Build the context compiler your work needs.
A context compiler turns a pile of documents into a set of dated, sourced statements — so you can ask what is true now, what was true then, and how we know. This repository is the kit; what it produces is a compiler of your own. Documents record events, they do not hold a view — so the answer to “why do we think that?” normally gets rebuilt by hand, every time. A compiler does that work once, as each source arrives.
One bootstrap. Two products.
The bootstrap creates and configures a compiler. The specialized compiler continuously compiles evidence into a package people and other agents can use.
-
01 · bootstrap
Create the compiler
Scaffold a clean repository containing the generic compiler core, profile extension contract, commands, validators, packaging, and tests.
./scripts/create-context-compiler.sh ../my-context
-
02 · specialize
Teach it the domain
A profile supplies the ontology, claim classes, controlled relationships, extraction guidance, schemas, outputs, and exception rules.
./scripts/use-profile.sh client-decision
-
03 · operate
Compile evidence repeatedly
New evidence updates only affected state. Users ask for briefs, deltas, decisions, evidence chains, historical views, and review exceptions in ordinary language.
Add this evidence and update my context. Prepare me for the Northstar meeting.
A specialization is explicit, inspectable, and replaceable.
The included client-decision profile demonstrates the contract. It defines clients, decisions, exposures, assumptions, risks, variables, speaker attribution, temporal state, supersession, contradiction, and evidence requirements without putting commodity logic into the generic core.
profiles/client-decision/ ├── profile.json ├── COMPILATION.md ├── schemas/ ├── vocabularies/ └── templates/
To create a different compiler, ask the coding agent to add a sibling profile with its own fixtures and acceptance tests. A project-decision compiler can understand owners, alternatives, constraints, and architectural supersession. A due-diligence compiler can define a different ontology without rewriting extraction, provenance, temporal logic, linting, or packaging.
The profile changes what the compiler understands. The core preserves how evidence becomes trustworthy context.
The daily workflow starts with evidence, not configuration.
After specialization, users normally do not edit claims or graphs. They drop evidence into EVIDENCE-INBOX/, read useful outputs in BRIEFS/, and inspect only exceptions in REVIEWS/.
-
morning
Prepare for a meeting
Add the latest emails, transcript, report, or spreadsheet. The workspace updates the context and saves a current decision brief.
Add these files and prepare me for the Northstar meeting.
-
after new evidence
See meaningful change
A semantic delta separates new, changed, superseded, unresolved, and unchanged-but-material state.
What changed for Northstar since August 1?
-
when challenged
Show why
Return the chronological evidence chain, speaker, role, source anchors, classification, confidence, and conflicting evidence.
Why do we think Northstar is concerned about Q1 availability?
People consume views. Systems consume the package.
- Meeting preparation — a concise current-state brief
- Change detection — a semantic delta rather than a document diff
- Accountability — an evidence chain for every material conclusion
- Historical reconstruction — what was believed at a previous date
- Maintenance — exception queues for contradictions, stale assumptions, and ambiguity
- Downstream automation — structured claims, decisions, relations, and provenance for another agent
The package is local, Git-friendly, portable, and verifiable. Every material claim can retain its source, timestamp, speaker, evidence anchor, classification, and confidence. Unsupported questions produce explicit uncertainty instead of a nearest-neighbour guess.
Use it as an AI workspace, not an IDE project.
The optional Context Workspace VS Code extension gives managed laptops a local sidebar for creating a compiler, adding evidence, preparing briefs, showing changes, explaining conclusions, reviewing exceptions, and opening exact evidence citations.
Add Evidence is one universal intake. Choose files or folders, drop them on the sidebar, paste text, paste one or many links, or use the inbox. Every regular file is preserved first. Known formats are extracted; unfamiliar content is retained and reported as needing attention rather than silently lost or falsely completed.
Create Specialization is conversational. Describe another kind of work in ordinary language. The agent asks only for missing domain meaning, generates the schemas and tests, previews supported, temporal, contradictory, and UNKNOWN examples, and requests explicit domain-owner approval before activation.
It uses the organization's existing GitHub Copilot chat surface. It stores no credentials, emits no telemetry, and keeps the evidence and compiled package in the selected local workspace.
./scripts/package-vscode-extension.sh code --install-extension dist/context-workspace-0.1.0.vsix
The extension is optional. The filesystem compiler remains portable and works through AI chat, terminal workflows, CI, or another supported agentic environment. Successful validated updates create a scoped local Git checkpoint automatically; unrelated work is excluded and nothing is pushed.
5core slash commands — the generic lifecycle; 22 in all
495pages in the measured scale fixture
10binary retrieval checks, no LLM grader
51deterministic smoke checks, wired into CI
What is measured, and what is not.
Those numbers describe the machinery: that the pipeline runs, that citations resolve to real anchors, that a gate fails when they stop doing so. They are not a claim about answer quality.
The decision benchmark was executed for the first time on 2026-08-25. Against BM25 retrieval the compiled package answered more of the twelve graded cases. Against a capable agent given the raw folder it did not — at 24 sources or at 828. Provenance precision is currently unscoreable: three of the seven metrics are defective and return zero for every arm.
Create your first specialized compiler.
You need an AI-enabled workspace such as VS Code with an AI assistant. The included profile provides a working client-decision specialization; the generic default remains available. The generated compiler opens through a simplified workspace that hides implementation machinery.
git clone https://github.com/FrancyJGLisboa/context-compiler-bootstrap bootstrap ./bootstrap/scripts/create-context-compiler.sh ./northstar-client-context cd ./northstar-client-context ./scripts/use-profile.sh client-decision
Then open AI-WORKSPACE.code-workspace. Choose, drop, paste, link, or inbox your evidence, and tell the AI:
Add this evidence and update my context. Update my context, then prepare a brief for Northstar Feeds.
Current boundary: this is an LLM-assisted development kit, not a no-code generator. Creating a genuinely new specialization still requires an AI-assisted development session. Daily operation uses AI chat and familiar folders; commands, schemas, and Git remain available as advanced controls.