The factory's per-issue agent used to receive a resolver-picked set of skills pasted in full into every system prompt, with a token budget that silently truncated whatever didn't fit. It now gets progressive disclosure:

resolver ──► load:      workflow skill only ──► full text in system prompt
         └─► suggested: marked in index    ──► agent decides

skill index (one line per skill: name — description)
         └─► read_skill({ skill })            ──► SKILL.md + reference filenames
             read_skill({ skill, reference }) ──► one reference document

• Skill sources are exactly two: .agents/skills-orchestrator/ (factory workflow skills, still front-loaded) and packages/boxel-cli/plugin/skills/ (the domain library, on demand). The root .agents/skills/ copies (boxel-development, ember-best-practices) are no longer read.
• The keyword resolver is demoted to a hint: its picks render as (suggested) markers in the index instead of gating what the agent can see.
• Realm-lifecycle skills (realm-sync, file-ops, realm-history, profile, boxel-environment) are excluded from the index — their write commands would fight the orchestrator's workspace→realm sync — and the index preamble tells the agent not to run realm-mutating boxel commands skills mention.
• --skills-dir <path> replaces the on-demand library with an operator-curated folder: nothing bundled is indexed or readable, no exclusion filtering, fail-fast at startup if the folder yields no skills. Front-loaded workflow skills are unaffected.

Verified on a staging run: the implementation-issue prompt shrank to one front-loaded skill plus an 18-line index, and the agent made 12 targeted read_skill calls (the core skill plus 11 of its 32 references, each pulled right before the work it covers) with all validators green in a single iteration.

🤖 Generated with Claude Code