SiftCoder Field Guide¶
SiftCoder is a Claude Code plugin that gives your terminal a memory. It remembers the work you and Claude do together — the files you touched, the bugs you chased, the patterns you settled on — and quietly hands that context back to Claude on later sessions so you stop re-explaining your own project.
It also brings two things you don't usually get out of the box: a local LLM offload path so summarisation and embeddings can run on Ollama instead of your Anthropic budget, and a Salesforce domain pack with skills for Apex, Flow, LWC, schema migrations, packaging, and security review.
This guide is the long-form companion to the README. The README answers what is this and how do I install it. This guide answers how does it actually think, what should I do with it on a Monday morning, and where does it break.
Where to start¶
If you have never installed it, go straight to the Quickstart. You will be running and capturing your first events in roughly five minutes.
If you have it installed but it feels mysterious, read the Mental model chapter. It is the shortest path from "this tool exists somewhere on my system" to "I know exactly what it is doing and why."
If you are trying to decide whether to adopt it, the When to reach for SiftCoder page is honest about the tradeoffs. It tells you when SiftCoder pays for itself and when a regular Claude Code session is fine on its own.
If you are already using it and you want to go deeper, the Operations and Cookbook tabs are where the real material lives — daily workflows, backfill recipes, multi-agent refactors, Salesforce migrations, and the kind of thing you actually do at work.
How this guide is organised¶
The guide has six tabs across the top. They are ordered roughly by how you will encounter them as a user.
Quickstart — get it installed and capturing in under ten minutes. Everything is verified end-to-end with copy-pasteable commands and what you should see if it worked.
Foundations — the mental model, the architecture, how memory is stored and retrieved, how namespaces work, what the hooks do. Read these once and the rest of the tool stops feeling like magic.
Operations — what you do day to day. Backfill, drain, prune, watch the web UI, troubleshoot when something seems off, keep costs down.
Reference — every CLI command, every slash command, every skill, every agent, every config knob. Searchable, exhaustive, cross-linked. This is the chapter you grep, not the one you read in order.
Salesforce — the domain pack. Apex patterns, schema migrations, deploy and rollback, security review, LWC debugging, org architecture review. Skip this tab if you don't work in Salesforce.
Cookbook — opinionated recipes for real workflows. Onboarding a new repo, capturing a debugging session you'll want back next week, doing a multi-agent refactor without losing your mind, migrating a Salesforce org piece by piece.
Project — the changelog, contributing guide, security policy, and licence. Where the project itself lives and how to participate.
A note on tone¶
This documentation is written by a human. There is no template-stamped marketing copy, no "comprehensive solution" or "leverage" or "robust framework." Where something is good, it says so. Where there is a sharp edge, it warns you. Where a decision was made for a reason that is not obvious, it tells you the reason.
If a page reads like it was generated by a model, that is a bug — please open an issue and we'll rewrite it.
What's actually in the box¶
A persistent memory daemon running per-workspace over a Unix domain socket. A WAL so writes are crash-safe. A SQLite event log with summaries and embeddings. Hybrid retrieval that blends BM25 and vector similarity through reciprocal rank fusion. Time decay so old context gradually loses weight. Backends for Ollama (local), Anthropic (cloud), and the Claude Code sampling API. A web UI for inspection. A transcript replay path so a freshly installed SiftCoder can hydrate itself from your existing Claude Code history. Hooks that capture tool calls automatically, redact secrets, and enforce file-write boundaries. A whole catalogue of skills and agents for code review, debugging, planning, refactoring, and Salesforce work.
The rest of this guide goes through each of those, one at a time, with the reasons behind the design choices and worked examples of how to use them.