grain-kit 0.1.2

CLI-first workflow toolkit for structured AI-assisted software development.

Grain

Grain is a CLI-first workflow system for structured AI-assisted software development.

It gives you:

• canonical docs as source of truth
• working docs for sequencing and planning
• task packets as the execution unit
• explicit review and close loops
• minimal-context execution instead of broad repo dumping
• more useful work per agent context window by reducing drift, retries, and unnecessary rereads

Grain is not a coding model by itself. It is the workflow and file system that external agent CLIs operate against. It exists in part to help agent-CLI users avoid burning through token windows on broad, repetitive, underspecified conversations.

---

Who It Is For

Grain is for developers and technical operators who:

• use agent CLIs such as Codex or Claude Code
• want inspectable markdown files instead of opaque orchestration
• want scoped task execution, review, and closure
• want less context drift and lower token waste than ad hoc prompting
• regularly hit context or usage limits before useful work is complete

---

Quickstart

New project

# 1. Install
uv tool install grain-kit

# 2. Create a repo and initialize
mkdir my-project && cd my-project
git init
grain init

After grain init your repo has:

docs/canonical/    ← source-of-truth decisions (you fill these in)
docs/working/      ← backlog, current task, open questions
docs/runtime/      ← workflow rules, manifest, adapter profiles
tasks/             ← one folder per task packet
prompts/           ← workflow entrypoints for your agent CLI

# 3. Open prompts/workflow.onboard.new.md in your agent CLI
#    Fill in the Project Context section at the top
#    The agent generates your canonical docs, backlog, and first task

# 4. Check what to do next
grain workflow next

Output looks like:

workflow next: ok
  phase             1
  active_task_id    P1-T01-TASK-0001
  next_action       task_execute
  recommended_prompt  prompts/task.execute.md

# 5. Run prompts/task.execute.md in your agent CLI — it executes the task
# 6. Run prompts/task.review.md — review the work
# 7. Run prompts/task.close.md — close the task and move to the next one
# 8. Repeat from step 4

That is the full loop. grain workflow next always tells you exactly where you are and what to do next.

---

Existing project

# 1. Install (if not already done)
uv tool install grain-kit

# 2. Scaffold Grain into your existing repo
cd your-existing-project
grain onboard

Output looks like:

onboard: ok
  root              /your-existing-project
Created:
- docs/canonical
- docs/working
- docs/runtime
- tasks
- prompts
- docs/canonical/product_scope.md
- docs/canonical/architecture.md
- docs/working/backlog.md
- docs/working/current_focus.md
- docs/working/open_questions.md
Skipped:
- (any files that already existed are listed here)

# 3. Open prompts/workflow.onboard.existing.md in your agent CLI
#    The agent scans your codebase, fills in the DRAFT docs, and records
#    gaps as open questions — review everything before treating it as canonical

# 4. grain workflow next — start the workflow loop from there

---

Core Idea

Grain separates work into layers:

• docs/canonical/
  • source-of-truth decisions
• docs/working/
  • current plan, backlog, questions, proposals, metrics
• docs/runtime/
  • execution rules, manifest, context-loading rules, model profiles
• tasks/
  • one packet per task
• prompts/
  • stable workflow entrypoints for agent CLIs

---

Adapter Inventory

Grain uses a contract-driven adapter model to tune workflow behavior for different domains.

The workflow loop is the same for every domain. Adapters change context selection, validation hints, and review focus — not workflow law.

Official Adapters (Supported Today)

• code_adapter
  • Python, Rust, backend services, CLI tooling
• frontend_adapter
  • TypeScript, JavaScript, React, Storybook, Tauri UI

Official Adapters (Supported)

• docs_adapter
  • Markdown documentation systems, Word documents (.docx), PDF files
• spreadsheet_adapter
  • Excel (.xlsx/.xls) and CSV spreadsheet workflows

Custom Adapters

Users may define custom adapter profiles locally for any domain.

Examples of valid custom adapter domains:

• DevOps workflows — VPS provisioning, deployment setup, service configuration
• System administration — reverse-proxy setup, firewall and SSH hardening, backup planning, monitoring
• Local ops — machine-admin repos, dotfile management, workstation automation
• Containerization — Docker/Compose setup, rollback procedures
• Content pipelines — editorial systems, publishing workflows
• Any operational domain expressible through repo artifacts and task packets

Custom adapters follow the same contract as official adapters. Add them to docs/runtime/adapter_profiles.md and declare them in the manifest. No plugin system or marketplace is required.

Future Possibilities

• shareable community adapter profiles
• adapter-aware template sets per domain

If you are managing content with Grain, docs_adapter is the natural direction because it fits markdown-first editorial, knowledge-base, and documentation workflows. If you are managing infrastructure or operations work, define a devops_adapter or local_ops_adapter with the file patterns and validation hints relevant to your domain.

---

Recursive Build Principle

Grain is intended to be used to build Grain itself, then to build Assay on top of it.

Assay is a planned independent verification layer. It will check work produced through Grain — test coverage, correctness signals, drift detection — without being part of the execution loop itself. The intended end state is: structure the work with Grain, execute the work with an agent CLI, verify the result with Assay.

That recursive use is deliberate. It is one of the main ways the product is validated in real work:

• if Grain cannot manage its own build cleanly, its workflow claims are weak
• if Grain reduces token waste, drift, and retries while building itself, that is stronger evidence than a synthetic demo
• if Assay can later verify work produced through Grain, the loop becomes: structure the work, execute the work, verify the result

Recursive building is validation, not proof of universal fit. Grain still needs to work outside its own repo shape and outside its creator's habits.

---

Installation

Requirements:

• Python 3.11+
• uv (recommended) or pip

Recommended: uv tool install

uv tool install grain-kit

This installs grain into uv's global tool path. No virtual environment needed.

Alternative: pip install

pip install grain-kit

Local source install (development or testing)

uv tool install --from . grain

Fallback using a local venv:

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Verify

grain --version
grain init --help

Expected output:

• grain, version <x.y.z>
• Usage: grain init [OPTIONS]

---

Updating

uv tool install

uv tool upgrade grain-kit

pip install

pip install --upgrade grain-kit

Local source install

uv tool install --from . grain --force

or for a venv-based install:

pip install -e . --upgrade

---

Troubleshooting

If grain is not found after uv tool install:

• run uv tool dir --bin and add that directory to your PATH
• macOS/Linux: export PATH="$(uv tool dir --bin):$PATH"
• Windows PowerShell: $env:Path = "$(uv tool dir --bin);$env:Path"

If you have a Python version mismatch:

• check: python --version (Grain requires Python 3.11+)
• install with a specific interpreter: uv tool install --python 3.11 grain

If venv conflicts cause unexpected behavior:

• check: which grain (or where grain on Windows) to confirm the active path
• reinstall: uv tool uninstall grain-kit && uv tool install grain-kit
• for local source testing, use isolated env vars:
  • point UV_TOOL_DIR, UV_CACHE_DIR, and HOME at temp directories

---

Contributing / Feedback

Found a bug, have a feature request, or want to share feedback?

Open an issue at: https://github.com/Diwata-Labs/Grain/issues

When reporting a bug, include:

• the command you ran
• the output or error message
• your Python version and install method (uv tool install or pip install)

---

Should You Use Grain For Your Machine?

Yes, sometimes.

Good use cases:

• managing dotfiles or local automation as a project
• maintaining workstation setup scripts
• organizing home-lab, local tooling, or machine-admin workflows
• managing content repositories, docs sites, and markdown-first knowledge bases
• treating your local environment as an inspectable system with tasks, review, and change history

Bad default:

• treating your entire home directory or whole machine as one Grain project

Better pattern:

• create a dedicated repo such as:
  • local-ops
  • machine-admin
  • personal-systems
• use Grain inside that repo

Reason:

• keeps scope bounded
• keeps context smaller
• keeps tasks coherent
• avoids mixing unrelated files and projects into one authority tree

---

How New Users Should Start

Use the README as the entrypoint and the prompts as the workflow engine.

There are two starting modes:

1. new project
2. existing project

New Project

1. create or enter the project repo
2. run grain init
3. open prompts/workflow.onboard.new.md
4. fill in the Project Context
5. paste that prompt into your agent CLI
6. let the agent generate the initial docs, manifest, backlog, and open questions
7. review those generated docs before treating canonical content as approved

Optional onboarding-aware init flags:

grain init --primary-adapter code_adapter --secondary-adapter frontend_adapter --bootstrap

• --primary-adapter sets the default adapter context for onboarding
• --secondary-adapter can be repeated for additional adapters
• --bootstrap creates a starter task packet and initializes docs/working/current_task.md

Compatibility note:

• prompts/workflow.init.md is kept as an alias for users who still invoke the old onboarding name.

After onboarding, use the normal loop:

1. prompts/task.plan.next.md only when a task must be selected or split
2. prompts/task.execute.md
3. prompts/task.review.md
4. prompts/task.close.md

Existing Project

1. run grain onboard inside the existing repo (or pass a path)
2. open prompts/workflow.onboard.existing.md
3. paste that prompt into your agent CLI
4. the agent reviews the scaffold manifest, scans the codebase, and generates draft canonical docs
5. review all generated docs before treating them as canonical — all output is marked # DRAFT

Optional flags:

grain onboard                  # scaffold into current directory
grain onboard /path/to/repo    # scaffold into a specific path
grain onboard --dry-run        # preview what would be created without writing
grain onboard --format json    # get a machine-readable manifest of created/skipped files

grain onboard is additive only — it never overwrites existing files. Any file already present is skipped and listed in the manifest.

---

Agent CLI Usage

Grain is designed to be used from an agent CLI.

The basic pattern is:

1. run the Grain CLI when filesystem scaffolding or validation is needed
2. run a prompt from prompts/
3. let the agent update the repo files
4. continue through the structured loop

Recommended Stable Prompt Surface

Phase planning:

• prompts/phase.plan.next.md
• prompts/phase.review.md
• prompts/phase.review_and_close.md

Task planning and execution:

• prompts/task.plan.next.md
• prompts/task.execute.md
• prompts/task.review.md
• prompts/task.close.md

Project bootstrap:

• prompts/workflow.onboard.new.md — new project onboarding
• prompts/workflow.onboard.existing.md — existing project adoption
• prompts/workflow.init.md (compatibility alias for new-project onboarding)

Recommended Daily Loop

Use this loop continuously inside the active phase:

1. task.plan.next — only when the next task must be selected, split, or added
2. task.execute
3. task.review
4. task.close

Do not run task.plan.next every time if a ready task already exists.

Recommended Planning Loop

Use this less often:

1. phase.plan.next
2. phase.review or phase.review_and_close at phase boundaries

---

Customization

Grain is meant to be customized to the project it is managing.

Users should customize:

• canonical docs for the project domain and scope
• working docs for sequencing, backlog, and open questions
• adapter selection and adapter profiles
• model strategy and agent preferences
• onboarding outputs for new or existing projects

Users should try to keep stable:

• the core workflow loop: plan → execute → review → close
• file-backed workflow state
• authority boundaries between canonical, working, runtime, and task layers
• explicit review and change-proposal gates

Good customization:

• adapt Grain to Python, Rust, React, docs, spreadsheets, local-ops, or mixed projects
• add project-specific constraints, risks, adapters, and backlog structure

Bad customization:

• silently bypass review
• hide state outside the repo
• collapse canonical and working docs into one layer
• make every project invent its own incompatible workflow loop

In short:

• customize the project model
• keep the workflow model disciplined

---

Minimal CLI Reference

# Setup
grain init                                    # initialize a new project
grain onboard [path]                          # scaffold into an existing repo (additive)
grain onboard --dry-run                       # preview scaffold without writing

# Docs and validation
grain docs validate                           # validate doc manifest and structure
grain task validate --id TASK-####           # validate one task packet

# Task management
grain task create --title "..."
grain task list
grain task show --id TASK-####
grain task status --id TASK-#### --status done
grain task next                               # which task to work on
grain task prepare --id TASK-####            # assemble context prerequisites

# Workflow automation
grain workflow next                           # next legal step + blockers
grain workflow next --format json             # machine-readable for agent CLIs
grain workflow run                            # execute one step, stop at gates
grain workflow loop                           # run full execute→review→close cycle
grain workflow loop --supervision gated       # stop at review/close gates (default)
grain workflow loop --supervision supervised  # approve every action
grain workflow loop --dry-run                 # preview loop plan without execution

# Context assembly
grain context build --id TASK-####
grain context build --id TASK-#### --format json

# Orchestration
grain orchestrate scope --scope "..."
grain orchestrate plan --scope "..."
grain orchestrate accept --plan <id>          # mark a plan accepted for loop ordering

# Adapter inspection
grain adapter list
grain adapter show --id <id>

# Prompts
grain prompt show                             # recommended prompt for current state

In normal Grain usage, the prompts drive most of the workflow and the CLI handles scaffolding, validation, and command surfaces.

---

Important Rules

• do not treat prompts as canonical truth
• do not edit canonical docs silently
• do not execute multiple unrelated tasks in one packet
• do not let review and close perform backlog planning implicitly
• if prompt or workflow-contract docs change mid-conversation, restart the relevant agent conversation

See:

• docs/runtime/PROJECT_RULES.md
• docs/runtime/docs_manifest.yaml

---

Current Product State

• v1 core workflow complete (Phases 1–5 closed) — init, docs, task, context, model, review commands
• Phase 6 closed — adapter system foundation (code_adapter, frontend_adapter)
• Phase 7 closed — new-project onboarding flow (grain init with adapter selection and starter-packet bootstrap)
• Phase 8 closed — workflow automation runner (grain workflow next/run, grain task next/prepare, grain phase next, grain prompt show, machine-readable JSON contract)
• Phase 9 closed — orchestration service (grain orchestrate scope/plan, grain adapter list/show, OrchestratorPlan domain model)
• Phase 10 closed — structural intelligence (tree-sitter extraction, knowledge graph, graph-assisted context selection)
• Phase 11 closed — distribution and global install (pip install grain-kit, uv tool install grain-kit, PyPI publish workflow)
• Phase 12 closed — automated workflow loop (grain workflow loop, supervision levels, per-step logging, grain orchestrate accept)
• Phase 13 closed — existing project adoption (grain onboard, CodebaseScanner, draft canonical doc generation, workflow.onboard.existing.md prompt)
• Phase 14 closed — document and spreadsheet adapters (SpreadsheetExtractor for xlsx/csv, DocsExtractor for docx/md, PdfExtractor for pdf with graceful degradation)

Workflow loop guardrails:

• grain workflow loop --steps N sets a hard loop-step limit
• grain workflow loop --dry-run previews planned actions without mutating state
• supervision levels:
  • supervised: operator approval before each action
  • gated: automatic run, stops at review/close gates (default)
  • autonomous: minimal stops, unverified automation (Assay will provide future independent verification)

See docs/working/current_focus.md and docs/working/implementation_plan.md for active phase detail.

---

Roadmap

v0.2.0 — Semantic Enrichment

• Semantic similarity — embedding-based similar-task detection, doc-to-task matching, and duplicate/overlap identification. All outputs labeled as inferred, not authoritative.
• Ranking layer — deterministic scoring across graph distance, semantic similarity, authority level, and packet-local priority. Applied to context selection and next-task suggestion.
• Assay — independent verification layer that checks work produced through Grain. Closes the loop: structure the work, execute the work, verify the result.
• Homebrew tap — brew install support for macOS users.

These features depend on an embedding infrastructure decision (local model vs. external API) that has not yet been made. v0.2.0 planning begins after that decision is resolved.

Longer term

• Community adapter profiles — shareable domain-specific adapter configurations
• --prefix flag for grain onboard — adopt Grain into repos with existing docs/ structures without conflict
• TUI interface — optional terminal UI for operators who prefer visual workflow navigation