claude-code-orchestra: Skill for Claude Code

Context Loader Skill

Purpose

Load canonical shared context from .claude/ plus project-owned design documentation so every agent runtime uses the same source files.

When to Activate

ALWAYS - This skill must run at the beginning of every task to load project context.

Workflow

Step 1: Resolve the Read Plan

Run load_context.py to get a deterministic read order instead of a hand-maintained file list, so the plan never drifts from what actually exists on disk:

python3 .claude/skills/context-loader/load_context.py [--task-libraries name,name]

Pass --task-libraries (comma-separated) when the task names specific libraries; the script matches them against .claude/docs/libraries/ and folds any hits into the read order.

The JSON reports {ok, read_order, rules, state, design, progress, libraries, missing, unreadable, warnings}. read_order is the exact ordered list of repo-relative paths to read, in this order:

1. the rule files in .claude/rules/ (coding principles, delegation, dev environment, language, security, testing, tiers, CLI execution, Codex delegation, and any newly added rule file);
2. .claude/STATE.md — the active main agent and current working blocks;
3. PROGRESS.md — the rolling record of the latest five checkpoints, and the session-to-session continuity /feature reads first;
4. .claude/docs/DESIGN.md — architecture decisions and constraints;
5. any matched library docs.

Step 2: Read Everything in read_order, Then Surface Gaps

Read each path in read_order. Then check the rest of the report:

• missing — canonical files that do not exist at all (e.g. .claude/docs/DESIGN.md, PROGRESS.md). Report these; do not silently proceed as if they were empty.
• unreadable — files that do exist but could not be read or decoded. This is a filesystem or encoding problem, not an un-bootstrapped repository: do not respond by suggesting /init.
• warnings — design.placeholder is tri-state: true means .claude/docs/DESIGN.md is absent or still the uninitialised /init template (its "Background & Purpose" section is empty), false means real prose, and null means the marker heading is gone so the question cannot be answered. progress.entries: 0 means PROGRESS.md exists but holds no checkpoint entry.
• libraries.matched vs libraries.files — matched is what read_order included for the current task; files is every doc that exists. If a library relevant to the task isn't in files at all, its documentation simply doesn't exist yet.