AgentClash Hub

Purpose

Give coding agents maximum context to run AgentClash evals through the CLI and guide humans to the right web UI pages — without reading the AgentClash source repository.

Use When

• A user asks to evaluate agents, run evals, compare models, or use AgentClash for the first time.
• You need to pick the right downstream skill before acting.
• You need hosted defaults, UI links, or the end-to-end eval workflow in one place.

Do Not Use When

• A narrower skill already matches (e.g. only CLI auth repair → agentclash-cli-setup).
• The task is only to edit AgentClash product source code in the monorepo.

Environment

Use production unless the user explicitly runs a local stack:

export AGENTCLASH_API_URL="https://api.agentclash.dev"
agentclash auth login --device
agentclash link
agentclash quickstart

Install the CLI: npm i -g agentclash or see /docs-md/getting-started/quickstart.

Portable bundle install (copy skills to another agent host without the integration command): export them with agentclash skills export --dir ./bundle --host <agent>. Canonical source repo: https://github.com/agentclash/agentclash

Procedure

1. Load this hub to pick the next skill.
2. Run agentclash quickstart if CLI readiness is unknown.
3. Follow dependency order for setup → pack → run → review → regression → CI.
4. Send the user to the matching UI page when they need a visual surface.

Drive The CLI Like An Agent

When you (a coding agent) drive AgentClash non-interactively, lean on the machine contract:

• Ask for machine output. Pass --json (or -o json) on every command. Data goes to stdout; human progress and diagnostics go to stderr — parse stdout only.
• Discover the surface first. agentclash schema --json returns the full command tree, flags, and documented exit codes. Prefer it over scraping --help prose.
• Branch on structured errors, never on prose. On failure the CLI prints a JSON envelope: error.code is a stable machine key — branch on it; error.next_step is the remediation to act on; error.details carries specifics (quota limit/used/remaining, reset_at, plan_key).
• Gate readiness on agentclash doctor. It exits non-zero when the environment isn't ready. agentclash quickstart --json is informational and exits 0 even when ready:false — use doctor as the hard gate in CI/agent flows.
• Stay headless. Set AGENTCLASH_TOKEN, AGENTCLASH_WORKSPACE, and AGENTCLASH_API_URL in the environment so no command needs an interactive prompt.

End-To-End Eval Workflow (CLI)

1. agentclash-cli-setup              → auth, workspace, doctor
2. agentclash-quickstart             → readiness checks + next command
3. agentclash-runtime-resources-setup → provider, model alias, runtime profile, secrets
4. agentclash-agent-build-author     → build spec + ready build version
5. agentclash-agent-deployment-setup → deployment ID for runs
6. challenge-pack skills             → plan, YAML, validate, publish pack
7. agentclash-eval-runner            → eval start / run create / follow / sessions / series
8. agentclash-scorecard-reader       → rankings, scorecards, replay, artifacts
9. agentclash-compare-and-triage     → baseline, compare latest/gate, replay triage
10. agentclash-regression-flywheel   → promote failures, suite-only reruns
11. agentclash-ci-release-gate       → CI manifest + gate (optional)

Optional branches (load when the workflow applies):

• agentclash-multi-turn-operator     → human takeover during multi_turn runs
• agentclash-dataset-workflows       → dataset eval, gate, traces, regression sync
• agentclash-prompt-eval-playground  → prompt-eval YAML + playground experiments
• agentclash-agent-harness-setup     → E2B coding-agent harness tasks and suites
• agentclash-workspace-admin         → org/workspace CRUD and membership (teams)
• agentclash-security-evaluation     → client-side security stress harnesses

Human-friendly shortcut after setup:

agentclash quickstart
agentclash eval start --follow
agentclash baseline set
agentclash eval scorecard
agentclash compare latest --gate
agentclash replay triage

Skill Dependency Order

Read skills in this order when multiple apply:

1. agentclash-hub (this file)
2. agentclash-cli-setup
3. agentclash-quickstart
4. agentclash-runtime-resources-setup
5. agentclash-agent-build-author
6. agentclash-agent-deployment-setup
7. agentclash-challenge-pack-planner
8. agentclash-challenge-pack-yaml-author
9. agentclash-challenge-pack-input-sets
10. agentclash-challenge-pack-tools-sandbox
11. agentclash-challenge-pack-artifacts
12. agentclash-challenge-pack-scoring-validators
13. agentclash-challenge-pack-llm-judges
14. agentclash-challenge-pack-validation-publish
15. agentclash-eval-runner
16. agentclash-scorecard-reader
17. agentclash-compare-and-triage
18. agentclash-regression-flywheel
19. agentclash-ci-release-gate
20. agentclash-agent-harness-setup
21. agentclash-multi-turn-operator
22. agentclash-dataset-workflows
23. agentclash-prompt-eval-playground
24. agentclash-workspace-admin
25. agentclash-security-evaluation

To author or change skills, browse the web catalog at https://www.agentclash.dev/docs/agent-skills — it is documentation, not an installable skill.

Each skill folder name matches its name in frontmatter. When a skill lists Related Skills, load those before mutating remote state.

All Skills In The Catalog

Skill folder	When to load
agentclash-hub	First — workflow map and UI links
agentclash-quickstart	Readiness checks and suggested next command
agentclash-cli-setup	Install, auth, workspace, doctor
agentclash-runtime-resources-setup	Provider accounts, models, runtime profiles, secrets
agentclash-agent-build-author	Agent build specs and build versions
agentclash-agent-deployment-setup	Create/select deployments
agentclash-challenge-pack-planner	Plan a pack before YAML
agentclash-challenge-pack-yaml-author	Write pack YAML
agentclash-challenge-pack-input-sets	Cases and input sets
agentclash-challenge-pack-tools-sandbox	Tools and sandbox policy
agentclash-challenge-pack-artifacts	Assets and artifact refs
agentclash-challenge-pack-scoring-validators	Validators
agentclash-challenge-pack-llm-judges	LLM judges
agentclash-challenge-pack-validation-publish	Validate and publish
agentclash-eval-runner	Start and follow evals, sessions, series
agentclash-scorecard-reader	Interpret results
agentclash-compare-and-triage	Baselines, compare, replay triage
agentclash-regression-flywheel	Promote failures to regression suites
agentclash-ci-release-gate	CI/CD gates
agentclash-agent-harness-setup	E2B coding-agent harness tasks, suites, failure review
agentclash-multi-turn-operator	Human takeover turns in multi_turn packs
agentclash-dataset-workflows	Dataset eval, CI gate, traces, regression sync
agentclash-prompt-eval-playground	Prompt eval YAML and playground experiments
agentclash-workspace-admin	Org/workspace CRUD and membership administration
agentclash-security-evaluation	Security pack stress-run and vault harnesses

Nested folders: agent-build-skills/ and challenge-pack-skills/ mirror the table rows above.

Product UI — Where To Send The User

Base URL: https://www.agentclash.dev

User goal	UI path
Sign in / account	https://www.agentclash.dev
Docs home	https://www.agentclash.dev/docs
Quickstart	https://www.agentclash.dev/docs/getting-started/quickstart
First eval walkthrough	https://www.agentclash.dev/docs/getting-started/first-eval
Agent skills (web catalog)	https://www.agentclash.dev/docs/agent-skills
CLI reference	https://www.agentclash.dev/docs/reference/cli
Challenge packs guide	https://www.agentclash.dev/docs/guides/write-a-challenge-pack
Multi-turn packs	https://www.agentclash.dev/docs/challenge-packs/multi-turn
Interpret results	https://www.agentclash.dev/docs/guides/interpret-results
CI/CD gates	https://www.agentclash.dev/docs/guides/ci-cd-agent-gates
Workspace runs (after login)	App dashboard → Runs list
Live run events	Run detail page while status is running
Scorecards & comparisons	Run detail → scorecard / ranking views after completion

When you create a run via CLI, tell the user:

Open https://www.agentclash.dev and navigate to your workspace runs, or search for run ID <RUN_ID> after signing in.

AgentClash Concepts (30-Second Model)

• Challenge pack — versioned eval workload (cases, scoring, tools policy).
• Input set — which cases run in a given eval.
• Agent build / deployment — the agent under test (model + runtime + tools).
• Run — one execution of pack × input set × deployments.
• Eval session — repeated runs (eval start --repetitions N or run series create).
• Scorecard — structured results, comparisons, release gate input.
• Baseline bookmark — workspace default run/agent for compare latest.
• Regression suite — promoted failures for suite-only reruns.

Expected Output

After loading this skill you can name the next skill, 1–3 CLI commands, and the UI page the human should open.

Failure Modes

• Skipping agentclash-cli-setup when auth or workspace is unset → commands fail with workspace errors.
• Running evals before pack publish → no runnable pack version.
• Using localhost API URL by mistake → empty workspace or auth failures against the wrong backend.

Safety Notes

• Confirm before production-scale evals, publishes, or CI runs that spend budget.
• Never paste tokens, secrets, or customer data into chat.
• Prefer agentclash doctor and read-only list commands before writes.

Report Back Format

Hub loaded: yes
Next skill: <skill-folder-name>
CLI status: <auth/workspace/doctor summary>
UI for user: <https://www.agentclash.dev/...>
Next commands: <1-3 commands>

Related Skills

Load all skills listed in Skill Dependency Order as needed; start with agentclash-cli-setup if CLI is not configured.

Related Docs

• /docs-md/agent-skills
• /docs-md/agent-skills/agentclash-hub
• /docs-md/guides/use-with-ai-tools
• /docs-md/getting-started/quickstart
• /docs-md/getting-started/first-eval