Testing Heterogeneous Agents

This project skill extends acceptance with one scenario: proving that every server-advertised official model completes through each supported external CLI agent and LobeHub's server-default provider binding.

It does not replace Acceptance. Use Acceptance for the plan, approval gate, evidence contract, immutable rounds, publishing, and teardown. Use .agents/acceptance/PROJECT.md for LobeHub's Electron launch, auth, CDP, and multi-instance commands. This skill owns only the compatibility-matrix semantics and its executable harness.

Manual Invocation Only

The user invokes this skill with /testing-heterogeneous-agents in Claude Code or $testing-heterogeneous-agents in Codex. Do not automatically load or execute it as part of development, debugging, acceptance, or release tasks, and do not schedule it. A reference from another skill is not permission to invoke it. Manual invocation still requires scope and cost approval before live requests.

Scope

The matrix covers:

• Claude Code, Codex, Grok Build, Kimi Code, Pi, and TRAE;
• models returned by the live getServerDefaultHeterogeneousCapability response;
• the real Desktop renderer → IPC → CLI → official relay path;
• the ingress selected by each agent (anthropic-messages or openai-responses);
• a unique marker round trip plus authoritative relay, provider, and model observations.

It does not cover user-defined providers, user-managed API keys, general model quality, tool-use quality, or ordinary chat-agent execution. Add a separate scenario instead of silently widening this matrix.

Harness

HARNESS=.agents/skills/testing-heterogeneous-agents/scripts/official-smoke.mjs

# Discover the live matrix and installed CLI versions. Makes no model calls.
node "$HARNESS" list --json

# Narrow discovery when validating one change.
node "$HARNESS" list --agent claude-code,codex --model MODEL_ID --json

# Execute selected or all cells through the real official provider.
node "$HARNESS" run --confirm-live --topic-id TOPIC_ID --cdp 9222

Filters, topic id, and non-default CDP port are optional when the attached Electron window already supplies the intended scope.

Exit codes:

Code	Meaning
0	Every selected cell passed
1	At least one selected cell failed
2	No failures, but at least one cell was blocked
3	Electron, authentication, topic, or capability preflight failed

Run Workflow

1. Load the parent contract and project adapter

Read the acceptance skill, .agents/acceptance/PROCESS.md, and .agents/acceptance/PROJECT.md. Follow their approval and teardown rules. This scenario makes billable requests and creates server operation records, so never infer permission to run it from a request to inspect or list the matrix.

2. Discover before spending

Run list --json first. It reads the live capability response and probes only already-installed CLI binaries with version/help flags. Use its exact cells as the plan; do not maintain a second hard-coded model list.

Report missing CLIs as blocked. Do not install, update, or sign in to an external CLI during this workflow. To suppress Claude Code's background update paths even during detection, start or restart the Electron instance with:

DISABLE_AUTOUPDATER=1 DISABLE_UPDATES=1 \
  .agents/acceptance/scripts/electron-dev.sh start

3. Establish the Electron surface

Use the Electron section of .agents/acceptance/PROJECT.md to attach to an already-running signed-in instance or start one for this run. Confirm an active personal topic or pass --topic-id. Never trigger an interactive OAuth flow.

4. Confirm scope, cost, and history

For each live run, present the discovered agents/models and obtain the Acceptance plan approval before passing --confirm-live. Run sequentially; do not parallelize cells because the CLIs share profiles and provider quota.

For repeated manual runs of the same compatibility scope, reuse one stable Acceptance subject so every execution becomes another immutable round in the same history. Previous runs do not authorize another live run.

5. Execute once

Run the selected matrix with --confirm-live. By default each CLI receives an isolated workspace inside the report directory. Pass --cwd only when a real project workspace is part of the compatibility claim.

The harness starts each cell idempotently in the renderer and polls it with short host-side evaluations. Do not replace this with one long agent-browser eval: long evaluations can be retried by the driver and duplicate live requests.

6. Inspect and publish the round

The harness writes an Acceptance-compatible result.json, one text evidence artifact per cell, and a structured matrix table under:

.records/reports/<timestamp>-heterogeneous-official-provider-smoke/

Unless --report-dir selects another location. Inspect the report and evidence; do not infer success from the process exit code alone. Publish it with the clean production lh acceptance run ingest environment required by .agents/acceptance/PROCESS.md. Confirm that every matrix cell has its required text evidence and return only the stable Acceptance URL.

7. Diagnose failures at the owning layer

Use each failed cell's evidence to identify the first broken boundary:

1. CLI launch or authentication;
2. request ingress and metadata normalization;
3. official relay/provider/model selection;
4. raw CLI output;
5. heterogeneous adapter or terminal mapping.

Load the heterogeneous-agent skill for raw trace, adapter, IPC, persistence, or terminal-event diagnosis. A model refusing one ingress but passing another is a request-compatibility signal, not proof of an adapter bug.

Safety Invariants

• run requires --confirm-live; never bypass this gate.
• Never invoke installers, updaters, or sign-in flows.
• Never read custom provider credentials for this matrix.
• Claude Code cells receive both update-disabling environment variables.
• Missing binaries are blocked, not repaired automatically.
• Cells run sequentially to limit quota and profile contention.
• The harness listens to direct IPC and does not persist chat messages.
• Stop only the Electron instance started by the current run.

Harness Self-Test

The self-test uses a fully stubbed agent-browser; it launches no browser, external CLI, or model request:

bash .agents/skills/testing-heterogeneous-agents/scripts/official-smoke.test.sh