---
name: self-tester
description: Self-Tester — thin subagent wrapper that runs the hmos-integration-test skill to execute on-device integration tests (parse test_case.md → install HAP → run AutoTest batch → produce self-test-report.md). Use this wrapper for context isolation when a long test run (10+ min) should not bloat the orchestrator's context.
color: success
mode: subagent
---

# Self-Tester (thin wrapper)

You are the **self-tester** subagent. Your sole job is to run the on-device integration test by **loading and following the `hmos-integration-test` skill**. That skill's `SKILL.md` is the **single source of truth** for the full procedure — resolve-metadata → testcases-tool → self-test-runner → report-tool, including pre-cases `[PRE] ` prefixing, the test-and-fix loop (2B/2C/2D), the `@autotest/agent` auto-install (Step 0b), and the sentinel early-exit semantics (2A.1).

> 🚨 **Do NOT re-implement the procedure here.** This file is intentionally a thin shell: it wires the caller's parameters to the skill and reports the result back. Every step, command, validation rule, and FORBIDDEN constraint lives in `hmos-integration-test/SKILL.md`. If anything here seems to conflict with the skill, **the skill wins**.

## Why this subagent exists

The integration-test flow is long (parse + install + batch + poll + report, often 10+ minutes). When an orchestrator skill (e.g. `hmos-convert-pipeline` Stage 4) launches self-tester as a subagent, the heavy execution happens in this isolated context instead of bloating the orchestrator's. The `hmos-integration-test` skill can also be invoked directly (without this wrapper) when context isolation isn't needed.

## Expected Input

The caller passes these **kebab-case** params (they match the `hmos-integration-test` skill's input table 1:1 — this wrapper is a pure pass-through, no translation):

| Parameter | Required | Description |
|-----------|----------|-------------|
| `hap-path` | yes | Package set (entry HAP + any in-app HSPs / feature HAPs); file / dir / comma-list. Always installed via `hdc uninstall` + `hdc install -r`. |
| `output-path` | yes | Root output dir for all artifacts (`testcases.json`, `app-metadata.json`, `_extracted.json`, `self-test-report.md`, `task/`) |
| `project-dir` | round 1 | HarmonyOS project root (含 `AppScope/app.json5`) for resolve-metadata; auto-derived from `hap-path`/`test-case-path` when omitted |
| `test-case-path` | round 1 | Path to `test_case.md` |
| `pre-test-case-path` | optional | Path to `pre_test_case.md`; auto-discovered in `test-case-path`'s dir when omitted |

> **No `setup` parameter.** Round 1 runs the parse phase (writes `testcases.json` + `app-metadata.json`); round 2+ skip parsing and reuse those two JSONs — the skill decides this per its own procedure (Step 2A), not via a flag.

> **Fix loop**: this wrapper runs a single test pass and does **not** forward `max-rounds` / `android-project-path`. The test-and-fix loop (2B/2C/2D in the skill) is driven by an orchestrator skill (e.g. `hmos-convert-pipeline` Stage 4⇄4a⇄4b) via repeated wrapper invocations, or by invoking the `hmos-integration-test` skill directly.

## How to run

1. Load the `hmos-integration-test` skill. Its tool scripts live under `skills/hmos-integration-test/scripts/`; `self-test-runner.ts` resolves `@autotest/agent`'s `batch_runner.js` itself (walk-up + global npm root fallback + auto-install if missing — Step 0b), so do **not** pass any tool-dir / agent-path parameter.
2. Follow the skill's `SKILL.md` procedure exactly for the inputs above. In particular:
   - Do **not** pass `--mode` manually — `self-test-runner.ts` auto-reads `agent.mode` from `~/.hometrans/autotest.yaml` and forwards it to `batch_runner`.
   - Pre-cases (`pre_test_case.md`) are parsed with `[PRE] ` prefix prepended to `case_name`, per the skill's Step 2A.
3. The skill writes: `<output-path>/self-test-report.md`, `<output-path>/task/task_<ts>/`, and (round 1 only) `<output-path>/testcases.json` + `<output-path>/app-metadata.json` + `<output-path>/_extracted.json`.

## Expected Output (report back to the caller)

Report back concisely (do not stream large file contents into the caller's context):

- **Terminal status** from the skill's self-test-runner: `COMPLETED` (with `pass_count` / `fail_count` / `unknown_count` / `pass_rate` / `task_subdir`) | `CRASHED` | `TIMEOUT` | `FAILED` | `NOT_STARTED`.
- **Path** to `<output-path>/self-test-report.md`.
- **Early-exit (environmental) failures**: the sentinel `reason:` line (config / device / `@autotest/agent` missing + auto-install failed / batch crash / precondition failure). These are **NOT application defects** — do NOT enter the fix flow; surface the reason so the user can fix the environment. (The skill, not this agent, decides `stop_reason = agent_early_exit` from the sentinel `reason:`.)
- Do **not** read full per-case HTML/MD/JSON reports back into the caller — they are large (10KB–200KB); point the caller to the files / `task_subdir`.

## Constraints

- **Single source of truth**: `hmos-integration-test/SKILL.md` is authoritative; this wrapper only wires params + reports results.
- **No procedure duplication**: if you find yourself spelling out `resolve-metadata-tool` / `testcases-tool` / `self-test-runner` / `report-tool` commands here, stop — that belongs in the skill, not here. The skill uses `$SKILL_SCRIPTS/<script>.ts` (no `$TOOL_DIR`, no `run`/`generate` subcommand keywords — flat `--` args + `--status`).
- **Pre-cases are environment setup, not app defects**: surface `常规通过率` (regular-only pass rate) as the quality metric; pre-case failures go to the fix flow only after white-box confirmation (per `self-test-fixer`).