promptwright/promptwright.config.example.yaml at main · sahajamit/promptwright · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
# Promptwright Configuration
# This file controls runtime settings for the application
# Edit this file or use the app settings to customize behavior

browser:
  # Whether to run browser in headless mode
  # true = headless (no visible browser window, faster execution)
  # false = headed (visible browser window for watching/debugging)
  headless: true
  # Browser automation backend
  # playwright-mcp = Rich tool integration via MCP protocol (default)
  # playwright-cli = Token-efficient CLI commands via bash
  automationMode: playwright-cli

# Orchestrator settings
orchestrator:
  # Model to use for the orchestrator agent
  # Leave empty or omit to use Copilot's default model
  # Models with reasoning effort support: claude-opus-4.6, claude-opus-4.6-fast
  model:
  # Reasoning effort level for models that support it
  # Options: low | medium | high | xhigh
  # Leave empty to use the model's default
  reasoningEffort:

# Per-agent model overrides (optional)
# agents:
#   pw-mcp-agent:
#     model: "claude-sonnet-4-5-20250514"
#   api-test-agent:
#     model: "claude-sonnet-4-5-20250514"

# ---------------------------------------------------------------------------
# Custom model provider (BYOK - Bring Your Own Key)  ★ login-less mode ★
# ---------------------------------------------------------------------------
# Promptwright uses GitHub Copilot purely as a bundled *harness*. When you set a
# provider below (or the PROMPTWRIGHT_PROVIDER_API_KEY env var), the app runs
# WITHOUT any GitHub Copilot login, license, or global `copilot` install - the
# model comes from your own cloud key. Pick ONE of the presets below.
#
# You can configure the provider here, OR launch purely from environment
# variables (no YAML needed) — handy for zero-cost local models:
#   export PROMPTWRIGHT_PROVIDER_TYPE=openai
#   export PROMPTWRIGHT_PROVIDER_BASE_URL=http://localhost:11434/v1
#   export PROMPTWRIGHT_PROVIDER_MODEL=qwen2.5-coder:7b
#   export PROMPTWRIGHT_PROVIDER_API_KEY=sk-...    # optional; not needed for Ollama
#   export PROMPTWRIGHT_PROVIDER_BEARER_TOKEN=...  # optional; takes precedence
# Config-file values win; env vars fill any gaps.

# --- Preset 1: OpenAI-compatible (OpenAI, most gateways) ----------------
# provider:
#   type: openai                   # default; works for any OpenAI-compatible API
#   baseUrl: https://api.openai.com/v1
#   apiKey:                        # or PROMPTWRIGHT_PROVIDER_API_KEY
#   model: gpt-4o
#   displayName: "OpenAI GPT-4o"

# --- Preset 2: Moonshot / Kimi — verified for agentic browser automation -
# OpenAI-compatible; Kimi K2.5 emits proper structured tool calls and handles
# the full Playwright-CLI agentic loop (open -> goto -> snapshot -> act).
# provider:
#   type: openai
#   baseUrl: https://api.moonshot.ai/v1
#   apiKey:                        # or export PROMPTWRIGHT_PROVIDER_API_KEY
#   model: kimi-k2.5
#   displayName: "Kimi K2.5"

# --- Preset 3: Local Ollama — zero cost, fully offline (no key needed) ---
# Run any model locally with `ollama pull <model>` then point Promptwright at it.
# IMPORTANT for *agentic* use (tool calls): the model must emit structured
# OpenAI tool_calls AND stay coherent under the agent's full toolset. In testing,
# small local models struggled — qwen2.5-coder:7b returned tool calls as plain
# text (not executed), and gemma "thinking" models returned empty content (and
# their content:null tool turns are rejected by Ollama as
# "invalid message content type: <nil>"). For agentic browser work prefer a
# strong instruct model (larger Llama/Qwen instruct) or a hosted provider like
# Moonshot above. Simple chat works with most local models.
# provider:
#   type: openai                   # Ollama exposes an OpenAI-compatible API
#   baseUrl: http://localhost:11434/v1
#   model: llama3.1:8b             # any locally pulled instruct model id
#   displayName: "Ollama (local)"

# --- Preset 3: Azure AI Foundry -----------------------------------------
# provider:
#   type: azure
#   baseUrl: https://my-resource.openai.azure.com
#   apiKey:                        # or PROMPTWRIGHT_PROVIDER_API_KEY
#   azureApiVersion: "2024-10-21"
#   model: gpt-4o                  # Azure deployment name
#   displayName: "Azure GPT-4o"

# --- Preset 4: Anthropic (Claude) ---------------------------------------
# provider:
#   type: anthropic
#   baseUrl: https://api.anthropic.com
#   apiKey:                        # or PROMPTWRIGHT_PROVIDER_API_KEY
#   model: claude-sonnet-4-5
#   displayName: "Claude Sonnet"

# Optional advanced fields (any preset):
#   bearerToken:                   # alt to apiKey; or PROMPTWRIGHT_PROVIDER_BEARER_TOKEN
#   wireApi: completions           # openai/azure only: completions | responses
#   wireModel:                     # wire name if it differs from `model`
#   headers:                       # custom HTTP headers
#     X-Custom-Header: value
#   maxPromptTokens:               # override model default
#   maxOutputTokens:               # override model default
#
# Note: AWS Bedrock is not a native provider type. Use an OpenAI-compatible
# gateway (e.g. LiteLLM) with the OpenAI preset above. Native support tracked
# as a follow-up.

# Optional: Override Copilot CLI path
# Use when Copilot is installed in a non-standard location not in PATH.
# Can be either:
#   - Full path to executable: C:\custom\bin\copilot.exe
#   - Folder path (app resolves copilot/copilot.exe inside it): C:\custom\bin
# copilotCliPath: