v0.2.5

Added

• Deferred tool loading: routed models no longer receive every tool
  definition on every request. A typical setup sent 153 tool definitions
  per call — grafana's 56, pentest-ai's 50, the multi-agent surface, and
  more — repeated on every turn, because the API is stateless. Codex now
  advertises a single tool_search tool; the model searches when it needs
  something and the matches arrive activated on the next request — 17
  entries per call instead of 153, measured on a live setup. The proxy
  plays the Responses backend's part of the round-trip, so MCP servers and
  multi-agent tools work through it unchanged.
• Real context windows in the picker and in Codex's catalog. Every
  model showed the same conservative 128K tag — up to 8x below the real
  limit — and that is the number published into Codex's catalog, so the
  agent compacted conversations far earlier than it needed to. Fetch
  models now learns the real window from the provider's own catalog when
  it publishes one, enriches from the public models.dev catalog when it
  doesn't, and remembers what it learned per model.
• OpenCode Go presets. The low-cost Go subscription is the same
  opencode.ai gateway under a different path, with the same dialect split
  as Zen. A Go key only gets a 401 on the Zen endpoint, so picking Zen
  with a Go key was a dead end.

Fixed

• Routed models can finally use MCP servers and multi-agent tools. The
  translator only forwarded one of the five tool shapes Codex sends and
  silently dropped the rest: of the 23 tool entries in a real request, 12
  survived. Gone with the dropped ones were the entire multi-agent
  surface, apply_patch, and every configured MCP server — which is why
  this presented for months as "routed models can't use MCP" rather than
  as an error.
• Spawned agents now receive their task. The multi-agent toggle wrote
  a flag Codex does not read for the surface it promised, so the model had
  no spawn tool and did everything itself; and when spawning did work, the
  child agent's task arrived in a shape the translator discarded, leaving
  it with environment and instructions but nothing to do. The toggle now
  writes the flag Codex actually reads — restart Codex after flipping it,
  it reads these flags at process start — and the task body reaches the
  child.
• Tool-using turns no longer fail against strict providers on macOS.
  The desktop app interleaves messages around parallel tool calls in a way
  the translator used to split apart, so strict upstreams rejected every
  tool-using turn on a Mac while the same setup worked on Windows.
• Applying or removing the Codex integration no longer dead-ends on
  installs left by older versions. A config written before the
  managed-block markers, or one whose ending marker the Codex desktop app
  dropped, made both apply and remove fail silently. Both shapes are now
  detected by ownership and migrated, and failures surface as errors
  instead of silence.
• The step that asks your shell where Codex lives now actually asks an
  interactive shell. zsh is the macOS default and only reads .zshrc for
  interactive shells — .zprofile and .zlogin cover login ones — so a
  login-only probe returned nothing on the setup it was meant to rescue,
  which is where most people put their PATH. Measured on a machine whose
  PATH lives in .zshrc: the login-only probe found nothing, the
  interactive one found the CLI. 0.2.4 still worked there because the
  known-locations fallback caught it; this makes the shell step carry its
  weight, which is what matters when Codex is installed somewhere unusual.
  The probe is bounded by a deadline so a slow shell profile cannot hang
  the screen that waits on it.