Plugin Request: Implement Argument Normalizer Plugin (Native)

[crivetimihai]

Feature Request: Implement Argument Normalizer Plugin (Native)

Implement a native (in‑process) plugin named ArgumentNormalizer that stabilizes prompt/tool inputs by normalizing:

• Unicode (NFC/NFD/NFKC/NFKD) and control characters
• Whitespace (trim, collapse internal spaces, normalize CR/LF, optional blank-line collapse)
• Optional casing (none/lower/upper/title)
• Numeric dates to ISO 8601 (YYYY-MM-DD) with day_first/year_first ambiguity handling
• Numbers to a canonical form (strip thousands separators; . as decimal)

Default mode should be non-blocking (permissive), returning modified payloads when changes occur. Target hooks: prompt_pre_fetch, tool_pre_invoke.

Problem Statement

Incoming args vary widely in normalization (Unicode forms, whitespace, casing), and frequently include ambiguous numeric dates or locale-specific number formatting. This leads to:

• Prompt template mismatches and brittle tool args
• Lower PII/regex filtering accuracy due to noisy inputs
• Unnecessary retries and inconsistent behavior across environments

Proposal

Add plugins/argument_normalizer/argument_normalizer.py implementing a Plugin subclass with:

• Config (ArgumentNormalizerConfig) for toggles and strategies (Unicode, whitespace, casing, dates, numbers)
• Per-field regex-based overrides (field_overrides) to tune normalization by key path (e.g., user.name, items[0].title)
• Safe recursive normalization for dict/list structures
• Non-blocking results with modified_payload and minimal metadata; no violations by default

Recommended ordering: run before PII filtering so detectors see stabilized inputs.

Scope

In scope:

• prompt_pre_fetch and tool_pre_invoke hooks
• Unicode normalization, whitespace cleanup, optional casing, numeric date and number normalization
• Per-field overrides and conditions to target specific prompts/tools

Out of scope (for this issue):

• Resource hooks; advanced locale-aware date parsing libraries; schema validation; rate limiting

Configuration (Example)

- name: "ArgumentNormalizer"
  kind: "plugins.argument_normalizer.argument_normalizer.ArgumentNormalizerPlugin"
  description: "Normalizes Unicode, whitespace, casing, dates, and numbers in args"
  version: "0.1.0"
  author: "Mihai Criveti"
  hooks: ["prompt_pre_fetch", "tool_pre_invoke"]
  mode: "permissive"
  priority: 40
  conditions: []
  config:
    enable_unicode: true
    unicode_form: "NFC"
    remove_control_chars: true
    enable_whitespace: true
    trim: true
    collapse_internal: true
    normalize_newlines: true
    collapse_blank_lines: false
    enable_casing: false
    case_strategy: "none"  # none|lower|upper|title
    enable_dates: true
    day_first: false
    year_first: false
    enable_numbers: true
    decimal_detection: "auto"  # auto|comma|dot
    field_overrides: []

Ordering Guidance

• Argument Normalizer should precede PII Filter. Suggested priorities:
  • ArgumentNormalizer: 40
  • PIIFilterPlugin: 50

Acceptance Criteria

• Native plugin class ArgumentNormalizerPlugin implements prompt_pre_fetch and tool_pre_invoke
• Config supports Unicode, whitespace, casing, dates, numbers, and field_overrides
• Non-blocking behavior by default; returns modified_payload when changes occur
• Recursive normalization works for nested dict/list args
• Unit tests cover unicode/whitespace, numbers (comma/dot locales), dates (day_first), casing, and nested structures
• README under plugins/argument_normalizer/ documents behavior, config, overrides, and examples
• Docs mention in llms/plugins-llms.md and ordering note
• Example config entry included in plugins/config.yaml (commented or sample)

Tasks

1. Implement plugin in plugins/argument_normalizer/argument_normalizer.py
2. Add unit tests in tests/unit/.../argument_normalizer/test_argument_normalizer.py
3. Write plugin README with examples and tuning tips
4. Update docs: add to llms/plugins-llms.md (Built‑in Plugins) and note ordering
5. Add example configuration to plugins/config.yaml
6. Validate with make doctest test and run targeted pytest selection

Risks & Mitigations

• Over-normalization (e.g., changing intended casing): mitigate via field_overrides and disabled enable_casing by default
• Ambiguous dates: controlled by day_first/year_first; default conservative transformations
• Locale edge cases for numbers: decimal_detection with explicit comma|dot override when needed