Workflow

/workflow

Run the full workflow for:

$ARGUMENTS

This command is the end-to-end executor.

/opusworkflow is the default outer route for normal build/plan work and for mutating specialist work in this harness. It reuses this /workflow lifecycle while adding the Claude/Opus planner-reviewer plus MiniMax-M2.7-highspeed executor split. Use plain /workflow only when the user invokes it explicitly, the provider split is unavailable, or the work should stay in one local supervisor loop.

When a specialist route changes files, keep /opusworkflow as the outer route and record the specialist as inner_contract=agentfactory|hiveworkflow|parallel|defineicp|deepretaste|demo|visualizeworkflow.

/digestflow is a sibling route that reuses this same inline lifecycle after an external report-intake prelude. Normal /workflow behavior stays unchanged when no external research reports are supplied.

Non-Negotiable Contract

• Finish the task in this command whenever it is feasible.
• For file-changing work, follow this order: deep research -> code audit -> plan -> SPEC.md -> Spec QA -> execute -> verify.
• Do efficacy-first deep research for every task before planning or execution.
• Audit the current codebase before planning or writing SPEC.md.
• Run hard-gate introspection after code audit and before freezing the plan.
• Synthesize a concrete plan before writing SPEC.md.
• Record an Agent-Native Estimate before writing, replacing, or reusing

SPEC.md. The estimate must be agent-native wall-clock by default, not a bare human-equivalent calendar estimate.

• After SPEC.md is created, updated, or intentionally reused, run /specqa

before implementation. Spec QA is the SOTA/currentness and requirements quality gate for the active spec, not a post-implementation review.

• Run hard-gate introspection after implementation and before closeout.
• Record a parallel-aware metacognitive route before research planning for

file-changing work: task class, capacity evidence, effective parallel budget, chosen route, required evidence, confidence threshold, and why the full parallel ceiling was or was not used.

• Treat worker or subagent summaries as claims until the parent workflow

verifies them with command evidence. For harness-governance changes, run the effectiveness smokes that apply: scripts/harness-scorecard.sh --json, scripts/hook-smoke.sh, scripts/codex-run-smoke.sh, and scripts/parallel-plan-lint.sh --fixtures.

• When a workflow uses /parallel run artifacts, aggregate

.taste/parallel/{run_id}/packet-dag.json, ownership.json, and worker-results/*.json with scripts/parallel-aggregate.sh before treating worker outputs as accepted evidence.

• Keep implementation surgical: smallest sufficient implementation, no speculative abstractions, no drive-by refactors, and a changed-line trace back to SPEC.md.
• Do not stop after planning.
• Do not tell the user to manually run /autoplan, /parallel, /sprint, /verify, or /ship.
• Do not rely on nested custom-skill chaining as the primary execution path.
• Keep plain /workflow autonomous. Do not insert a mandatory visualization

approval pause into this command. When the user explicitly wants to see and approve the intended product or operator experience before implementation, route that request to /visualizeworkflow instead.

Reason: In real Claude Code sessions, nested custom skills may complete their own turn and return control to the user before the rest of the chain runs. For /workflow, execute the phases inline with Claude Code tools so the full flow actually completes.

You may still use built-in Claude Code tools, shell commands, subagents, and optional specialist skills as reference material when useful, but /workflow itself owns the whole lifecycle.

No shortcut exceptions for file-changing tasks:

• Do not say Research: skipped.
• Do not say Code Audit: skipped.
• Do not say SPEC.md: not needed.
• Do not treat "trivial", "tiny", "single-file", or "local-only" as exceptions.
• If you are about to skip one of these phases, stop and complete the missing phase instead.

Research brief is mandatory:

• If mcp__MiniMax__web_search is available in the tool list, you MUST use it before planning, explaining, auditing, or editing.
• Use mcp__MiniMax__web_search as the primary external research tool when it is available.
• Fall back to Claude Code WebSearch only if the MiniMax MCP is unavailable.
• Deep research must follow the repo’s effectiveness-first deepresearch protocol rather than a generic search fan-out: start with a collaborative research plan, run an iterative search -> read -> refine loop, maintain a source ledger, resolve conflicting evidence, and do follow-up research before freezing the plan.
• Treat MAX_PARALLEL_AGENTS as a ceiling, not a quota.
• Choose an effective research budget based on the number of distinct questions that materially affect the plan.
• If the task is purely local and does not depend on current external facts, a concise local-only research brief is acceptable, but you must say why no external search was needed.
• Build a research brief before any code audit synthesis, spec creation, or code changes.
• Write a workflow artifact for file-changing tasks so the reasoning trail is inspectable.
• Hard-gate introspection must name likely mistakes, cite evidence checked, downgrade confidence when warranted, and block closeout when unresolved findings remain.
• Re-check any concrete library, framework, API, or error details again right before editing if the plan depends on them.
• If the plan depends on Claude product behavior, run the /claudeproduct

protocol inside the research brief first. Separate current official Anthropic/Claude docs from local harness contracts before freezing the plan.

Phase 0: Taste Gate

1. Check whether taste.md and taste.vision exist in the project root.
2. If they do not exist:

• stop before research, audit, planning, or edits
• tell the user that fresh repos must start with /tastebootstrap
• do not ask the bootstrap questions here
• do not write taste.md or taste.vision from inside /workflow
• resume /workflow only after the kernel is defined

1. Read taste.md and taste.vision.
2. Rehydrate current working state if it exists:

bash scripts/state.sh status 2>/dev/null || true

Treat .minimaxing/state/CURRENT.md as a compact continuity handoff, then reconcile it with live git status, SPEC.md, and the latest workflow artifact before editing.

1. Recall memory for the task:

bash scripts/memory.sh recall "$ARGUMENTS" --depth medium 2>/dev/null || echo "Memory recall skipped"
bash scripts/memory-eval.sh --summary 2>/dev/null || echo "Memory freshness: unavailable; using local truth surfaces"

1. Summarize:

• relevant taste principles
• relevant working-state continuity and any stale assumptions to refresh
• relevant recalled memories
• memory health/freshness status and whether local truth surfaces are the

fallback

• an alignment score from 0 to 10

1. If the task clearly conflicts with taste, pause only to get an explicit alignment decision from the user.

If memory is degraded, disabled, or unavailable, continue from local truth surfaces (SPEC.md, workflow artifacts, docs, git status, tests, and source files) and say that memory could not be trusted as the primary context source. Do not claim memory captured everything unless a memory event trace or durable memory artifact proves the relevant entry exists.

Phase 1: Route

Choose the route from user intent:

| Intent | Workflow Behavior | |--------|-------------------| | build, implement, create, add, refactor, optimize, migrate | run full research → audit → plan → spec → execute → verify → closeout flow | | fix, debug, investigate | research first, audit the relevant code path, then reproduce/fix/verify; create SPEC.md if files change | | audit, analyze, understand, deepresearch, webresearch | inspect deeply, report findings, make fixes only if the user asked for them | | deepretaste, retaste, detect intent, bootstrap taste and ICP, SOTA customer research for taste | route through /opusworkflow with inner_contract=deepretaste when files may change; preserve /deepresearch as the general research route when findings will not mutate taste | | define ICP, ideal customer profile, tailor taste to customer, update taste.md or taste.vision from ICP | route through /opusworkflow with inner_contract=defineicp when files may change; keep proposal-first unless explicit apply approval is present | | Claude product, Claude Code, Claude.ai, Anthropic API, connectors, plugins, skills, hooks, MCP, subagents, plan availability, limits, setup | route product facts through /claudeproduct before generic research; continue into /workflow only if files change | | remote control, rc, continue Claude Code from phone, claude.ai/code local session | route to /remote-control; use native Claude Code RC commands only (/remote-control, /rc, claude --remote-control, claude remote-control) and never build a custom control server | | opus solo, opusolo, all opus, opus for everything, max effort opus | route through /opusolo when the operator explicitly wants the all-Opus premium sibling of /opusworkflow; keep the same gates and runtime-identity proof rules | | goal mode, native /goal, /goal , keep working until, bounded continuation | route to /goal-mode for static readiness, Goal Assist copy-paste templates, and no-shadow diagnostics; use native /goal manually only around already-known failed gates with owned scope, forbidden paths, transcript evidence, stop bounds, and parent verification, never as /workflow, /parallel, or /verify replacement | | spec QA, spec review, specification QA, SOTA spec audit | route to /specqa; review the active SPEC.md after creation/update/reuse and before implementation, using current webresearch when SOTA or time-sensitive facts matter | | parallel, mode parallel, dense workflow, orchestrate subagents, split across instances | route through /opusworkflow with inner_contract=parallel for file-changing packet execution; run the /parallel eligibility audit and use packets only when capacity, ownership, and verification pass | | hive, hive mind, coordinated agents, swarm, multi-agent synthesis | route through /opusworkflow with inner_contract=hiveworkflow for file-changing hive work only when roles, blackboard, dissent, ownership, capacity, and verification pass; otherwise downgrade to /workflow or /parallel fallback | | explain | inspect and explain directly | | review | review directly | | qa | run focused validation directly | | visualize, visual approval, show me first, approve before implementation | route to /visualizeworkflow for approval-first planning; after approval, continue implementation through /opusworkflow with inner_contract=visualizeworkflow; route to /visualize when the user only wants a comprehension artifact | | SPEC.md requests generated images, edited images, hero assets, UI mockups, sprites, thumbnails, or product raster assets | keep the normal /workflow contract, but route the asset generation lane through Codex skill codex-imagegen under .agents/skills; use Codex subscription/ChatGPT auth when available, never API-key image generation unless the user explicitly changes billing route |

Default to the full build flow when the task changes files.

If a normal /workflow task has high visual or product-experience ambiguity, you may mention that /visualizeworkflow exists for approval-first execution, but continue autonomously unless the user asked for that approval pause.

For dense work, automatically consider /parallel even if the user did not name it. Use it only when the eligibility audit proves independent packets, clear ownership, host capacity, and a verification path. Downgrade to normal inline /workflow when those conditions fail.

For hive-like work, automatically consider /hiveworkflow when the request needs coordinated roles, a blackboard, dissent/conflict handling, synthesis, and verified learning. Prefer /parallel when disjoint execution packets are enough; prefer /hiveworkflow only when judgment breadth and role coordination are part of the value. For file-changing hive work, run it under /opusworkflow as inner_contract=hiveworkflow. Downgrade to local /workflow if role map, blackboard, ownership/locks, capacity, or verification cannot be written clearly.

Before freezing the route for file-changing work, perform a compact metacognitive route check. Use scripts/parallel-capacity.sh --json when available, treat MAX_PARALLEL_AGENTS, Codex max_threads, and hardware ceiling as ceilings, and compute the smallest useful effective budget:

effective_metacognition_budget =
  min(MAX_PARALLEL_AGENTS, codex_max_threads, hardware_recommended_ceiling,
      independent_questions_or_packets, supervisor_capacity, verifier_capacity)

Do not fill the parallel pool for theater. Do not claim max agents means max quality or linear speedup.

The metacognitive route is not an introspection pass. It must appear in the workflow artifact before ## Research Brief, and the later ## Introspection section must still record the required /introspect pre-plan and /introspect post-implementation gates for file-changing work.

Workflow Artifact

For file-changing tasks, create a durable workflow record before writing SPEC.md:

mkdir -p .taste/workflow-runs
STAMP="$(date +%Y%m%d-%H%M%S)"
WORKFLOW_ARTIFACT=".taste/workflow-runs/${STAMP}-workflow.md"

This artifact is the inspectable audit trail for the run. It must be created before SPEC.md and updated as phases complete.

Required section order:

# Workflow Run: [task]

## Task
## Taste Gate
## Metacognitive Route
## Research Brief
## Code Audit
## Introspection
## Plan
## Agent-Native Estimate
## SPEC Decision
## Spec QA
## Plan Mode Approval
## Execution Notes
## Verification Evidence
## Outcome

Keep it concise, but do not skip sections. For non-file-changing analysis tasks, this artifact is optional.

Required content inside the sections:

• ## Metacognitive Route must record the task class, capacity evidence,

effective parallel budget, chosen route, evidence required, confidence threshold, and why the full parallel ceiling was or was not used. It must explicitly state when the route downgrades from /parallel to local /workflow because packets overlap or verification cannot aggregate results. It must not claim that /metacognition replaces or satisfies required /introspect gates.

• ## Research Brief must record the investigation mode, collaborative research plan, effective research budget, iterative search -> read -> refine loop, source ledger, conflicting evidence, and any follow-up research required before planning.
• ## Introspection must record at least pre-plan and post-implementation entries for file-changing work, plus after-test-failure or pre-push entries when those triggers occur.
• ## Plan must record any delegated packets, their owners, and their dependencies when parallel execution is likely.
• ## Agent-Native Estimate must record estimate type, execution topology,

capacity evidence, effective lanes, critical path, agent wall-clock, agent-hours, human touch time, calendar blockers, confidence, and any human-equivalent baseline as secondary only.

• ## Spec QA must record /specqa decision, requested reviewer model,

proven reviewer model if any, spec_qa_model_identity_status, current webresearch source ledger when SOTA/time-sensitive claims matter, critical finding count, improvement suggestions, artifact paths, and whether execution is allowed.

• ## Plan Mode Approval must record the `/opusworkf

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: waitdeadai
• Source: waitdeadai/minmaxing
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.