What problem does it solve? Changes to an agent harness (prompts, skills, commands) are often asserted to be faster or cheaper without proof. This Skill measures one named phase of an oddyssey run — preflight, drive, observation, or the whole run — under opencode, claude, or copilot, against a baseline measured on main just before the work, so a harnessing change is proven rather than claimed. ## Core Features & Use Cases - Phase-level measurement: measure_phase.py launches a headless CLI run, watches for the phase's end marker (k6 drive, dispatch pattern, or a custom regex), stops the run, and writes a record with run id, timing, and command form. - Controlled A/B sampling: run_samples.py runs alternating base/after sample chains (ABBA), syncs deployed scopes into a fake HOME, resets the local stack's data between samples, and journals each sample's outcome. - Run analysis: analyze_run.py reads the run's own logs, transcripts, or events to report commands, turns, generation time, tokens, cost, silent gaps, and the behaviors a harnessing change should remove (authored scripts, stack resets, redundant questions, --help calls). - Use Case: Before opening a PR that changes anything on the path of /odd-observe, /odd-verify, or /odd-status, run two samples per side against a fresh main baseline and quote turns, tokens, cost, and wall clock per phase. ## Quick Start Ask the agent to measure the preflight phase of an oddyssey run under copilot against main using this skill's measure_phase.py with a drive-mode mission on the local stack.
