stacklok / toolhive / 42001429620

73%

Build Type

push

github

Committed by web-flow

Commit Message

Leave user-owned Bedrock keys alone in llm setup (#7051)

* Leave user-owned Bedrock keys alone in llm setup

A setup without --bedrock-compat, and teardown, removed Claude Code's
and Codex's Bedrock keys even when ToolHive never wrote them, so a plain
"thv llm setup" deleted a user's own Codex model and turned web search
back on. Each tool record now says whether its setup wrote the Bedrock
keys, and setup and teardown remove them only then. The record is per
tool because "thv llm config set" can change the persisted flag without
touching any tool's files. Records from older versions fall back to the
persisted flag and are pinned to it before the flag changes.

Also address the rest of the review on #7035:
- Leave openai.gpt-oss-* --models entries to other tools (Pi) instead
  of refusing Codex, the way nvidia.* and qwen.* entries already are.
- Recognise Codex model IDs case-insensitively, and refuse a mixed-case
  pin with its lowercase form instead of a misleading error.
- Return an error, not a warning, when Bedrock settings fail for a
  client named with --client.

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

* Split two llm setup functions under gocyclo limit

The previous commit's Bedrock key bookkeeping pushed
applyLLMGatewayKeys and configureDetectedToolsWithDiscovery to a
cyclomatic complexity of 16, one over the linter's limit of 15, and
failed CI lint. Move the spec value resolution and the Anthropic base
URL computation into their own helpers. Behavior is unchanged.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Coverage Stats

98 of 105 new or added lines in 8 files covered. (93.33%)

5 existing lines in 2 files now uncovered.

93736 of 128844 relevant lines covered (72.75%)

851.18 hits per line