Open Scaffold — AI skill for Claude Code

The problem

You pay frontier prices for review, status checks, and "where were we" because nothing cheaper can be trusted. Cheaper models guess; when a chat ends, the work's memory dies with it; the next session reconstructs from a scrollback buffer and invents what it can't recover.

What it does

Open Scaffold keeps a repo-native work record — git-tracked, observed-fact files about what your agents did — and turns it into three things:

• Record (ambient). Extracted from observed facts — transcripts, receipts, test results — costing the working model nothing. osc capture --from claude-code|codex reads a finished session into a record with no worker cooperation. Add a plan and evidence files to check claims against intent; feedback and lessons carry forward instead of being relearned.
• Handoff. osc handoff compiles the record into a budgeted, secret-redacted packet so the next reader — a fresh session, a smaller model, another vendor's agent, or a teammate — resumes from truth instead of re-deriving or inventing it.
• Review and gate. osc review reports plateaus, failing criteria, and requirements worth questioning; osc gate turns that into a retry authorization with stop authority outside the worker. Any file-reading model can be the judge. Fails