interflux

interflux

Multi-agent review, research, and model evaluation engine for Claude Code.

What this does

interflux is the review infrastructure behind Clavain's /flux-drive and /flux-research commands. Point it at a document or codebase and it runs a scored triage to figure out which domain agents are relevant, slices the content so each agent only sees what it needs, injects project-specific knowledge, and synthesizes the results.

There are 12 review agents (7 technical: architecture, safety, correctness, performance, quality, game design, user/product; 5 cognitive: systems thinking, decision quality, human systems, adaptive capacity, sensemaking) and 5 research agents (best practices, framework docs, git history, learnings, repo research). The triage step means you don't pay for all of them on every review — only the relevant ones fire.

The engine also powers flux-research, which dispatches research agents in parallel and synthesizes their findings, and flux-review, which generates review agents across a spectrum of semantic distance for deep analysis.

FluxBench

FluxBench is the closed-loop model evaluation system. It scores candidate models against ground-truth test fixtures, qualifies them for production use, monitors for quality drift, and manages a challenger slot for evaluating new models alongside the existing fleet.

The pipeline: qualify → score → sync → drift → challenger. Each script is independently testable and connected through shared contracts — qualification_run_id flows through the entire pipeline, qualified_baseline is write-once to prevent ratchet erosion, and concurrent writes are protected by flock (fd 200 for JSONL, fd 201 for registry).

Installation

First, add the interagency marketplace (one-time setup):

/plugin marketplace add mistakeknot/interagency-marketplace

Then install the plugin:

/plugin install interflux

Usage

Review a document or codebase:

/flux-drive docs/plans/my-feature.md

Research a topic with parallel agents:

/flux-research "best practices for SQLite WAL mode in Go CLI tools"

Deep multi-track review (adjacent → orthogonal → esoteric):

/flux-review docs/brainstorms/my-idea.md

Generate project-specific review agents:

/flux-gen

Explore ideas through progressively distant knowledge domains:

/flux-explore "agent orchestration patterns"

Run FluxBench qualification against test fixtures:

scripts/fluxbench-qualify.sh <model-slug> --mock --fixtures-dir tests/fixtures/qualification/

Architecture

agents/
  review/          12 review agents (fd-architecture, fd-safety, fd-correctness, ...)
  research/        5 research agents (best-practices, framework-docs, git-history, ...)
  fluxbench-discover.md   Weekly model discovery agent spec
skills/            flux-drive (frontmatter: flux-engine), flux-review (frontmatter: flux-review-engine)
commands/          7 commands (flux-drive, flux-research, flux-review, flux-gen,
                     flux-explore, flux-agent, fetch-findings)
config/flux-drive/
  budget.yaml              Token budgets, agent dropout, challenger slots
  agent-roles.yaml         Role → model tier mapping (planner/reviewer/editor/checker/challenger)
  model-registry.yaml      Candidate models, FluxBench scores, qualified baselines
  fluxbench-metrics.yaml   9 metric definitions (5 core gates + 4 extended)
  fluxbench-thresholds.yaml  Calibrated thresholds (overrides metric defaults)
  domains/                 12 domain profiles with research directives
  discourse-*.yaml         4 discourse modes (fixative, Lorenzen, Sawyer, topology)
scripts/
  fluxbench-score.sh       Scoring engine — 9 metrics, bipartite finding matching, flock writes
  fluxbench-calibrate.sh   Derives thresholds from fixture runs (p25 conservative)
  fluxbench-drift.sh       Drift detection with hysteresis and correlated-drift protection
  fluxbench-sync.sh        Store-and-forward sync to AgMoDB (idempotent, no git commits)
  fluxbench-qualify.sh     Qualification runner — fixtures → score → promote/reject
  fluxbench-challenger.sh  Challenger slot lifecycle (select, evaluate, status)
  discover-models.sh       Model discovery via interrank/AgMoDB
  estimate-costs.sh        Token cost estimation with slicing discount
  generate-agents.py       Project-specific agent generation from domain detection
  flux-agent.py            Agent dispatch engine
tests/
  test_fluxbench_*.bats    24 bats tests across 5 test files
  fixtures/qualification/  5 ground-truth fixtures (null-check, SQL injection, naming,
                             race condition, API design)
  test-budget.sh           Budget config validation
  test-findings-flow.sh    Finding sharing flow tests
hooks/
  session-start.sh         Budget signal reading + FluxBench model awareness