[anhnh2002]

Parent and repository overview prompts asked API models to read child pages without providing file-reading tools or the page contents. Responses such as “I'll start by reading…” followed by file paths were saved as finished Markdown and skipped on subsequent runs. Module agents could also finish without writing their required page.

This change supplies immediate-child documentation within a 32,000-token input budget, validates overview content before saving, and allows one correction attempt shared with Mermaid validation. Reruns automatically repair empty pages and the specific reading-preamble/path-only failure while preserving other existing prose. Successful replacements are atomic; failed recovery retains the original page.

Module agents now check for their required file. API agents receive up to two reminders in the same conversation before failing explicitly; subscription agents fail explicitly when the file is absent. Delegated module reservations remain resumable, and failures cannot silently pass after independent modules finish.

Fixes #134.

Validation:

• 576 tracked Python tests passed locally, including regression cases for bounded context, both layouts/backends, retries, recovery, nested dependency restoration, and usage accounting.
• 37 Mermaid parser fixtures passed.
• Ruff lint and formatting checks passed for all changed Python files.
• All GitHub CI checks passed: Test, Lint, and Mermaid syntax (run).
• Guard tests also passed locally with both PydanticAI 1.0.6 and 2.31.1.
• Two pre-existing, ignored local clustering tests fail against unchanged main as well; they are not part of the tracked CI suite and were left untouched.

No new dependencies, CLI flags, configuration fields, or backend interfaces. Child excerpts may be truncated to preserve the input budget; a real provider/pipeline rerun remains useful to confirm documentation quality.