v0.24.0

• hooks: guard one-craft-per-session - refuse a second medium load (promptctl-routing-rat.1) (#11)
• ticket(craft): draw destination-vs-mechanism on the boundary axis, so a precise output isn't stripped (#10)
• evals(judge): an optional reference-anchored judge tier, trusted only after human agreement (#9)
• evals(compare): repeated runs expose the harness's own noise floor (#8)
• evals(compare): one command that runs a task across arms and reports which did better (#7)
• evals(run): a single scored run joining one task and one configuration (#6)
• evals(configs): a configuration format that names a skill body by git ref, or none (#5)
• evals(tasks): a task-spec format pinning repo, commit, task text, and a machine criterion (#4)
• evals(isolation): prove Opus by reading the account banner, not by asking the model (#3)
• evals(driver): complete turns whose replies exceed the visible pane (#2)
• evals(driver): a tmux turn-driver that refuses to emit a bad turn (#1)
• chore(ci): install agent code-review action
• skills(prompt): green language, no reflexive absolutes, integrate-don't-append; one craft per session may be held directly
• skills(code): restore parse-dont-validate as a first-class law; make the boundary exception structural
• chore: track lit integration files, ignore local research scratch
• skills: add project-local laws skill governing laws-skill edits
• skills(prompt): recast far-end device craft as reveal-don't-fog
• skills: drop the "verifiable acceptance criterion" bullet from dispatch skills
• style: normalize em-dashes to hyphens across docs, skills, and scripts
• skills(ticket): replace frontmatter description with PLACEHOLDER
• evals(isolation): isolated interactive session that proves Opus + no CLAUDE.md leak
• skills(ticket): keep subagent-dispatch prompts clean of orchestrator solutioning
• skills(prompt): add proportion principle carried by the orchestra metaphor
• skills(ticket): distinguish requester-imposed constraints from agent-invented mechanism