v1.8.13

Released v1.8.13.

Sub-agents were born blind — judgment-day judges reviewing a Next.js project had no idea React 19 patterns existed, fix agents ignored project conventions, PR creators skipped the project's template. This release fixes that with the Skill Resolver Protocol.

Compact Rules

The skill-registry now reads each SKILL.md at build time and generates pre-digested rule summaries (5-15 lines per skill). At delegation time, the orchestrator matches skills on two dimensions:

• Code context — file extensions and paths the sub-agent will touch (.tsx → react-19, typescript)
• Task context — what actions it will perform (create PR → branch-pr, review → framework skills)

Matching compact rules are injected directly into the sub-agent prompt as ## Project Standards (auto-resolved). No one reads SKILL.md files at runtime. The cost is pre-computed once, then read cheaply at every delegation.

Feedback Loop

Sub-agents now report skill_resolution in their return envelope — injected, fallback-registry, fallback-path, or none. If the orchestrator sees anything other than injected, it re-reads the registry immediately and self-corrects for all subsequent delegations. This prevents silent degradation after compaction wipes the orchestrator's skill cache.

judgment-day with Project Standards

Pattern 0 resolves skills before launching judges. Both Judge prompts and the Fix Agent prompt now receive the matched compact rules, so adversarial reviews check against project-specific standards — not just generic best practices.

Global Delegation Rule

New rule in the global config: "When editing 3+ files for the same change, delegate the batch to a sub-agent." Applies always, not just in orchestrator mode — prevents context inflation from inline multi-file edits.

Install: brew update && brew upgrade gentle-ai