[BrianKWork]

Why

Agents sometimes refuse the task when the prompt lists skills. In about 15% of runs, Cursor treated the # Skills list (skill names, descriptions and SKILL.md paths in the user turn) as a prompt injection and declined to continue. Agents trust skills that arrive through their own skill support, such as a plugin, so codegen-runner wants to deliver skills that way and leave them out of the prompt's # Skills list.

Today the only way to leave them out is to rebuild this library's Skills section text and cut it out of Result.Prompt. That breaks silently whenever the prompt format changes. This PR lets the library do the rendering instead.

What changed

• RenderPrompt(r *Result, opts ...RenderOption) (string, error) re-renders a prompt from an existing Result without running discovery again.
  • WithSkillFilter(keep func(skills.Skill) bool) leaves filtered skills out of the prompt. Result.Skills keeps the full list.
  • With no options, the output equals Result.Prompt.
  • It reuses the existing rules: no rules and no kept skills gives task-only output, and # Skills is dropped when no skills remain.
  • It returns ErrRuleSourcesMismatch when RuleSources and Rules don't line up.
• Result.RuleSources carries the rule source paths the renderer needs.
• skills.Skill gains two fields, both tagged xml:"-" so the prompt XML is unchanged:
  • Plugin: the skill's plugin frontmatter value.
  • SearchPath: the local search path the skill was found under.
• WithSkillDiscovery(bool) discovers skills even when bootstrap is off. Rules and bootstrap scripts stay off, so a caller can recompute skills on every run.
• buildPrompt now calls a pure renderPrompt, so Run's output is byte-for-byte unchanged.
• The README API reference is updated.

Tests

• TestRenderPrompt_NoOptionsMatchesRun
• TestRenderPrompt_SkillFilter: keep some; keep none with rules (matches the no-skills render byte for byte); keep none without rules (task-only output)
• TestRenderPrompt_RuleSourcesMismatch
• TestSkillDiscovery_PluginAndSearchPath
• TestWithSkillDiscovery

Checks run locally:

• go test ./... passes.
  • The root package test reads ~/.agents/skills. It fails on a machine whose home folder has a skill with list-valued metadata, and passes with a clean HOME.
• go vet and deadcode -test are clean.
• golangci-lint run --new-from-rev=main reports 0 issues.
• Existing findings in files this PR doesn't touch:
  • goimports wants to reformat enumerate_test.go.
  • unparam flags taskparser/markdownparser.go:157.

🤖 Generated with Claude Code