v0.4.5

Fixed

• MCP tools and the built-in browser now work in Codex / ChatGPT desktop app with non-native models (#21) — Codex wraps MCP server tools (and the built-in browser, which is itself exposed as an MCP tool) in a proprietary {type:"namespace"} envelope. relay-ai already flattened these into callable function tools for the model, but sent the model's tool call back to Codex as a flat name instead of splitting it back into the {namespace, name} shape Codex's own dispatcher requires, so every call was rejected with unsupported call: .... Also added translation for tool_search (deferred/lazy-loaded tools), additional_tools turn-local tool definitions, and type:"custom" tools (e.g. apply_patch). Credit to bharat2808/codex-ollama-proxy — a community proxy for Codex + Ollama that solved this same class of problem first and whose approach to flattening/splitting namespace tools informed this fix.
• Fatal error: remote compaction v2 expected exactly one compaction output item no longer crashes Codex / ChatGPT app sessions — Codex's "remote compaction v2" requires the response to contain exactly one output item of type compaction; relay-ai was replying with a normal reasoning+message turn, which Codex rejected outright. relay-ai now asks the model for a plain-text summary and returns it wrapped as the single compaction item Codex expects, decoding it back into readable context on later turns.
• Compaction no longer fires prematurely in the ChatGPT desktop app — a single large tool result (e.g. a browser snapshot) could push a session past relay-ai's auto-compact threshold within a couple of turns. Raised model_auto_compact_token_limit from 55% to 90% of the model's context window.

Added

• Untruncated Codex request/response trace dump — relay-ai codex/codex-app --trace now writes full, unclipped /v1/responses request and response bodies (tools array, every input item) to ~/.relay-ai/logs/codex-body-dump.jsonl, for diagnosing Codex tool-shape and compaction issues without reproducing twice.

Documentation

• Updated the MCP known-limitation notes in README.md/docs/CODEX.md to reflect the fix above, corrected stale auto-compaction figures, and documented a new known limitation: image_gen (image generation) calls an OpenAI-only image backend relay-ai doesn't implement, and most registry models can't generate images regardless — no workaround planned.