fix(remote-sandbox): harden host workspace resumes (#5922)

## Thinking Path

> - Paperclip orchestrates AI agents through a control plane while
adapters execute work in local, remote, or sandboxed runtimes.
> - Remote sandbox execution depends on a strict host-versus-remote
workspace boundary: the host prepares/restores files, while the adapter
command runs inside the sandbox cwd.
> - Jannes' PR #5823 identified host-side failure modes that were not
covered by replacement PR #5822.
> - Persisting a remote pod cwd in session params could poison the next
host heartbeat resume and make Paperclip inspect or upload system temp
roots.
> - Plugin sandbox providers also need a narrow way to receive
model-provider API keys without exposing the full server environment to
every plugin worker.
> - This pull request ports the host-side fixes from #5823 in the
current codebase style, with focused regression coverage.
> - The benefit is safer remote sandbox resumes and plugin worker
environment handling without broadening core plugin privileges.

## What Changed

- Persist host workspace cwd, not remote sandbox cwd, in `claude_local`
session params while retaining remote execution identity metadata.
- Reject saved session cwds that point at system roots before heartbeat
falls back to agent home workspace.
- Skip sockets, FIFOs, devices, and other non-file entries during
workspace restore snapshot capture/comparison.
- Pass a small model-provider API-key allowlist only to plugins
declaring `environment.drivers.register`.
- Added focused regression tests for remote Claude session params,
unsafe session cwd detection, plugin worker env filtering, and non-file
snapshot entries.

Credits: ports host-side fixes from Jannes' #5823.

## Verification

- `pnpm vitest run
packages/adapter-utils/src/workspace-restore-merge.test.ts
server/src/services/session-workspace-cwd.test.ts
server/src/__tests__/claude-local-execute.test.ts
server/src/__tests__/plugin-database.test.ts` (25 passed, 7 skipped by
existing embedded-Postgres host guard)
- `pnpm --filter @paperclipai/adapter-utils typecheck`
- `pnpm --filter @paperclipai/adapter-claude-local typecheck`
- `pnpm --filter @paperclipai/server typecheck`

## Risks

- Low risk: changes are scoped to remote sandbox/session metadata,
workspace snapshot filtering, and plugin worker env setup.
- Sandbox-provider plugins now receive only the explicit model-provider
key allowlist; any provider needing another key name will need a
deliberate allowlist update.

> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.

## Model Used

- OpenAI Codex, GPT-5-based coding agent, tool-enabled local code
execution and repository editing.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>

packages/adapters/claude-local/src/server/execute.ts

@@ -89,6 +89,15 @@ interface ClaudeRuntimeConfig {
   extraArgs: string[];
 }
 
+export function claudeSessionCwdMatchesExecutionTarget(input: {
+  runtimeSessionCwd: string;
+  effectiveExecutionCwd: string;
+  executionTargetIsRemote: boolean;
+}): boolean {
+  if (input.executionTargetIsRemote || input.runtimeSessionCwd.length === 0) return true;
+  return path.resolve(input.runtimeSessionCwd) === path.resolve(input.effectiveExecutionCwd);
+}
+
 function buildLoginResult(input: {
   proc: RunProcessResult;
   loginUrl: string | null;
@@ -591,7 +600,11 @@ export async function execute(ctx: AdapterExecutionContext): Promise<AdapterExec
   const canResumeSession =
     runtimeSessionId.length > 0 &&
     hasMatchingPromptBundle &&
-    (runtimeSessionCwd.length === 0 || path.resolve(runtimeSessionCwd) === path.resolve(effectiveExecutionCwd)) &&
+    claudeSessionCwdMatchesExecutionTarget({
+      runtimeSessionCwd,
+      effectiveExecutionCwd,
+      executionTargetIsRemote,
+    }) &&
     adapterExecutionTargetSessionMatches(runtimeRemoteExecution, runtimeExecutionTarget);
   const sessionId = canResumeSession ? runtimeSessionId : null;
   if (
@@ -853,7 +866,7 @@ export async function execute(ctx: AdapterExecutionContext): Promise<AdapterExec
     const resolvedSessionParams = resolvedSessionId
       ? ({
         sessionId: resolvedSessionId,
-        cwd: effectiveExecutionCwd,
+        cwd,
         promptBundleKey: promptBundle.bundleKey,
         ...(executionTargetIsRemote
           ? {

packages/adapters/claude-local/src/server/index.ts

@@ -1,4 +1,4 @@
-export { execute, runClaudeLogin } from "./execute.js";
+export { claudeSessionCwdMatchesExecutionTarget, execute, runClaudeLogin } from "./execute.js";
 export { listClaudeSkills, syncClaudeSkills } from "./skills.js";
 export { listClaudeModels } from "./models.js";
 export { testEnvironment } from "./test.js";