docs: fix consensus quorum, Gemini transport and flow diagram drift
Pull request #298, reviewed by the code-reviewer expert on or:qwen3-8-max, or:gpt-6-1-sol and or:gemini-3-8-flash. The change is antonbabenko/deliberation/pull/298.
converged after 2 rounds
Recorded 2026-11-18 on Deliberation 4.6.1, run e2456d79, tool consensus-step. Reviewer text is verbatim, so it can disagree with itself and with the verdict line the engine read.
The prompt
Question
Should this change ship as is? Check that every claim it adds matches the code, regressions, missing tests and doc drift.
Sent to every reviewer
Full text, verbatim (31,353 characters)
Code review: pull request #298 of the Deliberation repo (merged as 4bf190b on master), "docs: fix consensus quorum, Gemini transport and flow diagram drift".
Question: should this change ship as is? Check that every claim it adds matches the code below, regressions, missing tests and doc drift (AGENTS.md: a user-visible behavior is not done until every surface that describes it is current - README.md, TECHNICAL.md, SETUP.md, docs/tool-guide.md, docs/dev/*, config/config.schema.json, and commands/ plus skill prose).
PR description (verbatim):
- State the real convergence rule: at least consensus.quorumFloor responding reviewers (default 2, capped at the selected panel size), every one approving, no accepted critical issue, host APPROVE.
- Compare pages: Gemini defaults to a GEMINI_API_KEY; agy is opt-in.
- consensus.maxWallMs also applies to the consensus-step path.
- Redraw both consensus flow diagrams without the removed Stage 2 cross-review step and re-render the PNGs at the same sizes.
The diagram sources (assets/consensus-flow*.html) and their PNGs are left out below; they changed 252 lines of SVG markup and two binary files.
=== Diff (text files) ===
diff --git a/README.md b/README.md
index ea91881..ed21224 100644
--- a/README.md
+++ b/README.md
@@ -74,8 +74,8 @@ on their own:
- Cross-vendor arbitration. GPT, Gemini, Grok, local models (Ollama, LM Studio), and any
OpenAI-compatible model through OpenRouter (400+ more, including Qwen, Kimi, and DeepSeek)
review the same input. The arbiter is your host agent or any configured provider. For either
- arbiter, a round converges only when at least one external reviewer answered and every
- external reviewer that answered approves.
+ arbiter, a round converges only when at least `consensus.quorumFloor` external reviewers
+ answered (default 2, or the whole panel if it is smaller) and every one that answered approves.
- A blind first verdict. In `/consensus` your agent records its own verdict before it sees the
panel, so it cannot reshape its opinion to match the others.
- Disagreement that stays visible. Every dismissed issue keeps a written reason in the report,
@@ -428,7 +428,7 @@ This section uses the Claude Code command names, and "Claude" means the host age
`/consensus` is the heavy one. Same parallel dispatch, but with a peer-review round and a multi-round loop that stops only when the models agree. The cost: the orchestrator (Claude) writes the review prompt, casts a vote, decides which objections are real, and runs the loop. Left alone, that setup can quietly rubber-stamp its own plan. Four guards stop that.
-
+
[See the detailed diagram with bias guards and per-model flow](assets/consensus-flow.png)
@@ -436,7 +436,7 @@ The four guards:
- **Blind verdict.** Claude posts its own verdict (APPROVE / REQUEST CHANGES / REJECT) in a message sent *before* the one that calls the panel. The pre-commitment sits in the transcript, so Claude cannot reshape its opinion after seeing the others. The engine enforces this: the panel is not revealed until the blind verdict is recorded.
- **Peer review.** Each external model reviews the plan independently and returns a verdict plus categorized critical issues; Claude weighs them as the arbiter. The models vote, Claude adjudicates.
-- **No self-approval.** A round converges only when every responding external approves and at least one external actually answered. Claude's own approval never carries a round by itself. A provider that errors (an unconfigured Grok returning `missing-auth`, for example) drops out of the count instead of jamming the loop.
+- **No self-approval.** A round converges only when every responding external approves and at least `consensus.quorumFloor` externals actually answered (default 2; a panel smaller than that needs every member to answer). Claude's own approval never carries a round by itself. A provider that errors (an unconfigured Grok returning `missing-auth`, for example) drops out of the count instead of jamming the loop.
- **No silent dismissal.** Every critical issue that gets dismissed or deferred ships with a one-line reason in the final report, including the times Claude walks back one of its own blind objections. The engine rejects an adjudication that dismisses an issue without a reason.
The `/ask-*` commands carry a lighter version of the same rule. The external model only advises: Claude reads the output, applies its own judgment, and owns the synthesized answer. When the models agree, that is input, not a verdict.
@@ -452,7 +452,7 @@ Every delegate prompt also carries today's UTC date and a rule against calling a
2. **Panel review.** GPT, Gemini, Grok (and any configured OpenRouter delegates) review the plan in parallel and emit APPROVE / REQUEST CHANGES / REJECT plus categorized critical issues. The server parses each verdict.
3. **Arbiter adjudication + revision.** Claude reconciles the panel verdicts and its own blind verdict; for each critical issue it picks accept, dismiss (reason required), or defer, then revises the plan for the next round.
-The loop converges when at least one responding external approves, none reject, zero critical issues remain accepted, and Claude adjudicates APPROVE - so Claude cannot self-approve. It otherwise stops at `consensus.maxRounds` (default 5, configurable) as `unresolved`. The confidence label reflects how fast it settled (round 1 = high, 2-3 = medium, 4-5 = low).
+The loop converges when at least `consensus.quorumFloor` externals respond (default 2, capped at the panel size), every one of them approves, none reject, zero critical issues remain accepted, and Claude adjudicates APPROVE - so Claude cannot self-approve. It otherwise stops at `consensus.maxRounds` (default 5, configurable) as `unresolved`. The confidence label reflects how fast it settled (round 1 = high, 2-3 = medium, 4-5 = low).
A round only pays for voices that can still answer: a peer that fails 2 rounds running is dropped from the panel (reported once as `droppedProviders`, then never dispatched again), and the whole loop stops at `consensus.maxWallMs` (default 30 min) rather than starting another round. Both apply on every driver. See [TECHNICAL.md](TECHNICAL.md#circuit-breaker).
diff --git a/TECHNICAL.md b/TECHNICAL.md
index d1e4eb3..3eb6f87 100644
--- a/TECHNICAL.md
+++ b/TECHNICAL.md
@@ -76,9 +76,12 @@ Claude: "I found 3 issues..." (synthesizes, applies judgment)
the core state machine (`core/consensus-loop.js`). Each round: Claude commits a blind verdict,
the server fans out to the panel (`dispatch_peers`) and parses each voice's verdict + critical
issues, Claude adjudicates (accept/dismiss/defer, every dismiss carries a reason), then revises
-the plan. The loop converges only when at least one responding peer APPROVES, none REJECT, zero
-accepted critical issues remain, and Claude's adjudicated verdict is APPROVE - so Claude cannot
-self-approve. The cap is `consensus.maxRounds` (default 5).
+the plan. The loop converges only when at least `quorumFloor` peers respond, every responding
+peer APPROVES, none REJECT, zero accepted critical issues remain, and Claude's adjudicated
+verdict is APPROVE - so Claude cannot self-approve. `quorumFloor` is `consensus.quorumFloor`
+(default 2) capped at the size of the selected panel and never below 1, so a one-peer panel
+needs that one peer. An errored or unparsed review does not count as responding. The cap is
+`consensus.maxRounds` (default 5).
**Critical-issue taxonomy** (the closed set every critical issue is tagged with, parsed by
`parseReview` in `core/provider.js`):
@@ -1226,7 +1229,9 @@ The loop ends `unresolved` once it hits the cap without converging.
the global wall-time budget (ms) for the server-side provider-arbiter convergence loop (the
`consensus` tool). When the budget is spent the loop stops BEFORE starting the next round and
returns UNRESOLVED with `stopReason: "budget-exhausted"`; it never aborts an in-flight
-provider call. Does not apply to the host-driven `/consensus` (`consensus-step`) path.
+provider call. The host-driven `/consensus` (`consensus-step`) path applies the same budget:
+`dispatch_peers` ends the loop as UNRESOLVED with the same `stopReason` instead of starting
+another fan-out once the budget, counted from `init`, is spent.
- **Default.** `1800000` (30 min). A non-integer or non-positive value is ignored (no budget
is applied) without failing the config.
@@ -1448,7 +1453,9 @@ Key rules and invariants pinned here:
- **Every genuine error kind counts**, not just `timeout`. Transport errors (`network`,
`timeout`, `upstream`, `empty`) and unparseable outputs count toward the streak.
- **Quorum floor ($N \ge 2$).** Under `consensus.quorumFloor` (default 2), `checkConvergence`
- refuses convergence unless `responding.length >= quorumFloor`. When circuit breaks drop
+ refuses convergence unless `responding.length >= quorumFloor`. Both drivers cap the configured
+ value at the selected panel size at init (`Math.min(panel, quorumFloor)`, at least 1), so a
+ one-peer panel is still valid and needs that peer to respond. When circuit breaks drop
surviving active providers below the quorum floor, the loop terminates with
`stopReason: "quorum-lost"`.
- **Mid-session peer substitution / identity drift prevention.** In `consensus-step`, round 1
diff --git a/config/config.schema.json b/config/config.schema.json
index c1768f6..64206c9 100644
--- a/config/config.schema.json
+++ b/config/config.schema.json
@@ -169,7 +169,7 @@
"type": "integer",
"minimum": 1,
"default": 1800000,
- "description": "Global wall-time budget (ms) for the server-side provider-arbiter convergence loop (the `consensus` tool). When the budget is spent the loop stops BEFORE starting the next round and returns UNRESOLVED with stopReason \"budget-exhausted\"; it never aborts an in-flight provider call. Does not apply to the host-driven /consensus (consensus-step) path. Default 1800000 (30 min)."
+ "description": "Global wall-time budget (ms) for the server-side provider-arbiter convergence loop (the `consensus` tool). When the budget is spent the loop stops BEFORE starting the next round and returns UNRESOLVED with stopReason \"budget-exhausted\"; it never aborts an in-flight provider call. The host-driven /consensus (consensus-step) path applies the same budget, counted from init, before each dispatch_peers fan-out. Default 1800000 (30 min)."
},
"quorumFloor": {
"type": "integer",
diff --git a/docs/compare/codex-plugin.md b/docs/compare/codex-plugin.md
index 0cd61d8..91cacd6 100644
--- a/docs/compare/codex-plugin.md
+++ b/docs/compare/codex-plugin.md
@@ -13,7 +13,7 @@ Apache-2.0, latest release v1.0.6 (published 2026-07-08).
| | Codex plugin | Deliberation |
|-|--------------|--------------|
-| Models | Codex (OpenAI) | GPT through the Codex CLI, Gemini through the Antigravity CLI, Grok, OpenRouter models, local Ollama and LM Studio models |
+| Models | Codex (OpenAI) | GPT through the Codex CLI, Gemini through an AI Studio API key (Antigravity CLI opt-in), Grok, OpenRouter models, local Ollama and LM Studio models |
| Review | `/codex:review` (read-only, current changes or `--base <ref>`), `/codex:adversarial-review` (takes focus text) | `/deliberation:ask-all` (parallel answers from the panel), `/deliberation:consensus` (multi-round review until the panel agrees), eight expert personas |
| Hand off work | `/codex:rescue` lets Codex investigate and fix, `/codex:transfer` moves the session into Codex | GPT is advisory only. Only Gemini can edit files, through Deliberation's Gemini bridge (Claude Code plugin) |
| Background jobs | `--background`, `/codex:status`, `/codex:result`, `/codex:cancel` | Calls run in the foreground. `panel` plus parallel `ask-one` calls show per-provider progress |
diff --git a/docs/compare/council-of-high-intelligence.md b/docs/compare/council-of-high-intelligence.md
index d56d50f..3a576c0 100644
--- a/docs/compare/council-of-high-intelligence.md
+++ b/docs/compare/council-of-high-intelligence.md
@@ -66,7 +66,7 @@ stars, MIT, latest release v1.2.0 (published 2026-07-04), last push 2026-11-16.
| Shape | Skill and commands per host | MCP server, Claude Code plugin, native host plugins |
| Hosts | Claude Code, Codex, Gemini CLI, OpenCode | Any MCP host; native plugins for Claude Code, Codex, Cursor, Kiro, Antigravity, Gemini CLI, OpenCode |
| Personas | 18 analytical lenses | Eight engineering and writing experts |
-| Runs on | Native host subagents, `codex`, `gemini`, Ollama, NVIDIA NIM, Cursor's agent CLI | `codex`, `agy` (Gemini), Grok, OpenRouter, Ollama, LM Studio |
+| Runs on | Native host subagents, `codex`, `gemini`, Ollama, NVIDIA NIM, Cursor's agent CLI | `codex`, Gemini (API key, or `agy` opt-in), Grok, OpenRouter, Ollama, LM Studio |
| Independence | Blind first positions, then cross-examination | Independent answers; no debate round |
| Ending | Verdict with dissent, kill criteria, next step; split votes returned as split | Converged verdict or `unresolved` at the round cap; dismissals carry reasons |
| Best for | High-stakes decisions | Plans, diffs, security and prose review |
diff --git a/docs/compare/migrate-from-pal.md b/docs/compare/migrate-from-pal.md
index 22b86cf..92bb6fd 100644
--- a/docs/compare/migrate-from-pal.md
+++ b/docs/compare/migrate-from-pal.md
@@ -61,7 +61,7 @@ Code command, where one exists.
| `analyze` (code analysis) | `architect` or `code-reviewer` expert | Deliberation also has a tool named `analyze`, but it does something else: it reports latency, tokens and verdict agreement from your own past runs. |
| `refactor`, `testgen`, `docgen`, `tracer` | no equivalent | |
| `challenge` | no dedicated tool | `/consensus` has rules against an agent approving its own plan (see below), which covers a related problem. |
-| `clink` | no equivalent | Deliberation calls `codex exec` and `agy` for one answer each, not interactive CLI subagents. For a Codex subagent in Claude Code, see [the Codex plugin page](codex-plugin.md). |
+| `clink` | no equivalent | Deliberation calls `codex exec` (and `agy` when you opt in to it for Gemini) for one answer each, not interactive CLI subagents. For a Codex subagent in Claude Code, see [the Codex plugin page](codex-plugin.md). |
| `listmodels` | `panel` tool, `/deliberation:doctor` | `panel` lists the providers a fan-out would call and the ones that cannot answer right now. |
| `continuation_id` threads | `threadId` on the `gemini-reply`, `grok-reply` and `openrouter-reply` tools (Claude Code plugin) | GPT calls are single-shot. Threads live in memory and are lost when the server restarts. |
| Context revival ("continue with O3") | no equivalent | |
@@ -90,7 +90,7 @@ It never overwrites an existing file. In Claude Code, run `/deliberation:setup`
| PAL setting | Deliberation |
|-------------|--------------|
| `OPENAI_API_KEY` | Not read. GPT runs through the Codex CLI: `npm install -g @openai/codex`, then `codex login` (a ChatGPT subscription). A ChatGPT Business or Enterprise workspace can set `CODEX_ACCESS_TOKEN` instead. With neither a login nor that token, Codex uses `CODEX_API_KEY`. |
-| `GEMINI_API_KEY` | Not read. Gemini runs through the Antigravity CLI (`agy`) and its own login. See [Getting Started with Antigravity CLI](https://antigravity.google/docs/cli-getting-started). |
+| `GEMINI_API_KEY` | Same variable (a Google AI Studio key). Gemini calls the Gemini API with it by default. The Antigravity CLI (`agy`) is opt-in: `providers.gemini.transport` (`auto` / `api` / `agy`) picks the path, and with no key `auto` falls back to `agy`. Read the [terms note](../../SETUP.md#gemini-api-key-default-or-agy-opt-in) before you use `agy`. |
| `XAI_API_KEY` | Same variable. The Grok provider reads it (`providers.grok.apiKeyEnv`). |
| `OPENROUTER_API_KEY` | Same variable. Also set `providers.openrouter.enabled: true` and declare each model you want as a record in `models` (next section). |
| `CUSTOM_API_URL`, `CUSTOM_API_KEY`, `CUSTOM_MODEL_NAME` (Ollama, LM Studio) | `providers.ollama.apiBase` (default `http://localhost:11434/v1`) or `providers.lmstudio.apiBase` (default `http://localhost:1234/v1`). Both are keyless by default; set `apiKeyEnv` to the name of a variable that holds a key. Declare models with `provider: "ollama"` or `provider: "lmstudio"`. |
@@ -168,20 +168,21 @@ Claude Code goes through `/deliberation:setup`.
- **Writes.** PAL providers can return code for your agent to apply. In Deliberation every
provider advises. Only Gemini, through the Claude Code plugin's Gemini bridge, can also edit
files (`workspace-write`). The standalone MCP server is advisory only.
-- **Login instead of API keys for GPT and Gemini.** Deliberation shells out to the `codex` and
- `agy` CLIs, so these calls use your existing CLI logins and plans. PAL calls the OpenAI and
- Gemini APIs with keys.
+- **Login instead of an API key for GPT.** Deliberation shells out to the `codex` CLI, so GPT
+ calls use your existing Codex login and plan. PAL calls the OpenAI API with a key. Gemini uses
+ `GEMINI_API_KEY` in both; Deliberation can also drive the `agy` CLI if you opt in.
- **Consensus.** PAL's `consensus` gives each model a stance and your agent combines the
answers. Deliberation's `/consensus` has your agent record its own verdict before it sees the
- panel. A round converges only when at least `consensus.quorumFloor` models answer (default 2),
- every one of them approves, no accepted critical issue is left, and the agent approves too. The
- agent cannot approve a plan on its own vote. For every issue the agent dismisses, the
- final report includes a written reason. The loop stops at `consensus.maxRounds` (default 5).
+ panel. A round converges only when at least `consensus.quorumFloor` models answer (default 2,
+ capped at the panel size), every one of them approves, no accepted critical issue is left, and
+ the agent approves too. The agent cannot approve a plan on its own vote. For every issue the
+ agent dismisses, the final report includes a written reason. The loop stops at `consensus.maxRounds` (default 5).
- **Disagreement stays visible.** `ask-all` returns each model's answer separately, and the
synthesis names where they agree and where they split.
-- **Repo context for HTTP models.** GPT and Gemini read the repository through their CLIs. Grok,
- OpenRouter and local models get a small bundle of repo files (AGENTS.md or CLAUDE.md,
- manifests, README) attached automatically. Turn it off with `orientation.enabled: false`.
+- **Repo context for HTTP models.** GPT, and Gemini on the opt-in `agy` path, read the repository
+ through their CLIs. Gemini on its default API key, Grok, OpenRouter and local models get a small
+ bundle of repo files (AGENTS.md or CLAUDE.md, manifests, README) attached automatically. Turn it
+ off with `orientation.enabled: false`.
- **Run history and metrics.** These features are opt-in: a debug log of latency, tokens and votes
(`debug.enabled`), saved runs (`sessions.persist`), and a local read-only dashboard
(`dashboard.enabled`).
diff --git a/docs/compare/pal.md b/docs/compare/pal.md
index e6c4475..078c9ac 100644
--- a/docs/compare/pal.md
+++ b/docs/compare/pal.md
@@ -24,7 +24,7 @@ As of 2026-11-17, from the GitHub API and PAL's README:
## Where PAL is the better fit
- **Provider breadth.** Azure OpenAI and DIAL have no Deliberation equivalent. PAL calls the
- OpenAI and Gemini APIs with keys; Deliberation needs the `codex` and `agy` CLIs for those two.
+ OpenAI API with a key; Deliberation reaches GPT only through the `codex` CLI.
- **Code from capable models.** PAL's `chat` can return complete code from capable models.
Deliberation's experts are prompted to advise; only Gemini, through the Claude Code plugin, can
edit files itself.
@@ -47,16 +47,19 @@ As of 2026-11-17, from the GitHub API and PAL's README:
stop that:
- The agent records its verdict before the panel answers, and the engine holds the panel back
until it does.
- - A round converges only when at least `consensus.quorumFloor` models answer (default 2), every
- one of them approves, no accepted critical issue is left, and the agent approves too.
+ - A round converges only when at least `consensus.quorumFloor` models answer (default 2, capped
+ at the panel size), every one of them approves, no accepted critical issue is left, and the
+ agent approves too.
- Every dismissed issue needs a written reason, and the engine rejects a dismissal without one.
The loop stops at `consensus.maxRounds` (default 5). PAL's `consensus` assigns stances (for,
against, neutral) and the agent combines the answers.
- **Disagreement you can read.** `/ask-all` returns each model's answer separately and the
synthesis lists where they agree and where they split.
-- **CLI logins instead of API keys.** GPT runs through `codex exec` and Gemini through `agy`, so
- those calls use your CLI logins and plans instead of separate API keys.
+- **A CLI login for GPT.** GPT runs through `codex exec`, so those calls use your Codex login and
+ plan instead of an OpenAI API key. Gemini uses a `GEMINI_API_KEY` by default; the Antigravity
+ CLI (`agy`) is an opt-in path (read the
+ [terms note](../../SETUP.md#gemini-api-key-default-or-agy-opt-in) first).
- **Live config.** Model records, routing and the OpenRouter block in
`~/.config/deliberation/config.json` take effect without a restart. Each record says whether it
joins `ask-all` and `consensus`, and which experts it serves.
@@ -75,7 +78,7 @@ As of 2026-11-17, from the GitHub API and PAL's README:
| Language | Python | JavaScript, no runtime dependencies |
| Install | `git clone` + `./run-server.sh`, or `uvx` | Claude Code plugin, `npx -y @antonbabenko/deliberation-mcp`, native host plugins |
| GPT | OpenAI API key | Codex CLI login |
-| Gemini | Gemini API key | Antigravity CLI login |
+| Gemini | Gemini API key | `GEMINI_API_KEY` (AI Studio); Antigravity CLI (`agy`) opt-in |
| Grok | `XAI_API_KEY` | `XAI_API_KEY` |
| OpenRouter | `OPENROUTER_API_KEY`, models in `conf/openrouter_models.json` | `OPENROUTER_API_KEY`, models as records in `config.json` |
| Local models | Custom endpoint (Ollama, vLLM, LM Studio) | Ollama and LM Studio providers; other endpoints through `providers.openrouter.apiBase` |
diff --git a/docs/dev/architecture.md b/docs/dev/architecture.md
index aff93d8..f5ebba3 100644
--- a/docs/dev/architecture.md
+++ b/docs/dev/architecture.md
@@ -68,7 +68,7 @@ shared by two drivers so there is ONE rules layer, not a Claude copy and a non-C
the prose. This is the transcript-visible host-arbiter path; the `consensus` tool is the
provider-arbiter path.
-The cap is `consensus.maxRounds` (config, default 5, clamped to 50; a per-call `maxRounds` overrides it). A configurable quorum floor `consensus.quorumFloor` (default 2, min 1) guarantees consensus cannot converge when surviving responding peers drop below the quorum threshold. A wall-time budget `consensus.maxWallMs` (default 1 800 000 ms, 30 min) stops the provider-arbiter `consensus` loop before the next round when the budget is spent, returning `stopReason: "budget-exhausted"`; the host-driven `consensus-step` path is not affected. The Codex provider is no longer unbounded: `core/providers/codex.js` caps each `codex exec` call to `CODEX_DEFAULT_TIMEOUT_MS` (600 000 ms), and a run killed by that timer reports `errorKind: "timeout"` from the kill flag rather than from a stderr substring (a SIGKILL'd codex often writes nothing, and an `unknown` can never trip the circuit breaker). Every provider's ceiling is configurable: `providers.defaults.timeout` covers all four at once, `providers.<name>.timeout` overrides one (`providers.openrouter.defaults.timeout` for OpenRouter), and a pinned alias's `models.<id>.timeout` still wins - the ladder resolves in `server/openrouter/config.js resolveProviders`, so the composition root just reads the resolved value. Both HTTP bridges keep the `AbortController` armed until the response BODY is read, so a slow body is a `timeout` rather than an unbounded run. `callProvider` retries once - and only once - on `network`, `rate-limit` (waiting for the upstream's `Retry-After`, clamped to 30s), or `empty`; it does not retry timeout or application errors.
+The cap is `consensus.maxRounds` (config, default 5, clamped to 50; a per-call `maxRounds` overrides it). A configurable quorum floor `consensus.quorumFloor` (default 2, min 1; both drivers cap it at the selected panel size, so a one-peer panel needs that peer) guarantees consensus cannot converge when surviving responding peers drop below the quorum threshold. A wall-time budget `consensus.maxWallMs` (default 1 800 000 ms, 30 min) stops the provider-arbiter `consensus` loop before the next round when the budget is spent, returning `stopReason: "budget-exhausted"`; the host-driven `consensus-step` path applies the same budget before each `dispatch_peers` fan-out, counted from `init`. The Codex provider is no longer unbounded: `core/providers/codex.js` caps each `codex exec` call to `CODEX_DEFAULT_TIMEOUT_MS` (600 000 ms), and a run killed by that timer reports `errorKind: "timeout"` from the kill flag rather than from a stderr substring (a SIGKILL'd codex often writes nothing, and an `unknown` can never trip the circuit breaker). Every provider's ceiling is configurable: `providers.defaults.timeout` covers all four at once, `providers.<name>.timeout` overrides one (`providers.openrouter.defaults.timeout` for OpenRouter), and a pinned alias's `models.<id>.timeout` still wins - the ladder resolves in `server/openrouter/config.js resolveProviders`, so the composition root just reads the resolved value. Both HTTP bridges keep the `AbortController` armed until the response BODY is read, so a slow body is a `timeout` rather than an unbounded run. `callProvider` retries once - and only once - on `network`, `rate-limit` (waiting for the upstream's `Retry-After`, clamped to 30s), or `empty`; it does not retry timeout or application errors.
The `consensus` tool AND the host-driven `consensus-step` loop persist a session record on a
terminal transition - converged or unresolved (when `sessions.persist` is on, with the mode flag).
`consensus-step` uses an atomic `loopStore.take()` before the write so a terminal transition writes
=== Code the claims describe, at 4bf190b ===
core/consensus-loop.js, checkConvergence (shared by both drivers):
function checkConvergence(state) {
const results = state.results || [];
const responding = results.filter((r) => !r.isError && VERDICTS.includes(/** @type {any} */ (r.verdict)));
const quorumFloor = Number.isInteger(state.quorumFloor) && /** @type {number} */ (state.quorumFloor) >= 1
? /** @type {number} */ (state.quorumFloor)
: DEFAULT_QUORUM_FLOOR;
const quorumMet = responding.length >= quorumFloor;
const everyApprove = responding.length > 0 && responding.every((r) => r.verdict === "APPROVE");
const anyReject = responding.some((r) => r.verdict === "REJECT");
const acceptedCritical = (state.decisions || []).filter((d) => d.action === "accept").length;
const hostApprove = !!(state.hostVerdict && state.hostVerdict.verdict === "APPROVE");
const converged = quorumMet && everyApprove && !anyReject && acceptedCritical === 0 && hostApprove;
return {
converged,
verdict: converged ? "APPROVE" : "REQUEST_CHANGES",
nextAction: converged ? "finalize" : "revise",
quorumMet,
};
}
server/mcp/index.js, provider-arbiter driver (runConsensusAuto), quorum cap:
// Exclude the arbiter from the peer panel - never let it review its own output.
// A single distinct peer is a valid minimal panel (peer != arbiter).
const peers = selected.filter((p) => p.name !== arbiterP.name);
if (peers.length < 1) {
return { payload: { converged: false, verdict: null, confidence: "none", rounds: 0, opinions: [], arbiter: { mode: "server", provider: arbiterP.name }, warnings: warnings.concat(["consensus needs at least one peer distinct from the arbiter"]), error: "insufficient-peers" }, parts: null };
}
const configuredQuorum = Number.isInteger(cc.quorumFloor) && cc.quorumFloor >= 1
? cc.quorumFloor
: loop.DEFAULT_QUORUM_FLOOR;
const quorumFloor = Math.min(peers.length, configuredQuorum);
server/mcp/index.js, host-driven consensus-step, init:
const cfg = dispatchConfig() || {};
const cc = cfg.consensus || {};
const maxRounds = Number.isInteger(cc.maxRounds) && cc.maxRounds > 0 ? cc.maxRounds : undefined;
const { providers: allCandidates } = registry.selectForConsensus({ config: cfg, expert: args.expert || "", unhealthy: await unhealthyMap(providers) });
const configuredQuorum = Number.isInteger(cc.quorumFloor) && cc.quorumFloor >= 1
? cc.quorumFloor
: loop.DEFAULT_QUORUM_FLOOR;
const quorumFloor = Math.max(1, Math.min(allCandidates.length, configuredQuorum));
const originalPrompt = typeof args.prompt === "string" ? args.prompt : "";
server/mcp/index.js, consensus-step dispatch_peers, budget and quorum gates:
// Budget gate. Like runToConvergence this gates STARTING a fan-out and never
// interrupts one in flight, so a slow-but-good answer is always collected.
const budgetMs = wallBudgetMs(dispatchConfig());
if (budgetMs !== null && typeof cur.startedAt === "number" && Date.now() - cur.startedAt >= budgetMs) {
return terminateLoop(sid, cur, "budget-exhausted", dropped);
}
// Only claim the breaker when it is actually why the panel is empty; a config
// with no eligible providers is a different problem and must say so.
if (!selected.length) return terminateLoop(sid, cur, dropped.length ? "all-providers-circuit-broken" : "no-providers", dropped);
if (selected.length < cur.quorumFloor) return terminateLoop(sid, cur, "quorum-lost", dropped);
=== Other surfaces that describe the same rule, unchanged by this PR ===
commands/consensus.md (the /consensus driver; scripts/sync-hosts.js copies it into the generated host skills):
```
- If the response has `converged: true`, the loop is done - go to Final output (use the
returned `finalReport` + `confidence`). The engine converges only when at least one
responding peer APPROVED, none REJECTED, zero accepted critical issues remain, and
your adjudicated verdict is APPROVE - so your APPROVE alone never converges.
- Otherwise `status` is `await_revision` - continue.
server/mcp/README.md (the npm package README), line 99:
The orchestrator commits a blind verdict before seeing the panel, cannot reach consensus on
its own vote alone (at least one external must approve), and records a reason for every
dismissed issue. The loop stops on agreement or at `consensus.maxRounds` (default 5).
Round history
| Round | Blind verdict | Reviewers | After adjudication | Outcome |
|---|---|---|---|---|
| Round 1 5m 28s | REQUEST CHANGES | or:qwen3-8-max REQUEST CHANGES[4]or:gpt-6-1-sol REQUEST CHANGES[3]or:gemini-3-8-flash REQUEST CHANGES[5] | REQUEST CHANGES revised | Not converged, 0 of 3 approved |
| Round 2 5m 48s | APPROVE | or:qwen3-8-max APPROVEor:gpt-6-1-sol APPROVEor:gemini-3-8-flash APPROVE | APPROVE | Converged, 3 of 3 approved |
Round 1
Blind verdict. Before any reviewer answered, the host agent recorded REQUEST CHANGES with 2 issues of its own: Every claim the PR adds matches checkConvergence and both drivers, but it fixes the rule in four places and leaves the old one in the two surfaces the host and npm users read, so the drift it set out to remove is still there.
or:qwen3-8-max REQUEST CHANGES
This reply has no summary line. The full response is below.
Full response, verbatim
VERDICT: REQUEST_CHANGES - [correctness] `commands/consensus.md` still describes a one-peer quorum ("at least one responding peer APPROVED"), contradicting the implemented `consensus.quorumFloor` rule (default 2, capped at the selected panel size) and the updated README/TECHNICAL/schema text; this file drives `/consensus` and is propagated into host skills, so the agent-facing rule remains wrong. - [correctness] `server/mcp/README.md` still says "at least one external must approve", which conflicts with the quorumFloor rule this PR adds elsewhere and will mislead readers of the npm package README. - [scope] The PR claims to fix consensus doc drift but omits required user-visible surfaces under AGENTS.md; at minimum the `/consensus` command prose and generated host skill outputs must be updated/regenerated, and `docs/tool-guide.md`, `SETUP.md`, and `docs/dev/design-decisions.md` should be checked for stale one-peer or "maxWallMs does not apply to consensus-step" statements. - [ops] The diagram sources/PNGs and any generated host assets are not verifiable from the provided diff; before shipping, confirm the Stage 2 cross-review step is actually removed, PNG dimensions are unchanged, links/anchors (including the SETUP.md anchor used by compare pages) resolve, and `npm run sync:check` passes after the missing command/skill updates.or:gpt-6-1-sol REQUEST CHANGES
REQUEST CHANGES: the supplied code supports the updated quorum and budget claims, but the authoritative command instructions and published package README retain obsolete convergence descriptions. Confidence: high in this documentation inconsistency; Gemini transport claims and diagram correctness require the omitted implementations and assets.
Full response, verbatim
**Summary**: The revised quorum and wall-budget descriptions match the supplied implementation. However, the PR leaves contradictory convergence instructions in the command source and npm package README, so it does not complete its stated documentation fix. **Critical issues (must fix)**: - **[MEDIUM] Stale convergence rule remains in operational instructions** — `commands/consensus.md` and `server/mcp/README.md:99`. The command still says one approving peer and no rejections suffice, omitting both the quorum requirement and the requirement that **every responding reviewer approve**. For example, an APPROVE plus REQUEST_CHANGES does not converge in `checkConvergence`. The package README also retains the obsolete one-external description. Update both to the complete rule, including the panel-size cap, then regenerate host skills from the command source and run `npm run sync:check`. **Recommendations (should consider)**: - Verify Gemini transport selection and orientation-bundle behavior against their implementations; those implementations were not supplied, so these added claims remain **[unverified]** in this review. - Inspect the omitted diagram sources and rendered PNGs before merging. Their removal of cross-review and preservation of dimensions cannot be confirmed here. **Verdict**: REQUEST CHANGES. <SUMMARY> REQUEST CHANGES: the supplied code supports the updated quorum and budget claims, but the authoritative command instructions and published package README retain obsolete convergence descriptions. Confidence: high in this documentation inconsistency; Gemini transport claims and diagram correctness require the omitted implementations and assets. </SUMMARY> VERDICT: REQUEST_CHANGES - [correctness] Update commands/consensus.md and server/mcp/README.md to require the capped quorum and unanimous approval among responding reviewers, then regenerate affected host skills.
or:gemini-3-8-flash REQUEST CHANGES
Verdict: REQUEST CHANGES. Top risks: (1) Host arbiter confusion in `/consensus` due to stale convergence instructions in `commands/consensus.md` contradicting `quorumFloor`; (2) Public doc drift in `server/mcp/README.md` and `docs/tool-guide.md` violating `AGENTS.md` PR completion criteria; (3) CI break risk if generated host files drift from command sources. Confidence: High. Missing context: None needed; code and doc requirements are fully specified in repository rules.
Full response, verbatim
**Summary**: This PR accurately captures the `quorumFloor` and `maxWallMs` behavior in the files it touched, but it leaves significant doc drift across other required surfaces (`commands/consensus.md`, `server/mcp/README.md`, and `docs/tool-guide.md`), violating the repository's completeness rules in `AGENTS.md`. --- ### Critical issues (must fix) 1. **Stale convergence rule in orchestrator prompt and host skills** - **Location**: `commands/consensus.md:14` (and generated skills in `plugins/deliberation/`, `.agents/`, etc.) - **Why it matters**: `commands/consensus.md` explicitly instructs the host arbiter (Claude) that the engine converges when *"at least one responding peer APPROVED"*. In reality, `checkConvergence` enforces `state.quorumFloor` (default 2, capped at panel size). Under a multi-peer panel where only one peer responds with APPROVE and others error, the host will assume the loop converged, while the engine returns `converged: false` or terminates with `quorum-lost`. Per `AGENTS.md`, `commands/` and skill prose must be updated when behavior changes. - **Fix**: Update `commands/consensus.md` to state the quorum floor convergence rule (`consensus.quorumFloor`, default 2, capped at panel size) and run `npm run sync` to update synced host skills. 2. **Doc drift in published npm package README** - **Location**: `server/mcp/README.md:99` - **Why it matters**: The standalone MCP server documentation still states: *"cannot reach consensus on its own vote alone (at least one external must approve)"*. This leaves stale behavior on the primary package surface published to npm. - **Fix**: Update `server/mcp/README.md` to describe the `consensus.quorumFloor` requirement. 3. **Missing updates to host-neutral tool guide** - **Location**: `docs/tool-guide.md` - **Why it matters**: `AGENTS.md` mandates that no user-visible behavior or config rule is done until `docs/tool-guide.md` is current. `docs/tool-guide.md` is the host-neutral reference for MCP tools (`consensus`, `consensus-step`) and serves as an input to `scripts/sync-hosts.js`. It was omitted from the PR. - **Fix**: Document the `quorumFloor` cap and `maxWallMs` budget behavior on `consensus-step` in `docs/tool-guide.md`. 4. **Stale generated host files and sync check** - **Location**: Repository root (`npm run sync:check`) - **Why it matters**: `AGENTS.md` requires committing regenerated files alongside source changes. Once `commands/consensus.md` and `docs/tool-guide.md` are updated, failing to run `npm run sync` will cause `npm run sync:check` to fail in CI. - **Fix**: Run `npm run sync` after editing source Markdown files and commit the generated outputs. --- ### Recommendations (should consider) 1. **Misleading explanation of erroring providers in README** - **Location**: `README.md:439` - **Why**: The text states: *"A provider that errors (an unconfigured Grok returning `missing-auth`, for example) drops out of the count instead of jamming the loop."* While errored providers are filtered out of `responding`, under `quorumFloor` (default 2), an erroring provider in a 2-provider panel leaves only 1 responding peer, which blocks convergence and ultimately halts with `quorum-lost`. Stating that it does not jam the loop is misleading unless qualified by quorum availability. - **Fix**: Clarify that while an erroring provider drops from the responding count, convergence still requires at least `quorumFloor` responding peers. --- **Verdict**: REQUEST CHANGES `<SUMMARY>` Verdict: REQUEST CHANGES. Top risks: (1) Host arbiter confusion in `/consensus` due to stale convergence instructions in `commands/consensus.md` contradicting `quorumFloor`; (2) Public doc drift in `server/mcp/README.md` and `docs/tool-guide.md` violating `AGENTS.md` PR completion criteria; (3) CI break risk if generated host files drift from command sources. Confidence: High. Missing context: None needed; code and doc requirements are fully specified in repository rules. `</SUMMARY>` VERDICT: REQUEST_CHANGES - [correctness] commands/consensus.md line 14 was not updated and still instructs the arbiter that convergence requires only at least one responding peer to approve, contradicting core/consensus-loop.js and quorumFloor - [correctness] server/mcp/README.md line 99 was omitted from the PR and retains the stale "at least one external must approve" convergence rule - [correctness] docs/tool-guide.md was omitted from the PR despite AGENTS.md requiring it to be updated for tool behavior and config changes (quorumFloor cap and consensus-step maxWallMs) - [ambiguity] README.md line 439 claims an erroring provider drops out instead of jamming the loop, which is misleading under quorumFloor >= 2 where an erroring peer in a 2-peer panel prevents convergence and triggers quorum-lost - [ops] commands/consensus.md and docs/tool-guide.md source updates require running npm run sync to regenerate host artifacts and satisfy npm run sync:check
| Issue | Raised by | Ruling | Reason |
|---|---|---|---|
| correctness commands/consensus.md still tells the host "The engine converges only when at least one responding peer APPROVED, none REJECTED", the rule this PR calls wrong. | host agent | accept | All three peers raise it too; fixed in the follow-up commit, with regenerated copies. |
| correctness server/mcp/README.md (the npm package README) still says "at least one external must approve". It is the surface npm and registry users read, and the PR fixes the same sentence in README.md. | host agent | accept | All three peers raise it too; fixed in the follow-up commit. |
| correctness commands/consensus.md` still describes a one-peer quorum ("at least one responding peer APPROVED"), contradicting the implemented `consensus.quorumFloor` rule (default 2, capped at the selected panel size) and the updated README/TECHNICAL/schema text; this file drives `/consensus` and is propagated into host skills, so the agent-facing rule remains wrong. | or:qwen3-8-max | accept | Same as the host issue. |
| correctness server/mcp/README.md` still says "at least one external must approve", which conflicts with the quorumFloor rule this PR adds elsewhere and will mislead readers of the npm package README. | or:qwen3-8-max | accept | Same as the host issue. |
| scope The PR claims to fix consensus doc drift but omits required user-visible surfaces under AGENTS.md; at minimum the `/consensus` command prose and generated host skill outputs must be updated/regenerated, and `docs/tool-guide.md`, `SETUP.md`, and `docs/dev/design-decisions.md` should be checked for stale one-peer or "maxWallMs does not apply to consensus-step" statements. | or:qwen3-8-max | accept | Checked at 4bf190b: SETUP.md:433 still says maxWallMs applies to the consensus tool only, so it is fixed in the follow-up; tool-guide and design-decisions are already correct. |
| ops The diagram sources/PNGs and any generated host assets are not verifiable from the provided diff; before shipping, confirm the Stage 2 cross-review step is actually removed, PNG dimensions are unchanged, links/anchors (including the SETUP.md anchor used by compare pages) resolve, and `npm run sync:check` passes after the missing command/skill updates. | or:qwen3-8-max | dismiss | Verified at 4bf190b: PNGs keep 1000x420 and 1200x900, neither HTML source mentions Stage 2 or cross-review, and the new SETUP.md anchor matches its heading. sync:check after the command fix is covered by the accepted command issue. |
| ambiguity Stale convergence rule remains in operational instructions** — `commands/consensus.md` and `server/mcp/README.md:99`. | or:gpt-6-1-sol | accept | Same as the host issues; the follow-up states the full rule in both. |
| ambiguity in this review. | or:gpt-6-1-sol | dismiss | A parse fragment of the reply with no claim in it; the reply's real issues are the two entries ruled on above and below. |
| correctness Update commands/consensus.md and server/mcp/README.md to require the capped quorum and unanimous approval among responding reviewers, then regenerate affected host skills. | or:gpt-6-1-sol | accept | Same fix as the host issues. |
| correctness commands/consensus.md line 14 was not updated and still instructs the arbiter that convergence requires only at least one responding peer to approve, contradicting core/consensus-loop.js and quorumFloor | or:gemini-3-8-flash | accept | Same as the host issue. |
| correctness server/mcp/README.md line 99 was omitted from the PR and retains the stale "at least one external must approve" convergence rule | or:gemini-3-8-flash | accept | Same as the host issue. |
| correctness docs/tool-guide.md was omitted from the PR despite AGENTS.md requiring it to be updated for tool behavior and config changes (quorumFloor cap and consensus-step maxWallMs) | or:gemini-3-8-flash | dismiss | Checked docs/tool-guide.md at 4bf190b: it does not state the convergence rule anywhere, and it already lists budget-exhausted (consensus.maxWallMs) as a consensus-step stop reason, so nothing in it is stale. |
| ambiguity README.md line 439 claims an erroring provider drops out instead of jamming the loop, which is misleading under quorumFloor >= 2 where an erroring peer in a 2-peer panel prevents convergence and triggers quorum-lost | or:gemini-3-8-flash | accept | The PR edited this bullet: with a two-peer panel one errored peer blocks convergence and the breaker ends the loop quorum-lost. The follow-up rewrites the sentence. |
| ops commands/consensus.md and docs/tool-guide.md source updates require running npm run sync to regenerate host artifacts and satisfy npm run sync:check | or:gemini-3-8-flash | accept | Part of the command fix; the follow-up includes the regenerated copies. |
Why the round closed this way. All three peers voted REQUEST_CHANGES and named the same two stale surfaces as the host's blind verdict (commands/consensus.md and server/mcp/README.md). The host checked every other claim against the commit, accepted 11 issues, dismissed 3 with evidence, and sent the PR back with a follow-up commit.
revision Follow-up commit: full quorum rule in commands/consensus.md (plus regenerated copies) and server/mcp/README.md, maxWallMs scope in SETUP.md, errored-provider sentence in README; diagram and anchor checks recorded.
Round 2
Blind verdict. Before any reviewer answered, the host agent recorded APPROVE: With the follow-up commit every surface that states the convergence rule or the maxWallMs scope says what checkConvergence and both drivers do, so the PR can ship.
or:qwen3-8-max APPROVE
Verdict: APPROVE. The follow-up commit resolves the prior one-peer quorum drift: README, TECHNICAL, SETUP, config schema, and commands now match `checkConvergence` and the two driver initializations, and the consensus-step budget gate matches the updated `maxWallMs` description. Remaining risks are minor: the README errored-provider example assumes the default floor, and I could not re-run `sync:check` or inspect diagram sources from the supplied diff. Confidence: high for the documented consensus/budget behavior based on the provided code excerpts; it would drop if generated host copies or `docs/tool-guide.md` contained stale one-peer wording. No blocking issues found.
Full response, verbatim
**Summary**: The change now matches the supplied runtime behavior. `checkConvergence()` requires at least the effective quorum of non-error, parseable responding peers, all approving, no accepted critical issues, and a host/arbiter APPROVE. Both drivers compute the effective floor as the configured `consensus.quorumFloor` capped by the peer panel size, with a minimum of 1. The updated `maxWallMs` wording matches the `consensus-step` budget gate, which checks elapsed time from loop start before starting another `dispatch_peers` fan-out. The follow-up commit fixed the stale “one external approver” wording in `commands/consensus.md`, `server/mcp/README.md`, `SETUP.md`, and the README’s errored-provider explanation. No blocking correctness, security, performance, or ops issues found from the provided diff and code excerpts. **Critical issues**: None. **Recommendations**: - The README errored-provider example is correct for the default `quorumFloor: 2`, but could be read as universal. If touched later, qualify it as “with the default quorum floor” to avoid implying the same outcome when a user explicitly sets `quorumFloor: 1`. - If maintainers treat `AGENTS.md` as a user-facing product description, a future consistency pass could mention the Gemini API-key default there too; this is not blocking for this docs-fix PR. **Verdict**: APPROVE <SUMMARY> Verdict: APPROVE. The follow-up commit resolves the prior one-peer quorum drift: README, TECHNICAL, SETUP, config schema, and commands now match `checkConvergence` and the two driver initializations, and the consensus-step budget gate matches the updated `maxWallMs` description. Remaining risks are minor: the README errored-provider example assumes the default floor, and I could not re-run `sync:check` or inspect diagram sources from the supplied diff. Confidence: high for the documented consensus/budget behavior based on the provided code excerpts; it would drop if generated host copies or `docs/tool-guide.md` contained stale one-peer wording. No blocking issues found. </SUMMARY> VERDICT: APPROVE
or:gpt-6-1-sol APPROVE
APPROVE. The previously blocking quorum and maxWallMs documentation drift is addressed. One nonblocking ambiguity remains in the README’s default-quorum example. Confidence: medium; Gemini transport implementation, rendered diagrams, and successful check/sync output would raise it.
Full response, verbatim
**Summary**: The follow-up addresses the material documentation gaps identified in round 1. The convergence and host-driven budget descriptions match the supplied implementation; no runtime changes or new security/performance risks are introduced. **Critical issues (must fix)**: None. No blocking issues found. **Recommendations (should consider)**: - **MEDIUM — Qualify the two-peer example** — `README.md`, “No self-approval”: the new statement that one errored peer blocks convergence assumes the default `consensus.quorumFloor: 2`. With a configured floor of 1, the remaining approving peer can satisfy quorum. Prefix the example with “With the default quorum floor” to avoid presenting configurable behavior as unconditional. **Verification limits**: Gemini transport selection and orientation bundling cannot be independently checked against the supplied code. Diagram rendering, generated-file freshness, and test results are reported as verified, but their artifacts/output are not included here. These are confidence limits, not demonstrated defects. <SUMMARY> APPROVE. The previously blocking quorum and maxWallMs documentation drift is addressed. One nonblocking ambiguity remains in the README’s default-quorum example. Confidence: medium; Gemini transport implementation, rendered diagrams, and successful check/sync output would raise it. </SUMMARY> VERDICT: APPROVE
or:gemini-3-8-flash APPROVE
Verdict: APPROVE. Documentation and diagrams are fully synchronized with runtime consensus logic across all required surfaces (README, TECHNICAL, SETUP, schema, architecture, and commands). Confidence: high.
Full response, verbatim
### Summary The pull request and its follow-up commit bring documentation, config schema, command descriptions, and architecture docs into alignment with the consensus engine implementation (`core/consensus-loop.js` and `server/mcp/index.js`). All surfaces accurately document the capped `consensus.quorumFloor` convergence behavior, `consensus.maxWallMs` budget handling across both drivers, and the default API key transport for Gemini. ### Critical issues No blocking issues found. ### Recommendations - None. The follow-up commit resolved the doc drift in `commands/consensus.md`, `server/mcp/README.md`, and `SETUP.md`, including regenerated host bundles, heading anchors, and asset dimension verification. ### Verdict APPROVE <SUMMARY> Verdict: APPROVE. Documentation and diagrams are fully synchronized with runtime consensus logic across all required surfaces (README, TECHNICAL, SETUP, schema, architecture, and commands). Confidence: high. </SUMMARY> VERDICT: APPROVE
No reviewer listed a critical issue, so there was nothing to rule on.
Why the round closed this way. Every responding peer (3 of 3) voted APPROVE with no critical issues, and the host adjudicated APPROVE.
Outcome
CONVERGED in 2 round(s) (confidence: medium)
Still in the report
What the run did not fix stays listed.
- The new README errored-provider sentence assumes the default quorum floor of 2; with quorumFloor set to 1 the remaining approving peer can still converge, so the example should say "with the default quorum floor".
- Nothing stops the next drift: a test that reads the doc surfaces and fails on the old one-peer wording would be cheap.
- The rewritten README sentence names the two-peer case only; larger panels with several errored peers behave the same way.