mathmod-init

name	mathmod-init
description	One-time bootstrap for a new mathematical-modeling competition repository. Use it before real contest work to create or initialize a separate GitHub repository, copy the minimum Competition workspace, and expose the reusable reasoning, compute, writing, and figure Skills. Stop after initialization.

MathMod Init

Use this Skill once, before work on a new competition begins. The framework repository supplies reusable Skills and knowledge; the new contest repository holds that contest's data, context, code, runs, paper, visuals, and delivery artifacts.

Keep the framework checkout or GitHub repository available as the Skill, knowledge, and reusable-tooling source. When a Competition script accepts --root, point it at the separate contest repository.

No particular host is required. ChatGPT, Codex, ChatGPT Work, another capable agent, or a human may perform any part of the workflow. Host/Skill mappings below are convenience defaults only.

Short-form request

When the target repository already exists and the current host has GitHub write access, treat either of these as a complete initialization request:

用 54xkeee/mathmod-full-pipeline 的 mathmod-init 初始化 <owner/repo>：<赛事/年份/题号>；完成后停。

按 mathmod-init 初始化当前仓库：<赛事/年份/题号>；完成后停。

Do not require a longer ceremony when the repository and project identity are already clear. Use sensible placeholders for scientific facts not yet known. If the named repository does not exist and the current host cannot create repositories, report that single blocker and stop rather than inventing another deployment architecture.

Read first

1. Read the Competition README.
2. Inspect the workspace template.
3. Collect or infer from the request the competition name/year, problem identifier, target repository, framework source, and any known team goal. Keep unknown scientific facts explicit.

Initialize the contest repository

1. Use the explicitly named contest repository. If it already exists and repository-write tools are available, initialize it directly through those tools; a local clone is not required just for bootstrap.

2. Copy these files from competition/templates/workspace/ into the contest root:

   AGENTS.md
   CONTEXT.md
   context/model.md
   context/data.md
   context/evidence.md
   context/paper.md
   context/visuals.md
   current_run.json
   

3. Fill the project identity fields in CONTEXT.md, including the framework source used for reusable Skills/references/tools. Keep unknown scientific content as explicit placeholders until the statement, rules, and data have been checked.

4. Create other directories only when the work needs them:

   context/handoffs/
   raw/
   src/
   runs/probe/
   runs/formal/
   paper/
   visuals/data/
   visuals/output/
   submission/
   

5. Copy only the Competition workspace files above. Keep .mathmod/, the framework root .codex/config.toml, integrations/dsh/mcp/, and all other framework-only files in mathmod-full-pipeline.

6. Create the first meaningful Git commit and publish it when the current host has write access. If publication is unavailable, leave the initialized repository contents intact and report that limitation plainly.

The initial current_run.json is an empty formal pointer map. It gains a question entry only after a successful formal run is selected:

{
  "schema_version": 1,
  "kind": "formal",
  "questions": {}
}

Suggested starting Skills

Use Skills by work type, not by mandatory host:

Work type	Suggested Skill
Problem understanding, decisions, review	mathmod-reasoning
Data, code, experiments, verification	mathmod-compute
Writing, captions, document/delivery work	mathmod-writing
Figure planning, rendering, visual QA	mathmod-figure

Typical convenience mapping: ChatGPT often uses mathmod-reasoning; Codex often uses mathmod-compute and may use mathmod-writing or mathmod-figure; ChatGPT Work often uses mathmod-writing or mathmod-figure. These are not access boundaries. One capable agent may use all four task Skills. mathmod-figure remains part of M4, not an M5 module.

Register the lightweight Skills

Register in the selected contest repository, not the user Skill directory. Each independent Git repository needs its own local links; a registration in a parent collection directory does not cross repository boundaries. From the contest root, invoke the framework helper by its actual absolute path (substitute the recorded framework source):

python /absolute/framework/competition/scripts/register_skills.py
python /absolute/framework/competition/scripts/register_skills.py --check

On Windows the helper creates directory junctions; on other operating systems it creates directory symlinks. It is idempotent and never replaces a conflicting path. Registration targets the current directory's .agents/skills by default. It links the five professional Skills plus mathmod-entry and mathmod-autonomous. Preserve explicit --skills-dir for other intended local roots. Check the targets before reporting installation complete.

Keep these machine-local links out of scientific commits: add .agents/skills/ to the project ignore rules or local Git exclude. Record the framework source in CONTEXT.md so another host can re-register. This setup does not install the strict root Harness or MCP. New projects use mathmod-entry for ordinary tasks; only an explicit request to execute autonomously enters mathmod-autonomous. Initialization alone does not start a full solve.

Finish

Report the contest repository URL or name, initialized files, first commit when available, suggested Skills, and any missing problem/rule/data inputs. Then stop. Later work proceeds through the four task Skills, the contest repository's semantic homes and artifacts, and ordinary Git sync.