interloop

modernize priority P2 effort M wave 2

What it does: A single-skill, prose-only Claude Code plugin (skill: project-loops; no commands, hooks, MCP, or agents). It ships a SKILL.md router plus three reference docs (a 12-entry loop-catalog, a project-routing map, a synthesis-from-analysis contract), two templates (loop-draft, loop-pilot-report), one worked example, and structural pytest tests that only assert files/strings exist. Each loop is a 5-part discipline (trigger/action/proof/memory/stop condition) whose thesis is proof-oriented stop conditions: produce an inspectable artifact (demo, baseline, commit, runtime check, handoff) instead of claiming done on confidence. It also defines a draft->pilot->promote lifecycle for generating new loops from interflux/flux-review/incident/transcript analysis.

Rationale: The product thesis — “proof before narrative, stop truthfully, evidence not confidence” — is independently validated by 2026 SOTA (verification is the bottleneck; private-codebase tasks underperform headline benchmarks; controlled-blast-radius; benchmark skepticism). The writing is mature, coherent, and low-conflict. The reason it is modernize rather than keep: every loop’s proof and stop condition is enforced only by the model reading prose and choosing to comply. The plugin describes verification but ships zero deterministic enforcement — no hook, no verifier subagent, no machine-checkable evidence schema, no StructuredOutput contract for its own evidence table. That is precisely the failure mode SOTA flags (“generic unstructured NL critiques are largely ineffective for refinement — structured, tool-grounded feedback drives improvement”). So the plugin embodies the right doctrine in the exact implementation form the literature says is weak. A small enforcement layer turns it from advisory prose into a real gate.

SOTA gap: Tool-grounded / executable verification (ReVeal, arXiv:2506.11442; “From Code to Courtroom,” arXiv:2510.24367; Bash-validation arXiv:2506.11237): the loops assert “proof” and “stop condition” but provide no deterministic check — proof is whatever the LLM narrates. Secondary gap: deterministic-skeleton + LLM-island structured output (Pre3/IterGen + the StructuredOutput contract) — the loop-draft/pilot/evidence-table artifacts are freeform markdown with no machine-parseable schema, so promotion gates can’t be mechanically enforced. The draft->pilot->promote lifecycle is also a hand-authored stand-in for self-evolving skill libraries (SkillFoundry/SkillFlow, EvolveR) but never closes the loop with logged evidence.

Redundancy: Significant conceptual overlap, no code overlap. Its loops restate capabilities that already exist as executable skills elsewhere: intertest:verification-before-completion and clavain:verify (proof-of-done), clavain:quality-gates + intercheck:quality (gate enforcement), clavain:handoff (Agent Handoff Integrity loop), interhelm:cuj-verification / runtime-diagnostics + clavain:smoke-test (Runtime Reality, Scenario Streak loops), clavain:landing-a-change (Release Boundary), and intermix:matrix-eval / interlab (Benchmark Baseline). interloop is the meta-catalog framing over these but calls none of them — the cleanest modernization is to make each loop dispatch to the existing verifier/gate skill rather than re-describe it in prose.

Actions:

Add a verifier hook or thin verifier-subagent so each loop’s stop condition runs a real check (tests/build/runtime probe) and emits a machine-parseable result, instead of relying on the model to self-attest — grounds proof in executable signal per ReVeal/Bash-validation SOTA.
Convert the loop-draft, loop-pilot-report, and the worked-example ‘promotion evidence table’ into a schema-constrained artifact (StructuredOutput / deterministic-skeleton) so promotion gates are mechanically checkable rather than freeform markdown.
Wire loops to compose existing interverse verifiers (intertest:verification-before-completion, clavain:quality-gates, clavain:handoff, interhelm:cuj-verification, clavain:landing-a-change) instead of restating them as prose, resolving the redundancy and giving loops real teeth.
Close the draft->pilot->promote loop with logged evidence (e.g. via interspect/intersearch session history as a replay buffer) so promoted loops are backed by recorded outcomes, moving toward a self-evolving skill library rather than manual approval only.