Blog
MCP servers, AI coding agents, and software architecture from the enola team.
How a missed N+1 led to better TypeScript detection, lower extractor complexity, and 1.3–3.6× faster extraction with byte-identical output.
Across 238 Claude Code and OpenCode sessions, 4,375 Enola calls added ~4.1M tokens to context. See latency, tool costs, and source-reading estimates.
An introduction to Enola’s constraint miner: find dependency patterns, review exceptions with your team, and turn agreed boundaries into checks.
Validator feedback changes whether a coding agent can repair rejected work. Location, observed state, and valid alternatives make the next attempt better.
We ran enola's package metrics on its own 117 typed packages. Nine were flagged; one warranted structural work, and the useful changes produced byte-identical facts across 91 repositories.
A graph can tell you nothing references a symbol. Whether to delete it is a separate decision, and it needs context the graph does not hold. How we cleared forty symbols out of Enola with zero regressions.
A coding model never sees the whole repository or session. The harness selects its context, tools, persistent state, and feedback at every step.
Open an architecture snapshot, investigate findings and dependencies, compare a change with a baseline, and verify snapshot freshness and coverage.
Repository instructions guide coding agents before a change. Deterministic checks establish whether the result complied—and gates decide what happens next.
A worked example of PCI DSS and GDPR-inspired constraints that check code reachability and policy coverage across a Go module.
Eight benchmark scripts cover reproducibility, regression precision, memory, architecture labels, command contracts, agent behaviour, and changes between sweeps.
A structural condition, like a service reading another team's database directly, isn't technical debt by default. It becomes debt only against a counterfactual: what a reasonable alternative would save on the work ahead. What that means for pull requests, baselines, and CI policy.
Enola found a wide slash-command dispatcher. Tracing the metric revealed 89 call targets, 97 constructed targets, and one declaration edge.
A model generates decisions; the harness around it makes those decisions operational: tools, context, execution, permissions, and feedback. Where the term ends, how it differs from a framework or a prompt, and where enola fits in the loop.
About 3 KB per fact while generating, about 0.7 KB per fact once served, and a 5.7x gap between the two. What Enola costs in RAM from a small service to the Linux kernel, why it is written in Go, and what supporting twenty languages does to the bill.
Turn the session hook off when you want speed and keep the gate in CI, read the --explain report, write a failure policy that fits your language, manage baselines, and read enola log — with real command output from Gitea, Tokio, Excalidraw and a demo service.
A dependency graph records one relation from specific evidence, not a universal truth about your code. What nodes and edges actually mean, why cycles aren't always a defect, and what reverse traversal proves versus what it doesn't.
Harness engineering splits the controls around a coding agent into guides and sensors, computational and inferential. Mapping Enola onto that grid, including the three quarters of it Enola deliberately does not fill.
Enola can now record and explore how a codebase's architecture changed over time. enola log, show, diff, and blame turn architectural drift into evidence you can query, instead of a story you reconstruct by hand from old commits.
Relationships between repositories are rarely written down in one place. Why cross-repository dependency analysis is an evidence problem, not a search problem.
Tokens are how every model reads and generates text. What they are, how byte pair encoding splits words into pieces, and why some languages cost several times more than English.
Coupling hotspots, two cyclic dependencies wrapped around the app's most complex function, and a handful of auth-related god classes.
A syntax tree is not an architecture graph. A symbol index is not one either. They can be inputs into one, but they are not the thing itself — and that gap is where most of the real engineering work lives.
Most coding agents have one memory layer: whatever is in the context window right now. The full picture is four layers — and most teams are missing at least two.
Input, output, cache writes, cache hits: where token costs actually land in an agentic coding loop, and what it's worth to close the context gap before the agent's first tool call.
The extractor cache's version history isn't glamorous, but its comments read like a changelog of every place a naive parser gets architecture wrong.
Gather context, take action, verify. Of those three phases, "gather context" has no fixed procedure — it's done by searching and reading, building an approximation in every fresh session. That's the phase a deterministic graph changes.
Coupling hotspots, cyclic dependencies, god classes, complexity outliers — and a 56,308-node blast radius hiding inside one module. We walked through the exact review a staff engineer would do.
Most codebase-memory tools parse symbols and retrieve fragments. That is useful, but it is not the same as building an architectural model. The difference matters.
The true cost of onboarding isn't the paperwork — it's the three weeks a new engineer spends building a mental model of the codebase, and the three weeks of everyone else's time it drains.
The hidden cost of how AI coding tools understand your codebase. Every AI coding tool is an explore agent — and the exploration has a measurable cost in tokens, time, and money.