Benchmarks the full AI coding stack (tool, configuration, workflow, model) on 100 real-repo tasks.
Grade Claude Code or Codex harness design: extract promised rules, run an 8-task probe, and show which rules held, broke, or never fired.
The 2025 Stack Overflow Developer Survey shows 84% of professional developers using AI in their workflow, up from 76% the year before. Yet only 33% trust AI accuracy while 46% actively distrust it (survey.stackoverflow.co/2025/ai). METR's RCT of 16 experienced open-source maintainers found that AI tooling increased task completion time by 19%. Developers self-reported a 20% speedup, a 39-point gap between perception and reality (arXiv:2507.09089). Static issue benchmarks such as SWE-bench measure model ability in isolation. SWE-bench Pro (arXiv:2509.16941) addresses contamination at scale but still scores patches, not whether a workflow can ship.
AWB measures whether a configured tool+workflow combination can ship correct, regression-safe, low-burden changes against pinned real-world repositories. A bare configuration and a purpose-built setup can produce meaningfully different results with the same model. AWB benchmarks tool + configuration + workflow + model together on 100 tasks from real open-source repositories.
Related work measures complementary axes. HAL analyzes agent traces with LLM judges across 11 tasks at 2.5B-token scale. Artificial Analysis publishes harness comparisons that hold the model constant. SWE-bench and SWE-bench Pro score patches against real GitHub issues. LiveCodeBench addresses contamination with time-segmented contest problems. METR RE-Bench compares humans and agents in matched ML-engineering environments.
AWB makes two distinct contributions. First, a paired vanilla-vs-custom run isolates the configuration delta for the same model and reports Workflow Lift with a sign-test p-value. Second, deterministic trace rubrics use OpenTelemetry-aligned .trace.jsonl artifacts instead of LLM judges. See METHODOLOGY.md#related-work for citation details.
AWB now measures Codex CLI end to end. The demos are live recordings of a Codex static audit and a report regraded from an eight-task Codex probe.
- Native Codex execution:
codex exec --ephemeral --sandbox workspace-write --jsonstreams token usage, command exits, file edits, model provenance, and trace events into the same benchmark pipeline used by other adapters. - Codex harness checkup:
awb checkup --tool codex-cliinventoriesAGENTS.md,config.toml, hooks, rules, agents, skills, and plugins, then grades observable behavior from the probe traces. - Native usage economics: result, submission, and cost reports retain ChatGPT credits and show a dollar-equivalent estimate without replacing the original credit measurement.
- Codex workflow comparison: workflow export records the Codex model and
harness inventory, while paired A/B runs accept separate authenticated
CODEX_HOMEdirectories. - Honest file metrics: AWB measures the pre-agent snapshot against the agent patch, includes staged and untracked changes, and preserves partial timeout patches before cleanup.
- Measured validation: the bundled Codex harness measurement records 4/8 passing tasks, 100.0 verification discipline, 97.9 scope discipline, and 101.7399 recorded credits. The credit total is a lower bound because two timed-out turns did not emit final usage.
See CHANGELOG.md for the v1.6 checkup and earlier releases.
pip install awb
awb quickstart # verify your setup
awb checkup --static-only # instant free audit of your CLAUDE.md + hooks
awb checkup --tool codex-cli --static-only # instant audit of ~/.codex + repo AGENTS.md
awb warmup --fast-check # warm the 8 probe repos (one-time)
awb checkup # static audit + 8-task probe + rule integrity
awb checkup --from-run results/runs/<run_dir> # re-grade a saved probe for free
awb run --progressive --adaptive claude-code-custom # full suite with early exit + smart re-runs
awb run codex-cli --fast-check --yes # 8-task Codex harness measurement
awb gap # capability gaps (defaults to your last run)
awb leaderboard --readiness --explain # Production Readiness Score per toolRun this end-to-end against the published v1.4.0 fast-check baseline. Should finish in roughly 12 minutes (fast-check now runs parallel at -j 4; 12 min measured on a real 2026-07-24 run) for ~$4 of metered spend and produce a tweetable Workflow Lift number plus a capability profile.
pip install awb==1.7.0
awb quickstart # 1. verify environment
awb warmup --use-uv # 2. pre-build templates
awb run --fast-check claude-code-custom # 3. ~12 min at -j 4, real run
awb leaderboard --readiness --explain # 4. composite readiness score
awb trace grade results/runs/<run_id>/ # 5. behavior rubric scoresCompare against the published baseline at results/baselines/claude-code-custom-1.4.0-fast-check.json. Same task_set_hash means your numbers are directly comparable.
Speed levers: awb warmup caches workspaces for 10-30x faster setup, --fast-check gives a quick signal, --progressive stops early on weak tools, --use-uv swaps pip for uv, and checkup --from-run re-grades saved runs for free. See Execution Modes below.
Clone repo at pinned SHA
→ Run setup commands
→ Capture baseline lint/security counts
→ Execute tool with task prompt
→ Run test suite + partial credit rubric
→ Sigmoid-normalize 7 metrics
→ Produce weighted composite + capability profile
Each task starts from a fresh git clone at a pinned commit. Every tool gets the same prompt, the same timeout, and the same verification suite. Results are scored with sigmoid normalization so scores are never negative and never collapse at the boundary.
Security: AWB clones third-party repos and runs their setup/test code plus the AI tool with no sandbox. Treat task sets and their repos as trusted input and run in a disposable environment. See docs/SECURITY.md for the trust boundary and the planned per-task Docker isolation.
Seven dimensions, sigmoid-normalized with per-task baselines derived from difficulty:
| Dimension | Weight | What It Measures |
|---|---|---|
| Correctness | 55% | Pass/fail (60%) + partial credit rubric (40%) |
| Cost efficiency | 15% | Estimated USD per task |
| Speed | 10% | Wall-clock seconds vs. estimated task time |
| Code quality | 10% | Lint warning delta (pre vs. post) |
| Reliability | 5% | Pre-existing tests broken by the change |
| Security | 3% | New security issues introduced |
| Efficiency | 2% | Blend of iteration count and tokens-per-iteration |
Weight profiles (select with load_weight_profile(name)):
| Profile | Focus | Use When |
|---|---|---|
default |
Balanced | Standard evaluation |
correctness_focused |
70% correctness | Research-grade rigor |
production |
45% correctness, 20% cost, 10% reliability, 8% security | Shipping to users |
token_efficient |
25% cost, 15% efficiency | Tight API budgets |
rate_limited |
30% cost, 15% efficiency | Hitting TPM/RPM limits |
Sigmoid curve: score = 100 / (1 + exp(k * (value - baseline)))
- Optimal performance (excellent) → ~95
- Baseline performance (adequate) → ~50
- Above baseline → smooth decay, never negative
Difficulty-weighted aggregation: hard tasks count 2.5×, medium 1.5×, easy 1.0×. A tool that solves hard tasks beats one that only solves easy ones even if the easy-task count is higher.
Per-task baselines by difficulty:
| Metric | Easy | Medium | Hard |
|---|---|---|---|
| Cost optimal / baseline | $0.05 / $0.30 | $0.20 / $1.00 | $1.00 / $3.00 |
| Speed | 50% / 100% of estimated_minutes | same | same |
| Iterations | 3 / max_iters | 8 / max_iters | 15 / max_iters |
Real open-source repos, pinned to release tag SHAs. Setup runs in under 15 seconds via venv + pip.
| Category | Count | Easy / Med / Hard | What It Tests |
|---|---|---|---|
| bug-fix | 12 | 7 / 1 / 4 | Root cause analysis, test-first diagnosis, N+1 queries |
| feature-addition | 9 | 3 / 0 / 6 | Convention adherence, ambiguous requirements, Dockerfiles, TypeScript typing |
| refactoring | 11 | 5 / 2 / 4 | Multi-file consistency, O(n^2) optimization, CI/CD config, async migration |
| code-review | 9 | 4 / 2 / 3 | Security review (report-only), concurrency analysis, migration guides, OWASP |
| debugging | 10 | 7 / 0 / 3 | Performance profiling, regression bisection, stack trace diagnosis |
| multi-file | 7 | 4 / 0 / 3 | Merge conflicts, plugin systems, auth chains |
| legacy-code | 12 | 9 / 0 / 3 | SQLAlchemy 2.0 migration, 20-file codebase navigation, dead code removal |
| workflow | 30 | 9 / 12 / 9 | Completeness tracking, convention discovery, security methodology, context utilization, async safety, config extraction, test-driven implementation |
Repos used: FastAPI (74), httpx (17), Flask (4), Click (4), Starlette (1). All Python.
Task IDs:
BF-001–014 · FA-001–010 · RF-001–012 · CR-001–010 · DB-001–011 · MF-001–009 · LC-001–012 · WF-001–030
Each task maps to 1–3 capabilities, producing a radar chart of tool strengths:
| Capability | Tasks | What It Measures |
|---|---|---|
| code_comprehension | 45 | Understanding existing code before modifying |
| framework_knowledge | 36 | Knowing API patterns (Pydantic v2, async SQLAlchemy, etc.) |
| refactoring_discipline | 29 | Changing code without breaking behavior |
| bug_diagnosis | 27 | Structured root cause analysis, test-first diagnosis |
| multi_file_reasoning | 22 | Coordinating changes across multiple files |
| test_writing | 12 | Writing correct, meaningful tests |
| security_awareness | 10 | Identifying and fixing vulnerabilities |
| convention_adherence | 8 | Discovering and following project conventions |
| context_discovery | 5 | Reading project docs and config before editing |
| security_methodology | 5 | Applying security checklists systematically |
| completeness_tracking | 4 | Following all requirements, not stopping at 80% |
| cost_discipline | derived | Token efficiency across all tasks |
Example awb gap output:
Capability Profile
------------------
code_comprehension ████████████████████ 82.4 (n=27, conf=high)
framework_knowledge ████████████████░░░░ 68.1 (n=26, conf=high)
refactoring_discipline████████████████░░░░ 65.3 (n=23, conf=high)
multi_file_reasoning ████████████░░░░░░░░ 51.2 (n=20, conf=high)
bug_diagnosis ███████████████░░░░░ 63.7 (n=17, conf=med)
test_writing ██████████░░░░░░░░░░ 44.1 (n=8, conf=low)
security_awareness █████████████░░░░░░░ 55.8 (n=8, conf=low)
Systematic Patterns
-------------------
- Fails 70%+ of multi_file_reasoning tasks → consider multi-agent workflows
- Token spend on failed hard tasks: $4.20 → add early-exit heuristics
- No failures on easy tasks → baseline is solid
Top Suggestions
---------------
1. Enable subagent mode for tasks spanning >3 files (impact: high)
2. Add repo-level CLAUDE.md with architecture overview (impact: medium)
3. Use --think flag for debugging tasks (impact: medium)
AWB ships two Claude Code adapters that run the same model with different configurations:
| Vanilla | Custom | |
|---|---|---|
| Hooks | Disabled | Your full hook suite |
| Skills | Disabled | Your registered skills |
| Auto-memory | Disabled | Active |
| System prompt | Generic | Default (loads CLAUDE.md) |
Both use the same model, same API, same task prompts. The only difference is whether your workflow automation (hooks, skills, memory) is active. This isolates the contribution of workflow configuration from model capability.
When awb run executes both vanilla and custom (the default), it produces a Workflow Lift, a single number measuring how much your workflow configuration improves over the baseline:
Workflow Lift: +4.2 pts (p=0.031, significant)
Pass rate: vanilla 62% vs custom 68%
Wins: custom 8 / vanilla 3 / ties 69
Where your workflow helps:
bug diagnosis +12.3 pts (17 tasks)
multi file reasoning +8.1 pts (20 tasks)
security awareness +5.4 pts (10 tasks)
Where it hurts:
cost discipline -4.2 pts (100 tasks)
Biggest task-level differences:
BF-014 +40 (V=35 C=75)
LC-012 +15 (V=65 C=80)
The lift is computed per-task (configured score minus vanilla score), averaged across all tasks, and tested for statistical significance. Capability-level breakdowns show where your workflow configuration actually helps vs. adds overhead.
The 100 public tasks calibrate the instrument. The point of the instrument is your own stack: your repos, durable instructions, hooks, skills, and rules. The tuning loop, in the order it actually runs:
# 1. Free static audit: what does your harness promise, and does its
# structure hold up (hooks resolve, settings parse, docs match repo)?
awb checkup --static-only
# 2. One paid probe: 8 real tasks through your full harness. The report
# says which of your stated rules HELD, BROKE, or went UNTESTED.
awb checkup
# 3. Fix ONE thing (the report ranks fixes by estimated impact; a broken
# prose rule usually becomes a hook). Then re-grade the SAME run for
# free to re-measure anything that does not need new behavior:
awb checkup --from-run results/runs/<run_dir>
# 4. Turn rubric failures and weak capabilities into ready-to-paste
# CLAUDE.md or AGENTS.md snippets with task-level evidence.
awb gap --prescribe
# 5. Mine private tasks from your own merged PRs so the probe measures
# YOUR work distribution. No contamination: nobody trained on your
# repo's future.
awb task from-pr https://github.com/you/repo/pull/123 --out ./tasks
awb run claude-code-custom --tasks-dir ./tasks
# 6. Prove a config change helped: same adapter, two config dirs,
# paired sign test. Vibes do not survive this step.
awb ab claude-code-custom --config-a ~/.claude --config-b ./candidate-config
# Codex uses separate authenticated CODEX_HOME directories:
awb ab codex-cli --config-a ~/.codex --config-b ./candidate-codex-home
# 7. Keep it won: exit code 1 on drift, built for cron/CI, plus dollars
# per solved task before you standardize on a config.
awb drift results/runs/<run_id>_run1/ --baseline results/baselines/<ref>.json
awb cost results/runs/<run_dir>/The loop, recorded live against a real merged PR (xmpuspus/cloudwright#69, a production bug fix with its own regression test):
from-pr pins the pre-merge commit, overlays the PR's test files onto the old tree (tests exist, implementation does not), and writes a schema-valid task YAML with provenance stamped real_pr. Two caveats. First, the test-file overlay resolves objects through AWB's local mirror cache, then falls back to fetching from GitHub; if both miss, refresh the mirror with awb warmup --clear. Second, a mined task executes the PR's own test and setup code on your machine during benchmark runs, so only mine repos you trust (see docs/SECURITY.md).
awb run # all tools, all tasks, 3 runs (vanilla vs custom comparison)
awb run claude-code-custom # single tool
awb run -t BF-001 # single task
awb run --category legacy-code # filter by category
awb run --difficulty hard # filter by difficulty
awb run --capability bug_diagnosis # filter by capability
awb run --runs 1 --dry-run # preview without executing
awb run --resume # skip tasks with existing results
awb run --parallel -j 4 # run 4 tasks concurrently
awb run --adaptive # re-run near-miss tasks (60-99%) after initial pass
awb run --progressive # easy → medium → hard, stop early if pass rate too low
awb run --fast-check # 8 representative tasks, 1 run, parallel -j 4 (~12 min)
awb run --use-uv # use uv instead of pip for 10-30x faster installsAWB ships four execution modes tuned for different evaluation scenarios:
| Mode | Tasks run | Wall clock | Token cost | Use when |
|---|---|---|---|---|
| Full suite | 300 (100 × 3 runs) | ~3 hrs | ~$150 | Final evaluation, publishing results |
| Full + adaptive | ~180 | ~1.5 hrs | ~$100 | Standard workflow, strong tools |
| Progressive | ~150 on weak tools | ~1 hr | ~$40-75 | Unknown/mediocre tools |
| Fast-check | 8 | ~12 min (-j 4) | ~$4 | PR gates, iterating on config |
Fast-check (8 representative tasks, 1 per category, reports estimated full-suite score ± margin):
Progressive (easy → medium → hard, stops if easy pass rate < 40% or medium < 20%):
--use-uv (rewrites pip install → uv pip install for 10-30x faster installs):
awb checkup # static audit + 8-task probe + rule integrity
awb checkup --static-only # stage 0 only: free, instant, CI-safe
awb checkup --config-dir ~/.claude # which harness to grade (default ~/.claude)
awb checkup --tool codex-cli # inspect ~/.codex and run the probe through Codex
awb checkup --tool codex-cli --config-dir DIR # inspect and run an alternate CODEX_HOME
awb checkup --repo-dir . # repo whose CLAUDE.md/AGENTS.md also count
awb checkup --paired # add the vanilla arm, report Workflow Lift
awb checkup --tool codex-cli --paired --baseline-config-dir DIR # separate authenticated baseline
awb checkup --format json --yes # machine output (needs --yes, no prompt)
awb checkup --from-run results/runs/<dir> # full report from a saved run: free, instantStage 0 parses the harness files with zero model calls: structural checks (JSON/TOML parses, hooks resolve, documented commands match the repo) plus extraction of testable promises across 8 rule patterns, each tagged hook-enforced or prose-only. Stage 1 runs the fast-check probe in parallel and grades traces on the 6 deterministic rubrics. Codex JSONL events supply token usage, command exits, and file changes to the same metric and trace pipeline. The report leads with a verdict sentence, pillar scores, and the rule-integrity table (HELD / BROKEN / ENFORCED / UNTESTED per stated rule). Exit codes: 0 clean, 1 findings, 2 tool failure.
awb warmup # build templates for all 63 unique (repo, commit, setup) combos
awb warmup --fast-check # warm only the 8 fast-check probe repos (fastest first checkup)
awb warmup --dry-run # show combos without building
awb warmup --clear # reset template cache
awb warmup --use-uv # use uv for faster initial buildsWorkspace templates are cached at ~/.cache/awb/templates/. First build takes ~5 min; subsequent awb run invocations copy templates in ~2s instead of running pip install from scratch. Cuts ~55 min off a full benchmark run with 74 FastAPI tasks.
Analyzes results to produce a capability radar, failure classification, systematic patterns, and ranked improvement suggestions. Add --prescribe to turn trace-rubric failures and weak capabilities into concrete prescriptions for CLAUDE.md or AGENTS.md.
awb task from-pr <pr_url> --out ./tasks [--category bug-fix] [--difficulty medium] \
[--test-command "python -m pytest"] [--dry-run]Fetches the PR via the gh CLI, pins the pre-merge SHA, splits changed files into tests vs source, and generates a task YAML whose setup overlays the PR's test files onto the pre-merge tree. The generated file is validated against the schema (partial credit sums to 100) before it is written. Run private tasks with awb run --tasks-dir ./tasks.
awb ab claude-code-custom --config-a <dir> --config-b <dir> [--task BF-001] [--category bug-fix]
awb ab codex-cli --config-a <codex-home-a> --config-b <codex-home-b> [--task BF-001]Runs the same adapter over the same tasks twice, once per config dir (via CLAUDE_CONFIG_DIR or CODEX_HOME), then reports per-task deltas, mean lift, and a binomial sign-test p-value. Alternate Codex homes must already be authenticated. Both config hashes are printed for reproducibility.
awb drift results/runs/<run_id>_run1/ --baseline <run_dir_or_baseline.json> --threshold 5.0Compares mean score and per-task scores against a reference (a prior single run directory such as <run_id>_run1/, or a published awb/v2 baseline JSON). Exits 1 when the mean drops more than the threshold, 0 otherwise, so a cron job or CI step can alert on silent model or harness regressions. Warns when task-set hashes differ.
awb cost results/runs/<run_dir>/ [<more_run_dirs>...]Groups results by tool. It reports total spend, spend per solved task, wasted spend, and tokens per solve. Failed attempts count toward spend per solved task. ChatGPT-authenticated Codex runs keep native credits and show a dollar-equivalent estimate. A timeout without a final usage event makes the reported credits a lower bound.
Side-by-side comparison of two benchmark runs with significance testing.
Shows all registered tool adapters and their availability status.
Checks all 100 task YAML files against the schema, including partial credit sum-to-100 validation.
Displays full details for a specific task including repo, capabilities, and partial credit rubric.
Per-task score variance across multiple runs. Flags unstable tasks for prompt clarification or tighter verification.
Generates a static HTML site with Chart.js radar chart, CSV export, and historical run tracking.
Add --readiness to print the Production Readiness Score per tool to stdout. The score is a weighted composite of correctness (35%), regression-safety (20%), security (15%), review-burden (10%), maintainability (8%), cost (7%), and speed (5%), all normalized 0-100. Weighted for shipping safety rather than headline accuracy.
Every benchmark run writes a <task_id>_<tool>.trace.jsonl file using OpenTelemetry GenAI semantic conventions (gen_ai.client.operation, gen_ai.tool.use, gen_ai.usage.input_tokens) plus AWB-specific spans for shell commands (task.shell_command), file edits (task.file_edit), and test runs (task.test_run). awb trace grade <run_dir> reads each trace and scores four shipping disciplines on a 0-100 scale:
| Behavior | What it checks |
|---|---|
read_tests_before_edit |
Did the tool read a test file before its first edit? |
ran_verification_after_change |
Was a test run / pytest invocation issued after the last file edit? |
no_out_of_scope_edits |
Did edits stay within files_to_examine from the task spec? |
no_repeated_failing_command_loop |
Did the tool retry the same failing shell command 2+ times? |
Recalibrates task difficulty labels from empirical pass rates. Use --apply to write changes back to task YAMLs.
Recomputes task timeouts from empirical p95 wall-clock data. Use --apply to write changes.
| Command | Description | Demo |
|---|---|---|
awb quickstart |
Verify setup: tools available, tasks load | - |
awb export <run_dir> -o file.json |
Export results in submission format | - |
awb submit <file.json> |
Validate an external submission | - |
awb compare-submissions <a> <b> |
Cross-tool comparison with statistics | - |
awb migrate-results <old_dir> |
Convert v0.5.x results to v1.0 format | - |
awb workflow <subcommand> |
Export, validate, diff, or init descriptors | - |
awb --version |
Show version | - |
awb run --dry-run |
Preview tasks without executing | - |
Tasks live in awb/tasks/<category>/. Copy awb/tasks/_template.yaml:
id: BF-012
category: bug-fix
title: "Fix response_model silently dropping extra fields in FastAPI"
difficulty: easy
estimated_minutes: 15
languages: [python]
capabilities: [framework_knowledge, test_writing]
repo:
url: "https://github.com/tiangolo/fastapi"
commit: "628c34e0"
setup_commands:
- "python3 -m venv .venv && source .venv/bin/activate && pip install -e '.[all]'"
issue:
description: |
The endpoint's response_model silently strips extra fields...
files_to_examine:
- "fastapi/routing.py"
verification:
test_commands:
- "source .venv/bin/activate && python3 -m pytest tests/test_extra_fields.py -v"
partial_credit:
- criterion: "Uses Pydantic v2 ConfigDict"
points: 50
check: "grep -q 'ConfigDict' tests/test_extra_fields.py"
- criterion: "Tests pass"
points: 50
check: "source .venv/bin/activate && python3 -m pytest tests/test_extra_fields.py -v"
constraints:
max_iterations: 20
timeout_seconds: 1800Run awb validate to check your task before opening a PR. Full guide: CONTRIBUTING.md
| Adapter | Name | Status |
|---|---|---|
| Claude Code (vanilla) | claude-code-vanilla |
Full |
| Claude Code (custom) | claude-code-custom |
Full |
| Pi | pi |
Full |
| Gemini CLI | gemini-cli |
Full |
| Codex CLI (custom harness) | codex-cli |
Full: run, JSONL metrics/traces, checkup, A/B, workflow export |
| Cursor | cursor |
Planned |
| Aider | aider |
Planned |
| Windsurf | windsurf |
Planned |
| Copilot | copilot |
Planned |
Run awb tools to see which are available in your environment.
Implement the ToolAdapter ABC in awb/adapters/. The ABC includes four optional methods:
from awb.adapters.base import ToolAdapter, ToolResult
from pathlib import Path
class MyToolAdapter(ToolAdapter):
name = "my-tool"
display_name = "My Tool"
async def execute(self, prompt: str, workspace: Path,
max_turns: int = 20, timeout_seconds: int = 1800,
on_event=None) -> ToolResult:
... # on_event(event) callback for streaming token monitor; return False to abort
def check_available(self) -> bool:
...
def get_config_hash(self) -> str:
...
# Optional - implement to enable pre-flight auth checks
def supports_auth_check(self) -> bool: ...
def check_auth(self) -> tuple[bool, str]: ...
# Optional - implement to enable streaming metrics
def supports_streaming(self) -> bool: ...
def get_model_pricing(self) -> dict[str, float]: ...Register in awb/adapters/registry.py and add an entry point in pyproject.toml.
Anyone can share results using the submission format defined in results/submission-schema.json:
awb run --runs 3
awb export results/runs/<run_dir>/ -o my-results.json
awb submit my-results.json # validate locally
awb compare-submissions a.json b.json # compare with significance testingThe format captures tool version, model, hardware class, and per-task run results. Hardware classes (e.g., apple_m5_24gb, linux_x86_16gb) enable fair speed comparisons: speeds are only compared within the same tier.
- Confidence intervals via t-distribution (no scipy required for core scoring)
- Significance testing via sign test for paired tool comparison
- Integrity checks: contamination detection (completions <10s flagged), variance anomalies (identical times/tokens across runs)
- Weight profiles:
default,correctness_focused,production,token_efficient,rate_limited(seeawb/scoring/weights.yaml) - Stability metric: per-task
TaskStability(std_dev, score_range, is_unstable); high-variance tasks can be down-weighted in composite scoring - Token efficiency: sigmoid normalizer (optimal=2k tokens/iter, baseline=15k) blended 50/50 with iteration count in the efficiency dimension
End-to-end Codex CLI support: JSONL execution and traces, token budgets, model provenance, native ChatGPT credit accounting, Codex checkup and workflow export, config A/B support, resilient file metrics, measured validation, and live Codex demo GIFs.
awb checkup --from-run <run_dir> re-grades a saved run through the full report: zero adapter calls, zero spend, the primitive that makes iterative harness tuning cheap. Four tasks (FA-001, MF-001, RF-001, DB-001) no longer punish their own graded deliverables as out-of-scope edits; on recorded real-harness runs the scope pillar moved 77 to 96.4 under the corrected oracle. Adds demos/checkup.gif, recorded live.
Three fixes found running checkup against a real harness the day 1.6.0 shipped: CLAUDE_CONFIG_DIR pointing at the default ~/.claude no longer breaks macOS Keychain login detection (the 1.6.0 wheel's checkup could never pass preflight on subscription-authed Macs); the stream reader survives JSON lines over 64KB instead of silently starving the trace; auth failure messages name the real causes and quote the CLI. Promise extraction learned the "Read tests before code" phrasing.
The checkup release. awb checkup: free instant static audit (promise extraction across 8 rule patterns + structural checks) plus a parallel 8-task probe graded on 6 deterministic rubrics, reported as a verdict line, pillar scores, a rule-integrity table (HELD/BROKEN/ENFORCED/UNTESTED per stated rule), and impact-ranked fixes. Two new trace rubrics (context_discipline, tool_call_efficiency). Prescriptions cover all 11 capabilities with impact estimates. P0 fix: tool-less awb run --fast-check silently ran the full suite twice; it now forwards correctly, preflights auth before cloning, and defaults to parallel -j 4. --last-run plumbing across run/checkup/gap/cost/drift/trace grade.
awb run --dry-run no longer pays the adapter auth preflight (a live model call); previews print instantly. Adds the from-pr demo GIF recorded against a real merged PR.
awb task from-pr files fetch fixed: gh api -F fields switch GET to POST and GitHub 404s the files endpoint; per_page moved into the query string. Found by running the published wheel against a real merged PR.
Submission schema also accepts trace_summary: null, which exports write when zero traces were graded. Caught by the same fresh-venv smoke that caught 1.5.1's gap.
Submission schema accepts the v1.4.0 trust columns (readiness, trace_summary, per-run trace_grade); packaged and repo schema copies pinned in sync by a regression test. Caught by the v1.5.0 fresh-venv release smoke.
Harness tuning: awb task from-pr + awb run --tasks-dir (private tasks from merged PRs), awb ab (paired config A/B via CLAUDE_CONFIG_DIR), awb drift (baseline regression alerts with an exit-code contract), awb cost (cost per solved task), awb gap --prescribe (config prescriptions from rubric failures). Task-set hash now derives from the loaded tasks directory; --resume with --tasks-dir is refused to prevent cross-set contamination; Rich markup disabled on all prints carrying PR-derived text.
Real trace grading (tool_use blocks translated to FILE_EDIT/SHELL_COMMAND spans; span-less traces report null, never a fake 100), baselines with per-run trace_grade and readiness columns, -j N enables parallel mode on its own, Aider became a real adapter, exact-pinned runtime dependencies with freshness guard tests, security posture documented in docs/SECURITY.md.
Performance and token optimization release. 33-50% faster full runs, ~97% cheaper quick evaluations.
- Workspace template cache - ~55 min saved on full runs (74 FastAPI tasks no longer re-run pip install)
awb warmup- pre-build all unique workspace templates in parallel--use-uv- 10-30x faster pip installs via uv--progressive- easy → medium → hard execution, stops early if weak tool (50-80% token savings)--fast-check- 8 representative tasks, 1 run, ~15 min, ~$4 (97% cheaper than full suite)- Token budget enforcement -
max_input_tokens/max_output_tokensin task constraints, streaming kill switch - Streaming token monitor - Claude Code adapter parses stream events as they arrive
- Parallel partial credit - independent grep/file checks run via asyncio.gather; pytest stays sequential
- Adaptive timeouts - runs 2+ tighten timeout to
min(original, 2x run1_actual) - Richer RunCost - cache_read, cache_creation, thinking token fields
- Token efficiency in scoring - efficiency dimension blends iterations + tokens-per-iteration
- Two new weight profiles -
token_efficientandrate_limitedfor cost-sensitive evaluation - Token-aware gap analysis - cost-per-point outliers, cache hit rate patterns, token burn detection
- JSONL results - additive output format alongside per-file JSON for fast batch loading
- 184 tests (up from 135)
- Add Python 3.13 and 3.14 to CI test matrix and PyPI classifiers
- Sync README changelog with PyPI long description; update GitHub repo description (80 → 100 tasks)
Product audit fixes: 27 findings across observability, scoring, reliability, performance, and CLI safety.
- Observability:
--verboseflag, test output logging, captured partial credit output, specific exception handlers, integrity checks inawb run - Scoring:
SECURITY_METHODOLOGYcapability, signed lint delta, removed hardcodedMETRIC_WEIGHTS, timeout calibrator can increase, leaderboard uses per-task aggregate scoring - Reliability:
KeyboardInterrupthandling,load_singleNone guard,find_incomplete_runscans all_runNdirs, 600s setup timeout,return_exceptionsin gather, finally cleanup - Performance: bare-clone cache (
~/.cache/awb/clones/), cachedRunEnvironment/adapter, schema cache - CLI safety: confirmation prompt (
--yes),quickstartis env-only check, resolved paths,check_availableguard for stubs
- Add trustme to 4 real httpx repo tasks (BF-003, BF-011, BF-013, FA-005)
- Add trio to 16 httpx-based tasks (fixes silent pytest crash on Python 3.13+)
- Fix 4 verification bugs (FA-010, RF-012, CR-007, BF-003)
Older releases
See CHANGELOG.md for the full history (v1.0.0, v0.5.x, v0.4.x, v0.3.x, v0.2.x, v0.1.0).
- Methodology - Fair comparison principles, metric definitions, related work, known limitations
- Architecture - Module graph, data models, pipeline diagrams
- Contributing - Adding tasks, tools, and submitting results
- PyPI -
pip install awb
If you use AWB in research, cite it via Zenodo. The v1.7.0 archive has DOI
10.5281/zenodo.21753418. The concept
DOI 10.5281/zenodo.20361437 always
resolves to the latest release. Machine-readable metadata lives in
CITATION.cff and codemeta.json; the release
process is in docs/zenodo-doi.md.
@software{puspus_awb_2026,
author = {Puspus, Xavier},
title = {{AWB: AI Workflow Benchmark}},
version = {1.7.0},
year = {2026},
month = aug,
publisher = {Zenodo},
doi = {10.5281/zenodo.21753418},
url = {https://doi.org/10.5281/zenodo.21753418}
}MIT