Extracts bash logic from agent prompts into versioned scripts under scripts/, formalises shared-state JSON via schemas/*.schema.json, and adds three loop primitive slash commands. Safety prose (single-author rule, force-with-lease, suspicious-halt) stays inline in agent prompts where it belongs.
No new plugin dependencies.
Highlights
scripts/lib/— atomic state writes, telemetry, mistakes, flake classification, EXTERNAL_CONTENT delim helpersschemas/— 12 JSON Schema (draft 2020-12) files, one per shared-state file;SHARED_STATE.mdcites them instead of inlining JSONCscripts/profiler/—parse_workflows.sh,detect_smoke.sh,catalog_generated.sh(500-entry cap)scripts/scorer/—compute_score.sh,historical_blend.sh(fixes audit §18 paren bug),append_score.sh,record_outcome.sh,regen_cooldown.shscripts/orchestrator/—iteration_cap.sh(3/6/10 by LOC),append_fleet_log.sh,write_run_summary.sh(terminal-state oracle, audit §4),reputation_gate.sh(4 exit codes),audit_impact.shscripts/builder/—push_force_with_lease.sh,smoke_gate.sh,ci_gate.sh(token-level allow/deny + flake-aware),drift_linter.sh- New commands —
/contribute,/repo-finder,/contribute-loop./contribution-dashboardadds latest-loop section. - Plugin manifest — bumped to 0.5.0; no new
requires.pluginsentries.
Invariants preserved
- Single-author commits, no AI attribution
- bash 3.2 portability across all scripts
- Atomic JSON writes via temp + rename
- Force-with-lease only — never plain
--force, never to upstream - flock claim, cap rules, identity-pin verifier, suspicious-halt logic stay inline in agent prompts where they belong
Stats
- 91 files changed, +7,495/−999
- 23 versioned scripts, 35 bash unit tests under
tests/scripts/ - Safety canaries grew, not shrank:
force-with-lease×10,suspicious×27,halt×24
Install
/plugin marketplace add https://github.com/gaurav0107/superhuman
/plugin install superhuman@superhuman
/reload-plugins
PR: #9