skills/local-agent-delegation/SKILL.md (1)

38-51: 🔒 Security & Privacy | 🔵 Trivial

Flag shell-injection risk for any real implementation of this guidance.

The CLI examples interpolate $PROMPT directly into shell command strings (e.g., codex exec --json -C "$WORKSPACE" "$PROMPT"). This is fine as illustrative documentation, but if any actual tool implementation follows this literally with shell-string concatenation rather than passing arguments as an array/exec without a shell, untrusted prompt content could lead to command injection. Worth calling out so future implementers pass $PROMPT/$WORKSPACE as separate argv entries (no shell) rather than building a shell string.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@skills/local-agent-delegation/SKILL.md` around lines 38 - 51, The CLI
guidance examples in SKILL.md should explicitly warn that directly interpolating
$PROMPT or $WORKSPACE into a shell command string is unsafe for real
implementations. Update the CLI execution guidance near the examples to call out
shell-injection risk and instruct implementers to pass arguments as separate
argv entries without a shell, rather than concatenating strings, while keeping
the examples clearly documented as illustrative only.

src/user-config.ts (1)

95-103: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

devspace init --force doesn't refresh the seeded skill.

ensureDevspaceDefaultSkills never overwrites an existing local-agent-delegation/SKILL.md, even when the CLI is run with --force (the function ignores force entirely — see src/cli.ts Line 144). As the bundled skill content evolves in future releases, existing installs will silently keep stale guidance unless users manually delete the file. Consider threading a force/version-stamp option through so --force can refresh the seeded copy.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/user-config.ts` around lines 95 - 103, ensureDevspaceDefaultSkills
currently short-circuits when local-agent-delegation/SKILL.md already exists, so
seeded skills never refresh on devspace init --force. Update the
ensureDevspaceDefaultSkills flow to accept and honor a force/version-based
refresh signal from the CLI (via src/cli.ts) and overwrite the bundled SKILL.md
when refresh is requested, while preserving the existing no-op behavior when it
is not. Keep the fix centered around ensureDevspaceDefaultSkills,
devspaceSkillsDir, and the local-agent-delegation/SKILL.md write path.

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

src/skills.ts (1)

63-79: 🎯 Functional Correctness | 🔵 Trivial | 💤 Low value

Diagnostics aren't filtered alongside the excluded skill.

When experimentalLocalAgents is false, result.skills is filtered to drop "local-agent-delegation", but result.diagnostics is returned unchanged. If a collision or other diagnostic referencing that skill name is produced by loadSkills (e.g., a user-configured path duplicating the seeded skill), the diagnostic would still surface even though the skill itself is hidden — a confusing, orphaned reference to a skill the user can't see.

Optional: also filter diagnostics referencing the excluded skill

   return {
     skills: result.skills.filter((skill) => skill.name !== "local-agent-delegation"),
-    diagnostics: result.diagnostics,
+    diagnostics: result.diagnostics.filter(
+      (diagnostic) => !("name" in diagnostic) || diagnostic.name !== "local-agent-delegation",
+    ),
   };

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/skills.ts` around lines 63 - 79, In loadWorkspaceSkills, the
non-experimental branch filters out the "local-agent-delegation" skill from
result.skills but leaves related diagnostics intact. Update the return path in
loadWorkspaceSkills (and any helper logic it uses, such as
effectiveSkillPaths/loadSkills results) so diagnostics that reference the
excluded skill are filtered out too, keeping diagnostics consistent with the
visible skills list.