Loop · Penn ESE 5180

The tutorial, step by step

Every step from the talk, with the same screens and the same lines to copy. The Copy button beside each line puts it on your clipboard.

This is new and experimental, and some of it was put together the night before. A step may fail on your machine; the lessons hold even when one does. Tell us what broke.

The tutorial

Your turn: set up your own loop

Step 1 of 5

Install Hapimatic

In a terminal on your laptop: install your agent with curl -fsSL https://claude.ai/install.sh | bash and run claude once to sign in (Codex and Gemini CLI work too); install Hapimatic with curl -fsSL https://github.com/SolidPD/hapimatic-core/releases/latest/download/install-hapimatic-v1.sh | bash; then hapi server, which prints your access token, and hapi daemon start in a second terminal. Linux: run it as is. Windows: Ubuntu on WSL2 first, with two settings, the two-page guide. macOS: a build ships, untested by us; the safe path today is to skip to step 3, where your agent drives Logic 2 and the PPK2 directly

Step 1 of 5

On Windows? Ubuntu on WSL2

Windows owns every USB device and runs Logic 2 with its MCP server on port 10530, the PPK2's Python, the flash tool, and the browser; WSL2 Ubuntu runs Hapimatic on port 3006 and your agent, Claude Code, Codex, or Gemini CLI, which calls the Windows tools. .wslconfig: mirrored networking, no idle timeout

The bench the guide was tested on: the Saleae Logic, the PPK2 lit green, the test board on a breadboard, and the Windows laptop running WSL2, each marked

Step 2 of 5

Sign in, and start a session

Your laptop's terminal after hapi server's first start: NEW CLI_API_TOKEN GENERATED, the token, and Saved to: ~/.hapi/settings.json — The first hapi server prints it once: save it. It also lives in ~/.hapi/settings.json, and this line prints it again

jq -r .cliApiToken ~/.hapi/settings.json

Hapimatic's sign-in page with the access token pasted in, shown as dots — Open localhost:3006 in your browser, paste the token, and sign in

New Session's folder picker, with the starter repo, closed-loop-starter, in the list — New Session, then Browse: pick your copy of the starter repo and Select. Create, with Claude Code as the CLI Harness

The session's prompt box with the first prompt typed in: Read AGENTS.md and BENCH.md, then tell me what you need from me to close a loop on my board — Send it, and read the answer: what it needs from you to close a loop

Read AGENTS.md and BENCH.md, then tell me what you need from me to close a loop on my board

Step 2 of 5

Sign in to your AI provider in Hapimatic

Hapimatic's Add to Anthropic box, with Open sign-in link — Settings › Provider Accounts › Anthropic › Add account, then Open sign-in link

Hapimatic's three sign-in steps: open the Anthropic sign-in page; sign in, and if claude.com emails you a 6-digit code, enter it there on claude.com; click Authorize and copy the long authorization code — On claude.com: your email, then the 6-digit code it emails you, entered there. Then Authorize, and copy the long code

Hapimatic's Authorization code field, then Done — Back in Hapimatic: paste the whole code, including the part after #, and press Done. It works once

Five routes, ranked. 1 Claude Code: a Claude plan, the three steps above, or an Anthropic API key in the same box; tested most. 2 Codex: install it with OpenAI's installer before Hapimatic, the line below the routes; today, sign in with codex login in a terminal, because the in-app sign-in is being fixed. 3 An OpenRouter key: paste it in Provider API Keys, and it runs OpenCode on the model you choose; OpenCode's qualification is pending. 4 Gemini CLI: a Google AI Studio key, whose Flash models have a free tier; it passed the bench loop. 5 A local model: free and slow, experimental; alone it failed 0 of 8, and with a frontier model reviewing it passed 3 of 3

curl -fsSL https://chatgpt.com/codex/install.sh | sh

Codex: install it before Hapimatic, then codex login in a terminal

Install Codex with OpenAI's installer, before you install Hapimatic, so Hapimatic finds it; then check it
```
curl -fsSL https://chatgpt.com/codex/install.sh | sh
codex --version
```
If you use the systemd services, add this line under [Service] in both units (hapimatic-server.service and hapimatic-daemon.service), then reload and restart. If you run hapi server in a terminal, just restart it
```
Environment=PATH=%h/.local/bin:/usr/local/bin:/usr/bin:/bin
```
```
systemctl --user daemon-reload && systemctl --user restart hapimatic-server hapimatic-daemon
```
In ChatGPT: Settings › Security › enable device code login. A school workspace needs its admin to allow it; if it won't, use an OpenAI API key
Today: sign in from your terminal and finish the ChatGPT sign-in it opens; Hapimatic reads that login. Signing in from Hapimatic's Settings › Provider Accounts › OpenAI is being fixed
```
codex login
```

The in-app Codex sign-in is being fixed, and a full Codex session has not been tested end to end on this path. Source: the students' before-class sheet (docs/talks/2026-10-28-penn-ese5180/students-before-class-v1.md, step 3, on ese5180-v4) and Lane E's report on #5871 (2026-10-27): OpenAI's installer before Hapimatic, device code login in ChatGPT's Security settings, the systemd PATH line; the in-app sign-in fix is PR #5882, not in this deploy

Step 3 of 5

Connect your agent to the logic analyzer

Logic 2's Automation panel with MCP Server switched on at http://127.0.0.1:10530 — Logic 2: the Automation button at the bottom right, or Settings › Automation. MCP Server on: 127.0.0.1:10530. Screens: Saleae

A Hapimatic session's Terminal tab after running claude mcp add --transport http -s user logic2 http://127.0.0.1:10530: Claude Code answers that it added the HTTP MCP server logic2 to the user config — Your session's Terminal tab: paste, press Enter, and stay there for the reply. Codex: the second line

claude mcp add --transport http -s user logic2 http://127.0.0.1:10530

codex mcp add --url http://127.0.0.1:10530 logic2

Nordic's drawing of the nRF7002 DK's headers P4 and P3: the Logic 8's ground lead on P4's GND pin, channel 0 on P3's P1.06, the LED1 pin — nRF7002 DK: channel 0 on P1.06 (LED1), its ground lead on GND. Another board: any pin your firmware toggles. Drawing: Nordic

The session's prompt box with the step's prompt typed in: List my Saleae devices, then capture two seconds on channel 0 and tell me what you see — Back on Chat, send it. It lists your Logic, captures, and says what channel 0 did

List my Saleae devices, then capture two seconds on channel 0 and tell me what you see

Step 4 of 5

Connect the Power Profiler

Windows: nRF Connect for Desktop first (the driver), then the first block in PowerShell; Ubuntu cannot see USB. Mac: the second block

py -m venv $HOME\ppk2-venv
& $HOME\ppk2-venv\Scripts\pip install ppk2-api==0.9.2

python3 -m venv ~/ppk2-venv
~/ppk2-venv/bin/pip install ppk2-api==0.9.2

The silkscreen over the PPK2's connector, GND, VIN, VOUT, GND, beside a legend: VOUT to your board's supply pin, GND to its ground, VIN only in ampere-meter mode — Plug in its USB and check it is seen: it prints a port, such as ['COM6']. Mac: use ~/ppk2-venv/bin/python. Photos: Nordic

& $HOME\ppk2-venv\Scripts\python -c "from ppk2_api.ppk2_api import PPK2_API; print(PPK2_API.list_devices())"

The session's prompt box with the step's prompt typed in: Measure my board through the PPK2 for ten seconds, save it to current.csv, and give me the average and the peak — Back on Chat, send it. The agent records ten seconds and reports the numbers

Measure my board through the PPK2 for ten seconds, save it to current.csv, and give me the average and the peak

One rule: powering the board needs a teammate's yes; measuring doesn't — First fill in AGENTS.md's Power line: source mode at 3600 mV (DK, XIAO) or ampere mode (Nucleo). The agent asks before it powers the board

Step 4 of 5

Put the PPK2 on your board

Three vendor pictures with the pins marked. The nRF7002 DK, source meter at 3.6 V: pull the jumper cap off P23 (VBAT), PPK2 VOUT to P23 pin 1, the upper pin, and GND to P21 pin 1 (minus); that measures the nRF7002 Wi-Fi chip, and the nRF5340 on P22 needs SB16 cut and SB17 soldered first; flash with nrfutil or west through the onboard J-Link. The XIAO nRF52840, source meter at 3.6 to 3.8 V: flash it, unplug USB, then VOUT to BAT+ and GND to BAT−; about 2 to 5 µA asleep; flash by double-tapping reset and copying the .uf2. The Nucleo-WL55JC, ampere meter with USB powering the board: pull jumper JP1 (I_SoC), PPK2 VIN to the 3V3 side and VOUT to the STM32 side; with JP1 off the 3V3 side reads 3.3 V to GND; PPK2 GND to any board GND; flash with STM32_Programmer_CLI through the onboard ST-LINK — Pictures: Nordic Semiconductor, Seeed Studio, STMicroelectronics

Before you clip onto the nRF7002 DK, read the pin-1 marker on the silkscreen: the lower P23 pin is the DK's own VBAT supply. VOUT goes on P23 pin 1, GND on P21 pin 1 (−), and powering the board needs a teammate's yes

Step 5 of 5

Five steps, and a repo to start from

The five steps: 1 install Hapimatic, and save the token hapi server prints; 2 sign in to your AI provider, then start a session in the starter repo; 3 connect the logic analyzer, Logic 2's MCP server then claude mcp add; 4 connect the Power Profiler, ppk2-api in a venv, VOUT and GND to your board; 5 close one loop, paste prompts/feature-v1.md, fill in one requirement with a number, and prove it. Beside them, the starter repository on GitHub, closed-loop-starter, and what is in it: AGENTS.md, the instructions your agent follows; BENCH.md, one row per instrument; prompts/, a feature prompt and a low-power prompt; tools/, PPK2 power and capture and Logic 2 edge counts; examples/, the 2026-10-26 run

A scan code for the starter kit, everything from the bench run ready to clone: agent instructions, the PPK2 and Saleae tools, the prompts, and the worked example — github.com/SolidPD/closed-loop-starter

The tutorial

Your loop, from install to commit

The whole loop on one sheet, five rows. 1 Install Hapimatic (Windows: WSL2 first; Mac: skip to step 3): curl -fsSL https://claude.ai/install.sh | bash; curl -fsSL https://github.com/SolidPD/hapimatic-core/releases/latest/download/install-hapimatic-v1.sh | bash; hapi server, and hapi daemon start in a second terminal; it worked when hapi server prints your access token, which you save; it is also in ~/.hapi/settings.json. 2 Sign in and start a session: open localhost:3006 and paste the token; AI in this order: Claude Code, Codex, OpenRouter, Gemini CLI; Codex is installed before Hapimatic, then codex login; New Session in the starter repo; ask: Read AGENTS.md and BENCH.md, then tell me what you need from me to close a loop on my board; it worked when the session answers. 3 Connect the logic analyzer: Logic 2's Automation, MCP Server on 127.0.0.1:10530; claude mcp add --transport http -s user logic2 http://127.0.0.1:10530; CH0 to a pin that toggles, GND to GND; ask: List my Saleae devices, then capture two seconds on channel 0 and tell me what you see; it worked when it names your Logic and describes channel 0. 4 Connect the Power Profiler: on Windows nRF Connect for Desktop, then in PowerShell py -m venv $HOME\ppk2-venv and & $HOME\ppk2-venv\Scripts\pip install ppk2-api==0.9.2; on a Mac the same pinned package in a python3 venv; DK: VOUT to P23 pin 1, GND to P21 minus; XIAO: BAT+ and BAT-; Nucleo: across JP1; ask: Measure my board through the PPK2 for ten seconds, save it to current.csv, and give me the average and the peak; it worked when current.csv exists with the average and the peak. 5 Close one loop: the full prompt is prompts/feature-v1.md in the starter repo; Write the pass test first, then change, build, flash, measure on the Saleae and the PPK2, decide pass or fail, commit or revert, and report the numbers; it worked when there is a commit with the numbers in it, or a revert with the reason

Back to the links