Enhance Fullsend review agent with multi-model collaborative review strategy

[fullsend-ai-triage]

Context

Guannan Sun's review squad uses multiple models collaboratively on the same change, producing stronger reviews than any single-model approach. Evidence from #1525 and #3965 shows the current Fullsend review agent misses issues that multi-model review catches (e.g., cross-file impact analysis).

This issue tracks bringing that multi-model collaborative review capability into the Fullsend platform so it becomes a first-class feature — eliminating the need for an external pre-submit tool.

Goal

Enhance the Fullsend review agent so it can orchestrate multi-model collaborative review ("review squad" style) natively. When this reaches comparable quality, update contributor docs so the external review squad is no longer the recommended default pre-submit path.

Prerequisites

This work depends on multi-model / multi-runtime groundwork:

• Investigate ACP proxy chains for multi-runtime support with consistent security hooks #587 — Investigate ACP proxy chains for multi-runtime support with consistent security hooks
• Track opencode as a supported agent runtime #1260 — Track opencode as a supported agent runtime
• TranscriptHandler assumes file-based JSONL traces — needs format-neutral contract before second runtime #1935 — TranscriptHandler assumes file-based JSONL traces — needs format-neutral contract before second runtime

These must progress before the platform can host multiple models collaboratively.

Acceptance criteria

• Fullsend review agent supports multi-model collaborative review as a configurable strategy
• Review quality is comparable to the external review squad (validated against the gaps in Review agent misses cross-file impact analysis that Review Squad catches #1525 and Review agent should cross-reference sibling commands for codebase consistency when reviewing new CLI subcommands #3965)
• Contributor docs updated to remove the external review squad as the default pre-submit recommendation

Related

• Review agent misses cross-file impact analysis that Review Squad catches #1525 — review agent misses cross-file impact analysis
• Review agent should cross-reference sibling commands for codebase consistency when reviewing new CLI subcommands #3965 — review-agent gap the squad addresses
• Investigate ACP proxy chains for multi-runtime support with consistent security hooks #587, Track opencode as a supported agent runtime #1260, TranscriptHandler assumes file-based JSONL traces — needs format-neutral contract before second runtime #1935 — multi-runtime groundwork (prerequisites)
• Parent tracking: Recommend review squad pre-submit; enhance Fullsend review agent with multi-model strategy #6318

[fullsend-ai-prioritize]

RICE Priority Score: 0.25

Score breakdown

Dimension	Score	Reasoning
Reach	1.5	The review agent is used by contributors across the platform, so improving its quality benefits multiple teams and customers. The issue references gaps identified in #1525 and #3965, suggesting multiple users have encountered review quality limitations. Score of 1.5 reflects impact across multiple active user groups.
Impact	1	This is a meaningful quality improvement — the current review agent misses cross-file impact analysis and other issues that multi-model review catches. However, a workaround already exists (the external review squad tool), so this is not blocking any workflow. Score of 1 reflects a noticeable improvement with an available workaround.
Confidence	0.5	The issue is well-structured with clear acceptance criteria, specific prerequisite issues, and references to concrete review gaps. However, confidence is tempered because the scope depends on three prerequisite issues (#587, #1260, #1935) that must progress first, making timeline and feasibility uncertain. No labels, milestone, or customer-specific context to further inform scoring.
Effort	3	This is very complex work requiring architectural changes across multiple components: ACP proxy chains for multi-runtime support (#587), a new agent runtime (#1260), and a format-neutral transcript contract (#1935) — all prerequisites before the core multi-model orchestration can be built. Additionally, quality validation against known gaps and documentation updates are required. Score of 3 reflects the large scope, multiple dependencies, and architectural risk.

Formula: (1.5 x 1 x 0.5) / 3 = 0.25

[fullsend-ai-scribe]

Meeting update — 2026-08-27

Relevant to this issue: The team discussed a tiered model approach for agent workflows — using different models for different task types — which directly supports the multi-model collaborative review strategy described here.

• A tiered approach was recommended: high-level planning models (e.g., Fable) for architecture and planning, middle-tier models or Codex for code implementation, and cost-efficient models (e.g., Gemini 3.7 Flash, Groq) for code reviews and automation.
• Trade-off discussed between a single "workhorse" model (simpler, predictable cost) vs specialized model selection per task (better performance, higher complexity).
• Caution was raised against over-complicating development environments with excessive tool and model variety, which can distract from engineering tasks.
• Gemini 3.7 Flash was recommended for automation tasks due to cost efficiency, with potential stability trade-offs noted.
• Related: per-role model configuration was previously addressed (resolved in Add per-role models:/effort: to .fullsend/config.yaml with models.aliases remap #6529, Runtime parity backlog: deferred runtime/model UX items (pi Agent tool, per-role models, per-run overrides, model switching, other clouds) #6527).

Related PRs: PR #6590 (route trivial diffs to a cheaper model) implements the cost-tier routing concept discussed.

Unresolved: Whether to adopt a single workhorse model or invest in per-task model specialization remains an open question.

Meeting notes

[ascerra]

Dependency on auto-merge work

Linked from PR #2791 (ADR 0071 — Auto-merge). Before Fullsend advances an agent-driven auto-merge path, review quality needs to be strengthened and validated: an approve verdict becomes an authorization to merge rather than advisory feedback.

The work remains complementary: #2791 must still resolve its identity, credential-isolation, CODEOWNERS-scope, and merge-queue controls.