1 · Collect
Pull requests, design decisions, tickets, meeting notes, and agent sessions land as source files that are never edited.
Context engineering · Coding agents
Claude Code knows how to program. It doesn't know why your system is shaped the way it is, which decisions are settled, or which mistakes already hurt you. I built a knowledge layer that tells it, and proved the difference with blind tests on real tickets.
The Vault
Pull requests, design decisions, tickets, meeting notes, and agent sessions land as source files that are never edited.
A model turns them into linked pages. Plain code writes the files and checks that every citation resolves.
A session starts from a short index and opens deeper pages only when the task needs them.
When an agent edits a file that a team rule covers, the rule is handed to it right then.
97% leaner startup context: agents get a map, not the pages
13k lines of Python, 900+ tests
The proof
Each run started from the commit before the original work, with a snapshot of the Vault from that day and the answer removed. Blind judges scored every change against an answer key and 1,277 real review comments.
| Setup (24 runs each) | Accepted changes | Blocker / major bugs | $ per accepted change | Avg minutes |
|---|---|---|---|---|
| Opus, no Vault | 11 | 3 / 6 | $7.79 | 12.8 |
| Opus + Vault | 16 | 0 / 0 | $4.12 | 9.1 |
| Sonnet, no Vault | 7 | 3 / 13 | $4.70 | 10.5 |
| Sonnet + Vault | 8 | 1 / 9 | $3.31 | 8.2 |
0 of 8 → 9 of 16
Runs that pushed back on a ticket asking for a tool that already existed.
3.4 → 4.0
Fit with the codebase, the score that moved most (out of 5, Opus).
The ontology pipeline
Knowledge tends to be filed where it was learned, so a rule learned on one project never reaches the other repositories it applies to. I solved it at the system level with the first four steps of Jessica Talisman's ontology pipeline.
One ID and one sourced definition per term. A concept needs two pages about it before it counts.
Every page says what it's mainly about and, where it matters, when it applies. Checked on every write.
Repositories roll up to teams and platform groups, so a rule for a group reaches every repository in it.
Alternate names and known name collisions, so search finds a page by whatever name you use.
95% vs 76%
Agents read the right rule before their first edit, with the pipeline vs without, on rules filed far from the code they govern. Steps 5 and 6, a formal ontology and a knowledge graph, are skipped by design: the reader is a model reading prose.
Rules at the edit
A written rule only helps if the agent has it at the right moment. Each rule now arrives the moment the agent edits a file it covers, and agents correct course mid-edit.
0 → 6
of 8 implementations done to standard
0 → 5
of 8 accepted as is
Dogfooding
A report reads my session logs back and shows what context each session was actually given: searches, hits, pages opened, and rules delivered.
Each session links to the changes it produced and to any later fix on the same lines, so quality is measured, not assumed.
A blind test flagged a workflow step that doubled cost without improving results. Cutting it halved the cost.
Let's compare notes
I'm always up for comparing notes on context for coding agents and knowledge systems that keep up.