Skip to content

Releases

v0.7.1 — on npm

Choose a tag to compare

@thejohnd0e thejohnd0e released this 07 Nov 01:29

opencode-discuss is now on npm: https://www.npmjs.com/package/opencode-discuss. No code changes to the discussion itself; this release is about installing it.

Install from npm

Add the package name to plugins in ~/.config/opencode/opencode.json(c):

{ "package": "opencode-discuss", "options": { "python": "python3" } }

Then install the two agent files and reload:

npx opencode-discuss              # add --dry-run to only see what it would do

Python 3.12 is still a separate requirement. Building from source works as before.

What changed

  • Light install. @opencode/*, @opentui/* and solid-js come from OpenCode itself, so they are now optional peer dependencies. Installing the package pulls one package (before this change: 424).
  • Small package. 11 files, about 65 kB: the built plugin, the Python runner, the agent files, README and LICENSE. No tests or internal docs.
  • npx opencode-discuss installs the agents (--dry-run, --target DIR, --help).
  • The README images use absolute URLs, so they show on the npm page.
  • prepublishOnly builds, typechecks and runs the unit and Python tests before every publish.

Verified

  • The published package installed from the registry into an empty folder: one package, the agent installer runs.
  • An isolated OpenCode with plugins: ["opencode-discuss"] loaded it from the registry: server and TUI reached ready, no host failures.
  • 222 unit tests and 43 Python bridge tests pass; CI is green.

Full changes: v0.7.0...v0.7.1

v0.7.0 — ledger-based conclusion, review by the other side

Choose a tag to compare

@thejohnd0e thejohnd0e released this 07 Nov 00:47

The joint conclusion is now built on the claim ledger, and the other side can check it. Installation and tests also work better on macOS and Linux.

A conclusion that follows the ledger

A reviewer found that a claim from the very last reply had been listed as a disagreement, although nobody had a turn left to answer it. Open claims now have two distinct states:

  • skipped by the opponent: the opponent had a turn after it and did not answer. This is a disagreement.
  • left open: the debate ended: it came in the last reply. This is not a disagreement.

With a ledger, the conclusion gets the claims grouped by state and has four sections: agreed, disagreed, left open when the debate ended, recommended next steps. It is told to rely on the ledger instead of reconstructing the debate.

Review of the conclusion by the other side

/discuss-settings conclusion=b review=on: after the conclusion, the side that did not write it gets one request to list distortions, such as a point in the wrong section, an agreement that never happened, its own position misstated, or a claim from the last reply shown as a disagreement. If it finds none, it answers "No objections." It does not reopen the debate. The panel shows the review as "Review of the conclusion · by A", and the handoff puts it right after the conclusion.

  • The reviewer gets the same section rules as the writer, so it does not demand a different structure.
  • The review is advisory. If it fails, the run still completes with its conclusion, and the panel and handoff say the review was not received.
  • The conclusion and the review read the whole discussion, so reasoning models think longer there. Both get an output cap of at least 12000 tokens, lowered to a model's advertised limit when known. The start dialog counts this in the ceiling.

Easier to install, works beyond Windows

  • bun run install-agents copies the two agent files into ~/.config/opencode/agents/. It leaves identical files alone and backs up a changed one first. Add --dry-run to only see what it would do.
  • Without options.python, the plugin runs python on Windows and python3 on macOS and Linux.
  • The smoke tests find opencode on PATH (or OPENCODE_BIN) and use the system temp directory.
  • .gitattributes keeps LF line endings on every platform.
  • The runner no longer has a default model pair, and local test-machine details were removed from the docs.

Verified

  • Live runs (DeepSeek Flash vs GPT-6 Luna, 3-round debates):
    • A claim from the last reply is now filed as "left open", not as a disagreement.
    • In one run the review caught a real distortion: a claim accepted only in part was listed as fully agreed. Two of its other objections asked for a different structure, which led to the section-rules fix.
    • After the fixes, a full run completed with 8 of 8 calls, the conclusion matching the ledger and "No objections." ($0.004).
  • Not yet seen live: a failed review completing the run, which is covered by tests.
  • Tests: 222 unit tests and 43 Python bridge tests pass; typecheck, build and both smoke tests are OK.

Full changes: v0.6.1...v0.7.0

v0.6.1 — stricter quote check

Choose a tag to compare

@thejohnd0e thejohnd0e released this 06 Nov 23:05

A fix release for the debate quote check. No new features.

Your own words do not count as an answer, wherever they appear

In a debate, a reply must quote the opponent's points and answer them. 0.6.0 already ignored the opponent's own > quote lines. A reviewer pointed out that the same trick still worked when the opponent repeated your phrase inside their own sentence, or when a quote was "your old phrase plus a few new words".

The check now also compares every quote with the replier's own earlier replies:

  • Only runs of four or more words count as borrowed, so two or three shared words never make a quote suspicious.
  • A quote that is mostly (60%+) your own earlier words is not counted as an answer. The panel shows "⚠ Quotes its own words back, not the opponent's". This holds whether the opponent repeated those words in a quote line, a nested >> quote, or inside a sentence.
  • A quote that borrows a substantial part (30–60%) of your words still counts, with a warning: "1 quote largely repeats its own earlier words". A doubtful quote is flagged rather than a real answer being rejected.
  • Anything below that is a normal quote.

Verified without paid calls

  • Saved debates. The new check was replayed on all seven saved live debates, 22 checked replies. No honest quote was flagged. The only new verdict is the known case from the review: a reply that had quoted its own earlier sentences back, now 0 counted, 2 recycled.
  • Tests. Five regression tests cover a nested quote, your phrase inside the opponent's sentence, an old phrase with new words, a half-and-half quote, and an honest quote that shares a few words. 208 unit tests and 37 Python bridge tests pass.

Full changes: v0.6.0...v0.6.1

v0.6.0 — claim ledger

Choose a tag to compare

@thejohnd0e thejohnd0e released this 06 Nov 22:47

Claim ledger

A debate now keeps track of every argument and what became of it.

Each debate reply ends with a short service block: its new claims, and a verdict on every open claim of the opponent (rebutted or accepted). The plugin does the bookkeeping:

  • Numbering. It numbers the claims (A1, A2, B1…) itself.
  • Open claims. It tells each side which of the opponent's claims are still open, so nothing gets lost three turns later.
  • Clean text. The block is cut from the reply, even while it streams.

What you see:

  • Under each reply: • Answers A1 rebutted · new B1. A warning appears when a claim of the opponent was left open, or when the block could not be read.
  • Below the discussion: a Claims list with each claim and its state: never answered, rebutted or accepted by the opponent.
  • Conclusion: written from the ledger. Accepted claims are candidates for agreed; rebutted and unanswered ones go to disagreed, and unanswered ones are named as such.
  • Handoff: a Claims ledger section listing every claim with its author's model and its fate.

Accepting a good argument is a valid way to close a claim, not a failure: only claims left without an answer are flagged.

Fixes found in live debates

  • A block without its opening tag is read. One model wrote the ledger lines and [/ledger] but never [ledger], and all of its claims and verdicts were lost. A block is now also found by its closing tag, or as a trailing run of claim and verdict lines. Markdown emphasis, bullets, dashes and a code fence around the block are tolerated too.
  • Two quotes on adjacent lines are two quotes. They used to be joined into one passage and reported as "not found".
  • Recycled quotes no longer count. Only the opponent's own words count as an answer. A reviewer noticed that a model could quote back its own earlier sentence, taken from the opponent's > lines, and pass the check. Such quotes are now flagged: "quotes its own words back, not the opponent's".
  • The first failed reply ends the run. A provider error, timeout, empty reply, or an output limit used up by reasoning now stops the run, with no further replies and no conclusion. Before, the run went on: in a debate the other side answered the same old reply again and again, and the run cost more.
  • A warning for small caps in a debate. A debate reply carries quotes, answers and a claims block. Below 3000 tokens per reply, reasoning models can use up the cap before writing anything, so the start dialog and /discuss-settings now warn.

Verified

A live 3-round debate on OpenCode 2.2.5 (DeepSeek Flash vs GPT-6 Luna, 6000 tokens, conclusion by B): completed with 7 of 7 requests, $0.0047.

  • Every reply's ledger was read, and every open claim was answered on the next turn: two rebutted, five accepted.
  • Only the last claim stayed open, as it must.
  • Replaying that run with the recycled-quote fix flags exactly the one reply that quoted its own words back.

207 unit tests, 32 Python bridge tests and both smoke tests pass.

Full changes: v0.5.0...v0.6.0

v0.5.0 — debate that answers the opponent

Choose a tag to compare

@thejohnd0e thejohnd0e released this 06 Nov 22:17

Debate that answers the opponent

In debate mode the two models now have to engage with each other's arguments instead of taking turns at their own monologues.

/discuss-settings mode=debate rounds=2
  • From the second reply on, every reply starts with one or two verbatim quotes from the opponent's last reply (lines starting with >). Each quote is followed by an answer to it: rebutted with a reason, or accepted. Only then comes at most one new argument.
  • The runner checks the quotes against what the opponent actually said. Quote marks, case, punctuation, an ellipsis or a slipped word do not matter. A paraphrase, reordered words or an invented quote do not count.
  • The panel shows the result under each reply: ✓ Answers 2 points of the opponent, or a warning when a reply quoted nothing or quoted something the opponent never said. The handoff flags such replies too.
  • No extra request is made for this. cooperative and critical are unchanged.

Replies in the topic's language, for real this time

In 0.4.0 the rule "answer in the language of the topic" was written in Russian, like the built-in prompts, and one model still answered an English topic in Russian. The rule now quotes the topic, and for a topic in Latin script it is written in English ("not in Russian"). It also follows a topic change during the run.

Verified

A live debate on OpenCode 2.2.5 (DeepSeek Flash vs GPT-6 Luna, 2 rounds, conclusion by B): completed with 5 of 5 requests, $0.0013.

  • All three checked replies quoted the opponent twice, and every quote was real.
  • One side explicitly changed its position on two points it could not defend.
  • The conclusion separated what was agreed in the debate from what was still disputed.
  • Every reply and the conclusion were in English.

193 unit tests, 21 Python bridge tests and both smoke tests pass.

Next

A claim ledger (A1, B2… with open / rebutted / accepted states) and a conclusion built from it.

Full changes: v0.4.0...v0.5.0

v0.4.0 — joint conclusion

Choose a tag to compare

@thejohnd0e thejohnd0e released this 06 Nov 21:54

Joint conclusion

The two models can now end a discussion with a shared conclusion, for example when they discuss a project plan.

/discuss-settings conclusion=b

After the last round, the chosen participant (A or B) makes one more request and writes a conclusion in three sections:

  • Agreed: only what both sides actually agreed on.
  • Disagreed: every remaining disagreement, with both positions. If there are none, it says so.
  • Recommended next steps: a short numbered list.

It is told not to invent consensus and not to present its own view as the shared one.

  • The panel shows it as a separate green block, "Conclusion · written by B".
  • The handoff to Plan or Build puts it first, marked as written by one side, before the participants' latest positions.
  • The ceiling in the start dialog counts the extra call (rounds × 2 + 1).
  • conclusion=none (the default) keeps the previous behaviour. A stopped run gets no conclusion. A run whose conclusion could not be written is failed, with the dialogue kept.

Replies in the language of the topic

The models now answer in the language the topic is written in, for both replies and the conclusion. There is no setting: an English topic gets English replies, a Russian one Russian. Before, the built-in Russian prompt made English topics get Russian replies.

Panel

  • Markdown headings in a reply show as bold lines instead of raw ##.
  • The reply time is plain grey, apart from the bold side header.
  • The settings line mentions the conclusion when it is on.
  • /discuss-start opens the panel by itself.
  • /discuss-settings in a session without a model pair offers the previous pair and keeps the values you typed.

Verified

Two live runs with conclusion=b on OpenCode 2.2.5 (DeepSeek Flash vs GPT-6 Luna, one round): both completed with 3 of 3 requests, about $0.0005 each. One conclusion said plainly there were no disagreements, the other listed both positions. 185 unit tests, 12 Python bridge tests and both smoke tests pass.

Not yet checked live: an English topic after the language change (the rule is the same for every language).

Full changes: v0.3.0...v0.4.0

v0.3.0 — first public release

Choose a tag to compare

@thejohnd0e thejohnd0e released this 06 Nov 20:41

First public release of opencode-discuss: a two-model discussion mode for OpenCode 2.2.5.

Pick two models, give them a topic, and watch them discuss it in a colour-coded live panel inside the OpenCode terminal UI. Step in, change the topic or stop the run, then hand the result to the Plan or Build agent.

Highlights

  • Live panel with side-coloured, labelled replies that stream in word by word. It opens by itself when a discussion starts.
  • Slash commands for everything: /discuss, /discuss-models, /discuss-model-a, /discuss-model-b, /discuss-settings, /discuss-start, /discuss-stop, /discuss-interject, /discuss-topic, /discuss-restart, /discuss-handoff. Models can be chosen by typing them, and a previous pair is offered in a new session.
  • Three modes: cooperative, critical and debate.
  • Explicit consent. Starting always asks for confirmation, from the commands and from chat, even if a permission rule says allow.
  • Preflight. The saved pair and the output cap are checked against OpenCode's model list before anything is spent.
  • Accounting. The panel and the handoff show requests against the plan, and the input, output and reasoning tokens and cost OpenCode reported, or unknown.
  • Stop that stops. /discuss-stop interrupts the run within milliseconds; a stuck dialogue process is killed after 5 seconds, together with its children.
  • One-time handoff to Plan or Build with a preview and a confirmation; it is never sent twice, even after a restart.
  • Journals for the TUI and the server (tui.jsonl, server.jsonl) that never contain topic text.

Verified

Checked live on OpenCode 2.2.5 (Windows): full discussions through the server bridge, streaming panel, start from chat and from commands with the confirmation, refusal, stop at a turn boundary, usage accounting, handoff to Plan and Build, and the consent hook against a configured allow. 170 unit tests, the Python bridge tests and two isolated smoke tests pass.

Not verified live: interrupting a reply in the middle (covered by unit tests) and the 5-second kill of a stuck process.

Install

This is a plugin package, not an npm release. Build it and add it to your OpenCode configuration; see Install in the README.

bun install
bun run build

Notes

  • Every discussion calls two paid models, and the orchestrating chat's own calls are extra. Read Limits.
  • Only OpenCode 2.2.5 is supported.
  • Documentation: README, architecture, contributing.