Path/glob-based exclusions for file optimization and context tracking

Problem

I use JFrog Boost with repositories managed by cARL, where agent instruction files are adapter shims and .github/carl/** contains canonical durable governance artefacts.

Examples:

• .github/copilot-instructions.md
• CLAUDE.md
• AGENTS.md
• .cursorrules
• ANTIGRAVITY.md
• .cursor/rules/carl.mdc
• .agents/rules/carl.md
• .github/carl/**

I need Boost to leave these files outside file optimization and, where possible, context/bootstrap tracking.

Boost v0.13.24 currently supports [hooks] exclude_commands, but this operates on command names rather than file paths. File optimization can be disabled globally, but that unnecessarily disables optimization for every file.

Proposed solution

Add repository-aware path/glob exclusions, using either .boostignore:

.github/carl/**
.github/copilot-instructions.md
CLAUDE.md
AGENTS.md
.cursorrules
ANTIGRAVITY.md
.cursor/rules/carl.mdc
.agents/rules/carl.md

or configuration:

[files]
exclude_paths = [ ".github/carl/**", ".github/copilot-instructions.md", "CLAUDE.md", "AGENTS.md", ".cursorrules", "ANTIGRAVITY.md", ".cursor/rules/carl.mdc", ".agents/rules/carl.md",
]

An optional behaviour-scoped model would provide finer control:

[ignore]
optimize = [".github/carl/**", "CLAUDE.md", "AGENTS.md"]
track = [".github/carl/**"]
upload = [".github/carl/**"]

Expected behaviour

When an excluded path is encountered:

1. Its contents are delivered to the agent unchanged.
2. File optimization/conversion is bypassed.
3. It is optionally excluded from persisted context/bootstrap analytics.
4. It is optionally excluded from telemetry-derived metadata.
5.  boost doctor  reports active rules and their source.
6. Matching works consistently across Claude Code and GitHub Copilot hooks.
7. Repository-relative matching respects Boost’s existing config precedence.

Use case

cARL separates canonical governance from harness-specific projections:

•  .github/carl/**  — canonical durable governance and memory
•  CLAUDE.md ,  AGENTS.md ,  .github/copilot-instructions.md , etc. — generated adapter shims

These files are deliberately loaded as authoritative context and should not be transformed or treated as optimization candidates. Other project files should remain eligible for Boost optimization.

Current workaround

I have disabled  boost-files-optimization  globally:

[feature_flags."boost-files-optimization"]
user = false

This works but is much broader than necessary.

Prior art

I raised the equivalent requirement with Headroom:

• Issue: https://github.com/headroomlabs-ai/headroom/issues/1150
• Implementation PR: https://github.com/headroomlabs-ai/headroom/pull/1796

Their implementation uses a repository-level ignore file plus behaviour-scoped rules. A generic Boost implementation would benefit any repository containing generated, externally managed, sensitive, or governance-owned context files.

No implementation files or tests are named. Start by tracing the existing [hooks] exclude_commands handling and the boost-files-optimization feature flag, then compare the proposed .boostignore and [files] exclude_paths models. Done should include repository-relative path matching, unchanged delivery for excluded files, and consistent behavior across Claude Code and GitHub Copilot hooks.