/start-agent-work

Start work in this repository with the project coordination protocol.

Arguments

The user invoked this command with: $ARGUMENTS

Read PLAN.md and PROGRESS.md.
State the current goal, the next action, and any blocker that matters for the task.
Read docs/WORKARCHIVE.md when historical completed-work context, previous verification, or sample conversion evidence is needed.
Read only the additional source documents needed for the requested work.
If subagents are useful and the user explicitly asked for delegated agent work, choose the smallest set of .codex/agents/*.toml roles that covers the task.
For substantial implementation work, use the harness sequence: harness-planner-agent drafts the plan and contract, feature-generator-agent implements one agreed chunk, and evaluation-agent reviews the contract and completed work.
Do not implement converter code unless the user explicitly requests implementation.
After meaningful changes, update PROGRESS.md; update PLAN.md only when sequencing, decisions, ownership, or blockers change.
Archive completed work in docs/WORKARCHIVE.md when it no longer needs to stay in PROGRESS.md.
Run the smallest useful verification, check git status, and commit project changes while excluding samples/, outputs/, build/, dist/, generated installers, wheels, models, and other local payload artifacts.

Keep MinerU 3.1.0 as the only conversion engine.
Allow MinerU 3.1.0's CLI-internal temporary local mineru-api, but prohibit --api-url, remote APIs, router mode, HTTP client backends, and remote OpenAI-compatible backends.
Keep runtime processing local-only.
Keep samples/ out of commits unless the user explicitly requests otherwise.
Keep generated packaging, UI build, conversion output, wheelhouse, and model artifacts out of commits.
Prefer official sources for changing facts about Codex, MinerU, Python, uv, CUDA, or licenses.