Getting started
Working with a coding agent
Most commands on this site will be run by an agent, not typed by a person. This page is the path for that.
The third reading form
Plectis reads in three forms. The pages here are for people. The AI packet is for a chat assistant: one file carrying every public page, component record, and paper module, with instructions for answering from the public evidence (the reviewer brief is the navigable public source: paper LaTeX, assertion rows, glossary terms, per-problem Lean and Comparator paths, component records, and the site map). The repository itself is for a coding agent: an AI that works inside a folder of files rather than a chat tab, such as Claude Code, Codex, or Cursor. Opened inside the repository, the agent reads the source, runs the components, and checks the result records directly: the same loop this site describes, executed rather than narrated.
The path
No git knowledge is assumed, and nothing here needs typing into a terminal: the agent performs the clone and runs the commands itself.
- Install a coding agent. Claude Code, Codex, and Cursor each install from their own documentation; any agent harness that works in a local folder fits.
- Give it the prompt below.
- Ask in plain language from there.
My task is [describe the question or change]. Use https://github.com/wcook04/plectis for software components and https://github.com/wcook04/plectis-erdos for mathematics and Lean proofs. Clone the repository my task needs from current main; use both only if the task crosses them. Preserve any existing local work. Read README.md and the compact agent entry: AGENTS.override.md where present, otherwise AGENTS.md. Record the checkout commit and local changes. For mathematics, use python3 scripts/agent_entry.py --entry with my task, and --checkout --check-upstream to compare with public main. For software, follow QUICKSTART.md and run the quickstart check or selected component example, saving output under .microcosm/. Do the task and return source paths, exact commands and outcomes, and the remaining assumptions or unresolved step. Report anything you could not run as not established, never as false. A release tag is a frozen edition, not current main.Choose the repository for your task and start from its current main branch. Record the commit and preserve existing changes; release tags describe frozen editions. Follow the software agent quickstart or the mathematics agent quickstart. Each names the compact agent entry, commands and expected result.
After an agent says done
When an agent reports it is done, check the claim against the record it left. Ask what files the agent claims it touched, what validation output exists, whether the repository state changed before a commit claim, what blockers remain, and where the claim stops.
Clone https://github.com/wcook04/plectis and open it. Read AGENTS.md and README.md, then inspect the coding-agent residue path: Durable Agent Work Landing Replay, Agent Completion Faithfulness Audit, and Bridge Phase Continuity Runtime. Find the first completion or resume claim that would create false confidence if its evidence were missing.What the agent finds
The repository gives the agent the same material this site projects, in runnable form: the component source under src/, the fixtures each component runs against, validators it can execute, result records accumulating in .microcosm/, and the same scope limits the site prints. The command blocks across this site are written so an agent can execute them verbatim; the quickstart is the usual first request.
Same evidence rules
The agent works under the same boundary the packet states: claims stop where public records stop, and a scope limit bounds authority rather than describing depth. Asking the agent to take one claim from this site and verify it against the source is the intended use, not an adversarial one.