Avenir-Web

What this skill does

This skill operates Avenir-Web for reliable web-task execution and iteration.

Responsibilities:

• - run single tasks and batch tasks
• choose mode (headless / headed / demo)
• improve instruction quality before execution
• analyze run outputs and recommend the next best change
• execute one atomic action without strategy/checklist overhead
• read the current page by screenshoting it and asking the main model a question

Use this skill for requests like:

• - run a task on a website
• run a task list and summarize outcomes
• improve success rate with better instructions/config

Canonical entrypoints

Single task:
CODEBLOCK0

Atomic action:
CODEBLOCK1

Read page:
CODEBLOCK2

Batch:
CODEBLOCK3

Prefer these scripts over ad-hoc commands.

Quick usage example

Single task:
CODEBLOCK4

Batch:
CODEBLOCK5

Atomic action:
CODEBLOCK6

Read page:
CODEBLOCK7

Run modes

Notes:

• - if mode is missing, use INLINECODE6
• INLINECODE7 improves observability, not model intelligence

Mode selection:

1. 1. benchmark/batch -> INLINECODE8
2. visual debugging -> INLINECODE9
3. demo/control flow visibility -> INLINECODE10

Instruction design

INLINECODE11 should include:

1. 1. objective
2. constraints
3. completion condition

Template:

• - INLINECODE12

Keep it single-goal, specific, and verifiable.

Single-task workflow

Input:

• - INLINECODE13
• INLINECODE14
• optional mode, task-id, INLINECODE17

Steps:

1. 1. check environment and API key
2. validate instruction quality
3. run INLINECODE18
4. inspect outputs
5. report status + cause + next action

Recommended report fields:

• - INLINECODE19
• status: success / partial / INLINECODE22
• evidence summary
• one-line cause
• one recommended next step

Atomic action workflow

Use scripts/atomic_action.py when you need exactly one browser operation and do not want strategist/checklist generation.

Typical uses:

• - one click
• one type
• one goto
• one scroll

Properties:

• - disables strategy generation
• disables checklist generation
• executes exactly one action
• returns structured JSON with result, URL, screenshot path, and output directory

Read-page workflow

Use scripts/read_page.py when you want to inspect the current page by screenshot and ask the main model a direct question.

Properties:

• - opens the page
• captures a screenshot
• sends the screenshot plus page metadata to the main model
• returns structured JSON with the answer and screenshot path

Batch workflow

Task file schema

CODEBLOCK8

Required per task:

• - INLINECODE25
• INLINECODE26
• INLINECODE27

Config checklist (src/config/batch_experiment.toml)

• - INLINECODE29
• INLINECODE30
• INLINECODE31
• INLINECODE32
• INLINECODE33
• API key source

Batch execution

1. 1. validate JSON schema and config paths
2. choose mode and INLINECODE34
3. run batch command
4. summarize per-task outcomes
5. provide one global improvement recommendation

Recommended batch report fields:

• - total/completed/failed counts
• per-task status list
• recurring issue patterns
• one highest-impact next change

API requirements

Required credential:

• - OPENROUTER_API_KEY (preferred)

Resolution order:

1. 1. environment variable INLINECODE36
2. INLINECODE37 in TOML (fallback)

Rules:

• - never hardcode real keys in source files
• never print full keys in logs/outputs/reports
• fail fast if key is missing with an actionable message

Script usage rules

• - script-first: use repository entrypoints before custom commands
• non-interactive CLI only
• explicit flags and paths
• deterministic behavior preferred
• clear, actionable error messages

If adding helper scripts:

1. 1. place under INLINECODE38
2. use CLI flags (no prompts)
3. return stable, parseable summaries
4. document usage in this file

Environment checklist

Before running:

1. 1. Python environment available
2. dependencies installed (pip install -e src)
3. Playwright Chromium installed (python -m playwright install chromium)
4. API key configured
5. config/task paths valid

Output contract

Each run summary should include:

1. 1. execution metadata: run type, mode, task IDs
2. outcome: status and evidence summary
3. diagnosis: root-cause hypothesis
4. next action: one highest-impact recommendation

Boundaries

• - do not claim completion without evidence
• do not skip issue summary
• avoid large refactors before instruction/config fixes
• avoid interactive prompts in core workflow

One-line identity

Avenir-Web execution and reliability skill: mode selection + instruction design + run analysis + iteration planning.