[elie222]

Groq reasoning models could receive a developer-role instruction that their chat template rejected. Requests could also exceed an account's token allowance or reserve more output than its available OpenRouter credit, even when a smaller request could run. Switching to a Google model after a tool call could also replay history without the signature its API requires.

Use system instructions for Groq. Before any output, recover once from either of two specific provider rejections: rebuild a Groq request within its reported token ceiling using the existing context policy, or reduce an OpenRouter output reservation to its reported positive allowance. Keep adjustments local to the run/model, preserve saved settings and output ceilings, and share the original request deadline. Both primary and helper calls use this path. Empty balances, key limits, unrelated errors, local reservation failures, and partial text/tool output retain their existing behavior.

For transferred tool history, use Google's documented transfer marker only on foreign calls whose signatures Pi discarded. Preserve native signatures, unsigned native parallel parts, original history, and existing payload hooks; keep completed results so tools are not repeated. An offline SDK test checks the literal outgoing wire format for both Google-to-Google and cross-provider transfers.

Recovery emits model.context_limit.reduced or model.output_reservation.reduced with provider/model, run correlation, and numeric limits only. It records no prompts, responses, tool arguments, account identifiers, or credentials. No UI copy or layout changes.

Validation: offline outgoing-wire tests reproduce Groq's previous instruction role. Runtime, limit, protocol, and offline integration tests pass, including actual context reduction, subsequent tool turns, bounded retries, deadlines, local reservations, partial output, and credential isolation. Workspace type checks and lint pass with existing warnings. Independent reuse, quality, and efficiency review completed; shared policy construction was consolidated. All fixtures are synthetic and no paid inference was used.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
rakazo-www	Skipped		Oct 23, 2026 1:54pm UTC

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new actionable issues were found.

Summary

The PR adds bounded recovery for Groq request limits and OpenRouter output allowances. It also uses system instructions for Groq and preserves transferred Google tool history.

• Model requests recover from specific Groq and OpenRouter limits.
• Google requests keep transferred tool history usable.
• Groq requests send instructions as system messages.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Prepare model request] --> B[Send request]
  B --> C{Provider rejects before output?}
  C -->|Supported Groq limit| D[Reduce context for this run]
  C -->|Supported OpenRouter allowance| E[Reduce output reservation]
  D --> F[Retry once within original deadline]
  E --> F
  C -->|Other error or partial output| G[Keep existing error behavior]
  B --> H[Continue model and tool turns]
  F --> H

Loading

_{Reviews (3) · Last reviewed commit: "Use SDK JSON types in Google history fix..." · Reviewed by Greptile}

[github-actions]

Playwright screenshots

No new feature frames; gallery is suite-vs-main drift.

Open screenshot gallery · Dashboard · CI run

Updated for commit 13df367.