[ejams1]

Routine agent work previously risked loading a 408 KB plan and 293 KB across three operational skills. Reduce the plan to a 15 KB overview and those skill entry points to about 9 KB, retaining detailed procedures and dated evidence in focused references and removing the duplicated phase-5 status paragraph.

Add a concise CLAUDE.md entry point and a task-scoped session workflow with explicit file ownership, launch instructions, brief and handoff templates, and complete log capture. A read-only log reader bounds its entire JSON response by lines and characters and returns a byte continuation offset, including for huge single-line output. Search exclusions keep generated definitions out of routine discovery while handwritten callbacks and explicit reads remain available. Personal launch settings remain ignored and local.

Preserves phase model assignments, independent review, upstream-shaped source, parity requirements, and mutation safeguards. Native project skills remain discoverable; unrelated local plugin overrides are applied separately from the committed workflow.

Validation: all 12 skill frontmatters and original procedure preservation from the context change; documentation links and concrete search examples; synthetic log checks for independent line and physical character budgets, Unicode/malformed text, byte continuation, empty/huge logs, and failed-command capture; local plugin override inspection; git diff --check. No model session, project build, oracle, or game launch was needed.