Contacts Event

contacts-event — log a complex situation around a contact

User input: /contacts-event (natural language).

Where's my dunbar root?

Same resolution as contacts-sync / contacts-add:

1. $DUNBAR_ROOT environment variable
2. Current working directory if it has contacts.csv and at least one 0N_* circle folder
3. Ask the user

What this skill is for (and what it isn't)

contacts-add keeps a clean profile of a person — basics, how you met, interaction history. It explicitly forbids writing TBDs, risk flags, commentary, or audits of the person into that profile.

But real life produces situations: a contact gets tangled in something complex, multi-party, with accounts that contradict each other, and still moving — a dispute, a scam, a crisis, a family emergency. That carries a lot of inference, open questions, judgments about third parties, and sensitive detail. None of it belongs in the clean profile, yet it all needs a disciplined home. This skill is that home.

Two core principles

Principle 1 — Keep the two artifacts separate

• Contact profile (.md): stays clean, neutral, shareable. Health / legal exposure / judgments about others / to-dos do NOT go here.
• Event dossier (YYYY-MM-DD_.md, same folder): all the analysis, sensitive detail, hypotheses, and to-dos live here. An internal working file, not for outside eyes.

Principle 2 — Five-layer tagging (the heart of the method)

Tag every line in the dossier with how trustworthy it is:

| Tag | Meaning | Use | |-----|---------|-----| | 🟦 Fact / firsthand | the principal's own firsthand account, or objective records (bank/police records, original message screenshots) | hardest; safe to act on | | 🟨 Secondhand | relayed by a third party / family member | stays 🟨 forever — never promote to 🟦 | | 🟧 Hypothesis | a logic / assumption you and others constructed | mark it as a guess, not a fact | | 💬 Commentary | the agent's analysis, caution, counter-reading | helps judgment, isn't the verdict | | ⬜ To confirm | a gap you still need to close | the holes to plug before acting |

Why this is the valuable part: complex situations blow up when someone treats "what the mother said" as "what the principal said", or "what we guessed" as "what we verified", and then reports to the police / confronts someone on that basis — and it collapses on first contact. Tagging puts whose-version-counts and what's-still-empty out in the open, so a later reversal doesn't catch you having recorded a guess as a fact.

Flow

1. Locate the contact

``bash NAME="" grep -ih "^${NAME}," "$DUNBAR_ROOT/contacts.csv" ` If absent → run /contacts-add` first to create the person, then come back.

1. Create / update the dossier in the contact's folder. Two default artifacts:

• Event dossier YYYY-MM-DD_.md — five-layer tagging throughout.
• Chronology _chronology_YYYY-MM-DD.md — narrative + timeline table.

Keep updating the same files as new information arrives across conversations — don't spawn a new file each time.

1. Optional sub-artifacts (only when asked):

• First-person narrative — the situation written as a story (for a retrospective, or possible publication; scrub real identities before sharing anything).
• Evidence / action checklist — when there's a dispute or legal angle, a tickable list.

1. Cross-link: add one line to the profile's "interactions" section pointing to the dossier (a pointer, no analysis); link [[nickname]] back from the dossier.

1. Evolution discipline (when updating):

• Firsthand beats secondhand. When the two clash, the firsthand account wins; the secondhand one stays 🟨.
• Reversals leave a trail. When a fact is overturned, mark the old judgment "superseded" but keep it — the process trail is how you avoid repeating the mistake.
• Status banner. Maintain a one-line status (latest conclusion / reversal / progress) at the top of the dossier.

Event dossier template

#  · 

> **Logged**: YYYY-MM-DD ｜ **Status**: 
> **Sources**: 
> **Layers**: 🟦 firsthand · 🟨 secondhand · 🟧 hypothesis · 💬 commentary · ⬜ to-confirm
> ⚠️ Internal full record; contains sensitive detail. Any outward-facing version is made separately, with a different scope (see "Outward scope").

## 1. One-liner + current status
## 2. The ask / the dispute
## 3. Established facts (🟦, line by line)
## 4. Each side's account (🟦 firsthand / 🟨 secondhand, side by side — don't merge or reconcile)
## 5. To confirm (⬜, each with "why it matters")
## 6. Our hypothesis vs the other reading (🟧 one version + 💬 the counter-reading, side by side)
## 7. To-dos
## 8. Updates / reversal trail (append by date; mark superseded judgments, don't delete)

## Outward scope
- For your own lawyer / advisor: full and honest.
- For police / outside parties: only true, fact-focused; no false or misleading statements.

Chronology template

#  · chronology

> Layers as above. ★ If there's a reversal, note "the corrected version in §N governs; earlier sections kept as a process trail."

## Narrative: root cause → trigger → course → outcome
(each paragraph tagged with its layer, told as a connected causal story)

## Timeline
| When | What | Layer |
|------|------|-------|

## What's been superseded (on a reversal)
| Old version | Corrected |

## Still open / watch (⬜ + ⚠️)

Don't ❌

• Don't put the analysis / to-dos / third-party judgments / sensitive content into the clean profile (that's contacts-add's turf, with its own bans).
• Don't write a 🟨 secondhand claim as a 🟦 fact — however confident the person relaying it is.
• Don't delete an overturned judgment on a reversal; mark it superseded and keep it.
• Don't dress up a hypothesis (🟧) as verified fact; and don't record only your version — put the counter-reading next to it (💬).
• Don't invent to-dos; record only what the user actually said.
• For anything legal, label conclusions "directional, not formal legal advice" and don't impersonate a lawyer.

Example (abstracted, no real PII)

User input:

/contacts-event log the mess where Sam's flatmate supposedly extorted money from him

The skill does:

1. grep contacts.csv for ^Sam, → found → use Sam's folder.
2. Create 2026-01-20_flatmate-money-dispute.md (event dossier) + Sam_chronology_2026-01-20.md.
3. Put the family's relayed account (🟨 "the flatmate threatened him at knifepoint") next to Sam's own firsthand account (🟦 "there was no threat, I inferred it") — flag the gap, don't merge them.
4. Record "the flatmate is the aggressor" as a hypothesis (🟧), and put the counter-reading next to it (💬 "the flatmate may be an ordinary person who got dragged in").
5. Unknowns — the amount, the recipient's real name, who called the police — go under "to confirm" (⬜).
6. When later conversations establish the truth (a reversal) → update the top status banner, mark the old judgment superseded but keep it, add a "corrected version" section to the chronology.
7. Add one line to Sam's profile interactions pointing to the dossier; [[Sam]] back-links from the dossier.
8. If asked, produce a first-person narrative or an evidence checklist.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Astralune-ai
• Source: Astralune-ai/dunbar
• License: MIT

Install and usage instructions live in the source repository linked above.