Dashboard¶

A local web chat front-end for /pycemrg-build. Have a short conversation with Claude to nail down what you want to build, then hit Finalize plan to get a structured build document and a project scaffold you can copy.

Source: dashboard/.

Prerequisites¶

pycemrg-context installed (see Getting Started) so ~/.claude/commands/pycemrg-build.md and ~/.claude/pycemrg-context/ are populated.
Claude Code installed and on PATH (claude --version should work).
ANTHROPIC_API_KEY set in your environment (or however you've configured Claude Code auth).
Python 3.10+.

Setup & run¶

macOS / LinuxWindows (PowerShell)

cd dashboard
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python server.py

cd dashboard
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
python server.py

Then open http://127.0.0.1:5050 in a browser.

Using it¶

Type a project description and Send. Claude responds with a proposed decomposition, a trace table mapping each step back to your exact words, and (usually) 2–4 clarifying questions.
Reply. Push back on inferred steps, answer the questions, ask for alternatives. Iterate as many turns as you need. (Cmd/Ctrl-Enter sends.)
When you're happy, click Finalize plan. The server sends the __FINALIZE__ sentinel and Claude produces the full document: confirmed decomposition, trace table, build plan with [SUITE]/[PROJECT]/[GAP] tags, the scaffold, install commands, and any flagged gaps.

How it works¶

Each turn, the server reads ~/.claude/commands/pycemrg-build.md, appends the full conversation transcript so far, and pipes the result to claude -p running from ~/.claude/pycemrg-context/ (so the @-references in the prompt resolve). Claude's stdout becomes the next assistant turn and is rendered as markdown.

Sessions live in process memory only — restarting the server clears them. The server caps concurrent sessions at 50 and LRU-evicts past that.

Cost & exposure

Each turn is one claude -p call against the Anthropic API and uses your API key (a finalized plan typically costs a few cents; prompt caching helps). The server binds to 127.0.0.1 only — do not change this without thinking about who can spend your API budget.

Troubleshooting¶

Symptom	Cause / fix
"claude CLI not found"	Claude Code isn't on `PATH`. Run `which claude` (`where claude` on Windows).
"command file not found"	pycemrg-context isn't installed. Run `./install.sh` (or `install.ps1`) from the repo root.
Timeout after 120s	The prompt is doing too much. Narrow the message, or back up a turn and split your ask.
"unknown session id"	The server was restarted. Reload the page to get a fresh session.
Plan looks wrong	Compare against `/pycemrg-build` in a Claude Code session. If they differ, the dashboard is invoking the prompt wrong; if the same, the prompt template needs work.