Compare (side-by-side adapter A/B)

Audience: operators picking between adapters (Claude vs Codex vs Aider vs Gemini) on the same task, without burning two manual runs.

Overview

bernstein compare runs the same task spec in parallel against up to four adapters in isolated per-adapter worktrees, diffs the produced changes against the baseline workspace, and writes a JSON sidecar plus a Markdown summary.

The single-adapter bernstein run path is untouched. The compare runner is a separate, deterministic harness.

Source:

• src/bernstein/cli/commands/compare_cmd.py
• src/bernstein/core/orchestration/compare_runner.py
• src/bernstein/eval/telemetry.py (adds compare_run_id)

CLI

bernstein compare SPEC_PATH --adapters NAME[,NAME...]
  [--workspace DIR]      baseline snapshot; defaults to cwd
  [--role NAME]          role applied identically to all adapters; default backend
  [--seed N]             deterministic seed forwarded to adapters; default 0
  [--keep-worktrees]     keep per-adapter worktrees on disk after the run
  [--traces-dir DIR]     override JSON sidecar dir; default .sdd/traces
  [--no-sidecar]         skip writing the JSON sidecar

Cap is 4 adapters per run. Asking for more exits with 2 and an explicit error. Duplicates are rejected. The degenerate 1-adapter case is allowed for harness symmetry.

Adapter spawn is parallel (one thread per adapter); the runner itself stays single-threaded for clean cleanup semantics.

Outputs

• stdout - Markdown summary table (one row per adapter) plus per-file diff blocks.
• JSON sidecar - .sdd/traces/compare-<id>.json (overridable via --traces-dir); skipped with --no-sidecar.
• Telemetry - AgentTelemetry rows carry compare_run_id so downstream eval ingestion can group runs.

Exit codes

Code	Meaning
0	At least one adapter exited cleanly
Non-zero	All adapters failed (exit_code != 0 for every adapter)

Per-adapter failures are always rendered in the summary regardless of the overall exit code.

Examples

Two-adapter A/B:

bernstein compare ./task-spec.md --adapters claude,codex

Four-way bake-off, keep worktrees for offline inspection:

bernstein compare ./task-spec.md \
  --adapters claude,codex,gemini,aider \
  --keep-worktrees

Throwaway smoke run, no sidecar:

bernstein compare ./task-spec.md --adapters claude --no-sidecar

Worktree hygiene

Each adapter gets its own worktree under .sdd/runtime/compare/<compare_run_id>/<adapter>/. The runner clones the baseline workspace into the worktree with shutil.copytree, runs the adapter, computes a unified diff against baseline, and then removes the worktree unless --keep-worktrees is set.

BERNSTEIN_TRACES_DIR is honoured for the sidecar location.

Troubleshooting

unknown adapter: X. The adapter is not registered. Run bernstein agents to list registered adapters; install / register the missing one or drop it from --adapters.

--adapters cap is 4; got N. Trim the list. The cap is hard.

Adapter binary missing. The runner does not crash; instead it reports a per-adapter error row in the summary so the comparison still shows the rest of the bake-off.