verification-run

name	verification-run
description	Execute project-defined checks (build/test/lint) using verify.md in WORK_PATH or commands pointed to by AGENTS/DEVELOPMENT. Show outputs; optionally record evidence.md.

verification-run

Purpose

Run checks as defined by the project. No guessing. Never claim a check was run without actual command output.

Inputs

• WORK_PATH
• Context Map (especially pointers for how to run checks)
• verify.md (preferred), evidence.md (optional)

Outputs

Default: no files written. Optional (only if user approves):

• Create/update ${WORK_PATH}/evidence.md

Internal loop

1) Select target and level

Ask (unless already known):

• Target: platform name(s) or Custom
• Level: Quick / Full

2) Resolve command source (strict priority)

1. ${WORK_PATH}/verify.md if present
2. Verification section in the nearest AGENTS.md — this is the canonical per-repo source; it defines two rings: static checks (no device) and dynamic checks (device required)
3. Commands in DEVELOPMENT.md / repo docs (via Context Map)
4. User-provided commands (ask user to paste exact commands)

If commands are still unknown:

• Recommend using verification-plan skill to create verify.md
• If user refuses, proceed only with user-provided commands.

3) Build a Run Plan (before executing)

Create a short plan:

• command list (in order)
• working directory (if needed)
• expected effect (1 line each)

Run discipline (based on check ring):

• Quick checks (linters, compile — fast, no special setup): run immediately without asking APPROVE.
• Full checks (instrumentation tests, emulator/device, network): ask APPROVE before running.

4) Execute and report

• Run commands sequentially.
• On first failure: stop by default, show output + exit code.
• Ask whether to continue remaining commands or stop.

5) Optional evidence

Ask: “Record results to evidence.md?” If yes:

• Prepare a compact evidence entry:
  • timestamp
  • target + level
  • commands executed
  • results (pass/fail + exit codes)
  • short failure summary if any

Completion criteria

• Requested checks executed (or explicitly skipped)
• Outputs summarized in chat
• Evidence recorded only if user approved

Return control to Fabricator.