💡 Gemini Engine Model Currency Refresh: Migrate Deprecated gemini-2.0-flash References

[github-actions[bot]]

Summary

engine.sh still configures gemini-2.0-flash as the Gemini triage model and the Copilot engine's rubber-duck model, but Google deprecated gemini-2.0-flash on February 18, 2026, with full shutdown on June 1, 2026. This means the Gemini triage tier and the Copilot cross-engine rubber-duck review have been silently broken for 12 days. Upgrade to Gemini 2.5 Flash (or 3.5 Flash) and add Gemini model chains for rate-limit resilience.

Market Signal

Google deprecated Gemini 2.0 Flash on February 18, 2026 and shut it down on June 1, 2026. Gemini 2.5 Flash is the current cost-efficient tier at approximately $0.15/$0.60 per MTok (with a 50% batch discount to $0.075/$0.30). Gemini 3.5 Flash is the newest Flash model. Google also enforced mandatory spending caps across all billing tiers in April 2026 and prepaid billing for new accounts, changing the Gemini cost model significantly. For prompts exceeding 200K tokens, Gemini 2.5 Pro and 3 Pro input pricing roughly doubles.

User Signal

engine.sh references gemini-2.0-flash in three places: line 79 (Gemini triage model), line 109 (Copilot engine label), and line 113 (Copilot duck model). validate-engines.sh performs a billing probe using a gemini-2.0-flash REST call but only catches depleted credits, not deprecated models — the probe itself may be receiving HTTP errors that are not diagnosed as deprecation. The Copilot engine's DUCK_MODEL is set to gemini-2.0-flash, meaning cross-engine adversarial review has been silently degraded since June 1 whenever the Copilot engine is primary.

Technical Opportunity

The fix is bounded: update three model strings in set_engine_config(), add pricing rows to model-pricing.tsv, and optionally add a Gemini model chain (GEMINI_TRIAGE_MODEL_CHAIN) mirroring the Claude chain pattern for rate-limit resilience. validate-engines.sh could also add a model-availability probe (test the configured model ID, not just the billing state) to catch future deprecations proactively.

Specific changes:

• Line 79: gemini-2.0-flash → gemini-2.5-flash
• Line 109: Update label string
• Line 113: gemini-2.0-flash → gemini-2.5-flash
• model-pricing.tsv: Add gemini-2.5-flash and gemini-3.5-flash rows
• validate-engines.sh: Update the billing probe model from gemini-2.0-flash to the active model

Assessment

Dimension	Score	Rationale
Feasibility	high	~10 lines of shell changes + 2 pricing table rows; no architectural changes needed
Impact	med	Restores three-provider resilience; fixes Copilot duck review degradation; updates pricing accuracy
Urgency	high	The configured model has been dead since June 1 — 12 days of silently degraded fallback capacity

Adversarial Review

Strongest objection: validate_engines.sh already catches Gemini unavailability and marks GEMINI_AVAILABLE=false, so the system degrades gracefully. The cross-provider fallback (claude → copilot → gemini) simply skips Gemini, leaving two functional engines. This is a low-risk configuration issue, not a strategic concern.

Rebuttal: Graceful degradation is not the same as proper function. The project invested in three-provider resilience specifically to avoid single-provider dependency. If Gemini fallback is permanently disabled because the model reference is dead, the system has silently lost one-third of its fallback capacity. The fix is ~10 lines of shell and two pricing table rows — the effort-to-impact ratio is extremely favorable. Additionally, the Copilot duck engine's use of gemini-2.0-flash means cross-engine adversarial review quality has been degraded since June 1, reducing review diversity exactly when the review pipeline needs it most.

Suggested Next Step

Update scripts/engine.sh set_engine_config(): replace gemini-2.0-flash with gemini-2.5-flash in the gemini and copilot engine blocks (lines 79, 109, 113). Add gemini-2.5-flash and gemini-3.5-flash rows to scripts/lib/model-pricing.tsv. Update the billing probe model in validate-engines.sh. Optionally add a model-liveness check that tests the configured model ID, not just the billing state.