Route Information

Given a batch of discrete information items (each with an ID, raw text, and optional agent-proposed metadata), present routing recommendations to the user in clusters, get confirmation or correction, and execute the approved routings.

Workflow name: route-information

Execution: main

Steps

#	ID	Name	Type	Depends on
1	cluster-items	Cluster items for review by type/theme	reasoning
2	present-clusters	Present each cluster to the user for confirmation	reasoning	cluster-items
3	record-decisions	Record user decisions for each item	reasoning	present-clusters
4	execute-routing	Execute approved routings (consent-gated writes)	code	record-decisions
5	return-results	Return complete routing record	code	execute-routing

Step instructions

cluster-items

Group items by proposed_type (or by theme if types are mixed). The goal is to minimize the number of review interactions while keeping each interaction coherent.

Suggested clustering: - Admin/personal items together (usually quick decisions) - Research/technical items together (need domain context) - Unknown/low-confidence items together (need the most user input) - Proposed deletions together (batch confirm)

present-clusters

For each cluster, present: 1. The cluster theme/type 2. For each item in the cluster: - The raw text (verbatim, never modified) - The agent's one-line interpretation - The proposed action (route to X, delete, park with review date) - Confidence level and any staleness notes 3. Ask the user to confirm, correct, or skip each item

Example interaction:

## Admin items (3 items)

**[t_d44f12]** "Admin - VPN/taxes/new laptop stuff"
→ Proposed: task (inbox) — "Set up VPN, handle tax prep, research laptops"
→ Confidence: low — this bundles 3 things, and taxes may be done
What is this? Confirm / correct / skip?

**[t_c44d1a]** "Taxes - Varsha"
→ Proposed: task (inbox) — "Tax-related task involving Varsha"
→ Confidence: low — don't know who Varsha is or what the task is
What is this? Confirm / correct / skip?

**[t_7b2e09]** "Varsha - T4 / T4A"
→ Proposed: linked to t_c44d1a? — "Tax form from Varsha"
→ Confidence: low
What is this? Confirm / correct / skip?

The user might reply: "Varsha is the lab admin. I already got the T4A. Delete both of those. The VPN/taxes/laptop thing — VPN is done, taxes are done, laptop is still a live task."

record-decisions

Update each item with the user's confirmed routing:

{
  "id": "t_c44d1a",
  "action": "delete",
  "user_note": "Already completed — got T4A from Varsha"
}

{
  "id": "t_d44f12",
  "action": "split",
  "splits": [
    {"text": "VPN setup", "action": "delete", "reason": "done"},
    {"text": "Taxes", "action": "delete", "reason": "done"},
    {"text": "New laptop research", "action": "route", "destination": "tasks/master-task-list.md", "task_text": "Research and purchase new laptop"}
  ]
}

execute-routing

For each confirmed item, execute the routing:

• Task creation: Use mcp__work-buddy__wb_run("task_create", {"task_text": "...", "project": "...", "due_date": "..."}) which produces clean lines and creates a SQLite metadata record:

  - [ ] #todo <task text> #projects/<project> 🆔 t-<hex> 📅 <date if applicable>
  

  State and urgency are stored in tasks/task_metadata.db, not inline tags.
• Consideration creation: Create a new file in work/considerations/<project>/ using the consideration template frontmatter
• Project note append: Append to the relevant project file
• Park with review date: Create a consideration with status: parked and a decision_date set to the review date
• Delete: Mark as deleted in the routing record (the caller handles actual text removal)

All write operations are consent-gated. The first write triggers @requires_consent. Grant with a reasonable TTL (e.g., 30 minutes) to cover the batch.

return-results

Return the complete routing record as JSON — the caller uses this to know what was processed and what action was taken.

{
  "routed": [{"id": "t_7b2e09", "action": "route", "destination": "tasks/master-task-list.md"}],
  "deleted": [{"id": "t_c44d1a", "reason": "Already completed"}],
  "split": [{"id": "t_d44f12", "splits": [...]}],
  "skipped": [{"id": "t_x9y8z7", "reason": "User wants to think about it"}]
}

Context

What NOT to do

• Don't modify the raw text of any item — present it verbatim
• Don't auto-route without user confirmation, even at high confidence
• Don't create elaborate categorization hierarchies — use existing destinations
• Don't invent new filing locations the user hasn't established
• Don't pressure the user to decide on every item — "skip" is always valid
• Don't treat "unknown" as a problem — some items need the user's context, that's expected
• Don't batch-delete without showing each item — the user might spot something worth keeping

Related blindspots

• Insight hoarding: if the user parks everything and deletes nothing, name the pattern. But don't force deletion — that triggers avoidance
• Infrastructure displacement: if routing itself is taking longer than the items are worth, flag it. "We've spent 20 minutes on 3 items — should we batch the rest?"
• Organization skill gap: if the user can't decide where things go, the destination system might need simplification, not more process