inbox

Inbox: Batch Triage and Processing

Overview

Launches the inbox-organiser agent to process all notes in 0-inbox/. The agent clusters by topic, assesses maturity via the promote-gate skill, detects counter-arguments, and executes promotions autonomously. Merges and deletes require approval.

When to Use

• Inbox has accumulated notes that need triage
• After a series of /reflect sessions that deposited notes
• When the user asks to clean up, process, or triage the inbox

Provenance

This skill emits provenance events for pipeline observability. Run each Bash command silently.

At session start:

node "${CLAUDE_PLUGIN_ROOT}/scripts/provenance-emit.js" '{"agent":"inbox","skill":"inbox","action":"session-start"}'

At session end:

node "${CLAUDE_PLUGIN_ROOT}/scripts/provenance-emit.js" '{"agent":"inbox","skill":"inbox","action":"session-end","promoted":N,"deleted":N,"merged":N,"limbo":N}'

Per-note tracking is handled automatically by the PostToolUse hook.

Process

Step 1: Launch Agent

Launch the inbox-organiser agent with:

• vault_path: {{VAULT}}/
• scope: all (or topic:<name> if the user specified a topic filter)

The agent definition is at PLUGIN/agents/inbox-organiser.md.

Use subagent_type: "learning-loop:inbox-organiser" with the full prompt from the agent definition, or launch as a general-purpose agent that reads the agent file.

Step 1.5: Surface Librarian Observations for Inbox Notes

Before the agent returns, check the librarian queue for pending observations targeting inbox notes.

Read PLUGIN_DATA/librarian/queue.jsonl (where PLUGIN_DATA = CLAUDE_PLUGIN_DATA env or ~/.claude/plugins/data/learning-loop). Parse each line as JSON. Filter to items where status === 'pending', target starts with 0-inbox/, and task is one of: voice_flag, tag_suggestion, duplicate_flag.

If matches exist, include them as advisory context when presenting the agent's results, grouped by task type:

Librarian observations:
  Voice flags:
    "gmail multi daemon pull deduplication": Names a topic, not an insight. Consider retitling.
  Tag suggestions:
    "ginkgo biloba acute pk profile" → pharmacology, neuroscience
  Duplicate flags:
    "foo-claim.md" ↔ 3-permanent/foo-claim-original.md (similarity 0.93)
  ...

These are informational: the user decides whether to act on them during triage. Apply or dismiss them in /health --librarian.

Step 2: Handle Gated Actions

When the agent returns, it will list any actions needing approval (merges, deletes, and Bundle 2 NLI contradictions). Present all three categories in one block; one user response handles all of them. NLI contradictions accept per-item replies in the form a:1 b:3 c:skip (1=supersede, 2=qualify, 3=keep-both) or batched all:3. Execute approved actions in the order: deletes → merges → NLI resolutions → autonomous promotions.

--skip-nli flag: if the user invokes /learning-loop:inbox --skip-nli, pass the flag through to the inbox-organiser agent prompt as additional context. The agent will skip Step 3a.5 (NLI contradiction check) entirely and surface a note in the report: note: --skip-nli set; promotions ran without NLI contradiction checks. Useful when calibrating thresholds or after a known-noisy NLI run.

Step 3: Report

The agent returns a structured summary. Present it to the user.

Key Principles

• The skill is thin. All logic lives in the inbox-organiser agent and its _skills/.
• Promotions are autonomous. No approval needed.
• Destructive actions are gated. Merges, deletes, and fleeting archival need explicit user approval.
• Counter-arguments get promoted, not suppressed. Quality determines folder.
• Fleeting sweep runs after inbox. Archives promoted notes (2+ permanent refs) and stale project notes (0 refs, 60+ days old) to _archive/1-fleeting/.