RalphSAFE

Claude Code plugin that generates individualized knowledge systems from conversation. You describe how you think and work, have a conversation and get a complete second brain as markdown files you own.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: ralph
description: Queue processing with fresh context per phase. Processes N tasks from the queue, spawning isolated subagents to prevent context contamination. Supports serial, parallel, batch filter, and dry run modes. Triggers on "/ralph", "/ralph N", "process queue", "run pipeline tasks".
version: "1.0"
generated_from: "arscontexta-v1.6"
user-invocable: true
context: fork
allowed-tools: Read, Write, Edit, Grep, Glob, Bash, Task
argument-hint: "N [--parallel] [--batch id] [--type extract] [--dry-run] — N = number of tasks to process"
---

## EXECUTE NOW

**Target: $ARGUMENTS**

Parse arguments:
- N (required unless --dry-run): number of tasks to process
- --parallel: concurrent claim workers (max 5) + cross-connect validation
- --batch [id]: process only tasks from specific batch
- --type [type]: process only tasks at a specific phase (extract, create, reflect, reweave, verify, enrich)
- --dry-run: show what would execute without running
- --handoff: output structured RALPH HANDOFF block at end (for pipeline chaining)

### Step 0: Read Vocabulary

Read `ops/derivation-manifest.md` (or fall back to `ops/derivation.md`) for domain vocabulary mapping. All output must use domain-native terms. If neither file exists, use universal terms.

**START NOW.** Process queue tasks.

---

## MANDATORY CONSTRAINT: SUBAGENT SPAWNING IS NOT OPTIONAL

**You MUST use the Task tool to spawn a subagent for EVERY task. No exceptions.**

This is not a suggestion. This is not an optimization you can skip for "simple" tasks. The entire architecture depends on fresh context isolation per phase. Executing tasks inline in the lead session:
- Contaminates context (later tasks run on degraded attention)
- Skips the handoff protocol (learnings are not captured)
- Violates the ralph pattern (one phase per context window)

**If you catch yourself about to execute a task directly instead of spawning a subagent, STOP.** Call the Task tool. Every time. For every task. Including create tasks. Including "simple" tasks.

The lead session's ONLY job is: read queue, spawn subagent, evaluate return, update queue, repeat.

---

## Phase Configuration

Each phase maps to specific Task tool parameters. Use these EXACTLY when spawning subagents.

| Phase | Skill Invoked | Purpose |
|-------|---------------|---------|
| extract | /reduce | Extract claims from source material |
| create | (inline note creation) | Write the {DOMAIN:note} file |
| enrich | /enrich | Add content to existing {DOMAIN:note} |
| reflect | /reflect | Find connections, update {DOMAIN:topic map}s |
| reweave | /reweave | Update older {DOMAIN:note_plural} with new connections |
| verify | /verify | Description quality + schema + health checks |

**All phases use the same subagent configuration:**
- subagent_type: knowledge-worker (if available) or default
- mode: dontAsk

Subagents inherit the session model. Users running opus get opus quality on processing phases. Users running sonnet get sonnet everywhere. Fresh context per phase already ensures efficiency — every phase gets full capability in the smart zone.

---

## Step 1: Read Queue State

Read the queue file. Check these locations in order:
1. `ops/queue.yaml`
2. `ops/queue/queue.yaml`
3. `ops/queue/queue.json`

Parse the queue. Identify ALL pending tasks.

**Queue structure (v2 schema):**

The queue uses `current_phase` and `completed_phases` per task entry:

```yaml
phase_order:
  claim: [create, reflect, reweave, verify]
  enrichment: [enrich, reflect, reweave, verify]

tasks:
  - id: source-name
    type: extract
    status: pending
    source: ops/queue/archive/2026-01-30-source/source.md
    file: source-name.md
    created: "2026-01-30T10:00:00Z"

  - id: claim-010
    type: claim
    status: pending
    target: "claim title here"
    batch: source-name
    file: source-name-010.md
    current_phase: reflect
    completed_phases: [create]
```

If the queue file does not exist or is empty, report: "Queue is empty. Use /seed or /pipeline to add sources."

## Step 2: Filter Tasks

Build a list of **actionable tasks** — tasks where `status == "pending"`. Order by position in the tasks array (first = highest priority).

Apply filters:
- If `--batch` specified: keep only tasks where `batch` matches
- If `--type` specified: keep only tasks where `current_phase` matches (e.g., `--type reflect` finds tasks whose `current_phase` is "reflect")

The `phase_order` header defines the phase sequence:
- `claim`: create -> reflect -> reweave -> verify
- `enrichment`: enrich -> reflect -> reweave -> verify

## Step 3: If --dry-run, Report and Stop

Show this and STOP (do not process):

```
--=={ ralph dry-run }==--

Queue: X total tasks (Y pending, Z done)

Phase distribution:
  Claims:       {create: N, reflect: N, reweave: N, verify: N}
  Enrichments:  {enrich: N, reflect: N, reweave: N, verify: N}

Next tasks to process:
1. {id} — phase: {current_phase} — {target}
2. {id} — phase: {current_phase} — {target}
...

Estimated: ~{N} subagent spawns
```

---

## Step 4: Process Loop (SERIAL MODE)

**If `--parallel` is set, skip to Step 6 instead.**

Process up to N tasks (default 1). For each iteration:

### 4a. Select Next Task

Pick the first pending task from the filtered list. Read its metadata: `id`, `type`, `file`, `target`, `batch`, `current_phase`, `completed_phases`.

The `current_phase` determines which skill to invoke.

Report:
```
=== Processing task {i}/{N}: {id} — phase: {current_phase} ===
Target: {target}
File: {file}
```

### 4b. Build Subagent Prompt

Construct a prompt based on `current_phase`. Every prompt MUST include:
- Reference to the task file path (from queue's `file` field)
- The task identity (id, current_phase, target)
- The skill to invoke with `--handoff`
- `ONE PHASE ONLY` constraint
- Instruction to output RALPH HANDOFF block

**Phase-specific prompts:**

For **extract** phase (type=extract tasks only):
```
Read the task file at ops/queue/{FILE} for context.

You are proce

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.