minmaxing

agent
Security Audit
Warn
Health Warn
  • License — License: Apache-2.0
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Pass
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

The ultimate MiniMax 2.7 Highspeed Claude Code harness - 100 TPS, 204K context, GStack skills, 5-tier memory, 10x productivity.

README.md

minmaxing

Install

Pick the command that matches the folder.

Clean/new folder:

curl -fsSL https://raw.githubusercontent.com/waitdeadai/minmaxing/main/setup.sh | bash -s -- --minimax-key 'YOUR_TOKEN_PLAN_KEY'

Existing project or harness update:

curl -fsSL https://raw.githubusercontent.com/waitdeadai/minmaxing/main/setup.sh | bash -s -- --import-existing --minimax-key 'YOUR_TOKEN_PLAN_KEY'

Get your key from platform.minimax.io.

Optional Claude-only suggested install, no MiniMax key:

curl -fsSL https://raw.githubusercontent.com/waitdeadai/minmaxing/main/setup.sh | bash -s -- --mode opussonnet

After install, start Claude yourself when you are ready:

claude

Then use the definitive workflow command:

/opusworkflow "build or fix the thing"

Mental model:

  • /opusworkflow is the product command developers use day to day and the
    definitive route for this workflow.
  • It requests Opus 4.7 high/xhigh for planning/review when the account proves
    that model is available.
  • It runs /specqa after SPEC.md and before implementation, so the active
    spec gets SOTA/currentness QA before code starts.
  • It uses MiniMax-M2.7-highspeed as the bounded executor for bulk edits and
    repair loops.
  • It must drive to a verified result, partial result, or blocked repair path.
  • /opusminimax is the advanced engine underneath. Use it directly only when
    debugging provider split, packet, repair, or benchmark behavior.

MiniMax M2.7 Highspeed Claude Code 100 TPS 204K Context

Delegate execution. Keep judgment. Require evidence.

Setup adds a governed Claude Code harness where AI researches with an efficacy-first agent budget, audits the codebase, writes a concrete plan and SPEC.md, runs Spec QA, implements with clear ownership, and produces evidence before you trust the result.

Stars Forks Issues License


Open-Core Boundary

minmaxing is the open-source core for governed AI workflows. The public repo is
licensed under Apache-2.0 and includes the harness, skills, rules, AgentFactory
contracts, verification patterns, and safe example blueprints.

The private commercial moat is not included: REVCLI/Revis runtime code,
customer-specific Hermes agents, enterprise connectors, memory seeds, audit
logs, operational playbooks, tenant infrastructure, and managed service
delivery remain private unless explicitly published later.

Read the operating boundary before packaging or selling anything from this
repo:

  • OPEN_CORE_STRATEGY.md
    defines what is public, what stays private, and why the moat lives in runtime
    operation rather than prompt files.
  • COMMERCIAL.md explains the
    managed-service boundary and public-claims rules.
  • SECURITY.md defines safe
    vulnerability reporting and secret/customer-data handling.
  • TRADEMARKS.md keeps project
    names, certification claims, and managed-service branding separate from the
    Apache-2.0 code license.

Runtime Governance

Use minmaxing as a governed harness, not as an unverifiable autonomy claim. The
public core is safe to validate without secrets:

bash scripts/test-harness.sh
bash scripts/release-check.sh --static-only

Runtime checks are intentionally separate. Authenticated Claude Code smoke tests
require local credentials and must not run by default on public PRs:

RUN_CLAUDE_INTEGRATION=1 bash scripts/test-harness.sh

Choose the runtime profile deliberately:

  • Default in this repo: trusted-local bypassPermissions for the operator's
    solo workflow. This is intentionally fast and intentionally riskier: use it
    only on a private machine/repo where you accept local command authority.
  • solo-fast: tracked example of the same trusted-local speed posture.
  • team-safe: shared-work fallback with acceptEdits; start from
    .claude/settings.team-safe.example.json for teammates, reviews, clients, or
    any workspace where automatic tool permission is too much authority.
  • ci-static: no-secret static validation for public PRs and release checks.
  • ci-runtime: manual authenticated validation through
    .github/workflows/harness-runtime.yml.

The quickstart is here:

Public examples must use dummy data only:

The harness is effective when its gates produce evidence: SPEC.md,
Agent-Native Estimates, parent-verified worker outputs, aggregate verification,
memory health/freshness checks, and command-backed closeout.

Setup Commands

There are two setup commands and they do different jobs.

  • Clean/new folder: use this only in an empty folder where minmaxing can
    become the project scaffold.
  • Existing project or harness update: use this inside a real app/repo. It
    imports or updates harness-owned files only, skips project-file conflicts,
    preserves .env, README.md, SPEC.md, taste.md, taste.vision, app
    code, package files, and .git, and records imported file hashes in
    .minimaxing/import-manifest.tsv so future runs can update files it owns.

Both install commands land on the same simple UX: use /opusworkflow for
normal work. That means Opus 4.7 high/xhigh is requested for planning,
adversarial review, and final ship/no-ship judgment when runtime identity is
proven, while MiniMax-M2.7-highspeed is the bounded executor for bulk coding
and repair. It is not a magic guarantee that every external blocker disappears;
it is a hard closeout discipline: verified result, partial result, or blocked
repair path. /opusminimax remains the lower-level engine behind that route,
not another daily command developers need to choose. The installer configures
the ignored local MiniMax executor profile, keeps the Opus planner profile
provider-clean, uses the default trusted-local bypassPermissions posture, and
then exits. Open Claude yourself with claude after setup finishes, so install
failures, warnings, and conflict messages stay visible.

Inline token commands can land in shell history. That is the intentional
fast path for trusted solo work; environment-variable, hidden-input, key-file, and explicit
--mode minimax|opusworkflow|opusminimax|opussonnet override forms still exist in
bash setup.sh --help, but they are not the default path.

If an earlier install did not detect MiniMax, rerun the existing-project command
above from the project root. In /opusworkflow split mode, MiniMax is configured
in the ignored executor profile .claude/settings.minimax-executor.local.json;
it is not automatically added as a user-scope Claude MCP server.

That's it. Memory system, governed runtime profiles, and 38 skills are
configured.

Suggested Claude-Only Install

The standard mode above is still the recommended MiniMax-backed workflow. If
you want the whole harness without a MiniMax Token Plan, use the optional
Claude-only profile:

curl -fsSL https://raw.githubusercontent.com/waitdeadai/minmaxing/main/setup.sh | bash -s -- --mode opussonnet

Existing repo/updater form:

curl -fsSL https://raw.githubusercontent.com/waitdeadai/minmaxing/main/setup.sh | bash -s -- --import-existing --mode opussonnet

This prepares ignored local profiles for Claude Code opusplan, pins
claude-opus-4-7 for planning/judgment and claude-sonnet-4-6 for execution,
keeps MiniMax base URLs out of the Claude-only profiles, and leaves the normal
governance hooks, /workflow lifecycle, /introspect, /verify, and
bypassPermissions trusted-local posture in place. After install:

claude
/opussonnet "build or fix the thing"

Use this as a suggested alternative, not the default budget strategy. Runtime
Opus access still depends on your Claude account state; use /status or an
explicit runtime check before claiming Opus 4.7 actually planned a run.

You can also choose an explicit governed model profile per workflow without
changing the default:

bash scripts/opusworkflow.sh --task "build or fix the thing" --model-profile sonnet
bash scripts/opusworkflow.sh --task "build or fix the thing" --model-profile opus
bash scripts/opusworkflow.sh --task "build or fix the thing" --model-profile custom --planner-model claude-sonnet-4-6 --executor-model claude-sonnet-4-6

minimax remains the default. opussonnet, sonnet, opus, default, and
custom are explicit operator choices; static artifacts record the request, but
runtime identity still depends on the current Claude Code account/session.

Claude subscription auth is separate account auth. Run claude auth login once
if this machine is not already logged in; the setup command does not store or
fake your Claude subscription session.

Native Claude Code Remote Control

Claude Code has native Remote Control for continuing a local session from
claude.ai/code or the Claude mobile app. In this harness, /remote-control
is a readiness/troubleshooting skill, not the live activator, because the
project skill can shadow a native slash command with the same name.

Use this shell command to start the live native Remote Control server:

claude remote-control

For this trusted-local workspace, an explicit launch is:

claude remote-control --name ultimateminimax --permission-mode bypassPermissions

Keep that process running, then connect from https://claude.ai/code or the
Claude mobile app. This is different from claude --remote or Claude Code on
the web. Remote Control keeps the Claude Code process running locally, with
this repo's hooks, tools, project settings, and trusted-local
bypassPermissions posture. The harness intentionally does not create a custom
remote server, websocket bridge, or API-key control path.

Before using it, run the no-secret static doctor:

bash scripts/remote-control-doctor.sh --static --json

Remote Control requires claude.ai subscription login. ANTHROPIC_API_KEY,
CLAUDE_CODE_OAUTH_TOKEN, third-party provider auth, DISABLE_TELEMETRY, and
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC can block eligibility, so shared
project settings avoid those variables.

Shared settings are committed on purpose, but they are provider-neutral. .claude/settings.json contains governance hooks, deny rules, and the trusted-local bypassPermissions default. This is an explicit operator-speed choice, not a team-safety recommendation. Real credentials and provider identity belong in ignored local files such as .claude/settings.opusminimax-planner.local.json and .claude/settings.minimax-executor.local.json.

Fresh repos should start with /tastebootstrap. It asks the 10 kernel questions, writes taste.md + taste.vision, and gives /workflow explicit taste to follow before anything is built.

Note: Setup adds hardware auto-detection to ~/.bashrcMAX_PARALLEL_AGENTS is set automatically for new Bash sessions that source ~/.bashrc.

Windows Setup

Recommended: use WSL 2 for the smoothest experience.

Claude Code officially supports both:

  • Native Windows + Git Bash
  • WSL 2

minmaxing's installer is Bash-first and updates ~/.bashrc, so:

  • WSL 2 is the best fit if your project already lives in Linux tooling
  • Git Bash is the best fit if your project lives on native Windows

Option 1: WSL 2 (recommended)

Install Claude Code inside WSL, then run minmaxing there too.

  1. Install WSL 2 and open your Linux distro.
  2. Install Claude Code inside WSL:
curl -fsSL https://claude.ai/install.sh | bash
  1. Move into your project directory inside WSL.
  2. Run the clean/new-folder command or the existing-project/updater command from the top of this README.

If you use the optional OpenAI Codex plugin in WSL, the same Linux flow applies there too.

Option 2: Native Windows + Git Bash

Claude Code's official Windows docs say native Windows requires Git for Windows. Claude can be launched from PowerShell, CMD, or Git Bash, but this repo's setup script should be run from Git Bash because it is a Bash script and writes shell config to ~/.bashrc.

  1. Install Git for Windows.
  2. Install Claude Code from PowerShell:
irm https://claude.ai/install.ps1 | iex

Or install it with WinGet:

winget install Anthropic.ClaudeCode
  1. Install Python 3.11+ and uv/uvx:
winget install Python.Python.3.11
winget install --id=astral-sh.uv -e
  1. Open Git Bash and run the clean/new-folder command or the existing-project/updater command from the top of this README.

  2. Verify Claude Code:

claude --version
claude doctor

If Claude Code cannot find your Git Bash installation, Anthropic documents this setting:

{
  "env": {
    "CLAUDE_CODE_GIT_BASH_PATH": "C:\\Program Files\\Git\\bin\\bash.exe"
  }
}

Windows Notes

  • Native Windows does not support Claude Code sandboxing; WSL 2 does.
  • PowerShell support in Claude Code is rolling out progressively, but minmaxing setup is still Bash-based.
  • If uvx is already available, the installer will use it. If not, preinstalling uv on Windows avoids shell-specific installer edge cases.

The Problem

AI coding's #1 failure mode: building the wrong thing.

Without minmaxing:
"add user auth"
→ AI builds something that looks like auth
→ "that's not what I wanted"
→ AI rebuilds
→ Hours wasted

With SPEC-first:

"add user auth"
→ AI asks: "What happens when login fails?"
→ You: "show error message"
→ AI writes SPEC.md with exact behavior
→ AI implements to spec
→ AI verifies against spec
→ You accept or reject

What Makes minmaxing Different

This harness Other harnesses
Taste-first: /workflow starts from taste.md + taste.vision No taste — just "build X"
5-tier memory: captures durable decisions, patterns, and errors when hooks/CLI are healthy Tabula rasa every session
Central orchestrator: /workflow runs research → audit → plan → spec → execute → verify with hard gates Skills are isolated, no chaining
Health-checked capture: outcomes and reusable lessons can be logged to memory Manual documentation
SPEC-first: write spec before code Vague prompts, rebuild loops
Efficacy-first parallelism: use the right number of agents for the task Sequential one-at-a-time
Independent verification pass with metadata when isolation can be proved Casual self-checks
Research-first: /workflow drafts a plan, runs the repo’s effectiveness-first deepresearch search -> read -> refine investigation loop, and keeps an inspectable source ledger before planning or edits AI hallucinates best practices
Surgical diffs: vague requests become verifiable contracts, then every meaningful change needs a changed-line trace to SPEC.md Drive-by refactors, speculative abstractions, and "while I was here" churn

Taste is the kernel. /workflow starts by checking taste.md and taste.vision. In a fresh repo, define them with /tastebootstrap before anything is built.

Taste is NOT just frontend/aesthetic, and it is not a frontend/backend checklist. It's the project's operating kernel — design principles, experience direction, interface contracts, observability rules, architecture constraints, code style, intent, non-goals, and values. When /tastebootstrap asks its 10 kernel questions, it captures the vision and guardrails that make the whole system coherent.

Memory is durable when healthy. minmaxing uses working-state handoffs, flat-file audit notes, and SQLite/FTS5 retrieval so reusable decisions, patterns, and failures can survive beyond one chat. Run bash scripts/memory.sh health when memory quality matters.

Harness quality matters. The point is not to pretend the model became smarter. The point is to reduce verification debt: force research, write the contract, preserve context, inspect assumptions, and prove the outcome before trust.


Key Features

Efficacy-First Parallelism

Spawn up to the hardware-aware ceiling when the task genuinely has bounded,
independent packets. Supervisor decomposes tasks, workers execute in parallel,
supervisor verifies.

Supervisor (you/AI)
├── Worker 1 → File A
├── Worker 2 → File B
├── Worker 3 → File C
... (up to 10)
└── Supervisor verifies all

Auto-detects your hardware: 32GB+ and 8+ cores -> 10 agents, 16GB and 4+ cores -> 6 agents, 8GB and 2+ cores -> 3 agents. For a live profile, run bash scripts/parallel-capacity.sh --summary.

The important part is not hitting the ceiling. The important part is using the smallest agent budget that preserves fresh context, clear ownership, and a shorter critical path.

Planning Time Awareness

minmaxing estimates meaningful work in agent-native wall-clock terms before a
plan or SPEC.md is frozen. The required Agent-Native Estimate separates
elapsed agent wall-clock, total agent-hours, human touch time, calendar
blockers, critical path, and confidence.

Human-equivalent estimates are allowed only as secondary context. The harness
must not say "6 weeks" as the main estimate when the real plan is a 24/7
agent-native DAG with 5 effective lanes, or when the honest answer is
blocked/unknown because credentials, CI, deploy windows, or product decisions
are missing. Current local capacity evidence comes from
bash scripts/parallel-capacity.sh --json.

Parallel Mode

/parallel is the dense-work orchestrator. It runs a Parallel Eligibility Audit,
reads the hardware capacity profile, chooses an execution substrate, writes a
Packet DAG and Ownership Matrix, inserts Sync Barriers, aggregates Worker Result
Schema returns, then verifies the aggregate against SPEC.md.

Execution substrate selection is explicit:

  • local for one tight reasoning loop, shared files, low hardware, or high coordination cost.
  • subagents as the default for bounded same-workspace research, audit, implementation, and review packets.
  • parallel-instances only for large disjoint work where separate sessions or worktrees materially shorten elapsed time.
  • agent-teams only as opt-in experimental behavior when explicitly enabled and fallback is documented.

/workflow may auto-consider /parallel for dense work, but it downgrades when
independent packets, file ownership, host capacity, or verification are weak.
Main stays orchestrator; workers do not own taste, architecture, SPEC,
security, registry, or final verification decisions.

SPEC-First

File-changing tasks start with research, code audit, and a concrete plan. SPEC.md is the formal contract that comes out of that work, and AI implements to spec.

Even tiny local file-changing tasks should still produce a small SPEC.md when you invoke the full /workflow contract.

Spec QA

/specqa is the Spec QA Agent that runs after SPEC.md is created or updated and before implementation. It checks requirements quality, measurable success criteria, SOTA 2026/currentness, security and governance risk, verification readiness, and improvement suggestions.

Under /opusworkflow, Spec QA requests an Opus 4.7 high/xhigh reviewer when runtime identity is proven. The harness must not claim Opus 4.7 reviewed a spec unless /status, a sentinel, or a durable artifact proves it. When SOTA or other time-sensitive facts matter, Spec QA requires webresearched actual-time data and a source ledger before execution can proceed.

Plan Mode Auto-Approval

/opusworkflow now records a plan-mode checkpoint before execution. The default --plan-mode-policy auto sets plan_mode.auto_approval.status=auto_approved_when_gates_pass, which means implementation can start automatically only after the workflow has recorded the research brief, code audit, /introspect pre-plan pass, Agent-Native Estimate, SPEC.md decision, and /specqa execution-allowed decision.

This is workflow transition approval, not a shortcut. It does not replace native Claude Code Plan Mode, SPEC.md, /specqa, /introspect, /verify, runtime model identity proof, or /visualizeworkflow human approval. Use --plan-mode-policy manual when a human review should be required after the plan checkpoint.

Surgical Diff Discipline

minmaxing does not just ask the model to "be careful." It requires the smallest sufficient implementation, no speculative abstractions, no drive-by refactors, and a changed-line trace from meaningful diff hunks back to SPEC.md.

That means vague requests become verifiable contracts before code changes, and unrelated cleanup stays out of the diff unless the spec makes it necessary.

Active SPEC + Archive

SPEC.md stays as the active contract in the project root so /verify and fresh agents always know what to read first.

Before a new task replaces a non-reused active spec, minmaxing archives the previous one under .taste/specs/ with a descriptive task/outcome filename. Verified closeout archives the final spec too, deduplicated by content hash, so repeated closeouts do not create noisy copies.

Independent Verification Pass

Verification is not a vibe check. /verify reads SPEC.md, checks every success criterion, records evidence, and now asks the workflow artifact to capture executor/verifier metadata when isolation is known. If the run cannot prove a separate agent/model/workspace, it must say unknown instead of pretending.

Research-First

AI training data can be stale, and repo context can be incomplete. Current external facts get verified when they materially affect the plan; purely local tasks can justify a local-only research brief instead of doing fake research theater.

/workflow now treats a research brief as mandatory for all tasks, with the MiniMax MCP as the preferred source whenever current external facts matter. But the brief is not just a search tally. The workflow now uses the repo’s effectiveness-first deepresearch protocol: draft a collaborative research plan, run an iterative search -> read -> refine loop, keep a source ledger, challenge conflicting evidence, and do targeted follow-up research before freezing the plan. It still uses up to MAX_PARALLEL_AGENTS tracks, but only when the added tracks are distinct and plan-changing. For a purely local task, it can justify a local-only research brief instead of doing pointless external calls.

Agent Factory

/agentfactory is a governed workflow for creating Hermes agents, not a prompt template. It uses the same effectiveness-first spine as /workflow: taste gate, 12-question intent intake, deepresearch brief, runtime audit, capacity-aware runtime contract, manifest, least-privilege capability stack, Hermes SPEC.md, generated agent files, hard-gate introspection, independent verification, registry closeout, and memory integration.

A Hermes agent can operate one workflow, one department lane, or one bounded subsystem. A fleet can operate a larger business process only by composing narrow agents with explicit handoffs, not by giving one agent omnipotent company authority.

Agent Factory writes its own run artifact under .taste/workflow-runs/*-agentfactory.md and keeps the durable registry in hermes-registry.md. The secondary factory taste contract lives in hermes-factory.taste.md. Production status requires hermes.runtime.json, development_host_profile, target_runtime_profile, host_capacity_profile, capacity_binding, concurrency_budget, queue/backpressure behavior, degrade_policy, a passing kill-switch test, verifier metadata, source ledger, memory-coherence check, runtime evidence, and registry links to manifest, spec, verify, kill-switch, and runtime artifacts. The local scripts/parallel-capacity.sh profile describes the developer machine unless the target runtime is explicitly local; a cloud server, VPS, container host, CI runner, managed workflow runtime, or REVCLI fleet needs its own target runtime capacity evidence before an agent can be marked active. The dedicated scripts/agentfactory-smoke.sh stress test keeps the skill from regressing into a checklist by checking negative fixtures such as raw secrets, read-only agents with write actions, missing capacity budgets, untested kill switches, active operator_exception, and verifier overclaims.

For REVCLI/Revis-style products, /agentfactory treats Hermes as the role-scoped interaction/runtime shell, REVCLI/Revis as the policy and audit control plane, and Odoo or the configured database as the system of record. The repo includes REVCLI_HERMES_AGENT_MAP.md so generated agents map to concrete roles, approval gates, runtime evidence, kill switches, and closed-loop outcomes instead of broad “run the company” authority.

Permission Mode

  • Project default: provider-neutral trusted-local bypassPermissions with governance hooks and secret-read denies. Warning: this allows Claude Code to act without normal permission prompts, so use it only where you accept local operator risk.
  • solo-fast option: tracked example of the same trusted-local fast profile for personal repos where you want fewer prompts.
  • Team-safe option: copy .claude/settings.team-safe.example.json to your local settings and keep defaultMode at acceptEdits.
  • Definitive workflow command: use /opusworkflow for all ordinary
    mutating work. It is the product-facing command: Opus 4.7 high/xhigh
    plans/reviews when proven available, and MiniMax-M2.7-highspeed executes
    bounded packets. Closeout is verified, partial, or blocked-with-repair.
  • Model profile override: add --model-profile sonnet, --model-profile opus, --model-profile opussonnet, --model-profile default, or --model-profile custom --planner-model MODEL --executor-model MODEL when you intentionally want a different Claude model route for that run.
  • OpusMiniMax engine: use /opusminimax directly only when you need
    benchmark, repair, provider, or lower-level packet control. It is not the
    day-to-day command.
  • OpusSonnet option: use /opussonnet when you want the same governed harness without a MiniMax token, via Claude Code opusplan with Opus 4.7 planning and Sonnet 4.6 execution.
  • If you want even more guardrails, switch your local Claude session to plan before high-risk work.

OpenAI Codex Plugin

This repo now ships a project-scoped Codex config under .codex/config.toml plus focused Codex agents in .codex/agents/ so the official OpenAI Claude Code plugin can inherit sane defaults when you use Codex from inside Claude Code.

Research-backed take:

  • The best plugin for using Codex inside Claude Code is the official OpenAI openai/codex-plugin-cc.
  • It is explicitly built for Claude Code users, uses your local codex CLI plus Codex app server, and picks up user-level or project-level .codex/config.toml.
  • It does not force parallelism by itself. OpenAI’s Codex docs say subagents are only spawned when you explicitly ask for them, so effective concurrency comes from your prompt plus the project max_threads ceiling.
  • I found no official source saying this official plugin violates OpenAI terms. This is not legal advice, but the repo itself documents Claude Code installation and current OpenAI Terms/Usage Policies do not appear to prohibit this official integration when used compliantly.

What this repo config gives Codex:

  • Main Codex default: gpt-5.5 with medium reasoning and detailed reasoning summaries
  • Codex Memories enabled where supported, so useful local context can carry into future Codex sessions
  • Subagent ceiling: 10 via [agents].max_threads
  • OpenAI docs MCP: https://developers.openai.com/mcp
  • Repo-scoped skill: .agents/skills/codex-imagegen for SPEC.md image
    assets that should use Codex subscription/ChatGPT image generation rather
    than OpenAI API-key billing
  • Focused helper agents:
    • repo_explorer
    • reviewer
    • docs_researcher

Taste-First (Kernel)

Every /workflow invocation checks taste.md + taste.vision before doing anything. In a fresh repo, run /tastebootstrap first. /workflow is the executor, not the bootstrap interview.

5-Tier Memory (Persistent)

SQLite-backed memory that remembers across sessions:

  • Every session start/end logged (episodic)
  • Decisions, patterns, errors all stored (semantic, procedural, error-solution)
  • Causal graph tracks what caused success/failure

Compaction-Safe Working State

LLMs forget because live conversation context is lossy. minmaxing now keeps a compact task handoff in .minimaxing/state/CURRENT.md by default.

Claude Code hooks refresh it after each turn, snapshot it before /compact or auto-compact, record the compact summary, and rehydrate it on startup, resume, and post-compact. Durable lessons still go to SQLite memory; CURRENT.md is only for the active task: files in play, current phase, latest SPEC.md, workflow artifact, verification status, and next steps.

Temporal Anchor

Claude Code models do not know the present date from pretraining. minmaxing
injects a fresh local-system-clock anchor through .claude/hooks/time-anchor.sh
at session start and before each user prompt. The same anchor is written into
.minimaxing/state/CURRENT.md snapshots.

Use it directly when needed:

bash scripts/time-anchor.sh text

For "today", "latest", "current", "recent", "SOTA 2026", pricing, models,
provider behavior, laws, docs, benchmarks, schedules, or news, research must
resolve dates against that anchor and verify live sources. If live verification
is not available, the correct answer is insufficient_data, stale, or
unverified, not a confident pretrained-memory claim.

Central Orchestrator

/workflow owns the full lifecycle inline: taste gate, deep research, code audit, hard-gate introspection, plan, SPEC.md, implementation, verification, and closeout. For file-changing tasks it also leaves a workflow artifact under .taste/workflow-runs/ and archived specs under .taste/specs/ so the research plan, loop log, source ledger, audit, introspection, plan, and spec trail stay inspectable. Specialist skills still exist, but /workflow no longer depends on nested custom-skill chaining to finish the job.

/digestflow is the report-informed sibling route. It adds Report Intake before deep research so Gemini, NotebookLM, ChatGPT Deep Research, Perplexity, or similar reports become provisional evidence instead of hidden assumptions.


How the Workflow System Works

Taste as OS, Skills as System Calls

Think of minmaxing as an operating system:

┌─────────────────────────────────────────────────────┐
│                    /opusworkflow                    │
│        (Default Claude judgment + MiniMax execution) │
│                      /workflow                      │
│                  (Lifecycle Engine Underneath)       │
├─────────────────────────────────────────────────────┤
│  PHASE 0: TASTE CHECK [GATE]  ← taste.md + vision │
│  PHASE 1: ROUTE                                     │
│  PHASE 2: DEEP RESEARCH (plan -> search -> read -> refine) │
│  PHASE 3: CODE AUDIT                               │
│  PHASE 3.5: INTROSPECT [HARD GATE]                 │
│  PHASE 4: PLAN                                     │
│  PHASE 5: SPEC.md                                  │
│  PHASE 6: EXECUTE                                  │
│  PHASE 6.5: INTROSPECT [HARD GATE]                 │
│  PHASE 7: VERIFY                                   │
│  PHASE 8: CLOSEOUT                                 │
├─────────────────────────────────────────────────────┤
│            taste.md + taste.vision                  │
│                  (Kernel / OS)                     │
├─────────────────────────────────────────────────────┤
│  /opusminimax(engine) /digestflow /autoplan       │
│  /verify /ship /sprint /investigate /memory       │
│  /audit /council /qa /review /deepresearch        │
│  /webresearch /browse /introspect /codesearch     │
│  /overnight /align /agentfactory                 │
│              (System Calls)                          │
└─────────────────────────────────────────────────────┘

Taste is the kernel. Every operation checks against your taste.md and taste.vision first. If the kernel is missing, stop and define it with /tastebootstrap before execution. Taste covers the full project philosophy — design principles, architecture, code style, intent, non-goals, and values.

Skills are system calls. Each skill does one thing well. They are still useful directly, but /opusworkflow is the default daily entrypoint for mutating work and /workflow is responsible for finishing the underlying end-to-end lifecycle itself.

/opusworkflow is the definitive shell. It is the default top-level route for normal build/plan work and mutating specialist work: Opus 4.7 high/xhigh is requested for judgment checkpoints when proven available, and MiniMax-M2.7-highspeed handles bounded execution packets. The default plan-mode policy auto-approves the transition to implementation only when research, code audit, /introspect pre-plan, Agent-Native Estimate, SPEC.md, and /specqa pass, recording auto_approved_when_gates_pass in the run artifact. Specialist routes are recorded as inner_contract=workflow|agentfactory|hiveworkflow|parallel|defineicp|digestaste|deepretaste|demo|visualizeworkflow, and closeout must be verified, partial, or blocked-with-repair. /opussonnet is the optional Claude-only sibling for operators who want Opus 4.7 planning plus Sonnet 4.6 execution without MiniMax.

/opusminimax is the engine, not the product command. Use it directly for
provider split debugging, packet control, repair mode, or benchmark mode. For
normal product work, use /opusworkflow and let it call the engine.

/workflow is the lifecycle underneath. It routes tasks to the right phase, performs live research, audits the repo, synthesizes the plan, writes SPEC.md, executes the work, verifies output, and gates progression. Use it directly when you explicitly want one local supervisor loop or the provider split is unavailable.

Inside Phase 2, /workflow now follows the repo’s effectiveness-first deepresearch protocol instead of a generic search fan-out: it drafts a collaborative research plan, launches only the discovery tracks that matter, reads and refines in loops, records a source ledger including reviewed but not cited sources, pressure-tests conflicting evidence, and runs follow-up research before locking the plan.

Before research planning for file-changing work, /workflow records a compact ## Metacognitive Route that explains task class, capacity evidence, effective parallel budget, chosen route, evidence required, confidence threshold, and why the full parallel ceiling was or was not used. This is the steering record; it comes before ## Research Brief and does not satisfy later /introspect gates.

Before confidence is allowed, /workflow runs the repo’s hard-gate /introspect protocol. It names likely mistakes, checks assumptions, looks for counterexamples, compares implementation against SPEC.md, identifies missing verification, downgrades confidence when evidence is weak, and blocks closeout or push when unresolved findings remain.

Before closeout, /workflow also applies surgical diff discipline: smallest sufficient implementation, no speculative abstractions, no drive-by refactors, and a changed-line trace for the meaningful diff. This keeps autonomy from turning into "helpful" churn.

The 4 Execution Paths

When you say... /workflow routes to...
"build X", "implement Y" deep research → code audit → plan → SPEC.md → implement → verify → closeout
"fix Z", "debug this" deep research → code audit → plan → SPEC.md when files change → reproduce/fix → verify → closeout
"explain" deep research → inspect → explain
"refactor", "optimize" deep research → code audit → plan → SPEC.md → implement → verify → closeout
"audit this", "analyze" deep research → inspect → findings

5-Tier Memory System

Memory is layered and inspectable, not magic. It captures task state, durable decisions, reusable patterns, error fixes, and causal notes when the hooks and CLI are healthy:

Tier What When
Episodic Session start/end Every shell start/exit
Semantic Decisions & principles /council decisions, /align verdicts
Procedural Code patterns /codesearch findings, /sprint outcomes
Error-Solution Bugs & fixes /investigate fixes, /verify failures
Graph Causal chains What caused success/failure
Commit Log Git commit summaries Every git commit (auto-summarized)

Memory uses flat-file audit notes plus SQLite-backed FTS5 search when the Python memory CLI is available. Type bash scripts/memory.sh health to see whether the current repo is healthy, degraded, or disabled.

Commit auto-summarize: Every git commit triggers commit-summarize.sh via a git post-commit hook, generating obsidian/Memory/Stories/commits/{date}-{hash}.md with structured frontmatter and a brief summary. Also written to SQLite for agent retrieval.

Working State vs Memory

.minimaxing/state/CURRENT.md is the short-lived working state that survives compaction. It is generated by hooks and should be reconciled with live repo state before edits.

SQLite memory is the durable layer for reusable lessons. When a decision, pattern, error fix, or research finding should survive future tasks, log it with bash scripts/memory.sh add ... instead of stuffing it into CURRENT.md.


From 0 to 100: Fresh Folder

Step 1: Install

Run the clean/new-folder command from the top of this README.

Step 2: Define Your Taste

Claude opens at the end of setup. If you closed it, run claude.

Then bootstrap the repo kernel:

/tastebootstrap

What happens:

  1. /tastebootstrap checks taste.md + taste.vision → don't exist
  2. /tastebootstrap asks the 10 kernel questions
  3. You answer → taste.md + taste.vision are created
  4. The repo is now ready for /workflow

Then run:

/workflow "build a REST API for a todo app"

What You Get

  • Workflow Artifact.taste/workflow-runs/... with research brief, code audit, plan, and verification trail
  • SPEC.md — exact specification created from the researched plan
  • Spec Archive.taste/specs/... snapshots of completed or superseded specs
  • Implementation — parallel agents building simultaneously
  • Verification — independent evidence pass against spec, with isolation metadata when known
  • Closeout — local completion by default, remote push only when you explicitly ask for it
  • Memory — durable lessons logged to 5-tier memory when the health check supports it

After Bootstrap

Your taste is saved. From then on:

/workflow "add user authentication"

No taste questions. /workflow knows your principles, plans, builds, verifies, and finishes the task.

Integrate Into Existing Project

Drop minmaxing into any codebase:

Step 1: Install Into Existing Folder

cd your-existing-project

Then run the existing-project/updater command from the top of this README.

This imports or updates minmaxing harness files without overwriting your
existing code. If a harness file path already exists and was not previously
imported by minmaxing, setup leaves it untouched and prints a conflict warning.

Step 2: Define Taste for This Project

Claude opens at the end of setup. If you closed it, run claude.

/tastebootstrap

Answer the 10 taste questions about your existing project:

  • What are your design principles?
  • What's your intent?
  • What's in/out of scope?
  • What experience, interface, and system rules matter?

Step 3: Use /workflow on Your Codebase

Now you can use any workflow pattern:

Command What it does
/workflow "explain this codebase" Understand what you have
/workflow "audit this for security issues" Deep security + quality audit with risk-based parallel coverage
/workflow "refactor the auth module" Research → audit → plan → spec → implement → verify
/workflow "optimize database queries" Research → audit → plan → spec → implement → verify
/workflow "investigate why X is slow" Root-cause debugging with hypothesis testing
/workflow "add REST API to existing endpoints" Research-backed spec-first build respecting existing architecture

Key Difference: Existing vs New

Aspect Fresh Project Existing Project
SPEC.md Greenfield spec Required for file-changing work; pure analysis can skip
Taste Bootstrap fresh Define from existing code
/workflow "build" Full research-backed spec-first flow Add features, respect existing patterns
/workflow "explain" N/A Understand existing codebase
/workflow "audit" N/A Find issues in existing code
/workflow "refactor" N/A Improve existing implementation

The 36 Skills

Definitive Workflow Command

Use /opusworkflow for normal file-changing work. You are not choosing
between /opusworkflow and /opusminimax:

  • /opusworkflow is the product command and definitive workflow: Opus 4.7
    high/xhigh planner/reviewer plus MiniMax-M2.7-highspeed executor.
  • /opusminimax is the advanced engine: provider split, packets, repair mode,
    benchmark mode, and low-level debugging.
  • Is /opusworkflow better? For humans, yes. It is easier and safer because it
    wraps the same engine with the normal workflow, budget, evidence, and
    verification gates.
  • Does it always succeed? It always drives to a truthful outcome: verified,
    partial, or blocked with the next repair action. It must not fake success.
Skill What It Does
/tastebootstrap Fresh-repo bootstrap — asks the 10 kernel questions and writes taste.md + taste.vision
/workflow Underlying lifecycle and explicit fallback — drives research → code audit → plan → Agent-Native Estimate → SPEC.md → implement → verify → closeout (supervises an efficacy-first agent budget)
/opusworkflow Definitive workflow command for mutating work — Opus 4.7 high/xhigh planner/reviewer when proven available, plus MiniMax-M2.7-highspeed executor; auto-approves plan-to-execution only after gates pass; closes as verified, partial, or blocked-with-repair; supports explicit --model-profile overrides
/opusminimax Advanced engine behind /opusworkflow — use directly only for provider split, packet, repair, or benchmark debugging
/opussonnet Optional Claude-only mode — uses Claude Code opusplan, pins Opus 4.7 for planning/judgment and Sonnet 4.6 for execution, no MiniMax token required
/visualize Taste-to-artifact comprehension check — creates ignored visual, diagram, prompt, or narrative artifacts without implementation
/visualizeworkflow Approval-first workflow — drafts SPEC + visualization, stops at WAITING_FOR_VISUAL_APPROVAL, then continues only with --continue
/demo Governed recorded demo pipeline — produces product recordings with Playwright evidence, bilingual voiceover, captions, manifests, and safety gates
/digestflow External-report-informed workflow — digests 1-10 AI research reports as untrusted candidate evidence, then runs the full governed workflow
/digestaste Research-to-bootstrap text — digests Deep Research .md reports into a sanitized DigesTaste Bootstrap Packet for a new or existing project
/deepretaste Intent-to-ICP-to-taste bootstrap — detects product intent, uses /deepresearch for taste-driving evidence, defines ICPs, and bootstraps or proposes/applies taste changes through /tastebootstrap or /defineicp semantics
/defineicp ICP-to-taste evolution — defines primary/secondary/anti-ICPs with deepresearch, drafts taste.md + taste.vision changes, and applies only with explicit approval
/icpweek ICP week-in-the-life stress test — simulates Monday-Sunday real usage with ideal-user, CTO, and senior product-engineer lenses, then delivers the A-J product diagnosis
/align Validate idea against taste + vision. Gates /workflow on taste mismatch.
/audit Deep codebase audit with risk-based parallelism
/autoplan Generate SPEC.md with parallel execution and Agent-Native Estimate in mind
/agentfactory Create governed runtime-bound Hermes agents with manifest, hermes.runtime.json, capability stack, memory seed, verification, registry, and tested kill switch
/parallel Run hardware-aware whole-workflow parallel orchestration with packet DAG, ownership matrix, sync barriers, and aggregate verification
/metacognition Parallel-aware control plane for task routing, evidence-grounded reflection, confidence calibration, and verified learning
/claudeproduct Official-source answers for Claude, Claude Code, Claude.ai, Anthropic API, connectors, plugins, skills, hooks, MCP, and subagents
/remote-control Native Claude Code Remote Control route with a static doctor, no custom network control plane, and no API-key auth fallback
/specqa Spec QA Agent for every active SPEC.md: requirements quality, SOTA/currentness webresearch, Opus 4.7 identity-proof boundary, and improvement suggestions before implementation
/hive Governed multi-agent coordination with role map, blackboard, dissent, synthesis, and verified evidence
/hiveworkflow Full workflow mode for hive-coordinated planning, execution, aggregation, introspection, and verification
/sprint Run an ownership-safe parallel execution wave
/verify Check output against SPEC with an independent evidence pass
/review AI review + you decide
/qa Playwright E2E testing — Pass/Fail only
/ship Pre-ship checklist + rollback plan
/investigate Debug with 3-fix limit
/overnight 8hr session with 30-min checkpoints
/council Multi-perspective analysis
/deepresearch Deep multi-pass investigation with source ledgers and follow-up loops
/webresearch Current web/docs/API verification using the same effectiveness-first method
/browse Backward-compatible alias to /webresearch or /deepresearch
/introspect Hard-gate self-audit for likely mistakes, assumptions, missing verification, and confidence downgrades
/codesearch Search code by pattern
/memory 5-tier memory system — log decisions, search patterns

Parallelism: All skills that support parallelism treat MAX_PARALLEL_AGENTS, Codex max_threads, and hardware capacity as ceilings, not targets. /align remains single-threaded by design because taste alignment is sequential judgment.

Smart Autorouting

minmaxing has smart autorouting through /metacognition and /workflow.
Before file-changing work, the harness classifies the task, reads capacity
evidence, computes the smallest useful budget, and chooses the route with the
least coordination overhead that can still improve correctness.

The routing ladder is:

/opusworkflow as the definitive workflow command and daily default for mutating build/plan/specialist work
-> local /workflow when the hybrid provider split is unavailable or explicitly bypassed
-> /digestaste when a Deep Research .md should become goal/taste bootstrap text before fresh or existing kernel decisions
-> /deepretaste when product intent, ICP, and taste kernel need a SOTA-2026 research-backed bootstrap or retaste
-> /defineicp when the product kernel needs ICP research before taste changes
-> /opusminimax only for advanced engine work: provider split, packet, repair, or benchmark debugging
-> /opussonnet when the operator explicitly wants the optional Claude-only Opus 4.7 + Sonnet 4.6 route
-> /parallel when independent execution packets are enough
-> /claudeproduct for Claude, Claude Code, Claude.ai, API, connector, plugin,
   skill, hook, MCP, subagent, availability, limit, model, or setup questions
-> /remote-control when the operator wants native Claude Code RC from
   claude.ai/code or mobile without building a custom control plane
-> /specqa after SPEC.md and before implementation when the active spec needs
   SOTA/currentness, requirements quality, and improvement-suggestion review
-> /hive for read-only coordination, or /opusworkflow with inner_contract=hiveworkflow
   when coordinated roles, blackboard state, dissent, synthesis, and mutation
   materially improve the outcome
-> blocked when evidence, ownership, capacity, or verification is missing

Use this rule of thumb:

Pick When The Developer Should Expect
/opusworkflow You want the definitive command for ordinary or specialist mutating work: Opus 4.7 high/xhigh planner/reviewer when proven available, plus MiniMax-M2.7-highspeed for bulk implementation. One-command split setup, provider doctor, default executor concurrency 1, bounded packets, outer_route + inner_contract artifacts, plan_mode.policy=auto, auto_approved_when_gates_pass, outcome_policy=verified-partial-or-blocked-with-repair, parent verification, and no silent PAYG.
/opussonnet You want the whole governed harness without MiniMax for a repo, and you are okay spending Claude subscription or extra usage on execution. setup.sh --mode opussonnet, Claude Code opusplan, pinned claude-opus-4-7 + claude-sonnet-4-6, no MiniMax base URL, same hooks and workflow gates.
`/opusworkflow --model-profile sonnet opus default
local /workflow Explicit user override, provider split unavailable, one tight reasoning loop, one shared file, unclear ownership, or coordination would slow the work down. One supervisor does the governed lifecycle and records why the hybrid outer route was not used.
/digestaste You have one or more Deep Research .md reports and want bootstrap text for a goal, fresh repo kernel, or existing project retaste. Report Intake, no-persist report bodies, prompt-injection quarantine, claim/source/conflict ledgers, DigesTaste Bootstrap Packet, /tastebootstrap for missing kernels, and /defineicp proposal/apply semantics for existing kernels.
/deepretaste You need to detect product intent, define ICPs, and bootstrap or retaste the project kernel from research-backed customer evidence. /deepresearch remains general-purpose; /deepretaste uses it only for taste-driving evidence, then routes fresh kernels through /tastebootstrap and existing kernels through /defineicp proposal/apply semantics.
/defineicp You need to define the ICP or ICPs and tailor taste.md / taste.vision to that customer profile. Deepresearch plan, primary/secondary/anti-ICPs, source and claim ledgers, taste patch proposal, explicit apply approval, backups, hashes, validation, and rollback evidence.
/opusminimax You are maintaining the engine itself: provider split, packet control, repair mode, benchmark mode, or low-level routing evidence. Provider split doctor, Opus planner artifact, MiniMax executor packets, quota-aware concurrency, parent verification, and no benchmark overclaims.
/claudeproduct The question is about Claude, Claude Code, Claude.ai, Anthropic API, connectors, plugins, skills, hooks, MCP, subagents, availability, limits, models, or setup. Official Anthropic/Claude docs first, surface separation, source ledger, connector permission/trust caveats, confidence downgrade when current docs are missing.
/remote-control You want to diagnose Claude Code native Remote Control for an already trusted local harness session. /remote-control runs the harness readiness skill; claude remote-control starts the live native server; claude.ai subscription login; no custom server; no static runtime-proof claim.
/specqa A SPEC.md was created, updated, or reused before implementation. Spec QA Agent checks requirements quality, SOTA 2026/currentness source ledger, critical blockers, Opus 4.7 proof boundary, and concrete improvement suggestions before execution.
/parallel The work splits into independent packets with clear owned files/surfaces and aggregate verification. Packet DAG, ownership matrix, sync barriers, worker sidecars, parallel-aggregate.
/hive The task needs multiple perspectives but may not need a full file-changing workflow: research branches, adversarial review, planning alternatives, risk ranking, or synthesis. Queen/supervisor, role map, blackboard, dissent/conflict log, evidence-backed synthesis.
/hiveworkflow The entire implementation lifecycle benefits from hive coordination and packet execution: broad audit plus implementation, multi-surface build, high-stakes verification, or agent/fleet design. Use through /opusworkflow by default with inner_contract=hiveworkflow, plus hive artifact, hive-run.json, optional /parallel packets, hive-aggregate, /introspect, /verify.

Do not pick /hive because it sounds more powerful. Pick it when role
specialization and dissent improve judgment. Do not pick /parallel because
there are available lanes. Pick it when disjoint ownership and aggregation make
the critical path shorter or the evidence better.

If both seem possible, default to /parallel for execution throughput and to
/hive for judgment breadth. Use /hiveworkflow only when the task needs both.

Metacognitive routing: /metacognition steers work before execution by
classifying the task, reading capacity evidence, computing the effective
parallel budget, naming required evidence, and routing to /workflow,
/claudeproduct, /deepresearch, /parallel, /hive, /hiveworkflow, /agentfactory,
/verify, /introspect, or a blocked state. It is upstream steering, not a
substitute for /introspect; required introspection triggers still need
explicit blocker decisions. It does not depend on raw hidden chain-of-thought
and it rejects reflection without evidence. Use
bash scripts/metacognition-scorecard.sh --fixtures --json to prove the static
contract.

Claude product knowledge: /claudeproduct answers user and harness
questions about Claude product behavior from official current docs. It is the
right path for "how do we use X from Claude?", Claude Code configuration,
skills, hooks, MCP, subagents, Claude.ai Projects, Artifacts, connectors,
Research, web search, API/platform behavior, availability, limits, and setup.
It separates Claude Code, Claude.ai, Desktop, Mobile, API, and MCP connector
surfaces, includes connector permission/trust caveats, and never reads .env
or secrets for product-doc answers. Use
bash scripts/claudeproduct-scorecard.sh --fixtures --json to prove stale
memory answers and unsupported Claude claims are rejected.

Native Remote Control: in this harness, /remote-control runs readiness
and troubleshooting checks. Start the live native server with
claude remote-control, then connect from https://claude.ai/code or mobile.
Use the native Claude Code feature only; do not add a custom remote server,
websocket bridge, MCP control plane, or API-key fallback. Static harness
evidence is compatibility evidence, not proof that a live browser or mobile
session connected. Use
bash scripts/remote-control-doctor.sh --static --json and
bash scripts/remote-control-smoke.sh --fixtures before blaming the harness.

Spec QA Agent: /specqa runs after SPEC.md and before implementation in
the governed workflow. It blocks critical spec defects, requires current
webresearch source ledgers for SOTA 2026 or time-sensitive claims, writes
.taste/specqa/{run_id}/spec-qa.md plus spec-qa.json, and keeps Opus 4.7
reviewer claims tied to runtime identity proof. Use
bash scripts/specqa-smoke.sh --fixtures to prove the static contract.

Harness capability map: docs/harness-capability-map.md and
docs/harness-capability-map.json are generated from repo truth surfaces and
act as the canonical self-lookup index for minmaxing capabilities. They list
skills, route groups, rules, required script gates, static evals, hooks, and
Codex surfaces without reading secrets. Regenerate and verify them with
bash scripts/harness-capability-map.sh and
bash scripts/harness-capability-map.sh --check.

Hive coordination: /hive adds governed multi-agent coordination above
/parallel: a queen/supervisor, capability-based roles, visible blackboard,
dissent/conflict log, evidence-backed synthesis, and verification. Use
/hiveworkflow only when the whole file-changing lifecycle benefits from that
coordination. Hive reuses /parallel packet DAGs, ownership matrices,
sidecars, and aggregation for execution; hive runs also emit
.taste/hive/{run_id}/hive-run.json validated by artifact-lint and
hive-aggregate. Consensus never replaces
/introspect, /verify, /workflow, or command evidence. Use
bash scripts/hive-scorecard.sh --fixtures --json and
bash scripts/hive-aggregate.sh --fixtures to prove the static contract.

Visualization approval: /workflow stays autonomous. Use /visualize when you only want to see the model's understanding, and /visualizeworkflow when you want a draft spec plus visual or operational artifact to approve before implementation.

Codex image generation: When SPEC.md asks for generated or edited raster
assets, use the repo Codex skill $codex-imagegen from
.agents/skills/codex-imagegen. It is intentionally subscription-first:
Codex/ChatGPT image usage when the current Codex runtime exposes it, no
OpenAI API keys or API-priced fallbacks unless you explicitly change that
billing route. If the runtime cannot generate an image, the harness writes a
handoff prompt and marks the asset blocked rather than pretending a file exists.

Effectiveness gates: The harness is designed to steer LLMs away from lazy completion. Claude Code runtime hooks and local smokes reject destructive Bash, evidence-free closeout, failed-verification positive closeout, fake source ledgers, tests-passed claims without command evidence, unverified worker claims, shallow metacognition, stale Claude product answers, missing Spec QA, shallow hive consensus, and linear lane-scaling claims. The Stop hook uses Claude Code's intentional blocking path: a blocked closeout is repair feedback, not a crash. Positive closeout must cite commands or verification; read-only/audit closeout may cite files inspected or sources reviewed. "Tests not run", "unverified", or equivalent wording must close as partial/blocked rather than done. Use bash scripts/harness-scorecard.sh --json, bash scripts/metacognition-scorecard.sh --fixtures --json, bash scripts/claudeproduct-scorecard.sh --fixtures --json, bash scripts/specqa-smoke.sh --fixtures, bash scripts/hive-scorecard.sh --fixtures --json, bash scripts/hook-smoke.sh, bash scripts/codex-run-smoke.sh, and bash scripts/parallel-plan-lint.sh --fixtures to prove the first-slice gates.

Artifact sidecars: Markdown remains the human contract, but machine gates can consume minimal JSON sidecars for agent-native estimates, verification results, and worker results. Validate the local fixtures with bash scripts/artifact-lint.sh --fixtures.

Static harness evals: evals/harness/tasks and evals/harness/golden define local no-network eval tasks over the harness gates. Run bash scripts/harness-eval.sh --json or bash scripts/harness-eval-report.sh --run to score the current harness behavior.

Run metrics and session insights: bash scripts/run-metrics.sh --json summarizes local workflow/eval/Codex artifacts, and bash scripts/session-insights.sh --json flags unhealthy runs. Missing provider cost, token, ACU, or calibration data is reported as insufficient_data.

Runtime hardening: bash scripts/runtime-hardening-smoke.sh proves the
local trace ledger, hook mesh, worktree runner, scenario evals, learning loop,
and harness doctor without secrets or network access. See
docs/runtime-hardening.md for the operator surface.

Security profiles: this repo's default is trusted-local bypassPermissions for the operator's solo loop. solo-fast documents that posture, team-safe is the shared-work fallback, ci-static is no-secret static validation, and ci-runtime is isolated authenticated validation. Run bash scripts/security-smoke.sh after profile changes.

Release governance: Public harness work should pass bash scripts/release-check.sh --static-only. The static GitHub Actions lane runs without secrets; authenticated runtime checks are isolated in the manual/scheduled runtime lane.


Usage

The full workflow — one command:

claude
/workflow "build a REST API for users"

/workflow now owns the whole lifecycle inline: taste check → deepresearch / webresearch → code audit → plan → SPEC.md → implementation → verification → closeout.

Report-informed workflow:

/digestflow "implement auth hardening based on these reports" ./gemini-report.md ./notebooklm-notes.md

/digestflow starts with Report Intake. It treats external AI reports as untrusted candidate evidence, labels imported claims report-derived, quarantines prompt-like instructions, records contradictions, and then runs the repo's own deepresearch plus the full workflow before any implementation is trusted.

Research-to-taste bootstrap text:

/digestaste "bootstrap the operating kernel for this roofing ops agent" ./deepresearch-result.md

/digestaste is lighter than /digestflow: it turns Deep Research markdown
into a sanitized DigesTaste Bootstrap Packet with goal text, claim/source
ledgers, /tastebootstrap answers, draft taste text, existing-kernel proposal
handling, and verification gaps. It does not run implementation. Fresh kernels
route through /tastebootstrap; existing kernels route through /defineicp
proposal/apply semantics.

Direct skill invocation (advanced):

/autoplan "build a login system"   # Generate SPEC.md
/sprint                            # Execute with ownership-safe parallelism
/verify                           # Check against spec
/ship                             # Ship checklist

Direct invocation skips the orchestrator — use when you know exactly what you need.

Verification

Before you trust a setup or a prompt-contract change, run the repo checks:

bash scripts/test-harness.sh

If Claude is authenticated locally, also run the runtime smoke flow:

bash scripts/workflow-smoke.sh

The smoke test validates the real /workflow path in a temporary repo and accepts either:

  • a justified local-only research brief for a purely local task
  • or positive MiniMax MCP research when the task depends on current external facts

For /opusworkflow, prove the planner side separately before claiming Opus
runtime identity:

unset ANTHROPIC_API_KEY ANTHROPIC_BASE_URL ANTHROPIC_AUTH_TOKEN
claude auth login
bash scripts/opusminimax-doctor.sh --runtime
claude --model claude-opus-4-7 --settings .claude/settings.opusminimax-planner.local.json -p 'Reply exactly: OPUSWORKFLOW_AUTH_OK'

Current local runtime evidence recorded on 2026-05-06: Claude Pro auth and the
Opus 4.7 planner sentinel were proven by operator test. MiniMax live executor
packet runtime remains pending until an explicit tiny executor packet is run.

Codex in Claude Code

/codesearch is the local repo skill for searching code. /codex:* is the official OpenAI plugin namespace.

Install the official plugin

You can install it from the shell without opening a manual slash-command flow:

claude plugin marketplace add openai/codex-plugin-cc
claude plugin install codex@openai-codex
claude plugins list

Then open Claude Code and run:

/codex:setup

If Codex is not logged in yet:

codex login

What this repo configures for Codex

Once the repo is trusted, Codex loads .codex/config.toml:

  • model = "gpt-5.5"
  • model_reasoning_effort = "medium"
  • model_reasoning_summary = "detailed"
  • [agents].max_threads = 10
  • openaiDeveloperDocs MCP for official OpenAI/Codex docs

Deep-Research examples

The Codex plugin exposes reviews plus long-running rescue tasks. For comprehensive research-backed planning, /codex:rescue is the better fit because it is steerable.

/codex:rescue --background
Use subagents up to max_threads for a research-backed plan, but only for bounded independent packets.
Have repo_explorer map the code paths, docs_researcher verify external APIs and current docs, and reviewer challenge risks and rollback gaps before proposing the plan.
Task: refactor the auth subsystem without user-visible behavior changes.
/codex:adversarial-review --background
Challenge whether this implementation plan is the right one, question hidden assumptions, and look for rollback, race-condition, and test-coverage risks.

Notes:

  • /codex:review and /codex:adversarial-review are read-only.
  • /codex:rescue is the command to use when you want Codex to investigate, plan, and potentially fix.
  • Codex only fans out subagents when you explicitly ask it to.
  • The shared project standard is already --model gpt-5.5 --effort medium; use explicit CLI overrides only for temporary one-off experiments.

Requirements

Claude Code 2.1+ (npm install -g @anthropic-ai/claude-code); Opus 4.7 planner mode expects 2.1.111+
Python 3.11+
MiniMax API Key platform.minimax.io
Claude Subscription Required for subscription-backed Opus planner usage; verify with claude auth status --text plus the OPUSWORKFLOW_AUTH_OK sentinel
Node.js 18.18+ if you want the optional OpenAI Codex plugin
Codex Auth ChatGPT sign-in or OpenAI API key if you want the optional OpenAI Codex plugin

Auto-Detection

Harness auto-detects your hardware and sets agent pool:

RAM Cores Agents
32GB+ 8+ 10 (default)
16GB 4+ 6
8GB 2+ 3

Override in your shell or ignored local profile:

{
  "env": {
    "MAX_PARALLEL_AGENTS": "6"
  }
}

Folder Structure

minmaxing/
├── CLAUDE.md                    # Core instructions (for AI)
├── README.md                    # This file (for you)
├── OPEN_CORE_STRATEGY.md         # OSS core vs private moat operating boundary
├── COMMERCIAL.md                 # Managed-service and enterprise boundary
├── SECURITY.md                   # Vulnerability reporting and secret handling
├── TRADEMARKS.md                 # Brand, fork, and certification limits
├── NOTICE                        # Apache-2.0 attribution notice
├── setup.sh                     # One-command installer
├── taste.md                     # Project operating kernel — created by /tastebootstrap
├── taste.vision                 # Product intent + tradeoff contract — created by /tastebootstrap
├── AGENTS.md                    # Project instructions for Codex
├── REVCLI_HERMES_AGENT_MAP.md    # REVCLI/Revis runtime-ready Hermes agent portfolio
├── .codex/
│   ├── config.toml             # Project-scoped Codex defaults
│   └── agents/                 # Codex custom agents for research/review
├── .agents/
│   └── skills/
│       └── codex-imagegen/     # Codex subscription image asset skill
├── .claude/
│   ├── settings.json           # Provider-neutral governance config
│   ├── settings.opusminimax-planner.example.json
│   ├── settings.minimax-executor.example.json
│   ├── settings.opussonnet.example.json
│   ├── settings.sonnet-executor.example.json
│   ├── hooks/                  # Lifecycle hooks, including working-state rehydration
│   ├── skills/                 # 38 skills (system calls)
│   │   ├── workflow/           # Central execution engine
│   │   ├── opusworkflow/       # One normal Opus + MiniMax product route
│   │   ├── opusminimax/        # Advanced provider/packet engine
│   │   ├── opussonnet/         # Optional Claude-only Opus + Sonnet route
│   │   ├── visualize/          # Taste-to-artifact comprehension check
│   │   ├── visualizeworkflow/  # Approval-first workflow route
│   │   ├── demo/               # Recorded product demo route
│   │   ├── digestflow/         # External report intake + governed workflow
│   │   ├── digestaste/         # Deep Research markdown to bootstrap text
│   │   ├── deepretaste/        # Intent-to-ICP-to-taste bootstrap/retaste
│   │   ├── defineicp/          # ICP-to-taste evolution route
│   │   ├── icpweek/            # ICP week-in-the-life product stress test
│   │   ├── tastebootstrap/     # Fresh-repo taste bootstrap
│   │   ├── align/              # Taste gate
│   │   ├── audit/              # Deep codebase analysis
│   │   ├── autoplan/           # SPEC.md generator
│   │   ├── agentfactory/       # Governed Hermes agent generator
│   │   ├── parallel/           # Hardware-aware workflow parallelizer
│   │   ├── claudeproduct/      # Official Claude product knowledge router
│   │   ├── remote-control/     # Native Claude Code Remote Control route
│   │   ├── hive/               # Governed multi-agent coordination
│   │   ├── hiveworkflow/       # Full hive-coordinated workflow
│   │   ├── sprint/             # Ownership-safe parallel executor
│   │   ├── verify/             # SPEC compliance checker
│   │   ├── ship/               # Pre-ship checklist
│   │   ├── investigate/        # Root-cause debugging
│   │   ├── council/            # Multi-perspective synthesis
│   │   ├── qa/                 # E2E testing
│   │   ├── review/             # AI review + human sign-off
│   │   ├── deepresearch/        # Canonical deep investigation
│   │   ├── webresearch/        # Focused current web research
│   │   ├── browse/             # Backward-compatible research alias
│   │   ├── introspect/         # Hard-gate self-audit
│   │   ├── codesearch/         # Code search
│   │   ├── memory/             # 5-tier memory skill
│   │   └── overnight/          # 8hr session with checkpoints
│   └── rules/                  # Modular rules (spec, pev, quality, etc.)
├── scripts/
│   ├── memory.sh               # 5-tier memory CLI
│   ├── state.sh                # Compaction-safe working state CLI
│   ├── spec-archive.sh          # Active SPEC.md archive helper
│   ├── memory-auto.sh           # Session start/end hooks
│   ├── taste.sh                 # Taste system CLI
│   ├── start-session.sh         # Session initializer
│   ├── harness-capability-map.sh # Generated harness capability map
│   ├── remote-control-doctor.sh # Static native RC readiness doctor
│   ├── remote-control-smoke.sh # Native RC compatibility smoke gate
│   ├── parallel-capacity.sh     # Hardware-aware parallel budget profile
│   ├── parallel-smoke.sh        # Parallel mode production-contract smoke test
│   ├── agentfactory-smoke.sh    # Agent Factory production-contract smoke test
│   └── detect-hardware.sh       # Auto-detect agent pool
├── memory/                      # Python memory package (SQLite + FTS5)
│   ├── sqlite_db.py
│   ├── causal.py                # Causal graph tracking
│   ├── consolidation.py          # Memory consolidation
│   └── cli.py                   # Memory CLI entry point
├── obsidian/Memory/             # Flat-file memory (git-tracked)
│   ├── Decisions/               # Semantic tier
│   ├── Patterns/                # Procedural tier
│   ├── Errors/                  # Error-solution tier
│   ├── Stories/                 # Graph tier
│   │   └── commits/            # Commit summaries (auto-generated)
├── .minimaxing/
│   └── state/                  # Generated current-task handoff, snapshots, and events
└── .taste/
    ├── sessions/                # Episodic tier (daily JSONL)
    ├── workflow-runs/           # Research/audit/plan/verification artifacts
    ├── specs/                   # Archived completed or superseded specs
    └── taste.memory             # Append-only decision log

License

minmaxing is licensed under Apache-2.0.

The open-source license covers this public core. It does not grant rights to
private REVCLI/Revis runtime code, customer deployments, customer data,
production credentials, managed-service operations, or project branding beyond
the limited trademark language in Apache-2.0. See
COMMERCIAL.md and
TRADEMARKS.md for the
commercial and brand boundary.

Reviews (0)

No results found