Pipeline 2

Vault Path Resolution (mnemos)

Before any file operations, resolve the vault root:

1. Read .mnemos.yaml from the workspace root (the project directory where you are running)
2. Extract vault_path — this is the root for ALL vault operations
3. All paths in this skill (notes/, ops/, inbox/, self/, memory/, templates/) are RELATIVE to vault_path
4. If .mnemos.yaml does not exist or vault_path is not set, fall back to the current working directory

Example: if vault_path is /data/vault, then ops/queue/queue.json resolves to /data/vault/ops/queue/queue.json.

EXECUTE NOW

Target: $ARGUMENTS

Parse immediately:

• Source file path: the file to process (required)
• --handoff: output RALPH HANDOFF block at end (for chaining)
• If target is empty: list files in {DOMAIN:inbox}/ and ask which to process

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. Run the full pipeline.

Pipeline Overview

The pipeline chains four phases. Each phase uses skill invocation or /ralph for subagent-based processing. State lives in the queue file — the pipeline is stateless orchestration on top of stateful queue entries.

Source file
    |
    v
Phase 1: /seed — create extract task, move source to archive
    |
    v
Phase 2: /reduce (via /ralph) — extract claims from source
    |
    v
Phase 3: /ralph (all claims) — create -> reflect -> reweave -> verify
    |
    v
Phase 4: /archive-batch — move task files, generate summary
    |
    v
Complete

The pipeline is the convenience wrapper. /ralph is the engine. /seed is the entry point.

Phase 1: Seed

Invoke /seed on the target file to create the extract task, check for duplicates, and move the source to its archive folder.

How to invoke:

Use the Skill tool if available, otherwise execute the /seed workflow directly:

• Validate source exists
• Check for prior processing (duplicate detection)
• Create archive folder
• Move source from {DOMAIN:inbox} to archive
• Create extract task file
• Add extract task to queue

Capture from seed output:

• Batch ID: the source basename (used for --batch filtering in subsequent steps)
• Archive folder path: where the source was moved
• next_claim_start: the claim numbering start

Report: $ Seeded: {source-name}

If seed reports the file was already processed: Ask the user whether to proceed or skip. Do NOT auto-skip — the user may want to re-process with different scope.

Phase 2: Extract (Reduce)

Process the extract task via /ralph. This spawns a subagent that runs /reduce, extracting claims from the source and creating task entries in the queue.

How to invoke:

/ralph 1 --batch {batch_id} --type extract

Or via Task tool:

Task(
  prompt = "Run /ralph 1 --batch {batch_id} --type extract",
  description = "extract: {batch_id}"
)

After completion, read the queue to count extracted claims and enrichments:

Check how many pending tasks exist for this batch. The reduce phase creates 1 queue entry per claim and 1 per enrichment.

Report:

$ Extracted: {N} {DOMAIN:note_plural}, {M} enrichments
  Processing {total_tasks} tasks through the pipeline...

If zero claims extracted: Report the issue. For TFT sources, zero extraction is a bug — the source almost certainly contains extractable content. Ask the user whether to retry with different scope or skip.

Phase 3: Process All Claims

Count total pending tasks for this batch from the queue. Then process all of them through the full phase sequence.

How to invoke:

/ralph {remaining_count} --batch {batch_id}

Or via Task tool:

Task(
  prompt = "Run /ralph {remaining_count} --batch {batch_id}",
  description = "process: {batch_id} ({remaining_count} tasks)"
)

This processes every claim through: create -> reflect -> reweave -> verify. And every enrichment through: enrich -> reflect -> reweave -> verify.

Each phase runs in an isolated subagent with fresh context. /ralph handles all the orchestration: subagent spawning, handoff parsing, queue advancement, learnings capture.

Progress reporting:

The /ralph invocation reports progress per task. The pipeline relays this:

$ Processing {DOMAIN:note} 1/{total}: {title}
  $ create... done
  $ reflect... done (3 connections found)
  $ reweave... done (2 {DOMAIN:note_plural} updated)
  $ verify... done (PASS)

For large batches (20+ claims): /ralph handles context isolation automatically via subagents. The pipeline does NOT need to chunk — /ralph processes N tasks sequentially with fresh context per phase.

Phase 4: Verify Completion

After /ralph finishes, verify all tasks for this batch are done.

Check the queue: count tasks for this batch that are NOT done.

If tasks remain pending:

• Report which tasks are incomplete and at which phase
• Show the specific task IDs and their current_phase
• Suggest: "Run /ralph --batch {batch_id} to continue from where it stopped"
• Do NOT proceed to archive

If all tasks are done: Proceed to Phase 5.

Phase 5: Archive Batch

When all tasks for the batch are complete, archive the batch.

How to invoke:

/archive-batch {batch_id}

Or execute directly:

1. Move all task files from ops/queue/ to ops/queue/archive/{date}-{batch_id}/
2. Generate a batch summary file: {batch_id}-summary.md
3. Remove completed entries from the queue (or mark as archived)

The summary should include:

• Source file name and original location
• Number of claims extracted
• Number of enrichments
• List of created {DOMAIN:note_plural} with titles
• Any notable learnings from the batch

Phase 6: Final Report

--=={ pipeline }==--

Source: {source_file}
Batch: {batch_id}

Extraction:
  {DOMAIN:note_plural} extracted: {N}
  Enrichments identified: {M}

Processing:
  {DOMAIN:note_plural} created: {N}
  Existing {DOMAIN:note_plural} enriched: {M}
  Connections added: {C}
  {DOMAIN:topic map}s updated: {T}
  Older {DOMAIN:note_plural} updated via reweave: {R}

Quality:
  All verify checks: {PASS/FAIL count}

Archive: ops/queue/archive/{date}-{batch_id}/
Summary: {batch_id}-summary.md

{DOMAIN:note_plural} created:
- [[claim title 1]]
- [[claim title 2]]
- ...

If --handoff flag was set, also output:

=== RALPH HANDOFF: pipeline ===
Target: {source_file}

Work Done:
- Seeded source: {batch_id}
- Extracted {N} {DOMAIN:note_plural} and {M} enrichments
- Processed all claims through 4-phase pipeline
- Archived batch to {archive_path}

Files Modified:
- {DOMAIN:notes}/ ({N} new {DOMAIN:note_plural})
- ops/queue/archive/{date}-{batch_id}/ (archived)

Learnings:
- [Friction]: {description} | NONE
- [Surprise]: {description} | NONE
- [Methodology]: {description} | NONE
- [Process gap]: {description} | NONE

Queue Updates:
- All tasks for batch {batch_id} marked done and archived
=== END HANDOFF ===

Error Handling

Phase failure at any stage:

1. Report the failure with context (which phase, which task, what error)
2. Show the current queue state for this batch
3. Suggest remediation: "Run /ralph --batch {batch_id} to continue from where it stopped"
4. Do NOT attempt to continue automatically past failures

The pipeline is resumable. Queue state persists across sessions:

• /seed detects prior processing and asks whether to proceed
• /ralph picks up from the last completed phase (queue is the source of truth)
• /archive-batch verifies completeness before archiving

Seed failure: If /seed fails (file not found, duplicate detected and user declines), stop the pipeline entirely.

Extract failure: If /reduce extracts zero claims, report and stop. Do not proceed to an empty processing phase.

Processing failure: If /ralph fails mid-batch, the queue preserves state. Individual claims resume from their failed phase on next /ralph invocation.

Archive failure: If archiving fails, the claims are still created and connected. Only the organizational cleanup is missing — re-run /archive-batch manually.

Resumability

The pipeline is designed to be interrupted and resumed at any point:

State lives in the queue file. The pipeline reads queue state, not session state. This means you can interrupt, close the session, and resume later.

Edge Cases

No target file: List {DOMAIN:inbox}/ candidates, suggest the best one based on age and relevance.

Source already seeded: /seed detects this and asks the user. If they decline, the pipeline stops cleanly.

Large source (2500+ lines): /reduce handles chunking automatically. The pipeline does not need special handling.

No ops/derivation-manifest.md: Use universal vocabulary for all output.

Critical Constraints

never:

• Skip the seed phase (duplicate detection is important)
• Continue past a failed phase automatically
• Process claims inline instead of via /ralph subagents
• Archive a batch with incomplete tasks

always:

• Report progress at each phase boundary
• Verify all tasks are done before archiving
• Show the user what was created (list of {DOMAIN:note_plural})
• Suggest next steps if interrupted
• Use domain-native vocabulary from derivation manifest