Read .agent/AGENTS.md before any repository work.
Then load only the documents needed for the task:
- Architecture or domain work:
.agent/PROJECT_CONTEXT.mdand.agent/SECURITY_INVARIANTS.md. - Privy, Arc, The Graph, demo, release, or submission work:
.agent/SPONSOR_REQUIREMENTS.md. - Intent, payment, retry, worker, queue, job, invoice, settlement, or
reconciliation work: the
oneshot-idempotencyrepo skill and.agent/TEST_MATRIX.md. - Failure handling or reliability work: the
oneshot-failure-injectionrepo skill and.agent/TEST_MATRIX.md. - Any implementation, review, commit, push, or pull request:
.agent/IMPLEMENTATION_LOOP.md. - Demo/release sponsor claims: the
sponsor-qualificationrepo skill. - Handoff, milestone boundary, session end, or deliberate context reset:
.agent/context/README.mdand the current context record.
Canonical shared policy lives only in this file and .agent/. Tool-specific
files such as .claude/CLAUDE.md and .agents/rules/repository-policy.md are
thin adapters: they point here and must not duplicate or override policy.
Repository skills live in .agents/skills/. Personal workflow skills may help,
but they never replace OneShot policy or mandatory FreePi Gate A and Gate B.
If a required document cannot be read, stop and report the missing policy.