Lynchpin

Lynchpin joins personal data that normally lives in separate systems. It reads activity captures, terminal history, Git and GitHub data, browser and communication exports, health records, and machine telemetry. It turns them into reproducible DuckDB datasets, evidence graphs, analysis products, and bounded context packs for humans and agents.

Raw data stays in the system that owns it. Lynchpin builds read models with explicit coverage, freshness, and provenance, so a result can be traced back to the captures and exports that support it.

The project is built for one operator with a long-running local data lake. It is not a hosted service, and private data does not belong in the repository. The public tree contains reusable parsers, schemas, analyses, query tools, tests, and neutral fixtures. Private identities, vocabularies, classifications, exports, and generated narratives are supplied through external configuration.

Project overview | Documentation | Roadmap

What Lynchpin can answer

Lynchpin is useful when one question depends on several sources. Examples include:

• Which terminal commands preceded a commit?
• What was happening on the machine when a build or service failed?
• How did time, focus, and change volume differ across projects?
• Which evidence supports a claim about work, health, sleep, or system load?
• Which sources are missing, stale, or too weak to support a conclusion?
• What bounded context should an agent receive for a particular project or time window?

Source-specific tools remain authoritative. Lynchpin makes their records joinable without erasing source boundaries.

How it works

captures, exports, repositories, service ledgers
                         |
                         v
                typed source adapters
                         |
                         v
          canonical datasets and manifests
                         |
                         v
                       DuckDB
                 |                 |
                 v                 v
          evidence graph      analysis products
                 |                 |
                 +--------+--------+
                          v
                  context packs, CLI, MCP

A materialization planner determines which datasets are missing or stale, rebuilds them in dependency order, and records a manifest for each result. A coherent time window can then be promoted into DuckDB for SQL, graph construction, analysis, and context generation.

Refresh IDs connect graphs and reports to one specific DuckDB snapshot. Narrative output remains downstream of the evidence rather than replacing it.

Current sources

lynchpin.sources provides typed, lazy access to sources including:

• ActivityWatch application and focus events;
• Atuin commands and asciinema recordings;
• Git repositories, GitHub activity, code snapshots, and PR review details;
• browser history, bookmarks, clipboard, and communication exports;
• wearable health, sleep, media, and other provider exports;
• machine metrics, service state, benchmarks, and system generations.

Adapters preserve source-specific caveats. Missing observations do not become zero, and export-bounded datasets remain distinguishable from continuous capture.

Polylogue, Sinex, and Sinnix

Lynchpin reads a broad set of owner-native captures and exports.

The public repository does not currently include a general adapter for Sinex's PostgreSQL event store. Sinnix can deploy both projects on the same hosts, but shared deployment does not merge their storage or provenance rules. A future Sinex integration should arrive as an explicit typed source adapter with its own coverage and freshness contract.

Datasets and queries

The DuckDB layer provides ordinary SQL and stable readers for:

• commits, files, symbols, and review history;
• activity, focus, health, and communication signals;
• machine state, pressure, services, and experiments;
• evidence claims, graph edges, and source readiness.

It is a join and acceleration layer, not a second archive of raw data.

Evidence graphs and context packs

The graph layer connects projects, commits, files, terminal activity, focus spans, GitHub items, machine context, personal signals, and analysis claims.

Current uses include:

• project and day correlation;
• issue, pull request, and commit closure chains;
• file and symbol overlap;
• source-readiness and confidence matrices;
• chronological current-state timelines;
• Markdown and JSON context packs with explicit weak-evidence options.

Analysis

The analysis package covers both repositories and operator data:

• code velocity, change surfaces, dependency maps, refactor candidates, and cross-project comparison;
• work rhythms, focus fragmentation, workflow mechanics;
• health, sleep, mood, and other longitudinal signals;
• machine pressure, workload overlap, service behavior, experiments, and calibrated attribution;
• inventories of analysis artifacts and promotion of claims into evidence-backed products.

Canonical metrics carry a timeframe, unit or denominator, and artifact path. Interpretive reports are expected to retain those boundaries.

MCP

The MCP server exposes eight public tools. Each tool groups related actions instead of exposing every internal Python function.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗