router-base.md: router must ignore ~/agent-sessions meta-instructions, honor no-branch/no-commit briefs, and audit scalar numeric metrics only.
agent-base-protocol.md: domain agents do not commit by default; commits require explicit ALLOW AUTONOMOUS COMMITS.
codeflash-java-setup.md: if JMH exists, setup must positively check the harness and record attempted workarounds on failure.
codeflash-java-cpu.md: KEEP requires actual JMH/benchstat output from a command run in-session; ranges/extrapolations/ad hoc loops are invalid.
Close gaps where agents declared premature BLOCKED, committed keeps
without benchmark evidence, and skipped correctness probes on
data-flow refactors. Adds four coordinated behavior changes:
- Blocked-state drill-down protocol (shared/blocked-state.md):
wrapper exit codes are never root causes; 1-2 workarounds required
before any "blocked" record.
- Correctness probe patterns (shared/correctness-probe-patterns.md):
behavioral-equivalence probe rules + boundary-byte checklist;
mandatory before KEEP on data-flow refactors.
- Git operations boundary (agent-base-protocol.md): domain agents
may not create branches, push, or commit keeps without numeric
improvement_pct + real optimized_metric.
- Post-return keep audit (router-base.md): router downgrades any
keep row lacking benchmark evidence to blocked before session exit.
Java agents (codeflash-java.md, -deep.md, -cpu.md) gain:
- Session mode detection/handling (LARGE-SCALE / LIBRARY PRIMITIVE /
CROSS-FUNCTION REFACTOR / PLUGIN VALIDATION), forwarded verbatim
to every spawned subagent.
- Planning-leak guard on the router: first action must be a tool
call, not prose.
- Autonomous-mode section on -deep.md forbidding user-question
blocks in HANDOFF.md under AUTONOMOUS MODE.
- CROSS-FUNCTION REFACTOR mode section on -cpu.md: preserved-behavior
contract, committed property test, touched call graph.
- Correctness-probe gate (step 11a) before any KEEP on data-flow
changes; replaces implicit reliance on existing tests.