openarche

OpenArche

OpenArche is a harness-first Claude Code plugin for non-trivial engineering work. It keeps complex tasks from closing before planning, validation, review, and closeout are explicit.

Features

• Task grading: decides whether a task stays light or enters harness control
• Embedding-based routing: uses the configured embedding backend to separate plain questions from execution work without language-specific keyword lists
• Persistent sessions: materializes .openarche/sessions/<task-id>/state.json only when execution work actually starts
• Stage gates: keeps validation, review, and maintenance open until they are actually closed
• Context injection: adds current task state, gate reasons, and relevant local knowledge
• Knowledge recall: retrieves repository-local knowledge first, then global knowledge, through embeddings and link expansion
• Background closeout: captures reusable knowledge after the task stops

Installation

1. Add the marketplace entry:

/plugin marketplace add Done-0/openarche

2. Install the plugin:

/plugin install openarche

3. Reload plugins:

/reload-plugin

4. Run setup:

/openarche:setup

If you use local embeddings, the first successful run may need network access to download the model files. If you prefer API-backed embeddings, switch to remote in /openarche:config.

How It Works

1. Use Claude Code as usual.
2. Light tasks stay lightweight. Non-light tasks receive harness context first, and .openarche/sessions/<task-id>/state.json is materialized only for explicit execution work or, by default, after actual write activity starts.
3. OpenArche tells Claude Code why the task was gated, which stages are still open, and which local knowledge is relevant.
4. The status line and session state keep showing what is still open, so the task does not quietly close too early.
5. Validation, review, and maintenance stay open until the required evidence is recorded inside the current session state and evidence directory.
6. When the task stops, OpenArche closes out the session and queues reusable knowledge capture for that transcript.

Commands

Command	Description
/openarche:setup	Prepare OpenArche for automatic task interception, harness sessions, and optional knowledge import
/openarche:config	View or update grouped capability configuration
/openarche:plan	Produce an execution plan with objective, acceptance criteria, and explicit task coverage
/openarche:run	Turn a plan into an execution checklist
/openarche:validate	Define browser and task validation expectations
/openarche:observe	Define investigation queries and targets
/openarche:review	Drive self-review, local agent review, cloud agent review, and repair loops
/openarche:maintain	Run maintenance and drift-reduction workflows
/openarche:knowledge-search	Search reusable engineering knowledge
/openarche:knowledge-save	Persist reusable engineering knowledge explicitly
/openarche:knowledge-reindex	Rebuild the knowledge vector index after embedding model changes

Configuration

Config file:

<home>/.claude/openarche/config.json

• knowledge.embedding.provider: local or remote
• orchestration: auto-injection, deferred materialization, and explicit command policy
• knowledge.embedding.localModel: used only when the provider is local
• knowledge.embedding.remoteModel, knowledge.embedding.remoteApiKey, knowledge.embedding.remoteBaseUrl: used only when the provider is remote
• knowledge.retrieval: recall threshold, recall fanout, and injection budget
• knowledge.extraction: extraction model and capture concurrency
• execution: isolation strategy and base ref OpenArche records the isolation plan by default. It does not automatically create git worktrees or switch branches in your repository.
• validation.browser: required browser evidence
• observability: logs, metrics, and traces requirements
• review: self-review, local agent review, cloud agent review, repair loops
• maintenance: quality and drift cleanup

If the embedding provider changes, run:

/openarche:knowledge-reindex

Runtime Layout

Global OpenArche state lives under:

<home>/.claude/openarche/
├── config.json
├── state.json
├── capture-log.json
├── decision-log.jsonl
├── prototype-cache.json
├── index.json
└── knowledge/

Repository-scoped runtime state lives under:

<repo>/.openarche/
├── sessions/
│   └── <task-id>/
│       ├── state.json
│       └── evidence/         # created only after evidence is actually recorded
└── knowledge/
    ├── index.json
    └── <entry-id>.md