The previous assertion read the generated CLIProxy config.yaml directly via
getConfigPathForPort(). On the self-hosted CI runner this produced ENOENT even
though applyCliproxyRoutingStrategy returned 'config-only', indicating the
read path and write path diverged in the full test-suite context.
Verify persistence via loadUnifiedConfig() instead — that's the canonical
source mutateUnifiedConfig writes to, independent of regenerateConfig's
file-path resolution.
The SearXNG-to-DuckDuckGo fallback test inherited the caller's CCS_PROFILE_TYPE
via `...process.env`, causing the hook to take the 'native_default_profile'
skip path and produce empty stdout. All other subprocess spawns in this file
set CCS_PROFILE_TYPE to NEUTRAL_PROFILE_TYPE; this one was the outlier.
Deterministic locally when CCS_PROFILE_TYPE=default is set in the shell.
Use ComSpec-aware shell selection for escaped wrapper launches.
Apply it to the remaining Windows Claude launch paths.
Rewrite the Codex exec regression test to use scoped spies so the full suite stays isolated.
Align Windows .cmd/.bat launch behavior with the Codex detector shell contract.
This keeps escaped -c overrides intact through wrapper execution.
Adds regression coverage for the Codex runtime path and the shared Windows shell launch path.
- enforce strict tool_result ordering and pairing against assistant tool_use ids
- reject tool_result image payloads that cannot map to OpenAI tool messages
- preserve raw tool schemas on the /v1/messages proxy path instead of silently tightening them
- forward Anthropic tool_choice semantics and cover adaptive routing plus upstream payload checks
- capture the resolved local cliproxy config/auth paths before regenerating
routing strategy config
- thread explicit override paths into regenerateConfig so scoped config writes
stay deterministic during the full test suite
- add regression coverage for explicit config path overrides
- skip managed Claude browser attach when the default CCS browser profile
directory has not been created yet
- keep explicit env override flows failing fast so real misconfiguration still
surfaces clearly
- add regression coverage for default and settings-profile launches
- derive repo imports from each test file via file URLs
- avoid hardcoded paths in the repo
- launch child scripts with the current runtime for consistent local and runner behavior
- isolate tokens and session-tracker tests in child processes
- make child scripts resolve repo modules from the test location
- avoid machine-specific paths in the repo
- skip CLIProxy auto-update checks on runtime bootstrap paths
- fail fast when local startup needs a missing binary instead of attempting installs
- add regression coverage for dashboard limited mode and startup test isolation
Claude Code's marketplace auto-update uses a three-step atomic rename:
clone to <name>.staging → rename old dir to <name>.bak → rename staging
to the final name. On Windows, EPERM errors can interrupt the rename-dance
and leave a <name>.staging directory permanently on disk.
CCS scanned physical directories to discover marketplaces, so a lingering
.staging directory was registered as a bare { installLocation } entry in
known_marketplaces.json. Claude Code's Zod schema requires each entry to
also have `source` and `lastUpdated` fields, so the corrupt entry caused
the /plugin command to throw a validation error.
Fix (two layers of defence):
- discoverMarketplaceEntries: skip any directory whose name starts with
'.' or ends with '.staging'. This filters all hidden dirs including
.staging and .bak left behind by interrupted rename operations.
- buildMarketplaceRegistryContent: invert the loop direction. Instead of
iterating discoveredEntries and adding to the registry (which could
introduce bare entries for disk-only directories with no registry
record), iterate the merged registry and only keep entries that also
exist on disk. Disk-only directories are now silently ignored.
Adds two regression tests: one for .staging pollution, one for orphan
registry entries whose physical directory has been removed.