ab

---

name: ab description: Run a counterbalanced A/B knob block end to end and produce a report. Orchestrates the whole validation loop the other skills do piecemeal — for each arm it starts a clawback proxy with that arm's knobs and its OWN --state file (fresh SESSION KEY, no clawback-side carry-over), drives N turns of load through it (pty or replay driver), stops it, then analyzes the per-arm turn-logs and renders the charts. No cache-salt: production forwards none, so neither does the harness. Default block is A0 (passthrough baseline) vs A5 (recommended stack); the analyzer pools all passthrough turns into the baseline and compares each treatment arm against it, so the block must include A0 and may carry several treatment arms at once. Use this to run the full A/B runbook in one command. The pty driver spends real Anthropic tokens (one run per arm) — pin --model.

clawback A/B knob block (end-to-end orchestrator)

Run .skills/ab/scripts/ab_block.sh from the project root. It is the single command that ties the validation harness together: capture (separately, once) → this (drive both arms + analyze + plot). It runs one arm at a time:

1. start a clawback proxy with the arm's knob set + its own --state file + a shared --turn-log (via .skills/monitor's run_monitor.sh --detach);
2. drive --turns N of load through it (--driver pty = real claude via .skills/drive; --driver replay = HTTP fixture via .skills/replay);
3. stop the proxy;
4. repeat for the next arm;
5. analyze the shared turn-log (.skills/bench) and render charts (.skills/plot) into the --out directory.

Cache hygiene WITHOUT a salt (production fidelity)

A knob A/B is only honest if it measures what production does, and production forwards no salt. The harness used to prepend a per-arm --cache-salt block to system to force each arm cold — but that salt forced whole-body re-serialization and zeroed within-arm cache warming (salted arms measured ~0% warm vs a pristine A0 at ~98%), measuring a virtual system instead of the proxy. It is gone. Arms are isolated three production-faithful ways instead:

1. clawback-side: each arm gets its own --state file, so no arm RESUMES a prior arm's persisted SESSION KEYs.
2. Anthropic-side, byte-MUTATING arms: a treatment arm that mutates forwarded bytes (strip-ephemeral, 1h-TTL, mobile) already gets a distinct ANTHROPIC KEY, so it cannot read a prior arm's warm entries.
3. Anthropic-side, byte-IDENTICAL arms: keep-alive-only forwards the same real-turn bytes as passthrough, so per-arm --state does not change the content-addressed ANTHROPIC KEY. Two production-faithful guards, neither of which deletes a turn: (a) the analyzer reports cold-starts in their own "first" gap bucket, so a carry-over-warmed open is never pooled into the warm-loop buckets it's compared within; and (b) --cooldown-sec spaces serial arms past Anthropic's 5-min cache TTL so each arm's opening turn is a genuine cold-start (counted, exactly as a production user pays it after an idle), not a fake-warm read of the prior arm's cache — set ~330 for a byte-identical-arm headline. The analyzer's --warmup discard (drop each arm's first N turns) stays an opt-in diagnostic (default 0), not the standing guard: dropping opening turns also deletes legitimate production cold-start cost, so the default counts them.

This is driver-independent — the proxy forwards the arm's bytes the same way regardless of how the turn was produced.

How the analyzer pairs the arms

The proxy labels every turn passthrough or treatment; this script writes one turn-log per arm and labels it into the analyzer as knobProfile <arm>. The analyzer pools all passthrough turns into the per-bucket baseline, then compares each treatment knobProfile against that baseline. So:

• The block must include A0 (passthrough) for a savings % — it is the baseline every treatment arm is measured against.
• You can put several treatment arms in one block (e.g. --arms "A0 A2 A5"): each is labeled separately and compared against the shared A0 baseline in the same run.

Usage

# short L0 no-regression smoke, baseline vs recommended stack, on Haiku:
.skills/ab/scripts/ab_block.sh --profile L0 --turns 8 --driver pty \
  --model claude-haiku-4-5-20251001

# headline L2 (5-30 min gaps) block, >=200 turns/arm for a reportable %:
.skills/ab/scripts/ab_block.sh --profile L2 --turns 200 --driver pty \
  --model claude-haiku-4-5-20251001 --out runs/L2-headline

# clean HTTP replay arm (needs a CLAWBACK_OAUTH_TOKEN bearer + a captured
# fixture; never an API key):
CLAWBACK_OAUTH_TOKEN=… .skills/ab/scripts/ab_block.sh --profile L2 --turns 200 \
  --driver replay --fixture benchmark/fixtures/ccode.json

# keep-alive warmth test: does A1 stay warm across a 15-min idle on the
# native 5m TTL?
CLAWBACK_OAUTH_TOKEN=… .skills/ab/scripts/ab_block.sh --driver replay \
  --arms "A1 A0" --gap-sec 900 --turns 5 --out runs/keepalive-15m

Arms

• A0 passthrough baseline (forces all knobs off)
• A1 keep-alive only
• A2 1h-TTL only (nested-cache-control rewrite on — the moat)
• A3 1h-TTL + keep-alive
• A4 strip-ephemeral only
• A5 recommended stack (keep-alive + 1h-TTL + strip-ephemeral)

Each arm's knobs are set explicitly on the proxy command, not inherited from DEFAULTS, so the arm is the arm regardless of the operator's global CLAWBACK.md.

Options

• --profile L0|L1|L2|L3|L4 — inter-turn gap regime (default L0).
• --gap-sec N — fixed inter-turn gap in seconds; overrides --profile's jittered range (forwarded to both drivers). For a precise idle, e.g. the 15-min keep-alive warmth test above.
• --turns N — turns per arm (default 8; analyzer marks savings insufficient below 30/arm; target ≥200 for a reportable %).
• --arms "A0 A5" — space-separated arm list (default "A0 A5").
• --driver pty|replay — load driver (default pty).
• --model ID — pin claude's model for the pty driver (cost control).
• --fixture PATH — replay fixture (default benchmark/fixtures/ccode.json).
• --prompts PATH — pty prompt file (default benchmark/prompts/coding.txt).
• --settle-sec N — pty quiescence threshold (default 8).
• --host, --port — proxy bind (default 127.0.0.1:8787).
• --out DIR — output dir (default runs/ab-<timestamp>).
• --no-plot — skip chart rendering.

Spends real tokens

--driver pty runs the real claude binary once per arm against your own Anthropic limits. Keep --turns small and pin --model (e.g. claude-haiku-4-5-20251001) for plumbing / no-regression runs; reserve a larger, default-model run for the headline numbers. --driver replay also spends tokens but forwards a CLAWBACK_OAUTH_TOKEN bearer (never an API key) and is cheaper to script.

Outputs

In the --out directory:

• turns.<arm>.ndjson — one turn-log per arm (labeled into the analyzer as knobProfile <arm>; the passthrough arm pools into the baseline).
• proxy.<arm>.log — each arm's proxy stdout+stderr.
• state.<arm>.json — each arm's own proxy state file, so a restart never resumes the previous arm's persisted SESSION KEYs (the clawback-side carry-over guard described above).
• report.md, report.csv, summary.json, manifest.json — the analyzer outputs (.skills/bench).
• charts/ — the four §9 marketing charts (.skills/plot).

Read the headline with sed -n '1,40p' <out>/report.md.

Where this sits

Capture a fixture once with .skills/capture. This skill is the runner; .skills/{drive,replay} are the load arms it calls, .skills/bench is the analyzer, .skills/plot is the chart renderer. To run an arm by hand instead, see those skills.