Daino is an evaluation harness for AI coding-agent context changes. It helps you compare baseline vs candidate setups with reproducible runs, structured event logs, and statistical verdicts.

• Did this context change improve verifier pass rate?
• Did efficiency improve (tokens, turns, cost, wall time)?
• Did agent behavior improve (repeat reads, hook blocks, tool errors)?

flowchart LR
    A[Bundle Manifest\ncontext surface] -->|bundle_id hash| B[Run Execution]
    T[Task Fixture + Verifier] --> B
    B --> E[NDJSON Event Log\nappend-only]
    E --> R[Reduction Layer\nDuckDB + SQL]
    R --> C[Bundle Comparison\nA/B verdict + CI]
    C --> D{Decision}
    D -->|Promote| P[Champion Bundle]
    D -->|Hold/Reject| N[Next Candidate]

bun install
bun run reduce
bun run check
bun run compare -- demo00000001 demo00000002

Useful commands:

bun run autopsy -- <run_id>
bun run new-run -- --bundle <bundle_id> --task <task_id> --replicate 0

• README.md: quick intro, architecture, and essential commands
• DESIGN.md: statistical model, schema, and design rationale
• docs/WORKFLOW.md: day-to-day operating playbook
• docs/CONTRIBUTING.md: contribution expectations and PR hygiene
• docs/TASK-CONTRACT.md: template for new task fixtures and verifiers

• src/: core logic (bundle, reduce, compare, stats, autopsy, new-run)
• sql/: reduction and analysis SQL
• bundles/: bundle manifests by bundle_id
• tasks/: fixtures, verifiers, and metadata by task_id
• runs/: run event streams (curated sample fixtures committed)
• scratch-runs/: local scratch run output
• results/: generated analysis outputs

• Two demo bundles in bundles/
• One smoke task in tasks/T00-smoke/
• Two sample run streams in runs/

These exist so reduce/check/compare work out of the box.