Claude Code sends the system prompt as the top-level `system` field and,
separately, sends skill/plugin listings as `role: "system"` entries inside
`messages` (#1459 made the transformer accept those). `transform()`
unconditionally prepends the top-level field, so once both are present the
OpenAI-compat payload ends up with two `system` messages that are not
adjacent. `coalesceMessages` only merges consecutive same-role messages and
explicitly skips `system`, so it cannot fix this.
Strict OpenAI-compatible backends (LiteLLM among them) reject that shape with:
400 A 'system' message can only appear at index 0 of the messages array.
Add `hoistSystemMessages`, run before `coalesceMessages`, which extracts every
`system` message in encounter order and reinserts a single merged one at
index 0. Content-preserving, no behavior change when at most one system
message is present.
The same atomic-rename divergence occurs in the plugins subtree: a
plugin install inside a session rewrites plugins/installed_plugins.json,
replacing the instance-level symlink with a regular file. The per-launch
relink then discarded it, so the plugin was effectively uninstalled on
every relaunch while settings.json still marked it enabled — sessions
then fail with 'Unknown skill: <plugin>:<skill>'.
Move adoptDivergedFileContent to fs-helpers (avoids a circular import)
and apply it to file-type plugin entries before re-linking.
Claude Code saves settings.json atomically (temp file + rename), which
replaces the managed shared symlink with a regular file holding the
user's latest changes (see #57). The launch-time relink then deleted
that file without reading it, silently reverting plugin enables and any
other in-session settings change on every profile relaunch.
Adopt the diverged file's content into the canonical ~/.claude file
(with a .bak-ccs-adopt backup) before restoring the symlink, at both
the shared-level and instance-level reconciliation points.
Fixes#1681
github-actions[bot]