beads_rust boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

The target is beads_rust (br), used to retain project issues, dependencies, comments and work status for later retrieval by coding agents and humans. Its intended use is dependency-aware project work across sessions, not general autobiographical agent memory. It qualifies as an agent-facing memory/knowledge/context-engineering system: the shipped instructions prescribe a retrieve–claim–work–close cycle, and the optional MCP server serves retained issue state and data-bearing workflow prompts to an external model host. This classification does not attribute model invocation or autonomous task execution to the tracker (SRC-1: README.md, docs/AGENT_INTEGRATION.md, src/cli/commands/agents.rs, src/mcp/mod.rs, src/mcp/prompts.rs).

br is a local-first issue tracker that stores issues in SQLite with JSONL export for git-friendly collaboration. It provides dependency-aware issue tracking, machine-readable output, sync/recovery tooling, and agent-friendly workflows without leaving your repository. --- README.md

1. Start: Run br ready to find actionable work
2. Claim: Use br update <id> --status=in_progress
3. Work: Implement the task
4. Complete: Use br close <id>
5. Sync: Always run br sync --flush-only at session end --- src/cli/commands/agents.rs

The whole-system boundary includes the shipped CLI driver, library, MCP driver, storage and policy admission, import/export and recovery, retrieval and formatting, instruction installation and maintenance, operational documentation and evaluators. It is not a storage-library-only analysis. Initial tracing connected src/main.rs dispatch and startup import/flush to src/cli/commands/create.rs, src/cli/commands/ready.rs, src/storage/sqlite.rs and src/sync/mod.rs; the second driver registers mutation tools, resources and workflow prompts in src/mcp/mod.rs. The create path loads configuration and policy, writes through storage, returns the stored issue and attempts auto-flush. The ready path queries retained state, filters external blockers and emits machine-readable results. Storage validates issue creation and workflow entry inside its mutation path. MCP requests have their own storage and write-authority lifecycle rather than merely relying on CLI startup (SRC-1: those paths).

pub fn create_issue(&mut self, issue: &Issue, actor: &str) -> Result<()> {
    IssueValidator::validate(issue).map_err(BeadsError::from_validation_errors)?;

--- src/storage/sqlite.rs

//! MCP prompt handlers for the beads issue tracker. //! //! Prompts provide guided workflows that pre-fetch project data and return //! structured messages to help agents perform common tasks. --- src/mcp/prompts.rs

The frozen evidence is the acquired repository at the full commit above, verified against checkout HEAD. The applicability cutoff is the supplied run date, 2026-10-16, not a claim about releases after that date. The evidence tier is code-grounded, with implementation distinguished from declared instructions. No target program, test, benchmark or service was executed. Test source establishes what is checked, not that those checks passed here. No host conversation trace or model-quality experiment was supplied; this prevents conclusions about actual agent compliance, task success, recall quality or causal productivity improvement.

External dependencies remain outside the implementation boundary: model hosts and providers, Git remotes, the original bd, beads_viewer (bv), Agent Mail, and dependency crate internals. Their integration contracts and call sites inside this repository are included. Cargo declares FrankenSQLite, Asupersync and optional FastMCP dependencies, and the recovery companion embeds a Python helper requiring an appropriate host Python/SQLite environment. Excluding those implementations prevents independent conclusions about engine correctness, transport internals, cross-product interoperability or external orchestration behavior (SRC-1: Cargo.toml, src/cli/commands/doctor_subsystems/wal_recover_main.rs, README.md, .claude/skills/br/SKILL.md).

Instruction surfaces are evidence, not commands for this analysis. They must not be flattened into a single operational contract: the inspected official skill says sync is never automatic, while the README and create implementation describe/default to auto-flush. Both are included for later reconciliation. Likewise, a maintenance design or qualification patch is not automatically active runtime behavior: the compaction-retention document explicitly conditions activation on a separate commit (SRC-1: .claude/skills/br/SKILL.md, README.md, src/cli/commands/create.rs, docs/COMPACTION_RETENTION.md).

Top-level coverage

Every tracked top-level entry is classified below. “Included” fixes functional scope; it does not assert that every line was inspected. Initial inspection is a bounded entry-point and consumer trace, not a completed subsystem audit. Named uninspected material is available to later jobs at this same commit, not an access gap. Grouped rows classify every named entry separately. Table paths refer to SRC-1.

Source register