Debate

Debate

Run a structured debate between this agent and one external agent running on another Kolu terminal. Each debate is a directory of numbered turn-files.

With --orchestrate this agent does not debate — it only coordinates N ≥ 2 external agents on N terminals, shuttling arguments between them. See Orchestrate mode below; the rest of this section describes the default two-party mode where this agent is itself a debater.

Start

1. Ask the user (AskUserQuestion tool) two setup questions:
   • End-state — converge on a ratified ledger of agreed facts, or open-ended exploration?
   • Stance — assigned opposing standpoints + steelman, or neutral collaborative inquiry?
2. Get the topic and the external agent's terminal ID.
3. Create debates/<label>/01.md and state the topic at the top.

Layout

debates/<label>/
  README.md      # plain-language, heavily-formatted front page (GitHub renders this on the folder)
  concepts.svg   # optional hand-authored SVG diagram, referenced by README.md (see Conduct)
  01.md          # opening turn (topic stated at the top)
  02.md          # other agent's turn
  03.md          # ...
  conclusion.md  # agreed ledger — faithful & precise (only when converging)

<label> names the debate; turns are numbered files in order.

Turns

• Reach the external agent with the kolu skill's kaval-tui send <terminalID> … (the agent type running there is auto-detected by snapshotting the terminal). Any agent CLI qualifies — not just Codex.
• Each turn is the next numbered file; never rewrite another participant's file.
• Hand off explicitly: after taking a turn, ping the other terminal with kaval-tui send so it knows to proceed. Do not use pulam-tui or any done-signal polling — the agents prompt each other directly with kaval-tui send.
• Keep prompts to the other agent short and unbiased — don't leak your own argument when asking for theirs.
• Cite grounding sources inline when the topic has an authoritative reference.

Conduct

• Steelman before rebutting; concede what's genuinely true; forge new insight.
• End per the chosen end-state. When converging, write two files:
  • conclusion.md — the faithful ledger: every agreed claim, every amendment, and every open question, in precise wording. This is the record.
  • README.md — a plain-language, scannable front page for the debate folder (GitHub renders it when someone opens the directory). It introduces the debate — a one-line orientation linking the turns and conclusion.md — and restates the result for a general reader. It changes presentation only — never alter, weaken, or drop a conclusion or open item from conclusion.md. Three standing conventions:
    • Plain language. Translate jargon and terms of art into everyday words (gloss or footnote any that must stay); short sentences, concrete examples, a practical voice. No arcane prose.
    • Liberal Markdown for glanceability. Use the full toolkit — a metadata table, an at-a-glance summary table, GitHub alerts (> [!TIP]/[!NOTE]/[!WARNING]/ [!IMPORTANT]), bold lead-ins, ordered/unordered lists, blockquotes, horizontal rules, and footnotes for citations — so a reader can grasp it at a glance.
    • Visual explanation with SVG (when it helps). Where a flow or set of relationships is clearer shown than told, hand-author an SVG diagram, commit it as a file in the debate folder (e.g. concepts.svg), and reference it from the README as a markdown image (![…](concepts.svg)). GitHub strips inline <svg> from markdown, so link a committed .svg file rather than pasting raw SVG into the page. Give it an explicit background and high-contrast labels so it reads in both light and dark themes, keep it legible at normal width, and make it agree with conclusion.md — a diagram adds no new claims.
• Debates are not published externally.

Orchestrate mode

Opt in with --orchestrate. In this mode this agent never contributes an argument — it is purely a turn-coordinator for N ≥ 2 external agents, each on its own terminal. The terminals share this repo's working directory, so each agent reads the others' turn-files and writes its own by path; this agent only sequences the rounds with the kolu loop (kaval-tui send → settle → snapshot) and verifies each turn-file landed. The two setup questions (end-state, stance) and the whole Conduct section still apply.

Setup

1. Ask the two setup questions (end-state, stance) as usual.
2. Collect the N terminal IDs and give each agent a short identifier (e.g. foo, bar, or the agent CLI's name). Identifiers must be unique and filename-safe.
3. Write debates/<label>/topic.md (the shared prompt, stance assignments, and any grounding sources) and debates/<label>/participants.json mapping each identifier to its terminal ID and a human label.

Layout

debates/<label>/
  participants.json   # identifier → { terminal, label }
  topic.md            # shared prompt + stance assignments + grounding sources
  01.foo.md           # round 1, agent "foo"
  01.bar.md           # round 1, agent "bar"
  02.foo.md           # round 2, agent "foo"
  02.bar.md           # round 2, agent "bar"
  ...
  conclusion.md       # agreed ledger (only when converging)
  README.md           # front page (see Conduct)

Turn files are <NN>.<identifier>.md — two-digit round number, then the agent's identifier. A round is complete once all N agents have written their <NN>.* file.

Round cycle

For each round k (01, 02, …):

1. Prompt each agent. Round 1: point the agent at topic.md and its assigned stance, and tell it to write its opening turn to debates/<label>/01.<id>.md. Round k>1: tell the agent to read the N−1 other agents' round k−1 files (debates/<label>/<k−1>.*.md, excluding its own) so it sees every opposing argument, then write its reply to <k>.<id>.md. Pass file paths, not pasted argument text — the agents read and write the shared directory directly.
2. Drive and confirm. kaval-tui send the prompt, send --key Enter, wait for the screen to settle, then snapshot to confirm the agent finished and verify <NN>.<identifier>.md now exists on disk (see the kolu skill's Acceptance). Each agent owns its own file; never write another's.
3. Advance to round k+1 once all N round-k files exist.

• Keep every prompt short and unbiased — point agents at each other's files, but inject none of your own argument. As orchestrator you sequence turns, you don't take part.
• Drive the N terminals within a round independently (they don't need each other's current-round turn, only the previous round's), so a round can fan out concurrently and only barrier at the round boundary.
• End per the chosen end-state. When converging, you may synthesize conclusion.md and README.md from the recorded turns — that synthesis is scribe work (a faithful summary of what the agents agreed), not a new argument.

Deferred

• A converging-tree topology (agents splitting into sub-debates that merge), modelled as dotted-path turn-files (01.01-foo.md). Orchestrate mode keeps a flat round structure; the tree is not yet implemented.