stacklok / toolhive / 41804595430

73%

Build Type

push

github

Committed by web-flow

Commit Message

Configure Codex for Bedrock models in llm setup (#7035)

* Configure Codex for Bedrock models in llm setup

A gateway serving Bedrock's OpenAI-compatible API can't be used by
Codex as configured today: Bedrock ends the stream when a request
carries Codex's hosted web_search tool, only the closed GPT models
behind us./global. inference profiles serve the Responses API Codex
speaks, and Codex's background memory models default to OpenAI names
a Bedrock-only gateway does not serve.

With --bedrock-compat, Codex setup now pins Codex's tiers the way
Claude Code's are pinned: the newest sol model as the top-level model
(with model_context_window when the gateway reports one), terra as
memories.consolidation_model and luna as memories.extract_model, each
picked per family from the gateway listing or overridden by a
--models entry naming the family. It also sets web_search to
"disabled". The Codex helpers mirror the Claude Code ones step for
step, and per-client Bedrock discovery errors share one map.

Only --models entries naming an OpenAI GPT model count for Codex.
gpt-oss and unprefixed pins are refused, as is a listing with no
usable model: an error under --client codex, a warning and skip when
Codex was auto-detected. Other vendors' entries are left to the tools
they were meant for.

ToolHive owns these keys the way it owns Claude Code's Bedrock env
keys: a setup without Bedrock and teardown remove them whatever they
hold. Teardown, and a setup that turns Bedrock off, now warn which
keys they removed, for both Claude Code and Codex.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Split Teardown and model discovery for gocyclo

The Bedrock warnings this PR added to Teardown and discoverSetupModels
pushed both to a cyclomatic complexity of 17, over the linter's limit
of 15, failing the Lint Go Code check. Move the Bedrock discovery
warnings, the tool revert loop and the token purge notice into
helpers. Behavior and output are unchanged.

Co-Aut... (continued)

Coverage Stats

302 of 312 new or added lines in 4 files covered. (96.79%)

20 existing lines in 5 files now uncovered.

92575 of 127717 relevant lines covered (72.48%)

135.04 hits per line