# Bobbin Feature Backlog
> **Implementation status (2026-07-23, harding):** ➖ **N/A — this is a backlog, not a plan.** It is a
> high-level idea list for Phase 4 (Higher-Order Analysis); its items are tracked/implemented under their own
> plan docs and issues as they mature. Nothing to assess or label here as implemented/planned — left as the
> living backlog it is.
>
> **Shipped-item audit (2026-08-22, verified against the code):** most of this backlog has since landed.
>
> - **§1 Test Coverage Mapping — ✅ shipped**, cheaper than sketched: `src/index/coverage.rs` derives the
> test↔source map at query time from the stored coupling table (path-based `classify_file`, no
> `ChunkType::Test` variant or index-time schema change needed), surfaced as `bobbin coverage` and the
> `test_coverage` MCP tool. Structural `Tests` chunk edges exist separately (`src/index/test_edges.rs`).
> - **§2 Claude Code Hooks — ✅ shipped** (see `docs/plans/hooks-integration.md`): the `bobbin hook`
> family covers all six hook types with `hook install`/`uninstall`/`status` plus a post-commit git
> hook installer. Not built: the `bobbin mcp-config` JSON generator.
> - **§3 Semantic Commit Indexing — ✅ shipped**: `ChunkType::Commit`, `src/index/commits.rs` on the
> `ChunkSource` seam with per-repo watermark incremental (`last_indexed_commit:{repo}`), the
> `commit_search` MCP tool, and commit→source bridging in context assembly. Note commit chunks are
> deliberately filtered *out* of context injection (see calibration-sweep-results.md).
> - **§4 Cross-Repo Coupling — ✅ shipped, by a different algorithm**: `src/index/cross_repo.rs` links
> repos by **bead-reference co-occurrence** in commit trailers within a repo group — the time-window
> correlation this entry proposed was explicitly rejected as too noisy (ian ruling, bo-oqny). Served
> through `related` on CLI/MCP/HTTP with mandatory role filtering.
> - **§5 Refactoring Planner — 🟡 foundation only**: `ImpactMode::Deps` is now implemented
> (`src/analysis/impact.rs`), no longer stubbed; the planner itself (`bobbin plan`, `RefactoringPlan`,
> ordered modification lists) remains unbuilt.
High-level feature ideas for Phase 4: Higher-Order Analysis. These compose
bobbin's existing signals (embeddings, coupling, complexity, deps, refs, history)
into capabilities that are greater than the sum of their parts.
---
## 1. Test Coverage Mapping
**Idea**: Use git temporal coupling to infer which test files cover which source
files. If `auth.rs` and `test_auth.rs` have a coupling score of 0.9, that's
strong evidence of a test relationship. Flag source files with no coupled test
files as coverage gaps.
**Building blocks that exist**:
- `FileCoupling` data model with score, co_changes, last_co_change
- `MetadataStore::get_coupling()` queries coupled files sorted by score
- Context assembly already expands via coupling
- Git coupling analysis (`analyze_coupling()`) with frequency + recency scoring
**What's needed**:
- Test detection via AST analysis (not filename heuristics): tree-sitter can
identify `#[test]`, `#[cfg(test)]`, `def test_*`, `describe(`/`it(`, etc.
directly during parsing. Tag chunks as test code at index time.
- New CLI command: `bobbin coverage [file]` — show test files covering a source file, or show uncovered files
- Coverage gap report: source files with zero coupled test files
- `ChunkType::Test` variant for richer chunk classification
- MCP tool: `test_coverage`
**Scope**: Medium. AST-based test detection during parsing + coupling queries.
---
## 2. Claude Code Hooks / Tool Integration
**Idea**: Make bobbin a drop-in context provider for Claude Code and other AI
tools. Ship hook scripts and MCP config so users can wire bobbin into their
workflow with minimal setup. Context injection at the right moments — on prompt
submit, on file open, on diff review.
**Building blocks that exist**:
- MCP server with 8 tools, stdio transport (`bobbin serve`)
- JSON output from all CLI commands (search, context, grep, refs, etc.)
- Context assembly with budget tracking and coupling expansion
- HTTP thin-client mode for centralized deployments
**What's needed**:
- Pre-built Claude Code hook scripts:
- `PreToolUse` hook: inject relevant context before file reads/edits
- `PostToolUse` hook: update index after file writes
- `UserPromptSubmit` hook: enrich prompts with bobbin context
- `bobbin hook install` command to configure hooks in `~/.claude/hooks`
- `bobbin hook context` — streamlined context output tuned for hook injection
(compact, budget-aware, no chrome)
- MCP server config generator: `bobbin mcp-config` outputs the JSON block for
claude_desktop_config.json or .mcp.json
- Documentation: integration guide for Claude Code, Cursor, Windsurf, Continue
**Scope**: Medium-Large. Hook scripts are small but the UX design matters.
The MCP server already works — this is about making setup frictionless.
---
## 3. Semantic Commit Indexing
**Idea**: Index commit messages as searchable chunks alongside code. Enable
queries like "find the commit that added rate limiting" or "what changed around
the auth refactor". Commits become first-class searchable entities.
**Building blocks that exist**:
- `FileHistoryEntry` with message, author, timestamp, extracted issue IDs
- `parse_file_history()` already extracts commit data from git log
- Embedder can embed arbitrary text (commit messages work fine)
- LanceDB schema has `repo` column for multi-repo support
- `ChunkType` enum is extensible
**What's needed**:
- New `ChunkType::Commit` variant
- Commit indexing pipeline: parse git log → embed messages → store in LanceDB
- Schema extension: add `commit_hash`, `commit_author`, `commit_date` columns
(or encode in existing fields)
- Incremental commit indexing: track last-indexed commit, only index new ones
- CLI: `bobbin search --type commit "rate limiting"` (already works if chunks exist)
- Commit-to-file linking: store which files a commit touched for navigation
- MCP tool update: search tool already supports type filtering
**Scope**: Medium. The indexing pipeline is the main work. Search comes free
once commits are stored as chunks.
---
## 4. Cross-Repo Coupling
**Idea**: Detect temporal coupling across repository boundaries. When changing
the API schema in repo A always requires updating the client in repo B, surface
that relationship. Critical for monorepo-adjacent architectures and microservice
ecosystems.
**Building blocks that exist**:
- Multi-repo support: `--repo` flag, repo column in LanceDB, `get_all_repos()`
- Per-repo coupling analysis via `analyze_coupling()`
- SQLite metadata store for coupling + dependencies
**What's needed**:
- Extend SQLite coupling table with `repo_a` and `repo_b` columns
- Cross-repo coupling analyzer: given repos A and B, find files that change in
the same time windows (not same commits — these are different repos)
- Time-window correlation: commits within N minutes across repos suggest coupling
- CLI: `bobbin related --cross-repo file.rs` — show coupled files in other repos
- Dashboard view: `bobbin coupling-matrix` — show which repos are tightly coupled
- Config: register related repos and their git roots
**Scope**: Large. Requires new analysis algorithm (time-window correlation vs
same-commit co-change). Schema migration for SQLite tables.
---
## 5. Refactoring Planner
**Idea**: Given a refactoring target (rename symbol, move file, extract function),
generate an ordered modification plan using impact analysis + symbol refs + dependency
graph. "To rename `FooBar` → `BazQux`, modify these 14 files in dependency order."
**Building blocks that exist**:
- `ImpactAnalyzer` with transitive expansion (coupling + semantic signals, decay)
- `RefAnalyzer` with `find_refs()` (definition + usages with exact line numbers)
- `MetadataStore::get_dependents()` — what files import a given file
- `GitAnalyzer::get_diff_files()` — precise line-level change extraction
- Import resolution across 7 languages (Rust, TS, Python, Go, Java, C++)
**What's needed**:
- Implement `ImpactMode::Deps` (currently stubbed) to use dependency graph
- `RefactoringPlan` data structure: target, operation type, ordered file list,
per-file change descriptions, confidence scores
- Dependency-aware ordering: modify leaf dependents first, work toward root
- Operation types: rename symbol, move file, extract function, inline function
- CLI: `bobbin plan rename FooBar BazQux` — generate plan
- CLI: `bobbin plan move src/old.rs src/new.rs` — generate move plan
- MCP tool: `refactoring_plan`
- Dry-run mode: show what would change without making changes
**Scope**: Large. The algorithm for ordering modifications correctly (respecting
dependency direction) is the hard part. Symbol refs are FTS-based (~80% accuracy),
so plans should include confidence levels.
---
## Priority Assessment
| Claude Code Hooks | High | Medium | Strong (MCP + CLI exist) | P1 |
| Semantic Commit Indexing | High | Medium | Strong (git parsing exists) | P1 |
| Test Coverage Mapping | High | Medium | Strong (coupling exists) | P2 |
| Refactoring Planner | High | Large | Good (impact + refs exist) | P2 |
| Cross-Repo Coupling | Medium | Large | Partial (multi-repo exists) | P3 |
---
## Pending follow-up beads (file once central beads_bobbin schema migration clears)
> Captured 2026-06-26 by strider while `bd create` was blocked by the schema migration.
> **Now filed as beads** (maldoon cleared the schema same day via fresh-init v52 + reimport):
> bo-7kr (FTS verify/deploy, P2), bo-uu5 (PPR wiring, P3 pitch), bo-fdd (telemetry
> auto-association, P3 pitch), bo-bf5 (telemetry layers 2-5, P3 pitch), bo-q82 (ontology
> auto-discovery, P3 pitch). Kept here as a human-readable index.
- **FTS 500 bug** (GitHub scbrown/bobbin#21) — FIXED IN CODE on main (157a14a):
self-healing FTS index (rebuild+retry on error) + watch rebuilds FTS after
compaction. Root cause: missing/stale FTS index (swallowed create error +
compaction invalidation) → every FTS query 500'd. **Awaiting search.example redeploy**
(routed to dearing→malcolm/infra) to take effect + live verification; #21 left open.
- **PPR ranking — final wiring + tuning** — WIRED as of the 2026-08-22 audit:
`ppr_weight` lives in `ContextConfig` (`src/search/context.rs`) and `SearchConfig`
(`src/config.rs`, feature-gated default) with a `--ppr-weight` CLI flag, and
`compute_code_ppr` folds into the score-adjust chain (`context.rs` ~1223-1261);
`bobbin context` errors honestly when PPR is requested on a build without the
`knowledge` feature. Remaining: tune weight/seed-count against the eval harness. **P3.**
- **#9 telemetry Layer 1.5 — auto-association** — populate `bead_lineage` automatically:
post-commit hook / commit-message bead-ID detection during indexing (git `CommitEntry`
already carries trailers + files). Makes Layers 2-5 useful. **P2.**
- **#9 telemetry Layers 2-5** — bundle suggest/drift (freq over `bead_lineage`),
association mining, predictive priming, graph enrichment. Need bead-data accumulation.
GitHub #9 stays open until done. **P2-P3.**
- **#14 ontology auto-discovery (D5)** — mine embedding clusters + coupling communities
to suggest ontology nodes. Feeds from #9 telemetry. **P3.**
- **Incremental bead `updated_at` caveat** — RESOLVED 2026-06-26: maldoon reverted the
`issues.updated_at` mass-bump (622 distinct timestamps restored) when resetting the
central DB to a clean v40 baseline. No recency pollution remains.
- **bd schema migration root cause** (maldoon, 2026-06-26): migrations 0041-0052 require
`ALTER TABLE dependencies DROP PRIMARY KEY`, which crashes the connection on Dolt 2.1.8
(reproduced in isolation) — blocks bd and manual SQL. Fix path = fresh-init at v52 +
reimport (avoids the crashing DDL), pending Stiwi go (it's a DB replacement).
**HOLD all beads_bobbin writes until maldoon signals clear.**