skill-observer

Use an observation log that records corrections, patterns, and principles as numbered entries with status fields during task execution. Entries are later reviewed and applied to skill files, so insights persist across sessions instead of being lost.

Log feedback as observations during the session, then run a periodic review that cross-checks every open observation against every installed skill. Approved changes are integrated into the relevant SKILL.md sections and marked ACTIONED with a date.

No. The default behavior is log-and-defer: mid-task work produces observations only. Skill files are modified only during a weekly review, on explicit user request, or when a skill is actively producing wrong output.

Yes. It is adapted for opencode's skill system and reads skills from ~/.config/opencode/skills/. It works alongside opencode's workflow manager and specialist agents without overriding scheduling or requiring a separate skill-creator agent.

The skill enforces concurrency rules: backup before writing, re-read the live log and diff against the snapshot, use collision-safe numbering via grep checks, and verify each written entry appears exactly once after the write.

Skip applying when a new skill is proposed, when the change removes or substantially restructures content, when there is self-flagged uncertainty, or when two observations conflict. These cases are escalated for manual review instead.