SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsround-table

round-table

Agents & models

View repository

Updated 10-17 20:48

About this skill

/round-table — Round Table

"Many bodies, one soul" — so one agent can seat many souls at one table, let them argue, and still hand the decision back to the human.

/oracle-prism wears souls one after another, inline. A Round Table seats them together: each soul is its own subagent, hears the others, pushes back, and finally votes with a confidence it has to defend. Bounded by design: 4 rounds, never 5.

Where the shape comes from. A 7-agent, 4-round, confidence-voted meeting that Boy (Pinyo) ran inside his Oracle family. Only its parameters survived — 7 agents, 4 rounds, a confidence vote at the end, reserved for expensive decisions. Mother Oracle rebuilt the round prompts, the 1–10 scale and the aggregation rules from those parameters. So this skill is spiritually faithful, not forensically reproduced: if better prompts turn up, replace prompts/ — don't defend them.

Files

round-table/
├── SKILL.md
├── prompts/round1-open.txt … round4-vote.txt   one template per round, {{slots}}
├── scripts/table.ts                            inspect (early stop) + vote (aggregate)
└── examples/adopt-a-pattern.json               worked example — must give 37/50

SKILL_DIR is the base directory printed above this skill.

Step 0 — Is this a Round Table question?

Suggest /oracle-prism instead unless all three hold:

  1. A decision, not a topic — something will be done differently depending on the answer.
  2. Canonical options, 2–4 of them (--options). If the user gave none, propose them and wait for a yes.
  3. Stakes — being wrong is expensive or hard to reverse.

"Should we …" phrasing is not consent to convene. Ask first.

Then find where the minute will live:

# Find oracle root — git toplevel that has CLAUDE.md + ψ/
ORACLE_ROOT=$(git rev-parse --show-toplevel 2>/dev/null)
if [ -n "$ORACLE_ROOT" ] && [ -f "$ORACLE_ROOT/CLAUDE.md" ] && { [ -d "$ORACLE_ROOT/ψ" ] || [ -L "$ORACLE_ROOT/ψ" ]; }; then
  PSI="$ORACLE_ROOT/ψ"
elif [ -f "$(pwd)/CLAUDE.md" ] && { [ -d "$(pwd)/ψ" ] || [ -L "$(pwd)/ψ" ]; }; then
  ORACLE_ROOT="$(pwd)"
  PSI="$ORACLE_ROOT/ψ"
else
  echo "⚠️ Not in oracle repo (no CLAUDE.md + ψ/ at git root). Writing to pwd."
  ORACLE_ROOT="$(pwd)"
  PSI="$ORACLE_ROOT/ψ"
fi
SLUG="$(date +%Y-%m-%d)_<short-slug-of-the-question>"
TABLE="$PSI/lab/round-table"; mkdir -p "$TABLE" "$PSI/inbox/decisions"
MINUTE="$TABLE/$SLUG.md"; STATE="$TABLE/$SLUG.json"

Step 1 — Convene: seat the souls (3 to 8)

Quorum is 3 — two is a debate, one a monologue. Default to 5 when the user names no roster. Every seat needs a soul read from somewhere real — never invented:

Seat kind Where the soul comes from
A persona file any path the user gives (resonance_<name>.md, a character sheet, a SKILL.md) — read it whole
An Oracle in your fleet if maw is installed: R=$(maw locate <name> --path) → $R/ψ/memory/resonance/<name>.md, else $R/CLAUDE.md's identity section. Not found → refuse the seat and say so
A real person (multi-DNA) one line — name + the single habit they bring — written or confirmed by the human

Print the roster with each soul's source, the question, the options and the aggregator (default weighted). Gather every fact the souls may use into question_context now — there is no tool use during the rounds.

Create $STATE in the shape of examples/adopt-a-pattern.json (id, question, options, aggregator, agents[] with name, soul, rounds[]) and start $MINUTE with the roster. Flush both after every round, so one flaky subagent never costs the earlier rounds. Persona files may be private — the minute stays in your ψ, not in a post.

Step 2 — Rounds: one subagent per soul, all in parallel

For each round, fill that round's template from $SKILL_DIR/prompts/ and launch one Agent call per soul, all in a single message. Each prompt carries the full soul text, the question, the options and the transcript so far. Subagents debate from soul + context only — no tools.

Round Template Each soul gets
1 OPEN round1-open.txt soul, question, context — hears nobody
2 DEEPEN round2-deepen.txt + every round-1 position; must challenge one by name
3 CONVERGE round3-converge.txt + rounds 1–2, + 2–4 one-line disagreement axes that you extract
4 VOTE round4-vote.txt + the full transcript; binding, 5 is forbidden

Parse each reply as JSON. Malformed or off-option → re-prompt that soul once; a second failure empties the seat and the minute says so. Append {round, position, confidence, self_update} to that soul's rounds in $STATE, and the round to $MINUTE.

After rounds 2 and 3, ask the inspector — don't eyeball it:

bun "$SKILL_DIR/scripts/table.ts" inspect "$STATE"
Output Do
EARLY-STOP everyone agrees at ≥ 8 — go straight to round 4
SKIP-ROUND-3 nobody moved in round 2 — confirm no new evidence arrived, then vote
EXPERT-DEFERENCE one soul at 10, the rest ≤ 3 — ask the human whether to short-circuit
CONTINUE run the next round

Step 3 — Vote: the script counts, not the model

Write each soul's round-4 answer into final (vote, confidence, rationale) and dissent_note, then:

bun "$SKILL_DIR/scripts/table.ts" vote "$STATE"
Exit Meaning Do
0 decided print the decision block
2 RE-PROMPT — a 5, an off-option vote, a bad number re-ask only the named souls, vote again
3 ESCALATE write the deadlock in plain words; the human breaks it
1 bad input, or fewer than 3 souls fix $STATE

What the script enforces:

  • weighted (default) sums confidence per option · majority counts heads · unanimity-or-escalate needs every soul on one side at ≥ 7.
  • Ties break by highest single confidence, then fewer dissent notes, then the human — never an automatic pick.
  • Escalates under every aggregator when the winner holds < 60% of the weighted confidence, when 10s land on opposite sides, or when a dissent cites a value (honesty, safety, consent — override with values_keywords in $STATE). The human decides the hard ones.

Step 4 — The minute, and handing it back

Close $MINUTE with the vote table (soul · vote · confidence · rationale) and the script's block verbatim, e.g.:

DECISION: adopt-modified
Weighted confidence: 37/50 (74%) for adopt-modified
Top dissent: none (no opposing votes)
Top hedger: brand (6) — Compromise — would prefer as-is, but unity matters more

Write one line — the decision and the minute's path — to $PSI/inbox/decisions/$SLUG.md, so /recap finds it. Then:

# announce-mode → absolute path (no ψ/, no ~/, no $VAR, no ...).
# Use: echo "marker: $RESOLVED_PATH"  — bash substitutes; never print $VAR literally. See CONVENTIONS.md.
echo "🪑 Minute: $MINUTE"
echo "📥 Decision: $PSI/inbox/decisions/$SLUG.md"

And step out of the table:

  • Every seat was a lens. Write "the finance lens voted…", never "Finance decided…". The souls were worn by the convener; a seat is not that Oracle's or that person's real view. For a real view, ask them (/talk-to if you have a fleet).
  • A decision is advice. The human ships, merges and signs — the table only recommends.

Self-check

bun "$SKILL_DIR/scripts/table.ts" vote "$SKILL_DIR/examples/adopt-a-pattern.json"

Must print DECISION: adopt-modified and 37/50 (74%) and exit 0 — the result the reconstructed spec predicts for its worked example. Anything else means table.ts drifted.

Not in scope

No tool use mid-round · no soul summoning another agent · no human interjection mid-round · no round 5.