Skip to content

Parallel Code Review

A coordinator agent (claude) directs two specialist agents (codex and cursor) that run concurrently. The coordinator waits for both first-pass reports, identifies gaps, and sends targeted follow-up questions for another round — iterating until satisfied, then producing a final summary.

How it works

flowchart LR
    Co[codex] -->|"review-{id}.codex"| L[leader]
    Cu[cursor] -->|"review-{id}.cursor"| L
    L -->|"review-{id}.feedback"| Co
    L -->|"review-{id}.feedback"| Cu
    L --> S[final summary]

Key characteristics:

  • Parallel — codex and cursor run simultaneously; total time is dominated by the slower of the two, not their sum
  • Iterative — the coordinator drives multiple rounds of targeted follow-up questions
  • Specialist split — codex focuses on correctness/quality, cursor on security/architecture
  • Best for: fast feedback with active human-like coordination, especially on larger diffs

Compare with Sequential Chain Review, which chains three agents in strict order and is better suited to adversarial analysis where you want reviewers to explicitly challenge each other.

Config

Add to ~/.config/hive/config.yaml:

usercommands:
  CodeReview:
    help: "parallel multi-agent review: codex + cursor coordinated by claude"
    silent: true
    windows:
      - name: leader
        focus: true
        command: |-
          {{ agentCommand }} {{ agentFlags }} 'You are coordinating a code review of the
          current branch changes.

          Two specialist agents are running in parallel in this session:
          - codex  publishes findings to: review-{{ .ID }}.codex
          - cursor publishes findings to: review-{{ .ID }}.cursor

          Your workflow:
          1. Understand the changes under review:
             git diff main...HEAD   (branch commits)
             git diff               (unstaged changes)
          2. Wait for both agents to complete their first pass:
             hive msg sub --topic review-{{ .ID }}.codex --wait
             hive msg sub --topic review-{{ .ID }}.cursor --wait
          3. Read their findings. Identify gaps, contradictions, and areas needing deeper
             investigation.
          4. Send targeted follow-up questions to each agent:
             hive msg pub --topic review-{{ .ID }}.feedback "your questions"
          5. Wait for updated responses and repeat steps 2-4 until satisfied.
          6. Produce a final written summary of all findings.'

      - name: codex
        command: |-
          codex 'You are reviewing code changes for correctness and quality.

          Your workflow:
          1. Read all changes under review:
             git diff main...HEAD   (branch commits)
             git diff               (unstaged changes)
          2. Analyse for logic correctness, error handling, naming, test coverage, and
             API design.
          3. Publish your findings:
             hive msg pub --topic review-{{ .ID }}.codex "your findings"
          4. Wait for follow-up from the leader:
             hive msg sub --topic review-{{ .ID }}.feedback --wait
          5. Address the follow-up, publish an updated report to the same topic, then wait
             again.
          Repeat steps 4-5 until no further feedback arrives.'

      - name: cursor
        command: |-
          agent 'You are reviewing code changes for security and architecture.

          Your workflow:
          1. Read all changes under review:
             git diff main...HEAD   (branch commits)
             git diff               (unstaged changes)
          2. Analyse for security issues, architectural concerns, dependency risks, and
             design patterns.
          3. Publish your findings:
             hive msg pub --topic review-{{ .ID }}.cursor "your findings"
          4. Wait for follow-up from the leader:
             hive msg sub --topic review-{{ .ID }}.feedback --wait
          5. Address the follow-up, publish an updated report to the same topic, then wait
             again.
          Repeat steps 4-5 until no further feedback arrives.'

Shell quoting

The '"'"' sequences in the commands are the POSIX shell idiom for embedding a literal single quote inside a single-quoted string (end quote → literal ' → reopen quote). They appear as an apostrophe in the prompt text that the agents receive.

Template variables

Variable Value
{{ .ID }} Session ID (6-char alphanumeric) — used as a unique namespace for topics
{{ agentCommand }} The configured agent binary (e.g. claude)
{{ agentFlags }} Any extra flags from the agent profile

{{ .ID }} ensures topic names are shell-safe even when session display names contain spaces or slashes.

Usage

In the hive TUI, select your working session and press ::

:CodeReview

Three tmux windows open in the current session. The leader window is focused — it will display coordination activity and the final summary. The codex and cursor windows each start reviewing immediately in parallel; the leader blocks until it receives both first-pass reports on their respective topics.

Customisation

Swap the specialist tools — replace codex or agent in the window commands with any CLI reviewer.

Add a third specialist — add a fourth window, give it a unique topic (e.g. review-{{ .ID }}.perf), update the leader prompt to include it in the wait list and feedback loop.

One-shot mode — remove steps 4–5 from the specialist prompts and the loop from the leader prompt to do a single pass without follow-up rounds.