Triage Coordinator

Unified inbox processor - handles ALL content dumped into inbox with parallel subagents and single-table review.

Key design: Subagents persist proposals immediately via TaskUpdate. If session crashes at item 23 of 50, items 1-22 are saved and resumable.

Scope

Single-session workflow. Creates quick resource/meeting notes with distilled: false. Use /para-obsidian:distill-resource for progressive summarization.

---

Skill Architecture

This orchestrator spawns triage-worker agents, each self-contained with preloaded skills:

┌─────────────────────────────────────────────────────────────┐
│  Tier 1: REFERENCE SKILLS (Knowledge)                       │
│  - para-classifier      (PARA philosophy, emoji mapping)    │
│  - content-processing   (note creation, Layer 1, commit)    │
│  - content-sourcing     (URL routing, tool selection)        │
└─────────────────────────────────────────────────────────────┘
                            ↓ loaded by
┌─────────────────────────────────────────────────────────────┐
│  Tier 2: WORKER SKILLS (preloaded into triage-worker agent) │
│  - analyze-web         (analyze web content → proposal)     │
│  - analyze-voice       (analyze transcription → proposal)   │
│  - analyze-attachment  (analyze PDF/DOCX → proposal)        │
└─────────────────────────────────────────────────────────────┘
                            ↓ orchestrated by
┌─────────────────────────────────────────────────────────────┐
│  Tier 3: ORCHESTRATOR (This skill)                          │
│  - triage              (unified inbox processing)           │
└─────────────────────────────────────────────────────────────┘

Note: content-processing is the canonical pipeline for note creation, Layer 1 injection, and commit — shared by triage-worker and quick-create. create-resource and create-meeting are standalone skills for non-triage workflows.

---

Architecture Overview

Phase 1: Initialize (coordinator)
├── Scan inbox, detect VTT files, create tasks
└── Load vault context (areas, projects)

Phase 2: Enrich + Analyze + Create (subagents)
├── Route to correct analyzer based on item type
├── Parallel for YouTube, articles, X/Twitter (batches of 10)
├── Sequential for Confluence only (single browser)
├── CREATE notes AND inject Layer 1 content (but DO NOT delete originals)
└── Enriched content stays in subagent context

Phase 2.5: Coordinator Verification (coordinator) ← AUTOMATIC
├── Runs automatically after ALL Phase 2 subagents complete, before Phase 3
├── For each proposal with created != null
├── para_fm_set: stamp summary + source + classification fields from proposal
├── para_fm_get: verify all critical fields populated
└── Override haiku's verification_status with coordinator's assessment

Phase 3: Present & Collaborate (coordinator) ← USER CHECKPOINT
├── Render table with all proposals
├── **ASK USER** - accept/edit/delete?
└── User reviews and can modify area/project/title

Phase 4: Edit (only if requested)
└── Apply edits via para_fm_set or re-create

Phase 5: Execute (coordinator) ← AFTER APPROVAL
├── Delete/archive originals (only now!)
├── Apply any remaining edits
└── Cleanup tasks + report

Key insight: Subagents create notes but originals stay until user approves. This enables collaborative review while keeping content isolated.

See: architecture.md for diagrams.

---

CRITICAL: Context Isolation

The orchestrator MUST NOT read content. All content reading happens in subagents. Never call para_read from the coordinator — spawn a subagent instead.

Why this is a hard rule: 50 inbox items × 10k average tokens = 500k tokens. This overflows the coordinator's context and causes timeouts. With isolation, the coordinator handles only ~500 bytes per item (proposals), keeping total context under 25k tokens.

See context-isolation.md for rules, token math, and common mistakes.

---

Input Routing

Check $ARGUMENTS:

Input	Action
Empty or all	Full inbox processing
clippings / voice / attachments	Filter by type
"filename.md"	Single file mode (skip batching)
"filename.vtt"	Convert VTT first, then process

---

Pre-flight: Sync Voice Memos (automatic)

Voice memo sync runs automatically via dynamic context injection before the skill reaches the agent. This ensures freshly recorded memos are in the inbox before Phase 1 scans it.

Result: !cd ${CLAUDE_PLUGIN_ROOT} && bun src/cli.ts voice 2>&1 | tail -5

If the output shows failures (e.g., parakeet-mlx not installed), log a warning and continue — existing transcriptions in the inbox will still be processed.

---

Phase 0: Check for Resume

const tasks = TaskList();
const triageTasks = tasks.filter(t => t.subject.startsWith("Triage:"));

0.1 Classify Existing Tasks

const analyzed = triageTasks.filter(t => t.status === "in_progress");  // proposals saved
const pending = triageTasks.filter(t => t.status === "pending");       // need processing
const completed = triageTasks.filter(t => t.status === "completed");   // already done

0.2 Offer Resume

If existing triage tasks found:

Found existing triage session:
• 32 analyzed (proposals saved, notes created)
• 18 pending (need enrichment + analysis)
• 0 completed

Resume from where you left off? (y/n)

0.3 Resume Flow

If yes:

1. Skip Phase 1 — tasks already created
2. Phase 2 — process ONLY pending tasks (spawn subagents for unprocessed items)
3. Phase 2.5 — re-run verification on ALL items with created != null (idempotent)
4. Phase 3 — collect proposals via TaskGet loop (not response text), present review table
5. Phases 4-5 — proceed normally

If no → delete existing triage tasks and start fresh (Phase 1).

---

Phase 1: Initialize

1.1 Load Vault Context (via preflight subagent)

Spawn a haiku preflight subagent to gather vault context cheaply:

Task({
  model: "haiku",
  subagent_type: "general-purpose",
  description: "Triage preflight",
  prompt: "<preflight prompt with mode=triage>"
})

Use the prompt template from ../brain/references/preflight-prompt.md with $MODE = triage.

Parse PREFLIGHT_JSON:{...} from the response. Extract areas, projects, inbox_items, and stakeholders.

Fallback: If the subagent fails, fall back to direct MCP calls:

para_list_areas({ response_format: "json" })
para_list_projects({ response_format: "json" })
para_config({ response_format: "json" })

Extract stakeholders array from config (names, roles, companies for transcription speaker matching).

1.1.1 Stakeholder Bootstrap (if missing)

If config.stakeholders is empty AND inbox contains voice memos/transcriptions, offer to add stakeholders. Stakeholders enable speaker matching in voice memo analysis — without them, all speakers appear as "Speaker 1", "Speaker 2", and project inference is disabled.

Trigger condition: Empty stakeholders + transcriptions in inbox. Skip if: No voice memos in inbox (stakeholders only help with transcription processing). Impact of skipping: Voice memo classification still works, but accuracy is reduced for multi-speaker scenarios and project inference is unavailable.

See stakeholder-bootstrap.md for the interactive wizard flow (bulk paste, one-at-a-time, or skip).

1.2 Scan Inbox

para_list({ path: "00 Inbox", response_format: "json" })

For each item, extract metadata only (no content):

para_fm_get({ file: itemPath, response_format: "json" })

Extract from frontmatter:

• type → routes to analyzer skill
• source → URL for enrichment (if clipping)
• areas, projects → pre-filled values

DO NOT call para_read. Content analysis happens in Phase 2 subagents.

1.3 Handle VTT Files

CRITICAL: VTT files must be converted to transcriptions before processing.

For each .vtt file found:

cd ${CLAUDE_PLUGIN_ROOT} && bun src/cli.ts voice convert "<vtt-path>" --format json

Date handling for VTT:

• If --date flag provided in $ARGUMENTS, use it
• Otherwise, prompt user: "VTT files require a meeting date. Enter date (YYYY-MM-DD):"
• If user cancels or enters empty: skip this VTT file and continue with other items
• If date format is invalid: re-prompt with "Invalid format. Use YYYY-MM-DD:"
• Store date in task metadata as vttDate for Phase 2 subagent reference
• Store converted transcription path for processing

1.4 Categorize Items

Item Type (routes to analyzer skill):

Type	Detection	Analyzer Skill
clipping	type === "clipping"	analyze-web
transcription	type === "transcription"	analyze-voice
attachment	PDF/DOCX extension	analyze-attachment
vtt	.vtt extension	Convert first → analyze-voice

Source Type (for enrichment within analyzer):

See enrichment-strategies.md for the canonical routing table (source detection, tool selection, parallelization constraints).

CRITICAL - X/Twitter: Web Clipper captures only stubs. You MUST enrich via X-API MCP tools (x_get_tweet) regardless of clipping content.

1.5 Create All Tasks Upfront

// For each inbox item:
TaskCreate({
  subject: "Triage: ✂️ Article Name",
  description: "File: 00 Inbox/✂️ Article.md\nType: clipping\nSource: youtube",
  activeForm: "Analyzing article",
  metadata: {
    file: "00 Inbox/✂️ Article.md",
    itemType: "clipping",
    sourceType: "youtube",
    proposal: null  // Filled by subagent
  }
})

Task IDs are auto-generated. Store mapping: { taskId → file }.

1.6 Pre-load Template Fields

Call para_template_fields once per unique template type (usually just "resource", sometimes "meeting"). Pass results to subagents so they skip this call.

// For each unique proposed template:
para_template_fields({ template: "resource", response_format: "json" })
// → { validArgs, creation_meta: { contentTargets, dest, titlePrefix, sections } }

Store the results keyed by template name. Include in every subagent prompt (see subagent-prompts.md).

1.7 Present Summary

Found 50 items in inbox:

📋 By Type:
• 40 clippings
• 8 voice memos (including 2 converted VTT)
• 2 attachments

📋 By Enrichment:
• 35 parallel (YouTube, articles)
• 5 X/Twitter (parallel via X-API)
• 10 no enrichment needed

Starting subagent processing...

---

Phase 2: Parallel Enrich + Analyze + Create

Key insight: Each subagent handles enrichment, analysis, AND note creation. Content stays isolated in subagent context - only lightweight proposals flow back to coordinator.

CRITICAL: Subagents create notes but DO NOT delete/archive originals. Deletion happens in Phase 5 AFTER user review and approval.

2.1 Spawn Subagents

For all items (up to 10 per batch), spawn subagents in a single message for parallel execution. For inboxes >10 items, use batches of 10. Claude Code handles 7-10 parallel Task calls well for haiku subagents.

EXCEPTION: Confluence items must be sequential (single Chrome browser). Process these separately after parallel items complete.

Use the prompt template from subagent-prompts.md.

Pass these variables to each subagent:

• taskId, file, sourceUrl, itemType, sourceType
• areas, projects, stakeholders (from Phase 1)
• templateFields (from Phase 1.6 — pre-loaded template metadata)

MANDATORY batch mode flags: Instruct subagents to pass no_autocommit: true and skip_guard: true to para_create. These flags are required for parallel execution — without them, parallel subagents trigger git guard conflicts and unintended per-item commits. The coordinator bulk-commits once after all subagents complete (Phase 5).

Each subagent will: enrich content, analyze, create note (no commit), persist via TaskUpdate, and return PROPOSAL_JSON:{...}.

2.1.1 Model Selection by Item Type

Override the agent's default model based on content complexity:

Item Type	Model	Why
clipping	haiku (default)	Enrichment provides strong structured source content (articles, videos, threads). Haiku handles well.
transcription	sonnet	Transcriptions have ambiguous speakers, unclear meeting boundaries, and nuanced categorization (is this a standup or a 1on1?). Sonnet's reasoning handles this ambiguity better.
vtt (converted)	sonnet	Same as transcription — multi-speaker meetings need stronger judgment
attachment	haiku (default)	Document structure is usually clear (invoices, contracts have obvious patterns)

// Pass model override in Task call:
Task({
  subagent_type: "para-obsidian:triage-worker",
  model: itemType === "transcription" ? "sonnet" : undefined,  // undefined = use agent default (haiku)
  description: "Process: ...",
  prompt: `...`
})

2.2 Handle Confluence Separately

Confluence requires Chrome DevTools (single browser instance). Process sequentially AFTER parallel items:

// After all parallel subagents complete, process Confluence items one at a time
for (const confluenceItem of confluenceItems) {
  Task({
    subagent_type: "triage-worker",
    description: "Process: Confluence Page",
    prompt: `... same prompt with sourceType: "confluence" ...`
  })
  // Wait for completion before next
}

Note: X/Twitter items are now processed in the parallel batch (Section 2.1) using stateless X-API MCP tools.

Proposal collection happens in Phase 3. Subagents return PROPOSAL_JSON:{...} in response text.

2.3 Error Handling

Subagents may fail during enrichment (timeouts, 404s, rate limits) or return invalid proposals.

Key behaviors:

• Failed subagents leave tasks as pending (no TaskUpdate called)
• Invalid proposals get enrichmentFailed: true in metadata
• Phase 3 table flags failed items with status column
• User can Retry (R), Delete (D), or Skip (S) failed items

See architecture.md#error-handling for detailed error flows and recovery patterns.

---

Phase 2.5: Coordinator Verification

Rationale: Haiku triage-workers generate correct values in PROPOSAL_JSON but drop fields ~50% of the time when constructing para_create args. Haiku also falsely reports verification_status: "verified" when fields are empty. The coordinator has authoritative values from PROPOSAL_JSON and stamps them unconditionally — idempotent (if haiku got it right, overwriting with the same value does no harm).

This replaces trust in haiku's verification_status. Workers now set verification_status: "pending_coordinator" and skip post-creation verification.

Stamp Set by Template

Template	Fields Stamped	Source
resource	summary, source, source_format, resource_type, areas, projects	proposal + task metadata (sourceUrl)
meeting	summary, area, meeting_type	proposal
invoice/booking	(skip)	N/A

Verification Loop

For each proposal where created != null and template not in [invoice, booking]:

1. Build stamp_set from PROPOSAL_JSON fields + task metadata (sourceUrl for source field)
2. para_fm_set with skip_guard: true — stamp all critical fields unconditionally (no git guard conflicts during batch)
3. Verify from fm_set response — use result.attributes.after to check fields (no separate para_fm_get needed)
4. Check critical fields — if all populated → "verified", if any still empty → "needs_review"
5. Override haiku's verification_status with the coordinator's assessment
6. After loop: single para_commit() to commit all stamped changes

See execution-phases.md for full pseudocode and field mapping.

---

Phase 3: Present & Collaborate

Key insight: This is the collaborative checkpoint. Notes are already created, but originals still exist. User reviews proposals and can edit before we clean up.

Collect proposals from subagent response text (normal flow) or TaskGet loop (resume flow). Render table with all proposals. Ask user to Accept/Edit/Delete/Quit. Never skip the collaborative checkpoint.

See execution-phases.md for proposal collection code, table format, and checkpoint details.

---

Phase 4: Edit (If Requested)

Quick inline edits for area, project, title, or type. Show current values, prompt for changes, update task metadata.

See execution-phases.md for edit flow.

---

Phase 5: Execute (After User Approval)

Notes are ALREADY created by subagents (uncommitted in batch mode). Phase 5 handles:

1. Bulk commit - para_commit() once to commit all notes created by subagents (batch mode)
2. Check creation status - verify created and layer1_injected fields
3. Apply edits - re-create if title changed, para_fm_set if area/project changed
4. Delete/archive originals - route by itemType (see table below)
5. Handle failures - fall back to coordinator creation if subagent failed
6. Cleanup tasks - mark completed
7. Report - summary of all processed items

CRITICAL: Original Cleanup Rules

Route on itemType (source content), NOT proposed_template (output type). A transcription classified as a resource is STILL archived, never deleted.

itemType	Action	Why
transcription	Archive via para_rename to 04 Archives/Transcriptions/, then update resource note source to [[archived note]] via para_fm_set	Raw recordings have intrinsic value - NEVER delete. Resource note must link back to archived transcription
clipping	Delete via para_delete	Content captured in resource note
attachment	Delete inbox note via para_delete	PDF/DOCX stays in Attachments/
clipping (non-URL)	Keep in inbox	Thought/conversation clippings are manual review items. Detection: sourceUrl is empty or doesn't start with http

NEVER use para_delete on transcriptions. Always use para_rename to archive them.

See execution-phases.md for status matrix, code patterns, and report format.

---

Quick Reference

Task States

State	Meaning	What exists
pending	Created by coordinator in Phase 1, NOT yet processed by subagent	Original inbox file only
in_progress	Subagent completed: note created, proposal saved in metadata	New note + original inbox file (both exist)
completed	User approved in Phase 3, original deleted/archived in Phase 5	New note only (original cleaned up)

State flow: pending → (subagent enriches + analyzes + creates note + calls TaskUpdate) → in_progress → (user approves + coordinator cleans up) → completed

Resume Capability

If session crashes/quits:

1. Tasks with status: "in_progress" have notes created + proposals saved in metadata + originals preserved
2. Tasks with status: "pending" need full processing (subagent never ran or crashed before TaskUpdate)
3. Run /triage again → Phase 0 detects existing tasks → offers resume
4. On resume:
   • Skip Phase 1 (tasks already created)
   • Phase 2 processes ONLY pending tasks (re-spawn subagents)
   • Phase 2.5 re-runs on ALL items with created != null (idempotent — safe to re-stamp)
   • Phase 3 collects proposals via TaskGet loop (not response text) and presents review table
   • Phases 4-5 proceed normally

---

Classification Reference

For classification decision trees and emoji mappings used during analysis, see the para-classifier skill:

• Classification Decision Tree: @../para-classifier/references/classification-decision-tree.md
• Emoji Mapping (source_format): @../para-classifier/references/emoji-mapping.md

These provide the framework for determining note types and source_format values.

---

Worker Skills

The triage-worker agent has these skills preloaded:

Skill	Purpose	When Used
analyze-web	Analyze web clippings	itemType === "clipping"
analyze-voice	Analyze transcriptions	itemType === "transcription"
analyze-attachment	Analyze PDF/DOCX	itemType === "attachment"

Standalone skills (not used during triage — the worker handles creation inline):

Skill	Purpose	When Used
create-resource	Create resource note	Non-triage workflows
create-meeting	Create meeting note	Non-triage workflows

---

Completion Signal

After the final report in Phase 5, emit a structured completion signal so the brain orchestrator can parse the outcome:

• All processed: SKILL_RESULT:{"status":"ok","skill":"triage","summary":"Processed N items","processed":N}
• Partial failures: SKILL_RESULT:{"status":"partial","skill":"triage","processed":N,"failed":M}
• Empty inbox: SKILL_RESULT:{"status":"ok","skill":"triage","summary":"Inbox is empty"}
• User cancelled: SKILL_RESULT:{"status":"ok","skill":"triage","summary":"Cancelled by user"}

---

References

File	Content
proposal-schema.md	Canonical proposal structure, field names, confidence levels, TaskUpdate format
architecture.md	Flow diagrams, design rationale
context-isolation.md	Context isolation rules, common mistakes
enrichment-strategies.md	Tool selection by source, voice memo cases, constraints
execution-phases.md	Phases 2.5-5 detailed implementation
subagent-prompts.md	Analysis prompt templates
output-templates.md	Table format, actions
task-patterns.md	TaskCreate/Update API usage
stakeholder-bootstrap.md	Interactive stakeholder wizard