[anhnh2002]

Small modules and whole-parent wrappers create unnecessary documentation pages and downstream agent tasks. Add a configurable 1,000-code-token minimum and reject partitions that leave fewer than two viable children.

Changes

• Share delegation eligibility and size checks across Pydantic AI and caw. Count all indexed components in owned files, including methods omitted from entry-point lists.
• Coalesce undersized groups by directory proximity before creating pages or running child agents. Apply the same rule to clustering, structural fallback, and artifact insertion while preserving complete, disjoint file ownership.
• Report merged and rejected requests to the parent agent. A partition that collapses to one child stays in its parent; indivisible oversized files remain supported.
• Raise the default delegation base to 8,000 tokens across backend and CLI configuration. Subsequent delegation levels require 16,000, 24,000, etc. Keep the clustering threshold at 36,369 tokens; retain explicit overrides and existing cached trees.
• Expose --min-token-per-module through configuration, generation, and CLI help, with regression coverage for both backends and language-independent sizing.

Validation

• Full Python suite: 654 passed (.venv/bin/python -m pytest -o addopts= -q).
• Ruff check and format check passed for changed Python files; git diff --check passed.
• Fresh tinydb generation completed with 18/18 files covered and no undersized modules or single-child wrappers.
• Fresh cmark generation using DeepSeek V4 Flash, excluding build/: identical 152,232 code tokens, 403 components, and 785 dependency edges to the saved baseline. Baseline delegation base was 4,000; the new run used 8,000.

cmark metric	Baseline	New
Modules	60	27
Agent-created modules	36	8
Depth	5	4
Leaves below 1,000 tokens, counting complete file scopes	21	0
Single-child nodes	2	0
API requests	673	353
Recorded cost, USD	0.699	0.394

The new cmark run took 59m49s, covered all 51 files, and produced all 28 expected Markdown pages with no missing local Markdown links. The 12,992-token inline module and the 71,524-token scanner module remain leaves. One leaf lists 960 representative-component tokens but owns 1,061 tokens across its complete files, satisfying the minimum under the existing size calculation.

Remote Mermaid validation timed out during generation, so diagrams were not fully verified remotely. Source repositories and saved baseline trees were not modified; generated scratch outputs were removed.