SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsusing-axis

using-axis

General & other

View repository

Updated 10-16 01:35

About this skill

Using AXIS

AXIS (Agent Experience Index Score) is a synthetic testing framework for AI agents. This skill is the operator's guide: how to invoke the CLI, where AXIS writes files, and how to read the scoring output.

For authoring scenarios and axis.config.json, see the configure-axis skill.

When to use this skill

Trigger phrases include "run AXIS", "run the ask suite", "what does --profile do", "compare runs", "explain this score", "which scenario regressed", "set a baseline", "where does AXIS put its reports", "what does this dimension mean".

Refer to the framework's output as the AXIS Result. The acronym is Agent Experience Index Score.

CLI commands

The binary is axis (or npx @netlify/axis without a global install).

axis init

Scaffold axis.config.json and a sample scenario, then install AXIS skills via npx skills.

Flag Purpose
-s, --scenarios <path> Scenarios directory (default ./scenarios)
-a, --agent <names> Comma-separated agents (e.g. claude-code,codex)
--format <format> json (default), js, or ts
-f, --force Overwrite existing files
--no-skills Skip the automatic skills install

axis run

Execute every scenario against every configured agent in isolated workspaces, score the results, and write a report.

Flag Purpose
-c, --config <path> Config path (default discovers axis.config.{json,js,ts,mjs})
-s, --scenario <keys> Comma-separated keys with glob support (cms/*, hello-*)
-a, --agent <names> Comma-separated agent names with glob support (claude-code|*)
--scenarios / --agents Plural aliases for the two options above
-p, --profile <name> Apply a named overlay from the config's profiles map
--concurrency <n> Max parallel jobs (default 15)
--runs <n> Run each pair n times; must be odd (default 1, or settings.runs)
--failed [reportId] Re-run only failed pairs from a prior report (default latest)
--no-score Skip the LLM judges, write raw results only
--compare-baseline [name] Diff results against a saved baseline (default main)
--refresh-skills Force re-clone of remote skills cached under .axis/skills-cache/
--refresh-repos Force re-clone of repos cached under .axis/repos/ for copy actions
--json Emit JSON to stdout instead of the live TTY display
-v, --verbose Per-step logging
--debug Capture raw agent stdout into {agent}.debug.ndjson
-o, --output-dir <dir> Also write the report manifest to this directory

--failed can be combined with --profile (though not with -s/-a), but retry a report under the profile that produced it: profiles can rename agents and narrow the suite, so pairs from a different suite match nothing and the run exits non-zero having discovered no jobs. A repeated pair is retried in full, not run by run, so the retry report holds a complete sample.

--runs must be odd (1, 3, 5, up to 19): an even sample has no middle run, so the median falls between two runs, and at runs: 2 the representative selection degenerates to "always run 1". Even values error out rather than rounding.

--runs trades cost for score stability. Each extra run is a full agent execution plus its four judge calls, so --runs 3 costs roughly 3x the tokens and 3x the wall clock. Use it for baselines and CI gates, not for every exploratory run. See "Repeated runs" below for how the reported number is chosen.

If the config defines profiles, axis run runs the default suite and axis run --profile <name> swaps in that overlay: usually a different agent matrix and a different slice of scenarios. An unknown profile name is an error, not a fallback to the default suite, so a typo in CI cannot quietly run the wrong suite. -s and -a narrow within the active suite; neither can reach a scenario the suite excluded. The active profile is recorded in the report manifest as profile.

axis reports

View past runs. Three call shapes:

  • axis reports lists every report in .axis/reports/ (most recent first; -n to cap the list).
  • axis reports <reportId|latest> shows the manifest for one run, scored summary per scenario.
  • axis reports <reportId|latest> <scenarioKey> opens the per-run detail (transcript, audits, criteria). Use -a <agent> to filter when multiple agents ran the same scenario.

Flags: --json for machine output, --html to open the rendered report in a browser.

axis baseline

Manage saved baselines for regression detection. A baseline is a snapshot of one report's scenario scores under a name (default main).

  • axis baseline set [name] saves the latest report as a baseline.
  • axis baseline list lists saved baselines.
  • axis baseline show [name] prints the saved entries.
  • axis baseline compare [name] diffs the latest report against the baseline.
  • axis baseline delete [name] removes a baseline.

axis run --compare-baseline combines a fresh run with an immediate diff in one step.

Expected directories

Everything AXIS writes lives under .axis/ at the config directory (typically the project root). Source-of-truth files (config, scenarios, skills) are NOT under .axis/.

project/
├── axis.config.{json,js,ts}        ← config (you author this)
├── scenarios/                      ← scenarios (you author these)
│   └── hello-world.json
├── skills/                         ← optional local skills referenced by config
│   └── <skill-name>/SKILL.md
├── adapters/                       ← optional custom adapter modules
│   └── <name>.{ts,js}
└── .axis/                          ← AXIS-managed, safe to gitignore
    ├── reports/
    │   └── <reportId>/             ← e.g. 2026-06-12-205816
    │       ├── report.json         ← run manifest (summary, results[])
    │       ├── report.html         ← rendered HTML report
    │       └── scenarios/
    │           └── <scenarioKey>/
    │               ├── <agent>.json           ← per-run detail (pair that ran once)
    │               ├── <agent>.raw.ndjson     ← raw transcript
    │               ├── <agent>.sparse-index.txt
    │               └── <agent>/
    │                   ├── artifacts/         ← files captured by scenario.artifacts globs
    │                   └── run-<i>/           ← only when the pair ran more than once
    │                       ├── result.json    ← that run's detail
    │                       ├── raw.ndjson
    │                       ├── sparse-index.txt
    │                       └── artifacts/
    ├── baselines/
    │   └── <name>.json             ← saved baselines (default name: main)
    ├── remotes/                    ← cloned scenarios from remote git URLs
    ├── repos/                      ← repos cloned by `copy` actions (per repo + ref)
    └── skills-cache/               ← cloned remote skills

When asked "where is X", the answer is almost always here. Do not search the project tree blindly; jump to the path.

Scoring framework

Every run is scored on four independent dimensions, each 0-100, combined into a weighted composite.

Dimension Default weight What it measures
Goal achievement 0.4 LLM judge scores the run against the scenario's judge checks
Environment 0.2 Execution reliability of filesystem, shell, and network operations
Service 0.2 Execution reliability of external service interactions (APIs, MCP)
Agent 0.2 Decision quality across every tool call the agent made

Override weights in axis.config.json under settings.scoring_weights.

Important distinctions

  • Environment and Service evaluate execution reliability only. Did ls, cat, bash, fetch, MCP calls succeed cleanly? They do NOT judge whether the output was useful or task-fit.
  • The Agent dimension is decision quality across every interaction. Every tool call is an agent choice, even a plain ls. The agent judge audits every interaction, not just calls tagged as agent-y.
  • Speed is always heuristic (threshold buckets per category), never LLM-evaluated. Every other dimension uses an LLM judge.

Agent sub-dimensions

The agent dimension weights its own sub-dimensions:

Sub-dimension Weight What it captures
Necessity 0.4 Was the call needed at all? Redundant exploration tanks this hard.
Relevance 0.2 Did the call advance the goal?
Weight 0.2 Was the call's cost proportionate to its value?
Success 0.1 Did the call succeed?
Speed 0.1 Heuristic speed buckets

Env and Service sub-dimensions are simpler: success 0.7, speed 0.3, rest 0.

Composite formula

axisScore = goal * w_goal + environment * w_env + service * w_svc + agent * w_agent

Calibration

All dimensions use log-normal CDF mapping with median 0.5 and sigma 0.4:

Raw input Mapped score
0.5 50
0.8 88
0.985 96

Practical consequence: even flawless runs cap around 95-99. Treat 95+ as a top-band result. A clean 100 is structurally near-impossible across all four dimensions.

Reading a report

.axis/reports/<reportId>/report.json has:

  • version, reportId, timestamp, durationMs
  • summary: { total, completed, failed, averageAxisScore }
  • results[], one entry per (scenario, agent) pair

Each result entry contains:

  • scenarioKey, scenarioName, agentName
  • durationMs, exitCode, tokenUsage: { input, output, cacheReadInput }
  • score.axisScore (0-100 composite)
  • score.goalAchievement.{score, criteria[]} where each criterion has check, weight, score, rationale
  • score.environment.{score, dimensions, audits[]}
  • score.service.{score, dimensions, audits[]}
  • score.agent.{score, dimensions, audits[]} (audits every interaction, not just agent-tagged)

Each dimensions object has { success, speed, weight, relevance, necessity } mapped to 0-100.

Repeated runs

A pair configured with runs > 1 still occupies one results[] entry. Its score, durationMs, tokenUsage, and file describe the representative run: the real run whose composite sits nearest the median composite. Because run counts are odd, that run's score is the median, so the headline equals the median of the runs listed in runs[]. AXIS never averages composites into a headline, because an average has no transcript behind it to explain.

Three extra fields appear:

  • runCount -runs configured for the pair (absent when 1).
  • runs[] -one entry per run: runIndex, axisScore, dimensionScores (the four dimensions), the complete score, durationMs, tokenUsage, totalCostUsd, file, plus failed / withheld / representative flags. A withheld run publishes no score at all; a failed run publishes its zeros. The representative's score is the same as the pair's top-level one.
  • spread -axisScore: { median, min, max, mean, stdev } and representativeRunIndex, computed over successful runs only. Per-dimension variance lives in runs[].dimensionScores, not here.
  • reliability -{ succeeded, total, withheld }.

summary.total / completed / failed count pairs, with runsTotal / runsFailed alongside them for individual executions. A pair is completed when at least one run scored, failed only when none did.

Reading these correctly:

  • Quote the representative's axisScore as the pair's result, and quote spread when discussing stability. "86, with runs spanning 79 to 91 (sigma 5.1)" is the honest phrasing.
  • reliability.total excludes withheld runs, so 2/2 scored (1 withheld) means the agent succeeded on both runs it was actually measured on. Do not report a withheld run as an agent failure.
  • Never report a p value as if it were an effect size, or the reverse. "p=0.002" says the move is unlikely to be chance; "d=-3.3" says it is large. A move can be significant and trivial (tight runs, tiny gap) or huge and insignificant (scattered runs). Quote both.
  • A wide spread with a healthy median is a volatility finding, not a quality finding. Say so rather than reporting the median alone.
  • When the composite is steady but a dimension is not, compare runs[].dimensionScores across runs. "AXIS held at 84 across three runs, but goal achievement ranged 62 to 95" is a finding the composite alone hides.
  • Per-run detail lives at .axis/reports/<reportId>/scenarios/<key>/<agent>/run-<i>/result.json, but you rarely need it: every run's full score is already in report.json under runs[].score. The HTML report opens on the representative and lets you select any row to switch the breakdown to that run.

Diagnosing a low score

Look at which dimension dropped, then inspect:

  • Goal dropped: read goalAchievement.criteria[] and find entries with score < 10. The rationale says which check the agent failed.
  • Environment dropped: read environment.audits[] and find entries with success < 1. Common causes: command-not-found, missing files, network errors.
  • Service dropped: same as environment but for MCP / API calls. Audit entries point at the specific tool.
  • Agent dropped: check agent.dimensions.necessity first. If it's low, the agent made redundant calls. agent.dimensions.relevance low means calls were off-task.

Comparing against a baseline

Match by scenarioKey (variants like foo@bar are distinct keys). Subtract axisScore from the baseline entry's axisScore. The dimension that moved the most is the failure mode.

Each row reports two verdicts and you should quote whichever the question calls for.

The tolerance band is the simple view: did the representative value move further than the baseline's own measured spread. The significance test is Welch's two-sample t-test over both distributions, present only when baseline and current each ran the pair more than once, and reported per metric as d= (Cohen's effect size) and p=. Markers: ● significant with a large effect, ◐ significant but smaller, ○ not separable from noise.

The two can disagree, and that is not a bug. At three runs a side the test needs about 2.27 sigma where the band needs 2.00, so the band flags slightly more. The test only becomes the sharper instrument at higher run counts, where it tightens and the band does not.

Significance covers duration and tokens as well as the scores, and its verdicts are direction-aware: duration or tokens going down is an improvement, the same drop in a score dimension is a regression.

The exit code follows the band, not the test. summary.significant is a separate tally.

Deltas are judged against a per-row noise band, not a flat threshold. The band is max(1, 2 * stdev), where stdev is the spread the baseline measured across its own runs; a baseline captured from a single run has no measured sigma and falls back to the 1-point floor. Each comparison row reports the band it used, so a large delta counted as unchanged is explainable. A drop in reliability is a regression on its own, even when the score holds.

Citing numbers in analyses

Open the file and read it. Do not invent values.

  • "AXIS Result dropped from 84 to 53, a 31-point regression"
  • "Service success collapsed from 0.95 to 0.30"
  • "Agent necessity was 0.32, meaning roughly 68% of tool calls were judged unnecessary"
  • "5 of 7 service interactions returned errors per the audits[] entries"

Rules you must follow

  1. Refer to the framework's output as the AXIS Result, never "AXIS Score" (which reads as "score score"). The acronym is Agent Experience Index Score, never "eXperience".
  2. Do not use em dashes in any prose, comment, or analysis you author. Use a comma, semicolon, colon, parenthesis, or a new sentence instead.
  3. When asked to read a report, open the actual file. Do not paraphrase, do not guess at numbers. Cite values verbatim from report.json or the per-scenario detail JSON.
  4. Treat 95+ as the practical top band. Do not call a 95 result "merely good" or imply 100 is the realistic target; the log-normal calibration makes a clean 100 nearly impossible.
  5. Distinguish execution quality (Environment, Service) from decision quality (Agent). They measure different things; using them interchangeably is wrong.
  6. When the user asks where AXIS writes something, give the exact path from the directory map above. Do not search.
  7. Do not invent CLI flags. The full surface is listed above; if a user asks for something not listed, say so and suggest the closest documented option. --scenario/--scenarios and --agent/--agents are interchangeable on axis run, but note axis init --scenarios <path> is a separate option that sets where the scenarios directory is written.
  8. For a repeated pair, never present the median or the representative score as if it were the only number. Quote the spread alongside it, and keep withheld runs out of any reliability claim about the agent: a dead judge is a measurement failure, not an agent failure.
  9. Write paths relative to the project root. When you put an AXIS path into a runbook, analysis, plan, or any file you author, spell it as .axis/reports/<reportId>/report.json, NOT as an absolute path like /tmp/axis-xyz/work/.axis/reports/<reportId>/report.json or /private/var/.../.axis/.... Absolute workspace paths leak the agent's isolated temp directory and will not match what the user sees in their own checkout. The only correct form is the project-root-relative path.

Reference

  • Documentation site: https://axis.run
  • Scoring source: src/scoring/ in the netlify/axis repo (deep-eval.ts, category-score.ts, composite.ts, aggregate.ts)
  • Report writer: src/reports/writer.ts; per-run file layout: src/reports/paths.ts
  • Companion skill for authoring: configure-axis

Installing this skill

Use the skills CLI:

npx skills add netlify/axis --all

This installs every AXIS skill (configure-axis, using-axis) into every detected agent config directory. axis init runs this automatically; pass --no-skills to opt out.