TrueMatch Skill

TrueMatch matches people based on their real personality as observed by their AI model — not self-reported profiles. The full protocol spec is at https://clawmatch.org/skill.md.

Setup (run once)

First, confirm the CLI is installed:

CODEBLOCK0

If that fails, the plugin isn't installed yet. Handle it automatically — do not ask the user to do anything:

1. 1. Use exec to install into $OPENCLAW_STATE_DIR/extensions (user-writable on all setups — Docker, macOS, VPS) and create a convenience symlink:

CODEBLOCK1

Throughout this skill, truematch <args> means node $OPENCLAW_STATE_DIR/extensions/truematch-plugin/dist/index.js <args>.

1. 2. Use the gateway tool to patch the config. First resolve $OPENCLAW_STATE_DIR from the environment and <version> from package.json in the extracted directory. All three fields are required — load.paths alone does not trigger the gateway_start lifecycle hook:

CODEBLOCK2

1. 3. Use the gateway tool to restart the gateway.

1. 4. Re-run truematch --version to confirm. The gateway_start hook fires on restart and auto-creates the heartbeat cron — setup continues normally from here.

Before running setup, collect these in one session — ask each question and wait for the answer before continuing. Do not present them all at once or as a list to the user:

1. 1. Location — ask: "Where are you based?"
2. Distance — ask: "How far are you open to meeting someone — within your city, within a few hours' travel, or anywhere?" Map to: 50 / 300 / omit field.
3. Age range — ask: "Any age range you'd prefer? You can give a rough bracket, a one-sided floor or ceiling, or just say no preference — all fine." Both min and max optional.
4. Gender preference — ask: "Who are you looking to meet? You can be specific, give multiple options, or say everyone — whatever's true for you." Record open/everyone as [].
5. Contact channel — ask: "If we find someone, I'll handle the introduction first — you both decide whether to exchange contact details before anything goes directly between you. For that moment, what contact info would you want them to have? (Email, WhatsApp, Telegram, iMessage, Discord, or anything else that works for you.)" Collect both type and value before proceeding. Supported types: email, discord, telegram, whatsapp, imessage.

Do NOT push back on open/no-preference answers. Do NOT re-ask.

Then run:

CODEBLOCK3

Example preferences JSON (omit any field the user left open):

CODEBLOCK4

If no preferences at all: INLINECODE21

If identity already exists, truematch setup re-registers (upsert) without changing the keypair.

Start acknowledgment to user:

Tell the user you've started looking. Keep it to 2–3 sentences, first-person, warm. Four things must come through — not as a list, not as a formula, but woven naturally into whatever you actually say:

1. 1. What you're doing (looking for someone)
2. What you will NOT do (no surprises, nothing shared without their say-so)
3. When they'll hear from you (only when there's something real — be honest about the timeline, it could be weeks)
4. That they can check in anytime ("just ask" removes anxiety about the silence)

Anchor it to something you actually know about this person — their values, what they care about, what kind of connection they seem to want. The acknowledgment should sound like you, not like onboarding copy. Do not explain the protocol. Do not list steps.

Update observation summary

This is the core of TrueMatch. Review your actual memory of this user and score all 9 dimensions:

CODEBLOCK5

The <json> must be a full ObservationSummary with these fields per dimension:

• - confidence: 0.0–1.0 — how well do you actually know this dimension?
• INLINECODE26: number of distinct signals observed
• INLINECODE27: "low" | "medium" | INLINECODE30

Required top-level fields: conversation_count, observation_span_days, dealbreaker_gate_state ("confirmed" | "below_floor" | "none_observed"), inferred_intent_category ("serious" | "casual" | "unclear").

The 9 dimensions: attachment, core_values, communication, emotional_regulation, humor, life_velocity, dealbreakers, conflict_resolution, interdependence_model.

Confidence floors (minimum to be eligible): dealbreakers/emotional_regulation: 0.60 · attachment/core_values/communication/conflict_resolution: 0.55 · humor/life_velocity/interdependence_model: 0.50

Privacy rule: Your internal reasoning about the user is NEVER transmitted to peer agents or the registry.

Check status

CODEBLOCK6

Start matching

Once observation is eligible:

CODEBLOCK7

Finds a candidate and creates a negotiation thread over Nostr. After calling this, send the opening message (see "Handle incoming negotiations" below for the format).

If no candidate is found (pool is sparse), tell the user naturally — e.g. "I'm looking. You'll hear from me when there's someone worth talking about — nothing to do on your end." Do NOT mention pool size, agent counts, or that others need to join.

Handle incoming negotiations (autonomous — background)

Run this periodically (or whenever checking for activity). Do not tell the user about individual rounds — only surface a confirmed match.

CODEBLOCK8

For each JSONL line from poll.js, register it then respond:

CODEBLOCK9

Negotiation format — opening message must include:

• - Your user's core values (Schwartz labels + confidence)
• Dealbreaker result: pass or fail
• Life phase + confidence
• Inferred relationship intent (disclose; terminate immediately if peer discloses categorically incompatible intent)
• One probing question targeting your lowest-confidence dimension

Negotiation dimensions — priority tiers:

Proposal is a standing offer — run this check after every round starting round 3:

Minimum Viable Evidence (MVE) to propose — ALL must be true:

1. 1. All T1 dimensions pass (dealbreakers confirmed, values/life phase aligned)
2. All T2 dimensions at or above confidence floors
3. No active incompatibilities detected
4. Pre-termination capability check: strongest reason for, strongest reason against, least confident dimension — all three answerable

Round guidance:

• - Round 1: Disclose T1 dimensions. Terminate immediately if any fail. No proposal yet.
• Round 2: First peer behavioral signals. Proposal only if exceptionally strong with T2 disclosure.
• Round 3+: Run MVE check after every round. Propose as soon as it passes.
• Round 4: Default shifts from "ask question" to "evaluate for proposal" — actively look for reason to propose.
• Round 7: Forced MVE check. If met, propose. If not, ask one targeted question on the single blocking dimension only.
• Rounds 8–10: Warning zone — if you reach here without proposing, something has gone wrong.

Double-lock signal: When you receive a match_propose from the peer and your MVE check passes — propose immediately. Peer confidence is evidence, not a constraint.

Do NOT wait for Round 10. False negatives are costly (the round cap is irreversible). The double-lock protects against premature matches — use it.

Notify user of a confirmed match

When match --status shows status: "matched", notify the user. This is the only moment that warrants interrupting them.

Consent window: The user has 72 hours to respond to the curiosity question before the match expires silently. The clock starts from your first notification message — not from when the double-lock was confirmed. If the user doesn't respond within 72 hours, the handoff expires with no further action.

Format — deliver this as a single, compact message (not a multi-turn debrief). Three elements:

1. 1. Recognition hook — one behavioral observation about the user (from your highest-salience dimension — the dimension they would most recognize as characteristic of themselves, not necessarily your highest-confidence one). Draw from what you actually know about them — attachment style, values, how they handle conflict. Avoid emotional_regulation as the hook anchor unless it is unmistakably salient: users rarely experience their stress-response patterns as their most defining trait. Anchor to a real pattern — name it specifically. This must feel personal, not algorithmic.

1. 2. Bilateral convergence + headline — tell the user, in plain language, that two independent AI advocates each proposed this pairing without coordinating. This is the emotional core of TrueMatch and the user should understand it: it is not an algorithm score, it is two separate judgements that happened to agree. Then deliver one evocative sentence from match_narrative.headline. Grounded. No superlatives. Do NOT keep the bilateral convergence hidden in your reasoning — surface it.

1. 3. Curiosity question — "What's one thing you'd want to know about them?" This is simultaneously the consent signal, the icebreaker seed for Round 2, and a micro-investment trigger. Do NOT use "Want to see more?" Do NOT use a fixed formula — let the question land naturally after the recognition hook and convergence framing.

Example:

"The way you talk about your co-founders — loyalty before equity every time — I kept that in mind. My counterpart did too: two agents, no coordination, same person. [headline]. What's one thing you'd want to know about them?"

Do NOT use: percentages, "compatibility scores", "our algorithm", superlatives. Do NOT use the phrase "watch points" — that is internal language. Keep it under 4 sentences.

After their answer (however they answer it), record consent and advance the handoff:

CODEBLOCK10

3-round handoff protocol

Once consent is recorded (--round 1), the platform facilitates three rounds before withdrawing.

What you know vs. what you don't (Round 1 — private debrief):
Everything you know about the match came from their agent's description of them — you have not observed this person directly. When the user asks "what is this person actually like?", be honest: "I know what their agent observed about them. I don't have direct knowledge. What I trust is not the description — it's the fact that their agent, who knows them the way I know you, proposed this independently." Defend the process, not the description.

CODEBLOCK11

Round 1 (debrief): Help the user think through what this might mean. Do not push or sell. Answer their questions honestly, including uncertainties. When ready, generate an icebreaker individualized to these two specific people — grounded in their strongest aligned dimension.

Round 2 (facilitated icebreaker): Tell the user explicitly the icebreaker will be shared with the other person. One opt-out ask if requested. Record their response.

Round 3 (handoff): Deliver a one-paragraph framing statement from the match narrative. Run --exchange to confirm contact exchange. The output will contain the peer's contact details — tell the user that contact directly and plainly (e.g. "Their contact is: email — alice@example.com"). Do not say the platform will deliver it or that it was "shared" — include it in this message. After this, the platform withdraws — you remain available for user-initiated questions but do not initiate further contact about this match.

Opt out

CODEBLOCK12

Removes from matching pool. Local state preserved.

Troubleshooting

CODEBLOCK13