debate

One AI agent writes the code. A second agent — from a different company, in a different app — checks it. They talk by taking turns in two text files, and you can read every word.

Zero dependencies. Two files. One rule: nobody posts out of turn.

What is this?

You have Claude Code in one terminal. You have a second AI agent somewhere else — a different vendor, a different tool, maybe a different machine. You'd like one of them to write code and the other to review it, the way two developers review each other's pull requests. The bet this tool is built on is that an AI reviewer from the same vendor shares too much of the builder's training to be a real second opinion — that a second opinion is only a second opinion if it comes from somewhere else. That is a hypothesis, not a measured result. It is being tested in a pre-registered study, which is underway and has no results yet; when it has, this README will report them either way.

Problem: those two agents can't talk to each other. There is no shared API between vendors, and the AI subscriptions you already pay for only work inside each vendor's own app. So in practice you become the messenger, copy-pasting between two windows.

debate fixes this with the simplest thing that could work: a shared mailbox made of two text files in a folder both agents can reach (a git repo is perfect — the history becomes your audit trail).

• <channel>.channel.md is the conversation. Messages are only ever added, never edited or deleted, so it doubles as a complete record of who said what, when.
• <channel>.signal.json is the doorbell: five core fields that say whose turn it is and which discussion is open. Brokered cases also persist their phase, absolute deadline, and typed terminal result there.

<channel> is the channel's own name — debate init generates it once, as <label>-<NNNNN> (the label defaults to your repo's directory name; five random digits make two channels unable to collide). Because every file carries the name, several channels can share one folder, and a message can never land in the wrong project's record. (init also drops <channel>.debate.json next to them — party names, settings, and the project the channel serves. That one is configuration, not conversation; the mailbox is the two files above. Channels created by 0.3.x use the older fixed filenames CHANNEL.md/signal.json/debate.json — still fully supported; see debate migrate.)

One channel writer is the only code that writes to either file — whether called by the legacy debate post surface or the brokered controller — and it enforces the rules instead of politely asking: you can't post out of turn, you can't open a second discussion while one is open, and a runaway back-and-forth gets cut off by a message cap. A small scheduled job wakes whichever agent the doorbell points at. No server, no message broker, no API keys, no framework to adopt.

What a review looks like

After one round trip, the mailbox reads like this:

## MSG-12 | 2026-07-06T14:02:11+00:00 | from: claude | type: review-request | thread: retry-backoff | refs: retry-backoff@4e9f21c

Please review branch retry-backoff at 4e9f21c: adds exponential backoff to the
HTTP client. 14 new tests. The part I'm least sure about is the jitter range.

## MSG-13 | 2026-07-06T14:07:48+00:00 | from: glm | type: verdict | thread: retry-backoff | refs: retry-backoff@4e9f21c

APPROVE — checked out 4e9f21c and ran the suite myself: 87 passed. Verified the
backoff caps at 60s and jitter cannot go negative. Non-blocking nit: the retry
log line prints the attempt number twice.

## MSG-14 | 2026-07-06T14:11:02+00:00 | from: claude | type: close | thread: retry-backoff | refs: retry-backoff@4e9f21c

Nit fixed in 5a01d33, merged. Closing.

Every message has a sequence number, a sender, a type (review-request, verdict, fix-report, question, info, close), a thread name, and refs — the exact branch-and-commit it talks about, so claims are checkable. Note the reviewer re-ran the tests itself and said so. That culture is configured in the prompts; the format that makes it auditable is enforced by the tool.

Try it

pip install debate        # Python 3.10+, stdlib only — or just vendor the two modules

# Create a managed mailbox: two headless agents named claude and glm, plus you as supervisor.
# Prints the generated channel id, e.g. 'myproject-48213' (--label overrides the prefix).
# New channels record managed_version 1 and default to a 12-entry thread cap.
debate init --root ./collab --parties claude,glm --supervisor owner

# The builder asks for a review:
debate post --root ./collab --channel myproject-48213 --from claude --type review-request \
    --thread feature-x --refs feature-x@abc123 \
    --body "Please review commit abc123: ..."

# The reviewer answers (its own tool/app runs this after reading the thread):
debate post --root ./collab --channel myproject-48213 --from glm --type verdict \
    --thread feature-x --refs feature-x@abc123 \
    --body "APPROVE — verified: 27 tests pass at abc123."

# Whoever acted last closes the thread:
debate post --root ./collab --channel myproject-48213 --from claude --type close \
    --thread feature-x --body "Merged. Closing."

Try posting twice in a row from the same party: the tool refuses. That refusal is the protocol.

Running it unattended

There are two recorded managed versions. Version 2 is the brokered path for new isolated gates. Version 1 is retained so existing two-command channels keep working, but those agents receive the channel path and self-post with --from; it does not provide sender binding or context isolation.

Brokered managed version 2

Initialize the channel explicitly as brokered, then fill in watcher.brokered.example.json. Party names are arbitrary; the two adapters keys must exactly match the addressed channel.

debate init --root ./collab --parties party-a,party-b --supervisor owner --brokered

# Use the generated id and a full, already-existing 40-character commit SHA in the config.
# This validates topology, cost mode, profile hashes and timing without invoking a model.
debate adapter-doctor --root ./collab --channel <id> --config watcher.json

# The controller snapshots the case and posts a neutral supervisor docket. The first-seat
# choice controls completion order only; neither sealed input contains the other result.
debate broker-open --root ./collab --channel <id> --config watcher.json \
  --thread feature-x --first-seat party-b --refs feature-x@<sha> \
  --body-file review-docket.md

debate watch-once --root ./collab --channel <id> --config watcher.json

# After a fix, update source_ref/docket files in the config and snapshot them before
# another seat runs. This supervisor entry does not steal or change the party turn.
debate broker-revise --root ./collab --channel <id> --config watcher.json \
  --thread feature-x --refs feature-x@<new-sha> --body "Revision ready after fixes."

Each profile records its provider/model, reasoning setting, CLI version, authentication mode, cost mode, permission policy and relationship to the artifact author. Exactly one independent seat is the minimum two-agent topology: the interactive author's fresh headless seat is honestly labeled an isolated author-affiliated self-review, while the opponent is independent. The interactive session never fills a turn itself. Two independent seats are the recommended three-agent topology, where the interactive author/controller is outside both debate seats. The core never infers either topology from names such as Opus, Codex, GLM or Kimi.

The model pair is configuration, not policy. A GPT-5.6 Sol author can, for example, use headless Opus 5 plus a smoke-approved GPT-5.6 Terra profile so neither reviewer is the interactive author. Another channel can use Opus/GLM, GLM/Kimi, or local models without a code change. A Kimi controller can oversee separate Opus/Codex and Opus/GLM channels, but each remains a two-seat debate with its own explicit channel id. Every example must record the author relationship and the requested and resolved model identities; changing any of those, the effort, permissions, authentication, or cost mode opens a fresh case.

adapter-doctor prints every seat's authentication and cost mode before it invokes anything and explicitly reports that the check incurred no charge. Subscription seats have no per-call API line item but still consume plan quota; API seats are metered by their provider's current input/output-token prices; local seats consume host compute. Debate does not estimate a dollar total because CLI prompts, cache discounts and provider prices vary. Confirm the printed mode and the provider's current price before the first smoke.

expected_runtime_model is optional because some CLIs do not expose a stable resolved ID. When omitted, Debate still records the returned runtime identity but cannot refuse silent model substitution; configure it whenever the adapter can report a falsifiable exact ID.

Before the neutral docket is posted, the controller creates two separate read-only exports of the complete tracked repository at the pinned commit. collab/, var/ and .git are separated; tracked project settings stay present as evidence but live settings sources are refused. Each seat gets a clean project-local HOME/cache/temp area, an allowlisted environment, a Git discovery ceiling, the immutable docket revision, and a controller-owned result path. Gitignored cited files such as a plan or watcher.json are materialized and hashed separately. Stdout/stderr are diagnostics only; the result must be schema-versioned JSON, may not contain sender, and is posted under the bound seat by the controller. broker-revise maintains a content-addressed revision chain and blocks invocation if a revision was only half-recorded; verdict provenance therefore never points only at a mutable gitignored filename.

Every verdict result also carries a typed decision of PASS or NO_PASS:

{
  "schema_version": 1,
  "entry_type": "verdict",
  "decision": "PASS",
  "body": "APPROVE - fresh export run: 412 passed.",
  "runtime_model": "resolved-model-id"
}

The case state advances through docket -> sealed -> reveal -> deliberation -> terminal. Both initial results are kept outside the shared record until they exist, then published by one atomic mailbox replacement with each private capture timestamp recorded in its reveal provenance. A crash after that replacement but before the doorbell update is repaired idempotently without duplicating either position. After reveal, each fresh seat sees only the current thread. Matching party votes close automatically as PASS or NO_PASS; PASS requires an agreeing author-independent seat. A supervisor verdict is visible context but never a vote. Thread-cap exhaustion closes NO_PASS, while adapter/retry/deadline failure closes ERROR; close_reason records why separately. The typed-close intent is persisted before its mailbox write, so the recurring scheduler can likewise distinguish and repair a close whose signal update was interrupted; unrelated mailbox-ahead anomalies still follow the normal fail-closed escalation path.

The runtime lives below <project>/var/debate/<channel>/, never .pytest_cache or another tool cache. adapter-doctor prints the unconstrained schedule estimate and the enforced whole-case deadline from the same timing calculation; adapter timeouts above 60 minutes or an absent deadline are refused. It also prints cost_mode before any future smoke can spend money. The absolute deadline spans sealed capture, reveal, deliberation, retries and process restarts. Every invocation is capped by its remaining budget, and an expired case is closed idempotently as ERROR / case-deadline-expired on the next scheduler tick.

Completed case exports are intentionally read-only and Debate never deletes provenance automatically. Project cleanup must first restore owner write permission within the exact completed var/debate/<channel>/ case directory, then remove only that validated directory.

This is strong protection against accidental contamination, not a claim that a same-user process is hostile-code safe. Read-only permissions, a clean environment, Git ceiling and canaries are mechanically checked; an isolation_mode: advisory profile can still read an absolute host path if the selected CLI/tool sandbox permits it. That includes the private sealed-submission state stored elsewhere below the same project-local case runtime: prompt separation does not stop a hostile same-user process from traversing parent directories. Use os-enforced only when an external sandbox actually denies those reads.

Managed version 1 compatibility

debate watch-once is one tick of a deliberately simple watcher. Put it on a schedule (cron, every minute): it checks the doorbell, prints any new messages to stdout — route that wherever you already look, a log file or a chat gateway — and, if it's an agent's turn on an open thread, starts that agent with a fixed, pre-written prompt from a config file:

{
  "state_path": "/somewhere/outside/the/channel/watcher-state-myproject.json",
  "commands": {
    "claude": ["claude", "-p", "{prompt}"],
    "glm": ["glm-agent", "{prompt}"]
  },
  "prompts": {
    "claude": "It is your turn on ./collab --channel myproject-48213. Read only the open thread with `debate read --root ./collab --channel myproject-48213`, act, post with the same explicit channel, then stop.",
    "glm": "It is your turn on ./collab --channel myproject-48213. Read only the open thread with `debate read --root ./collab --channel myproject-48213`, cite your own fresh evidence, post with the same explicit channel, then stop."
  },
  "debounce_seconds": { "claude": 0, "glm": 0 },
  "retry_seconds": 1800
}

Name the state file after the channel, not watcher-state.json. One channel gets one state file, and its stem is the channel's identity everywhere else: the watcher tags every log line with it, and the scheduler unit should be named after it too (debate-watch-<stem>). Since 0.4 the stem to use is the channel's own generated id — that is the first of the two edits debate migrate prints, and it makes units, state files, locks and journals all carry one identity. Two channels on one host that both take the generic default end up with colliding tags and colliding unit names, and telling their watchers apart goes back to reading /proc by hand — which is how a wrong-process kill happened here once.

*/3 * * * * cd /absolute/path/to/project && debate watch-once \
  --root /absolute/path/to/project/collab --channel myproject-48213 \
  --config /absolute/path/to/project/watcher.json

Agents run in the watcher's own working directory — cd to your project root before watch-once (as above), and under systemd or Task Scheduler set WorkingDirectory / "Start in" explicitly, or relative paths in your pinned prompts will resolve somewhere surprising.

When nothing changed, nothing runs — no model is invoked, no tokens are spent. A managed version 1 channel requires one command for each of its two recorded parties. If either is absent, or an open managed thread has no party turn, watch-status reports INVALID and exits 4; the watcher never represents that state as healthy or waits for a live human session. Configs without managed_version remain readable as legacy/manual history but must be reconfigured before managed unattended use. Managed version 2 instead requires exactly two brokered adapter profiles and refuses direct party posts. Headless seats normally use zero debounce.

Running to completion

Cron is for unattended operation. At the keyboard and just want the current review driven to its close? Run the same watcher in the foreground:

debate watch --root ./collab --channel myproject-48213 --config watcher.json --until-close

Same config, same safety rails — agents launch with stdin detached and a timeout (timeout_seconds, default 1800), a crashed or hung agent is reported and retried once, a stuck thread exits loudly (code 4) instead of spinning, and a kernel-enforced lock beside the watcher state file keeps a foreground watch and a cron watch-once from double-driving the same channel. debate status --stale-after 3600 exits 3 when a turn has been parked longer than an hour — put it wherever you already alert from.

One driver per channel: the scheduler OR a foreground watch, never both. watch holds that lock for its whole process lifetime, so while it runs, every scheduler tick is refused with exit 1 — and unless something reads exit codes, the channel just quietly stops being driven. Stop the scheduler while you drive a round by hand, or don't run watch at all. This is a real failure mode, not a hypothetical: a silent channel is indistinguishable from a quiet one until someone looks.

Is anything actually driving this channel?

debate watch-status --root ./collab --channel myproject-48213 --config watcher.json

Reports whether a watcher holds the lock — naming the holder's pid and working directory — when the channel was last ticked, and what it is waiting for. Exits 4 when a human should look. Run it before killing any debate process: ps cannot tell two channels' watchers apart, and killing the wrong one is a mistake that has been made here.

Housekeeping: the mailbox grows, agents shouldn't read all of it

The conversation file grows forever by design. Real numbers from the production channel this tool came from: 63 messages, 112 KB, in four days — an agent that naively reads the whole mailbox burns a quarter of its context window on history it doesn't need. Three commands keep that honest:

• debate read prints the open thread — an agent's working set is the open thread, never the whole file. --thread <slug> prints one thread (archives are searched too); --since <seq> prints only what's new. Put debate read in your agents' pinned prompts instead of "read the mailbox".
• debate compact is supervisor housekeeping, run occasionally: closed threads older than --keep-days (default 14) relocate verbatim to archive/<channel>-YYYY-MM.md, with a one-line index per thread in archive/<channel>-INDEX.md. Nothing is edited or deleted — the record moves house, and if your channel lives in a git repo, history keeps every byte anyway. --dry-run shows the plan first.
• debate post --verify-refs <repo> refuses a post whose name@sha citations don't resolve to real commits in that repo. This exists because of a real incident: a close message once cited a commit hash that was written down before the commit existed — wrong by construction, correction entry required. Machines are better at this check than authors are.
• debate verify reads the record back and reports whether it still agrees with itself: a repeated message number, or a mailbox holding entries the doorbell never rang for. Exits 0 when clean and 4 when something needs a human. A gap in the numbering is reported as information, not a fault — compaction relocates whole threads, so gaps are normal.

One channel carries one project

A channel records, at init, the absolute path of the repo it serves. From then on a post whose refs cite a commit (name@sha) that doesn't exist in that repo is refused, with the message naming both sides. This too exists because of a real incident: another project's code review was once conducted through this repo's channel — every turn-order rule passed, because no rule could object — and the interleaved record became unpublishable. The supervisor can --force a deliberate exception; channels created before 0.4 carry no binding and are not gated.

If a folder somehow ends up holding more than one channel, every command refuses and lists them; --channel <id> picks one. Guessing between channels is how a message lands in the wrong project's record, so nothing guesses.

Upgrading from 0.3.x: existing channels keep working unchanged. debate migrate --root <folder> renames one in place to the named layout — a pure rename, the record's bytes untouched (debate verify before and after proves it) — then prints the two edits you owe: the watcher's state_path stem and the scheduler unit name, which both take the channel's id. Writing new legacy-layout channels is no longer possible, and posting to existing ones is deprecated as of 0.5 — it still works, debate migrate is the supported path forward, and no removal date is promised.

What's enforced — and what isn't

Be precise about what this tool guarantees, especially before running agents unattended:

• Enforced, hard: turn order, one open thread at a time, reply types (verdict, fix-report) cannot open threads, message caps, the message format, and write ordering (the mailbox entry always lands before the doorbell rings, so a watcher can never read a half-written message). An agent that breaks these rules gets its post refused, not a warning.
• Broker-enforced for managed version 2: exact party/profile binding, at least one author-independent seat, a full pinned source export with no reachable parent Git store, immutable docket/profile/config hashes, clean environment and project-local runtime, controller-owned sender, schema-validated typed results, sealed paired reveal, deadline recovery, and automatic PASS/NO_PASS/ERROR terminal transitions.
• Advisory, soft: everything an agent does outside the mailbox. "Don't push to main", "don't touch the config" — if those live in a prompt, you are trusting the model to comply. The tool can force when an agent speaks. It cannot force what the agent says to be true.

That second bullet is not theoretical. The one time our unattended fallback fired in production, the agent did everything right — and also repeated a stale fact from the channel history as if it were current, because it never re-checked the repo. The fix is cultural and cheap: require verdicts to cite fresh evidence (commit hash + test count), keep merges in human hands, and give unattended sessions their own working copy. The full story is in the case study.

Rules we learned the hard way

Each of these is encoded in the tool or the shipped watcher, and each one was paid for:

1. Check for an open thread, not just the turn field — after a close, the turn field means nothing.
2. Invoke once per doorbell change — an agent that produced no reply gets one timed retry. Version 1 then escalates; version 2 records and closes ERROR. Two agents in a refusal loop would burn money forever; the cap and absolute deadline are the brakes.
3. Drive every managed turn — both parties have headless commands and normally zero debounce; a missing command is INVALID, never a human fallback.
4. The watcher's memory lives outside the shared folder — its state file is not part of the record and never collides with a fresh clone.
5. The supervisor can speak at any time without taking a turn — the human interjecting never breaks the agents' alternation.
6. The mailbox is the record — if it didn't happen in the channel file, it didn't happen. Corrections are new messages, never edits.
7. A brokered seat never self-posts — it receives no live channel path; the controller validates its result file and derives the sender from the configured seat.
8. Initial positions reveal as a pair — no party can anchor its first judgment on the opponent, while later disagreement deliberately becomes a real, current-thread debate.

Why not just…

• GitHub PRs and a review bot? Great if both agents live where your repos are hosted. debate needs no hosting service, works completely offline, and round-trips in seconds on a cron tick — no webhooks, no tokens. (If GitHub fits your setup, use GitHub.)
• A multi-agent framework (AutoGen, LangGraph, CrewAI)? Those orchestrate agents they own, inside one program. debate coordinates agents that nobody jointly owns — different vendors, different apps, different lifetimes — and leaves a human-readable paper trail as the primary artifact.
• A message queue? You'd be trading two greppable text files and git log for a broker you have to run. The paper trail is the point.

Limits, honestly

• Two parties by design. A review needs a builder and a reviewer; strict alternation between exactly two named agents (plus a supervisor who can always interject) is the feature. Getting N agents to agree is a different protocol.
• Polling, not push. The doorbell is made to be checked once a minute by cron. If you need sub-second latency, this is not your transport.
• The writer lock is advisory. post and compact serialize on a transient per-channel lock file (a crashed holder's lock is broken after 30 seconds), so two simultaneous posts cannot interleave — the second sees the first's thread open and is refused. But it only binds writers that go through the CLI. (compact's crash ordering can duplicate an entry across mailbox and archive; it can never lose one.)
• The record is tamper-evident, not tamper-proof. post refuses a message whose body or refs would forge an entry header — that is a real hazard, because quoting an earlier message is exactly how it happens by accident, not by malice. debate verify catches careless or automated tampering after the fact. But the mailbox is a plain text file: anyone who can write to it, and who uses the next message number correctly, produces a record that verifies clean. Detecting that needs per-entry signatures, which this tool does not have. Treat the record as an honest log among cooperating parties plus a guard against accidents — not as evidence against a determined forger with write access.
• Young. Extracted from a working production setup, generalized, and tested — but read the code before trusting it; it's about 4,900 lines including the CLI and the broker, with 395 tests as of this writing.

Where this comes from

This is the setup debate was extracted from — provenance, not prescription. This is not a design exercise — it's the generalization of a channel that ran real code-review cycles between two commercial AI ecosystems:

• The builder seat: Claude Code, Anthropic's terminal coding agent, running Fable 5 (their strongest model tier), on a flat-rate subscription.
• The reviewer seat: a GPT-5.5 agent on Hermes, Nous Research's open-source agent harness, authenticated through an OpenAI Codex subscription. Hermes matters here: it is not a chat window but a full agentic environment with its own scheduler, its own subagents, and a Telegram gateway — its cron ran the watcher, and every channel message was mirrored to the supervisor's phone by the same infrastructure.

No API key existed anywhere in the system. Two subscriptions, each valid only inside its own app, collaborating through two files in a repo. A typical review round-tripped in about five minutes, most of which was the reviewer independently re-running the test suite.

One way to read that setup: an orchestrator conducting another orchestrator. The top-tier model doesn't just answer reviews — it writes the specs and test contracts, and the Hermes-side agent executes them inside its own 24/7 infrastructure, then the roles flip for review. In the best run of that shape, the stronger model authored a spec-and-tests contract, the Hermes agent implemented it, and the result — one round trip, about ten minutes — was a 137× speedup on the function under contract. That figure is one anecdote from one function, measured on that project's own workload; it is offered as a description of the shape, not as a benchmark of anything. debate is the baton between the two conductors, and the score everyone can read afterwards.

The same shape fits whatever pair of ecosystems you already run. The origin above is one example, not the design. This repo's historical channel used a human-driven Opus builder and headless GLM reviewer; that append-only record remains under collab/ as incident and 0.4 provenance, but its commandless-seat scheduler is retired. The fresh repository-unattended-02750 channel selects headless Opus/Codex only in local config and uses the same vendor-neutral broker shipped here. Its record includes the repository's own end-to-end sealed/reveal/automatic-close proof. The first end-to-end proof is MSG-11..14: independent Opus and Codex capture timestamps, one paired reveal, and automatic PASS with no live-session or human intervention. The record has grown past it since — the same machinery's first production acts were gating this repository's own merge and the v0.5.0 release itself, and those cases are entries in the same file. Earlier MSG-1..6 adapter-integration attempts remain visible as bounded ERROR closes rather than being rewritten. The pinned profiles were Claude Code 2.1.223 / claude-opus-5 high and Codex CLI 0.146.1 / gpt-5.6-terra high, both author-independent and subscription-authenticated; the final CLIs reported a $0.355168 Opus usage-equivalent and 43,729 Codex tokens. Those numbers consume subscription quota and are operational evidence, not a promise of zero cost. Branch-gate verdicts separately cite the reviewer's own full export run rather than author-pasted evidence.

A GLM + Kimi pairing works the same way anywhere (see examples/glm-kimi.md — both seats verified live), and a local open-weight model can hold either seat, beholden to no vendor. If it can read files and run a shell command, it can hold up its end of a review.

The name

Parliamentary, not adversarial: strict turns, one motion on the floor at a time, and everything said is on the record. (If you arrived from the "AI safety via debate" literature: this is not the formal debate game — it's review correspondence with teeth.)

License

MIT