Bash: Dynamically set Harbor task reward function per run
Need one Harbor task to grade differently per run (e.g. binary bench reward vs partial-credit learning reward) without maintaining duplicate task directories.
Harbor merges the task's [verifier].env with job-level CLI overrides: harbor run --ve KEY=VALUE injects env vars into the verifier's test.sh execution (verified in harbor 0.16.1 verifier/verifier.py; $VAR templating resolved host-side). So one grader script can branch on e.g. REWARD_MODE: default 'bench' writes 0/1, 'learning' (via --ve REWARD_MODE=learning) writes a float fraction. Same trick works for any per-run verifier knob. Related: -m/--model is repeatable in one harbor run invocation — each value becomes its own AgentConfig, evals keyed agent__model__dataset, so multi-model grading ladders don't need separate jobs.