Agent Review Workflow

How coding and autonomous agents discover and handle reviews on your behalf

Coding agents (Claude Code, Cursor, Windsurf, etc.) and autonomous agents (OpenClaw, Hermes) can manage the review process for you. When a curation triggers a review, the agent sees the pending changes and can approve or reject them — with your permission.

How agents discover reviews

There are two paths depending on whether the agent waits for curation to complete.

Inline notification (default)

When an agent runs brv curate without the --detach flag, it waits for the curation to finish. If any operations are flagged for review, the review summary is printed directly in the command output:

✓ Context curated successfully. (Log: cur-1775299176764)

⚠  1 operation requires review (task: 30555b0f-b593-4c01-af3f-503f84f5a6bc)

  [UPSERT · HIGH IMPACT] - path: infrastructure/caching/caching_strategy.md
  Why:   Update caching strategy: replacing Redis with Memcached
  Before: Unified caching strategy with in-memory (L1) and Redis (L2) layers
  After:  Replaced Redis with Memcached, 10-minute TTL, LRU eviction

  To approve all:  brv review approve 30555b0f-b593-4c01-af3f-503f84f5a6bc
  To reject all:   brv review reject 30555b0f-b593-4c01-af3f-503f84f5a6bc
  Per file:        brv review approve/reject 30555b0f-... --file <path> [--file <path>]

The output includes:

• The task ID
• Each flagged operation with its type, impact, reason, and before/after summaries
• The exact brv review approve and brv review reject commands

The agent reads this output, presents the changes to you, and asks for your decision.

Detached mode (--detach)

When an agent runs brv curate --detach, the command queues the curation and returns immediately without waiting for completion:

✓ Context queued for processing. (Log: cur-1775299231152)

Since the agent doesn't wait, it never sees the inline review notification. Instead, it uses brv review pending to check for unresolved reviews after the task completes:

# Agent runs curation in detached mode
brv curate --detach "Update: switching from Redis to Memcache for caching"

# Later, agent checks for pending reviews
brv review pending

# Agent acts on the results
brv review approve <taskId>

This is the standard pattern for MCP-based integrations, where the coding agent calls brv-curate (which uses detached mode internally) and then follows up with brv review pending.

The agent review loop

A typical agent-driven review follows this sequence:

1. Agent runs brv curate (or brv-curate via MCP)
       │
2. Agent checks for pending reviews
   (inline output or brv review pending)
       │
3. Agent presents changes to you:
   "This curation wants to update the caching strategy
    from Redis to Memcache. Approve or reject?"
       │
4. You respond ("approve", "reject", or give feedback)
       │
5. Agent runs brv review approve/reject <taskId>
       │
6. Done — changes are finalized or rolled back

Agent review in practice

Any agent with a ByteRover connector (skill, MCP, rules, or hook) follows the same pattern. The connector teaches the agent the brv review commands, so it can:

• List pending reviews with brv review pending
• Approve or reject with brv review approve/reject <taskId>
• Selectively approve/reject files with --file
• Use --format json for structured parsing

The agent reads the review summary, presents the changes to you conversationally, and asks for your decision before running the approve or reject command.

Example — OpenClaw handling a review:

Step 1: User asks OpenClaw to curate

Remember this: we're switching from Redis to Memcached for all caching.
New TTL is 10 minutes with LRU eviction.

Step 2: OpenClaw curates and discovers a pending review

• OpenClaw — I've curated the caching strategy update, but ByteRover flagged it as a high-impact change that needs your review.

  Operation	UPSERT · HIGH IMPACT
  File	infrastructure/caching/caching_strategy.md
  Why	Update caching strategy: replacing Redis with Memcached
  Before	Unified caching strategy with in-memory (L1) and Redis (L2) layers
  After	Replaced Redis with Memcached, 10-minute TTL, LRU eviction

  This changes your project's caching approach from Redis to Memcached. Should I approve or reject this change?

Step 3: User responds

Approve it

Step 4: OpenClaw approves the change

• OpenClaw — Done — approved infrastructure/caching/caching_strategy.md. The change is now finalized and eligible for push.

Detach mode reference

Scenario	How the agent discovers reviews
brv curate (no flags)	Review summary printed in stdout after task completion
brv curate --detach	Agent calls brv review pending to check after the fact
MCP brv-curate tool	Curation is queued (detached); agent calls brv review pending

NOTE

The --detach flag is useful when you don't want the agent to block while curation processes. The tradeoff is that the agent must explicitly check for reviews afterward using brv review pending.

TIP

If your agent runs unattended and you don't want to review changes at all, you can disable the HITL review log for the current project with brv review --disable. See Disable the review log for the full behavior.

Connector setup

To enable agent-driven reviews, your agent needs a ByteRover connector installed. The connector teaches the agent the brv review commands alongside brv query and brv curate.

# Install a connector for your agent
brv connectors install "<agent name>"

For setup instructions, see Coding Agent Connectors and Autonomous Agents.