Support alternative CLI backends (Gemini CLI) for cross-model second opinions

[korabeland]

Problem

12+ gstack skills hardcode codex exec as the cross-model second opinion engine:

• /codex, /review, /ship, /plan-eng-review, /plan-ceo-review, /plan-design-review, /design-consultation, /design-review, /office-hours, /autoplan

These calls use Codex-specific flags (-s read-only, -c 'model_reasoning_effort=...', --enable web_search_cached, resume <session-id>) that have no direct equivalent in other CLIs.

Users who don't have (or don't pay for) OpenAI Codex CLI but do have Google Gemini CLI installed currently have to maintain a translation shim to use these skills. The shim works for basic calls but can't fully translate:

• codex exec resume <session-id> (session continuity — no Gemini equivalent)
• -c 'model_reasoning_effort=...' (dropped silently)
• --enable web_search_cached (dropped — Gemini has built-in search but no equivalent flag)
• --json JSONL output mode (Gemini doesn't support structured trace output the same way)

Related

• office-hours: Phase 3.5 should fall back to Claude subagent when Codex is unavailable #463 addresses fallback to Claude subagent when Codex is missing, but only for /office-hours
• Several skills already have CODEX_NOT_AVAILABLE detection but handle it inconsistently (some skip silently, some fall back to Agent tool)

Proposal

Add a configurable "outside voice" backend in ~/.gstack/config.yaml:

outside_voice:
  backend: gemini  # or: codex, claude-subagent
  # backend-specific config inherited from each CLI's own auth

Then in skill files, replace direct codex exec calls with an abstraction that dispatches to the configured backend. Each backend adapter handles flag translation:

Codex flag	Gemini equivalent	Claude subagent
-s read-only	--sandbox	Agent tool (read-only by default)
-c 'model_reasoning_effort=...'	n/a (drop)	model param
--enable web_search_cached	built-in (drop)	WebSearch tool
resume <session-id>	n/a	SendMessage to existing agent
--json (JSONL traces)	n/a	parse agent output

This would:

1. Let users choose their preferred second-opinion CLI
2. Unify the inconsistent CODEX_NOT_AVAILABLE handling across skills
3. Make the "outside voice" pattern a first-class gstack concept rather than a Codex-specific feature

Workaround

A bash shim at ~/.local/bin/codex that intercepts codex exec and translates to gemini -p works for most cases but breaks on session continuity and JSONL output parsing.

Environment

• macOS (Apple Silicon)
• @google/gemini-cli@0.35.1 via npm
• gstack 0.13.0.0

[drandrewlaw]

I just had a real-world experience with this in an /office-hours session that shows why multi-model second opinions are valuable.

What happened: I ran Codex AND Gemini CLI as independent cold reads during Phase 3.5. They genuinely disagreed on a key premise (whether IoTeX was the right wedge chain for a DeFi monitoring product):

• Codex: "IoTeX is too small to learn from. Go to Ethereum/Base/Solana where there are users."
• Gemini: "IoTeX is the ONLY place to prove the physical+digital DePIN thesis. On Solana you're just another alert tool."

This forced a real decision that a single model wouldn't have surfaced. Three perspectives (Claude + Codex + Gemini) produced a genuinely better outcome than any single model alone.

How I ran it: After the existing Codex Phase 3.5 flow, I manually ran Gemini via gemini -p "$(cat prompt_file)" with the same structured context block, adapted for Gemini's different strengths. The integration was straightforward.

Key CLI differences I hit:

Feature	Codex	Gemini
Prompt flag	codex exec "prompt"	gemini -p "prompt"
Read-only	-s read-only	--approval-mode plan
Working dir	-C dir	No equivalent (run from cwd)
Output	Plain text stdout	Works fine with -p

I'd like to implement this. The resolver pattern in scripts/resolvers/review.ts is clean — adding generateGeminiSecondOpinion() alongside the existing Codex resolver is straightforward. The test infrastructure already exists (test/helpers/gemini-session-runner.ts, test/gemini-e2e.test.ts).

Proposed approach:

1. Add GEMINI_SECOND_OPINION resolver that auto-detects which gemini
2. User gets asked which model(s) to use when both are available (Codex, Gemini, or both)
3. When both run, add a cross-model synthesis section comparing all three perspectives
4. Respects the existing outside_voice config pattern proposed in this issue
5. Apply the same pattern across the 12+ skills that currently hardcode Codex

Refs: #772 (auto-detect), #215 (Gemini CLI support)