flux-point-studios/fluxpoint
Flux Point engineering harness: one Definition of Done, one Evidence discipline, two drivers. Loop mode runs a deterministic Stop-hook DoD gate over one agent's cycle; graph mode compiles a declarative WORK.md IR into verified multi-agent campaigns. The harness decides done, never the agent.
Adversarial semantic pass over the WORK.md IR — context scoping, tier-vs-stakes, prompt quality, hidden coupling — ending VERDICT SOUND or REWIRE. Structure is the compiler's job. Codex entry point for the fluxpoint graph-audit command.
Turn a campaign goal into a validated WORK.md IR — nodes with named contracts, verification tiers chosen by stakes, a budget ceiling — compiling clean and audited SOUND before it may run. Codex entry point for the fluxpoint graph-design command.
Compile WORK.md's IR to a Workflow script, execute it, record provenance and the Evidence row — targeted node repair and cached resume on partial failure, never a restart from zero. Codex entry point for the fluxpoint graph-run command.
Draft and stress-test a spec before implementation with researched options, recommended baselines and evidence. Use for spec-first design, grill-me requests, or the required decision pass before Flux Point loop and graph implementation.
Scaffold the Flux Point harness into the current repo — harness contract, WORK.md, outer loop runner, settings and .gitignore wiring — then tailor the harness and prove it green. Codex entry point for the fluxpoint init command.
Migrate a pre-1.0 repo onto the unified WORK.md contract, via the tested migrate script — plan, apply, verify, then finalize. Codex entry point for the fluxpoint migrate command.
Check that verification got stronger, not just greener — run the proof-strength ratchet, then the proof-auditor over the diff. Codex entry point for the fluxpoint proof-audit command.
Search the recall graph — lessons, decisions, counterexamples, and primitives — by meaning, identifiers, and graph proximity; or build/inspect the index. Codex entry point for the fluxpoint recall command.
Review a working change or PR with attack coverage, executable evidence, a composition pass and skeptical reproduction. Codex entry point for the fluxpoint red-team command.
Release a node parked on a human or a third party — record the proof that the work actually happened, so the campaign can resume on evidence rather than recollection. Codex entry point for the fluxpoint release command.
Report the harness state — gate counters, last harness verdict, WORK.md progress, recorded graph runs — and recommend the single next action. Codex entry point for the fluxpoint status command.
How to run Graph Engineering at Flux Point — deciding when a loop must become a graph, specifying multi-agent campaigns as a declarative IR in WORK.md with typed contracts and verification tiers, compiling it deterministically to a Claude Code Workflow script, and repairing runs with targeted resume. Use whenever the user mentions graphs, graph engineering, multi-agent work, orchestration, councils, judge panels, fan-out, swarms, wiring or organizing subagents, WORK.md, workflow scripts, or asks to parallelize verified work across agents, even if they never say "graph engineering".
How to run Loop Engineering at Flux Point — choosing between /goal, native /loop, /schedule Routines, and the outer scripts/loop.sh runner, writing verifiable completion conditions, and working with the Stop-hook DoD gate. Use this whenever the user mentions loops, Ralph, /goal, /loop, /schedule, routines, watching or monitoring a build, deploy, or transaction, autonomous or overnight runs, "keep going until green", iteration budgets, or asks the agent to work unattended on a task, even if they never say "loop engineering".
How to work with credentials at Flux Point — never reading a secret, passing its path to code that emits only public derivations, keeping a passphrase off the command line, filtering config dumps at the source, and the repr hazard that leaks a signing key through an f-string or a traceback. Use whenever the work touches a mnemonic, a wallet file, a signing key, a passphrase, a signed transaction or witness, a .env or an API token, whenever a permission rule denies a read the work actually needs, whenever an operator is about to hand-edit settings.json to unblock a session, or whenever custody of a key must be proved, restored or rotated, even if nobody says "secret handling".