Handoff MCP v0.31 through v0.32 are out. The theme is convergence control for multi-agent quality assurance.

The "Review That Never Ends" Problem

The session-loop's basic pattern is straightforward: AI agents write code, then other agents review and test it. But as automation deepens, inefficiencies that also plague human teams get amplified by parallel agent execution:

Late-arriving findings: Round 2 introduces entirely new categories of criticism, leaving developers chasing a moving target.
Everyone restarts: A single developer's failure triggers a full-team restart, re-running agents whose work was already accepted.
Endless rounds: Trivial findings trigger full review cycles, burning through tokens.

These are structural problems in any review process, but when each round costs real token budget across parallel agents, the need for control is acute.

All Findings Up Front

Starting in v0.31.1, reviewers and testers must surface every category of finding in the first round. Subsequent rounds are limited to verifying prior fixes and catching issues introduced by the rework itself.

This eliminates the "moving goalpost" effect and lets developer agents focus purely on fixing what was already identified.

Scoped Rework

Previously, any test or review failure restarted every developer agent. With v0.31.2's Owner-limited rework, only the developers whose tasks have findings are relaunched. Completed reports from unaffected developers are preserved, eliminating redundant re-execution entirely.

Reviewer Micro-fixes

When a reviewer finds something at the level of an indent fix or a typo, sending it back to a developer is overkill. v0.31.2's Reviewer micro-fix lets the reviewer agent apply small corrections directly, with test evidence attached.

A Micro-safe boundary keeps this power in check: file count limits apply, and changes to migrations, lockfiles, CI config, and authentication paths are forbidden.

Session Termination for Non-convergence

Sessions that exceed the maximum round count now file each unresolved finding as a backlog task and terminate cleanly. The problem record and next action are preserved without falling into an infinite loop.

Automatic Document Reconciliation

Code changes often require corresponding updates to specs and guides. The Doc reconciler agent detects documentation impacts from code changes and either reports the diff in audit (read-only) mode or applies it directly in apply (single-writer) mode. A safety rule prevents spec auto-justification — rewriting specs just to match changed code.

Correct Handling of Staged Implementation

When a feature is split across multiple tasks, intermediate tasks naturally have unwired parts. In v0.31.0, reviewers and testers now confirm that a downstream task and design document both name the staging as intentional before downgrading a finding. If that confirmation is missing, it remains a defect.

Other Changes

Finding closure states — Four terminal states (fixed, accepted_in_session, observed_out_of_scope, deferred_blocked) for structured finding outcomes
Follow-up creation gate — Prevents deferring blocking findings, enforces a 2-per-session limit, and checks for duplicates
Session repairer agent — Fixes individual findings with targeted verification scoped to affected_paths only
Stage telemetry — Workflow results carry per-agent metadata: round, stage, role, model, effort, verdict, crash status
Enhanced finding schema — affected_paths, repair_class, verification_commands, and doc_impacts fields
include_dependents flag — handoff_get_task optionally returns downstream tasks that depend on the queried task

Install / Update

npx handoff-mcp-server@latest

If you're using Handoff as a Claude Code plugin:

/plugin install alphaelements/handoff-mcp

See Handoff case studies