ldo

Design the interface a model works in — context blocks, labels, markers, command names, state — the way you'd design a dense console interface

Read the codebase cold for what accretion did to it — file bloat, comment sprawl, duplicated logic, decomposition candidates — then route findings into real fixes, not just a report

Set up the environment, implement the plan with tests, update user-facing docs

Configure LDO model routing — which model runs which role, and when the optional agents fire

Record a project contract — a rule that must hold regardless of task size, read only when a change touches the area it governs

Read the project's documentation cold, as a newcomer would, and find what has drifted — contradictions, stale claims, undefined jargon, wrong ordering

File a bug or observation about LDO itself — structured, secrets redacted, as a GitHub issue in the LDO repo

Write LDO's working instructions into the project CLAUDE.md so the agent self-routes real work through the pipeline without manual /ldo

Record an operational note (deploy quirk, environment gotcha) or a decision/mandate (why something was overridden or chosen) — not a rule, just a fact worth remembering

Read the codebase and produce an implementation plan with complexity and security ratings

Multi-source web research with cross-verification — decompose, search from several angles, verify claims, report confidence

Track and recover LDO pipeline runs across an interrupted session — check for runs left in progress, resume from cache instead of restarting cold

Review the diff and drive the running app to prove acceptance criteria hold

Threat-model an implementation plan before code is written — OWASP top 10, injection, auth, data exposure, supply chain

Ship the pipeline's changes — branch, commit, push, PR, optional squash-merge — with the review report as the PR body

Design and build terminal interfaces that earn the screen — keyboard-first, honest, degrading gracefully — in Textual (Python) or Ink (TypeScript)

Copy LDO directly into a project's .claude/ directory so it ships with the repo, for use without a local plugin install — e.g. claude.ai/code sessions that just clone a repo