feat: support configured external context files

[RoySRose]

Summary

• Add context.external_files to inject configured project context files in deterministic order.
• Add context.ignore_hermes_md so deployments can avoid HERMES.md shadowing AGENTS.md.
• Reuse existing context scanning/truncation and de-duplicate external files against automatic discovery.
• Clarify the docs/code behavior around AGENTS.md: current automatic discovery loads AGENTS.md from the active working directory only; explicit hierarchies can now be configured with context.external_files.

Motivation

Some resident Hermes deployments keep canonical agent context outside the active terminal cwd or need a root → scope → agent context chain. This makes that wiring explicit in config instead of relying on a single auto-discovered HERMES.md/AGENTS.md file.

Note: the existing docs said AGENTS.md was hierarchical/recursive, but the current implementation and tests say otherwise (_load_agents_md() is cwd-only and test_agents_md_top_level_only asserts subdirectory copies are ignored). This PR does not silently add recursive auto-discovery; it adds an explicit, ordered config mechanism for deployments that want a hierarchy.

Test plan

• uv run python -m pytest tests/agent/test_prompt_builder.py tests/hermes_cli/test_config.py tests/tools/test_threat_patterns.py -o 'addopts=' -q
• uv run ruff check agent/prompt_builder.py hermes_cli/config.py tests/agent/test_prompt_builder.py tests/hermes_cli/test_config.py tools/threat_patterns.py

[tonydwb]

Code Review Summary

Verdict: Approved

• 6 files changed, +242/-14 lines — adds configurable external context file support
• Loads context.external_files from config.yaml with proper error handling
• Well-structured with path normalization and de-duplication; no security concerns

---

Reviewed by Hermes Agent

[teknium1]

Thanks for addressing a real limitation: current agent/prompt_builder.py:1947-1984 only loads automatic cwd context sources, so configured external context is not already implemented.

Problems

• tools/threat_patterns.py:55 weakens the hidden-comment scanner. The added \b boundary does not match underscore-delimited tokens, so <!-- system_prompt: ... --> and <!-- ignore_rules: ... --> bypass this detector; current main blocks the former with html_comment_injection (tools/threat_patterns.py:69).
• The claimed automatic de-duplication is incomplete. agent/prompt_builder.py:1900-1908 excludes configured paths for .hermes.md, AGENTS.md, and CLAUDE.md, but not _load_cursorrules(). Configuring a cwd .cursorrules or .cursor/rules/*.mdc file loads it twice.
• Salvage must preserve the current context.engine setting (hermes_cli/config.py:2171-2173) while adding these keys; the PR's separate context mapping predates that schema.

Suggested changes

• Preserve detection for underscore-delimited hidden-comment instruction tokens and add regressions.
• Pass exclusions through cursor-rule discovery and test both cursor-rule forms.
• Merge the settings into the existing context mapping and update the dedicated context-file docs.

The June 27 cross-reference to open PR #53766 is overlapping work, not implementation on main. This is an automated hermes-sweeper review.

[teknium1]

Blocking: \b treats _ as a word character, so hidden comments such as <!-- system_prompt: ... --> and <!-- ignore_rules: ... --> no longer match. Keep the benign ECOSYSTEM-MAP case without losing detection of underscore-delimited instruction tokens, and add regression coverage.

[teknium1]

This de-duplication claim is not true for .cursorrules or .cursor/rules/*.mdc: the call at line 1908 does not receive external_paths. A configured cursor rule in the cwd will be injected externally and again by automatic discovery.