Newly spawned agents receive stale skill versions; skills_save changes are not injected

From 2026-10-08 evening until an app restart on 2026-10-10, every newly spawned agent in this project got stale skill text. This affected both Claude Code and Codex agents (about 12 trigger spawns checked). Every change saved on 2026-10-09 was missing from the injected skills.

• A skill saved on 2026-10-08 13:37 was injected in its current version.
• Three skills saved on 2026-10-09 via skills_save (10:02, 20:01, 20:03) were injected in their old version.
• skills_get and the local skills cache (updated 2026-10-09 22:03, not dirty) returned the new text the whole time.
• Both the project baseline skills and a trigger's defaultSkillIds were affected.
• Trigger prompts edited the same day were injected correctly.

Impact: for about two days, agents silently followed outdated rules.

Workaround: the trigger prompt tells agents to re-read their skills with skills_get.

Fix observed: the user opened and closed the Skills Library and then restarted the desktop app. Agents spawned right after the restart (2026-10-10 20:41) received the current skill text, matching skills_get line for line. It is unclear which of the two steps fixed it.

Possibly related to the previously fixed "skills missing after skills_save" issue (in-memory libraries not refreshed).

AgentsRoom diagnostics

• App version: 1.198.0
• OS: darwin 25.6.0 (arm64)
• Agent CLI: claude
• App language: system

Expected behavior

Skills injected into a newly spawned agent (Claude Code skills file or Codex developer_instructions) match the current Skills Library as returned by skills_get.

Actual behavior

For about two days, new agents received skill bodies from an in-memory snapshot that was apparently stale since 2026-10-08 evening. Skill updates saved via skills_save from agents over MCP were never injected until the desktop app was restarted.

Steps to reproduce

1. Update a project skill with skills_save, called from an agent over MCP.
2. Let a schedule or quota trigger spawn a new agent that has this skill, as a baseline skill or in defaultSkillIds.
3. Compare the injected skill text with skills_get: the old version is injected.
4. Restart the desktop app and spawn again: the current version is injected.