NEWS.md

In saber: Context Engineering for Large Language Model Agents

saber 0.8.0

Consolidates the 0.7.2.x development cycle. All previously exported functions keep their signatures and defaults. No dependencies were added to Depends or Imports, and the minimum R version is unchanged.

New features

• New context_manifest(), context_render(), and context_audit() provide explicit source routing, native-file suppression, exact deduplication, provenance, and named character/line budgets. Print methods and audit results expose metadata without source bodies. Native coverage only suppresses sources when its audience matches the current consumer. Source metadata retains requested_path alongside resolved and canonical paths. Native path aliases are deduplicated by canonical identity, preserving the first requested spelling.
• New src_symbols(): C, C++, Python, Rust, and JavaScript symbol index for any repository via tree-sitter (suggested bonsaisitter runtime + a grammar package per language: treesitter.c, treesitter.cpp, treesitter.python, treesitter.rust, treesitter.javascript). Mirrors symbols() with a lang column; exported marks definitions visible beyond their own file (non-static in C/C++, no leading underscore in Python, pub in Rust, export-wrapped in JavaScript). New default_src_exclude() lists directories skipped while scanning: the default_exclude() opt-outs (Documents and friends) plus dependency/build trees (node_modules, __pycache__, venv, build, dist, renv, target); hidden and *.Rcheck directories are always skipped.
• blast_radius() gains include = "src" to report C/C++/Python/Rust/JS callers from the target project's sources.
• New heartbeat(): cross-project git activity summary. Scans every repository under scan_dir, reports projects with commits in the lookback window (busiest first), and writes briefs/_heartbeat.md. The one-glance complement to briefing().
• New exports git_commit_count_since() and git_log_since(): the git count/log primitives behind heartbeat(), for downstream tooling that composes its own activity reports. git_log_since() supports an unbounded window (since_date = NULL) and an "iso" line format.
• New read-only skill_manifest() and skill_read() for package-owned and explicitly configured personal skills. Support flat and legacy layouts, folded descriptions, whole-package source precedence, qualified ids, and resource drift checks without dependencies or namespace loading.
• Ship saber's own skill under inst/skills/saber/. Registration and stable installed-package snapshots remain explicit consumer setup work.
• Add the agent-hygiene skill under inst/skills/agent-hygiene/: the layout contract for one shared instruction body and one flat skill hub, how each consumer deduplicates, and scripts/agent-hygiene.R, which audits instruction files, hub entries, agent roots, corteza roots, and Claude plugins, and creates or repairs the symlinks. It never removes a real file.

Behavior changes

• agent_context() now merges project CLAUDE.md and AGENTS.md instead of picking one. The file the consumer autoloads is still left out, and so is any heading-delimited section whose text the consumer already holds. Identical copies add nothing, and a drifted copy adds only the sections that differ, under a (sections not already loaded) heading. corteza, unknown consumers, and include_project = TRUE used to receive CLAUDE.md alone when both files existed; they now also receive the sections only AGENTS.md has. A single project file is returned as before. The comparison is verbatim, so differently worded files are both loaded in full.
• The session-start script now emits the project instruction file the named agent does not load natively: CLAUDE.md for Codex, AGENTS.md for Claude Code, and CLAUDE.md (else AGENTS.md) for corteza. Since 0.7.1 the script had passed include_project = FALSE, so a project with only one of the two names reached only one agent. Invocations without an agent name are unchanged.
• The session-start script accepts an optional trailing --native-shared flag. It suppresses shared preferences only when the native global entrypoint resolves to the same readable file, respecting Codex's override file. Without the flag, historical injection is unchanged.

Fixes

• Deduplicate symbolic-link aliases in project and downstream discovery, while preserving separate checkouts with identical package names.
• Fix JSON escaping for all representable control characters in hook output.
• The session-start hook now reads littler's argv as well as Rscript's command arguments, so explicit consumer names work with either launcher.

Documentation

• Document portable registration through the installed package, all-event SessionStart matching, and native shared-file links. Essential instructions should remain available without hooks; Claude Explore/Plan limitations are documented rather than treated as supported hook injection.

saber 0.7.2

Consolidates the 0.7.1.x development cycle.

Changes

• fn_graph() now accepts cache_dir, mirroring blast_radius() and symbols(). The default is unchanged. The example and tinytest pass tempdir() so R CMD check no longer leaves files under tools::R_user_dir("saber", "cache") (#32).
• briefing_git() no longer leaks a system2() "had status 128" warning when run against a non-repository (e.g. a worktree, an invalid .git, or a dubious-ownership directory). It now confirms the working tree with git rev-parse and suppresses the warning, returning empty silently (#33).
• The SessionStart hook only sources a local package's R/ when that package is saber itself, so a same-named function exported by another project on the load path is no longer picked up in place of saber's (#31).

saber 0.7.1

Changes

• Rebranded package as "Context Engineering for Large Language Model Agents".
• briefing() now emits output via message() instead of cat() for CRAN compliance.
• agent_context() examples use \donttest{} instead of \dontrun{}.
• agent_context() and the SessionStart hook now load memory reciprocally: Codex receives Claude Code MEMORY.md, while Claude Code, Corteza, and other non-Codex agents receive Codex memories.
• Codex hook setup docs now use [features].hooks instead of deprecated [features].codex_hooks.
• Added Depends: R (>= 4.4.0) and removed local %||% definition (now in base R).
• Added copyright holder person("cornball.ai", role = "cph") to Authors@R.
• Expanded acronyms in DESCRIPTION ("AI", "AST") per CRAN policy.
• Single-quoted file-name references in DESCRIPTION ('AGENTS.md', 'CLAUDE.md').
• Added ?saber package-level help page.
• README examples switched from r -e to Rscript -e for portability.
• Fix blast_radius() vignette scan crashing on Windows paths (backslashes were interpreted as regex backreferences).
• Replace em-dashes in roxygen @title lines with colons to keep the generated Rd files ASCII-clean.

saber 0.7.0

New features

• agent_context() assembles agent context from memory, instructions, and identity files.
• fn_graph(), pkg_graph(), and graph_svg() render interactive SVG call graphs.
• blast_radius() gains include parameter for scanning @examples and vignettes.

Improvements

• Expanded "AST" acronym in DESCRIPTION per CRAN reviewer feedback.
• briefing() gains agent parameter for multi-agent support.
• Session-start hook script accepts agent name as CLI argument.

saber documentation built on Nov. 10, 2026, 9:06 a.m.