Guides

Understanding the platform

• Capability catalog — generated. Every package, product and app with a "use this when" line and key exports, the proven runtime paths and their credentials, and the known gaps. The first stop before you build anything.

• Architecture — the tiers, the real dependency graph, why the DAG is enforced by CI, and the two time models store-sqlite offers.

• The binding pattern — how a product fixes a package's type parameters in one module so every call site stays unchanged. The single idea most worth reading before you adopt anything.

• Where code goes — mechanism in packages, pixels in titan-design, policy in products; how a product's own npm scope shrinks as its engine is ported; and what a port pull request must prove.

• Where logic goes — the four homes for pipeline logic: rule file, recipe, step and service, and when old-path code is a valid home.

• Package families — the packages grouped by the job they share (storage, command surfaces, the front-end kit, agent execution, landing a change, session mining, code audit, messaging), with a line and a link for each.

• Multi-harness contracts — the additive Claude/Codex contracts and the compatibility migration required before mixed-session ingestion.

Running the products

One guide per runnable unit. Each covers what the unit is for, prerequisites, the build, every command with an invocation, where state and logs live, and how it fails.

• Factory — titan-factory: serve on loopback port 7410 with /health, /rpc and /mcp, land, gate resolve, resume, the launchd service, and the config file.
• Shepherd — the factory's PR watcher end to end: register a pull request or a branch, phases, seat policy, hold and release, the merge evaluation, the MRG-AU-RV authority row, and what is not built yet.
• Set up the autonomous loop — what to create to run the build-and-merge loop: the release GitHub App, npm trusted publishing, rulesets, the factory service, seat policy and Claude Code grants.
• Session miner — titan-miner: index Claude Code and Codex transcripts, search, cluster failures, keep a playbook, serve over MCP and HTTP.
• Retrieval eval — retrieval-eval: mine query and label pairs from transcripts and score candidate retrievers.
• Code report — apps/codewatch: index a repo, browse the report live, or export one HTML file.
• Matrix hub — deploy/hub: the Tuwunel and Caddy compose stack the human queue runs on.

End-to-end stories

• Case study: the session miner — fifteen packages composed into one product. Follows a refresh and a search through the DAG, with real output.
• Case study: adopting registry and daemon — active-work replacing two hand-written modules with packages: about a thousand lines deleted, 1,321 tests unchanged, and the two regressions that nearly slipped through.

Working here

• Working in the repo — the scaffold, the DAG check, changesets, and how a release happens.