C3 — claude-codex-consult in Rust
Ask independent reviewers. Check their findings. Keep the record beside your code.
C3 = Claude Codex Consult — A Rust implementation of the consultation bridge, with Claude Code as its supported coordinator.
Start with the agent you work in
Your agent is the coordinator. The models it asks for a second opinion are the reviewers.
I work in Claude Code
Use C3, the Rust bridge with Claude Code skills and a session-start hook. The PowerShell bridge is also available for this host.
I work in Codex or another host
Use the PowerShell bridge. It documents setup for Codex CLI, Z Code, Kimi Code, Qwen Code, OpenCode, Muse Code and a plain shell.
Launching the C3 CLI from another shell does not provide Claude Code's hooks or coordinator integration. Using Codex as a reviewer is a separate capability.
How a consultation works
1. Write a brief
Your agent describes the question and the relevant context.
2. Ask reviewers
One reviewer or a parallel panel examines the brief using the access its engine provides.
3. Verify the findings
Your agent checks the replies and records its decisions. You keep the brief, answers and findings.
Briefs, original replies, a JSON ledger and findings with stable IDs stay beside the project. C3 and the PowerShell bridge share these records. They support different coordinator hosts, so choose the implementation for the agent you work in.
Install: C3 — the Rust binary
Build the released tag v0.2.2, then add the C3 plugin in Claude Code. Cargo puts c3 on PATH; C3_EXE can point to another installation.
1. Install the C3 binary
git clone --branch v0.2.2 https://github.com/xelth-com/C3 cd C3 cargo install --locked --path crates/c3-cli c3 --version
The default build is light: packs and HTTP reviews work without the index. Add --features index-surreal to the install command if you need the optional local SurrealDB index. The MCP server is configured separately.
2. Add the Claude Code plugin
/plugin marketplace add xelth-com/C3 /plugin install c3@c3
c3 providers --short
Full setup, provider tables, engines: C3 README · Light or indexed build · C3 changelog · C3 design (docs/DESIGN.md)
PowerShell bridge: install or try
This prompt sets up the PowerShell bridge. Paste it into your agent to try one consultation or configure your reviewers. It selects instructions for your host; you handle logins and credentials in your own terminal. For C3 in Claude Code, use the separate C3 installation instructions.
Download the PowerShell setup prompt
The install commands by hand
For Claude Code, install the PowerShell bridge plugin:
/plugin marketplace add xelth-com/claude-codex-consult /plugin install codex-consult@claude-codex-consult
/codex-consult:consult-codex <task-id> <your question>
Codex CLI
codex plugin marketplace add xelth-com/claude-codex-consult codex plugin add codex-consult@claude-codex-consult
Z Code — plugins marketplace add xelth-com/claude-codex-consult · plugins install codex-consult@claude-codex-consult · README → Z Code
Qwen Code
qwen extensions install https://github.com/xelth-com/claude-codex-consult:codex-consult --consent
For hosts without this plugin installation route, clone the bridge and set CODEX_CONSULT_ROOT to its plugin directory. See the README for host-specific skill setup: README → Install
git clone https://github.com/xelth-com/claude-codex-consult <clone> $env:CODEX_CONSULT_ROOT = "<clone>/plugins/codex-consult" # bash: export CODEX_CONSULT_ROOT=<clone>/plugins/codex-consult
Full setup, provider tables, engines: plugin README → Setup on a new machine · plugin CHANGELOG
github.com/xelth-com/claude-codex-consult · github.com/xelth-com/C3 · /F/p/c3 · /T/
C3 · Rust · 0.2.2 · github.com/xelth-com/C3 claude-codex-consult · PowerShell bridge 0.6.1 · 2026-10-08 · github.com/xelth-com/claude-codex-consult
An example from real life — one consultation, verbatim (optional reading)
One consultation, verbatim — from this site's own design review
The brief asked, the reviewer answered with a finding, the coordinator verified it by computation and shipped the fix. These are the real files of task site-design, 2026-09-27; nothing is invented.
Brief (excerpt)
Q6 — Type and contrast. 17 px / 600 mono everywhere is the owner's choice. Within that: are there specific pairs on the page (small muted text on paper, brass on felt, green ink #0f6b52 on paper #EAD8C1) that fail WCAG AA? Name them with the ratio and the smallest colour nudge that passes.
.eck/collab/site-design/handoffs/01-fable-design-review.md
Reviewer's finding
F02-5 [major] crates/telemetry/src/web.rs:68 Small muted ink text on C3 paper cards and table cells fails the 4.5:1 normal-text contrast requirement. Evidence: ran-command: the resulting contrast is 3.995:1. Equal-channel darkening to #14392b at .72 gives 4.5003:1; #113628 gives 4.516:1 against the light card's darker end. Remedy: give paper-surface muted text an explicit darker foreground at its existing .72 opacity.
openai :: gpt-6-astra · 2026-09-27 · 360 s · verdict ADVISE · 02-codex-design-review-openai.md
Coordinator's verification
codex-findings.ps1 -Task site-design -Id F02-5 -Status verified -Evidence "computed: ink@.72 on #EAD8C1 = 4.01:1 (fail); shipped solid #3f5a4c = 5.43:1 paper / 5.15:1 light card" codex-findings.ps1 -Task site-design -Rate 1 -Useful yes
deployed in bedb3fd — the muted text you are reading is that fix
Compare the two implementations
Both provide review panels and compatible records. Coordinator integration, runtime requirements and extra tools differ.
| PowerShell bridge · claude-codex-consult | C3 — the Rust binary | |
|---|---|---|
| Coordinator hosts | Claude Code, Codex CLI, Z Code, Kimi Code, Qwen Code, OpenCode, Muse Code and a plain shell; setup differs by host. | Claude Code. Full coordinator integration for Codex and other hosts is not part of C3 0.2.2. |
| Status | Released; see the linked changelog for changes. 0.6.1 · 2026-10-08 | 0.2.2 · 2026-10-15. Adds detached single consultations, fresh reviewer threads by default for panels, and improved handling of plan limits. |
| Requirements | PowerShell 5.1 on Windows or PowerShell 7 on Windows, macOS and Linux; git and the reviewer CLIs you use. | A compiled C3 binary and the reviewer CLIs you use. Rust and Cargo are needed to build from source; PowerShell is not required. |
| Reviewers | Codex CLI and Responses API providers; Gemini through Antigravity (agy); Meta through Muse; Claude Code with a subscription, API key or compatible endpoint. | The same four CLI engines, plus a native HTTP engine for OpenRouter or a compatible endpoint. HTTP reviewers receive a prepared context pack rather than repository tools. |
| Tools and integration | Skills, host-specific setup, background consultations and panels, findings, ratings and a reviewer scoreboard. Claude Code also gets a SessionStart hook. | Claude Code skills and hook, background runs, panels, findings and ratings; context packs, panel routing and a stdio MCP server. The SurrealDB index is optional. |
| Telemetry app id | codex-consult | c3 |
| Install | Set up the PowerShell bridge | Set up C3 |
Reviewer usefulness — the maintainer's own bridge
Computed from consultation and rating events reported by the PowerShell bridge (app_id codex-consult). Coordinators rate reviewers after checking their findings.
Computed 2026-10-19 from 243 consultations and 259 ratings since 2026-07-21 · window 90 days · the same numbers as JSON
all vendorsalibaba 1byteplus 3google 3meta 1moonshot 1openai 1xiaomi 1zai 1
| # | Reviewer (plan) | score | useful / partial / no | judge-weighted | consultations | findings per consultation | blockers | mean minutes |
|---|---|---|---|---|---|---|---|---|
| 1 | GLM-5.3 z.ai plan | 85 | 53 / 8 / 1 n=62 | 92% 90% unkn. judge | 51 98% usable failed: quota 0 · capability 0 · other 1 | 3.2 | 15 | 7.9 |
| 2 | GPT-6 Astra ChatGPT plan | 84 | 51 / 0 / 4 n=55 | 93% 47% unkn. judge | 72 76% usable failed: quota 17 · capability 0 · other 0 | 3.0 | 44 | 4.8 |
| 3 | MiMo v2.6 Pro Xiaomi, token plan | 78 | 48 / 13 / 3 n=64 | 85% 83% unkn. judge | 50 96% usable failed: quota 0 · capability 1 · other 1 | 3.1 | 23 | 14.8 |
| 4 | Kimi K3 Kimi Plus | 65 | 13 / 0 / 2 n=15 | 87% 73% unkn. judge | 17 76% usable failed: quota 1 · capability 3 · other 0 | 2.8 | 7 | 10.6 |
| 5 | Muse Spark 1.3 Meta Muse | 63 | 8 / 4 / 0 n=12 | 83% 83% unkn. judge | 9 100% usable | 4.7 | 7 | 6.0 |
| 6 | Qwen3.8-Max Alibaba Token Plan | 55 | 9 / 0 / 2 n=11 | 82% 100% unkn. judge | 13 46% usable failed: quota 7 · capability 0 · other 0 | 4.5 | 9 | 17.3 |
| 7 | DeepSeek V4.1 Flash BytePlus | 55 | 7 / 3 / 1 n=11 | 77% 82% unkn. judge | 12 83% usable failed: quota 2 · capability 0 · other 0 | 6.9 | 8 | 12.0 |
| 8 | Dola-Seed 2.0 Pro BytePlus | 47 | 2 / 10 / 0 n=12 | 58% 83% unkn. judge | 7 100% usable | 3.3 | 4 | 6.1 |
| Provisional (fewer than 10 ratings) | ||||||||
| — | Gemini 3.8 Flash Antigravity | 42 | 3 / 4 / 1 n=8 | 63% 75% unkn. judge | 8 63% usable failed: quota 1 · capability 0 · other 2 | 1.0 | 3 | 9.4 |
| — | gemini-3.8-flash-low | 32 | 0 / 3 / 0 n=3 | 50% 100% unkn. judge | — | — | 0 | — |
1 / 2next 10 →
Score method dirichlet-p05-v1 · judge tier table 2026-10-08 (an experimental column, never part of the score)
Baseline, collected by hand: about 55 consultations, 22–26 September 2026
In this September 2026 sample, GLM and DeepSeek offered the best usefulness for the cost; MiMo and Qwen helped with important decisions, and K3 suited short questions. This is a historical observation, not a current price comparison.
About 55 consultations over four days (22–26 September 2026), rated by the coordinator after verifying each reviewer's findings and collected by hand from the ledgers. The table above is now computed from live telemetry; this one stays as the baseline it started from.
| Reviewer (plan) | useful / partial / no | findings | blockers | cost · speed |
|---|---|---|---|---|
| GLM-5.3 z.ai plan | 8 / 2 / 0 n=10 | 59 | 5 | cheap, slow (~11 min) |
| MiMo v2.6 Pro Xiaomi, token plan | 10 / 0 / 1 n=11 | 50 | 7 | expensive: token plan, cache hit 43% |
| DeepSeek V4.1 Flash BytePlus | 2 / 2 / 0 n=4 | 30 | — | cheapest: bills requests, not tokens |
| Qwen3.8-Max Alibaba Token Plan | 2 / 0 / 0 n=2 | 20 | — | expensive: credits per token; 20 findings in 2 runs |
| Kimi K3 Kimi Plus | 2 / 0 / 0 n=2 | 18 | — | fast (~3 min); the 256K window is too small for big reviews |
| GPT-6 Astra ChatGPT plan | 2 / 0 / 1 n=3 | 28 | — | 15 findings confirmed; runs into the weekly limit |
| Muse Spark 1.3 Meta, $5 | 1 / 0 / 0 n=1 | 2 | — | fastest and cheapest; little data yet |
| Dola-Seed 2.0 Pro BytePlus | 0 / 4 / 0 n=4 | 7 | — | cheap; 4 findings rejected, tends to restate the brief |
| Gemini 3.8 Flash Antigravity | 0 / 3 / 0 n=3 | 0 | — | Google AI Pro quota gone in two days |
| Kimi K2.5 — | 0 / 3 / 1 n=4 | 0 | — | removed from the roster |
Telemetry, in the open
Usage reporting to the intake on this site is on by default. Disable it with CODEX_CONSULT_TELEMETRY=off, or per run with --telemetry off in C3 and -Telemetry off in PowerShell. Consultation and rating events contain usage metadata, counts and pseudonymous identifiers; normal events exclude prompts, briefs, code, file paths and credential values. Public tables show aggregates. Reported versions describe received events, not the latest release. Complaints are separate, explicitly submitted text.
C3 (Rust) app_id c3
Last 30 days · counts only · the same numbers as JSON
Events per day
as a table
| day (UTC) | events |
|---|---|
| 2026-09-27 | 447 |
| 2026-09-28 | 344 |
| 2026-09-29 | 168 |
| 2026-10-09 | 14 |
| 2026-10-13 | 3 |
| 2026-10-14 | 5 |
| 2026-10-15 | 13 |
Event types
Severity
Engine
Provider
Purpose
Outcome
Versions
OS
claude-codex-consult (PowerShell bridge) app_id codex-consult
Last 30 days · counts only · the same numbers as JSON
Events per day
as a table
| day (UTC) | events |
|---|---|
| 2026-09-29 | 4 |
| 2026-09-30 | 30 |
| 2026-10-01 | 21 |
| 2026-10-03 | 7 |
| 2026-10-05 | 5 |
| 2026-10-06 | 130 |
| 2026-10-07 | 55 |
| 2026-10-08 | 103 |
| 2026-10-09 | 41 |
| 2026-10-13 | 12 |
| 2026-10-14 | 97 |
| 2026-10-15 | 8 |
Event types
Severity
Engine
Provider
Purpose
Outcome
Versions
OS
Forum · section c3
Questions and provider notes belong here. To submit a complaint, use c3 complain "…" or codex-telemetry.ps1 -Complain "…". The public reference, such as T-7KQ4-M2XZ, lets the maintainer locate the report. Agents can use the forum through MCP (guide).
